Ground-based Validation of the MODIS Leaf Area Index Product for ...

Ground-based Validation 

of the MODIS Leaf Area Index Product 

for East African Rain Forest Ecosystems 

Der Naturwissenschaftlichen Fakultät 

der Friedrich-Alexander-Universität Erlangen-Nürnberg 

zur Erlangung des Doktorgrades 

vorgelegt von 

Tanja Kraus 

aus Nürnberg

Als Dissertation genehmigt 

von der Naturwissenschaftlichen Fakultät 

der Friedrich-Alexander-Universität Erlangen-Nürnberg. 

Tag der mündlichen Prüfung: 10. Oktober 2008 

Vorsitzender der Promotionskommission Prof. Dr. Eberhard Bänsch 

Erstberichterstatter Prof. Dr. Cyrus Samimi 

Zweitberichterstatter Prof. Dr. Roland Baumhauer

i 

Summary 

In the context of current global changes, tropical rain forests are undergoing significant alterations. Deforestation 

and degradation processes caused by human exploitation have already destroyed between 8 and 12 million km2 (i.e. 35-50%) of the original forest cover. Large amounts of CO formerly stored in vegetation and soil are thus 

2 

released and add to globally rising atmospheric CO rates. Regional and global climate change is enforced 

2 

as precipitation patterns, surface temperature and carbon uptake are modified. Several international initiatives 

(mostly under the framework of the United Nations Framework Convention on Climate Change, UNFCCC) call 

for action to stop or at least reduce deforestation and degradation in the tropics and thus emissions. Concrete 

mechanisms will probably be defined in a post-2012 climate agreement, the details of which are currently being 

discussed and negotiated among nations. 

One of the biggest challenges to the estimation of changes in forest cover is the monitoring of forest areas on a 

reliable, fast, cost effective and area-wide basis. Here, operational remote sensing techniques form a valuable 

data source for the scientific and political community since they permit repetitive and synoptic observations of 

vegetation cover. Changes in forest structure and dynamics can, for instance, be monitored through repeated 

measurement of remotely measured biophysical attributes. In contrast to discrete representations of land 

cover, biophysical variables alter continuously over space and time and may thereby reveal early ecosystem 

modifications. 

In this context, leaf area index (LAI) is one of the key biophysical variables. It characterises canopy structure 

and accounts for differences in phenological development, assimilation and biomass growth among plant 

species. Operational standard products of LAI derived from satellite data, as e.g. the MODIS LAI product, can 

thus contribute to the constant and repetitive monitoring of forest areas. 

However, a prerequisite for the use of operational satellite products is the evaluation of their accuracy in a 

process called validation. Validation is defined as the process of assessing by independent means the quality of 

data products. So far, there has been a clear lack of validation sites for the MODIS LAI product in tropical rain 

forest environments. Especially in Africa, no test sites have been available. 

Consequently, this thesis deals with the validation of the MODIS LAI product for two test sites in East Africa 

that were accessible within the framework of the BIOTA East Africa project. Ground-based measurements were 

performed in Budongo Forest (Uganda) and Kakamega Forest (Kenya) and upscaled to the spatial resolution 

of the MODIS LAI product based on high resolution satellite data of ASTER and SPOT-4. Finally the MODIS 

LAI product was validated with respect to its accuracy in representing landscape LAI and to its temporal 

consistency. 

First of all, existing methods of in situ measurements were reviewed concerning their applicability to tropical 

rain forests. Additionally, a representative and valid sampling scheme was elaborated based on recommendations 

of the CEOS-LPV and VALERI networks. LAI-2000 PCA and digital hemispherical photography were found to 

be suitable for ground measurements as they meet the demand of fast and area wide sampling. Yet environmental 

conditions in the tropics (especially with respect to prevailing radiation regimes) are not ideally suited to these

instruments, so that the measurements had to be analysed thoroughly with respect to different error sources. A 

correction method for LAI-2000 PCA data was developed for measurements made under direct radiation. Other 

errors, influencing both LAI-2000 PCA and digital hemispherical photography, could not be corrected, but their 

influence was quantified in terms of measurement precision. 

Theil-Sen regression was found to be suitable for the upscaling process, as its performance is robust even in the 

presence of measurement errors in field and satellite data. For Budongo Forest, where a better quality of in situ 

data could be retrieved than for Kakamega Forest, a thorough analysis revealed that different transfer functions 

had to be established for different forest stages. Interestingly, mean in situ LAI retrieved for intermediate and 

late forest stages was relatively similar, but structural differences mainly in the upper canopy lead to significantly 

different surface reflectances. Whereas Simple Ratio was found to perform best for early and intermediate forest 

stages (R2 of 0.94) a texture measurement (GLCM variance of ASTER band 4) retrieved better results for late 

forest stages (R2 of 0.71). Consequently, a high resolution LAI map could be produced for Budongo Forest 

e 

with a relative accuracy of 9%. This signifies only a slight degradation over field measurement precision. For 

Kakamega Forest inferior in situ data quality led to reduced quality of the high resolution LAI map. A regression 

model based on Reduced Simple Ratio was here used to estimate LAI for the whole test site. Overall this model 

yielded an R2 of 0.53 with a relative accuracy of 16% (RMSE of 0.8). 

Based on the high resolution LAI maps, the MODIS LAI product was then validated for the two test sites. The 

spatial validation for Budongo Forest revealed that it represented the up-scaled in situ LAI with an accuracy of 

0.53. This corresponds to a relative accuracy of 9%, which is identical to the accuracy of the high resolution 

LAI maps. For Kakamega Forest validation led to a comparatively low accuracy of the MODIS LAI product of 

1.5 (relative accuracy of 25%). This is most likely the result of inferior field data quality for this test site and the 

resulting degradation of accuracy in the upscaling process. 

The investigation of the temporal consistency of the MODIS LAI product was based on a time series analysis 

for the years 2000-2005. Results showed that the data was reliable and stable, but only if temporal interpolation 

was applied to bad data quality pixels. The observed variability in LAI (0.4 for intermediate and late forest 

stages) further corresponds to in situ measured seasonal LAI trajectories found by other studies for comparable 

semi-deciduous rain forests. Maximum LAI values are associated with the end of the rainy season, minimum 

LAI values with the dry season. It can thus be assumed that the MODIS LAI product responds correctly to 

biome-level LAI changes associated with interannual climate variability. 

This thesis shows that the MODIS LAI product represents the LAI of the two East African test sites with an 

accuracy that is comparable to the accuracy of field measurements. In addition, seasonal variations are captured 

correctly, but only if quality information for the MODIS LAI product is included in the analyses. 

Although structural changes in rain forests can only be monitored if absolute LAI values are affected (differences 

in in situ LAI of intermediate and late forest changes were not found to be significant for the test sites), the 

outcomes of this thesis can nevertheless help to improve knowledge of tropical rain forests and their response 

to climatic changes and human disturbances. The results may e.g. help to reduce predictive uncertainties in 

biophysical process models and serve as supplementary data for model calibration. 

ii

iii 

Zusammenfassung 

Im Kontext des globalen Wandels unterliegen tropische Regenwälder derzeit starken Veränderungen. 

Schätzungen zufolge haben Entwaldung und Degradationsprozesse bereits 8 bis 12 Millionen km2 (35-50%) der 

ehemaligen Waldfläche zerstört. Die Freisetzung großer, ehemals in der Vegetation und im Boden gespeicherter 

Mengen an Kohlenstoff trägt wiederum zum Anstieg der atmosphärischen CO -Konzentration bei und verstärkt 

2 

so den Treibhauseffekt. Modifizierte Niederschlagsmuster, Oberflächentemperaturen und eine veränderte 

Kohlenstoffaufnahme durch die Vegetation haben ihrerseits Einfluss auf das regionale und globale Klima. 

Mehrere internationale Initiativen haben daher im Zusammenhang mit der Klimarahmenkonvention der Vereinten 

Nationen, UNFCCC (United Nations Framework Convention on Climate Change), zum Handeln aufgerufen. 

Ziel ist es, die Abholzung und Degradation von Waldflächen in den Tropen zu stoppen oder wenigstens zu 

reduzieren. Konkrete Mechanismen werden wahrscheinlich in einer Klimavereinbarung nach 2012 („Post- 

Kyoto-Prozess“) definiert, deren Details momentan zwischen den Nationen diskutiert und verhandelt werden. 

Eine der größten methodischen Herausforderungen um die Veränderungen in der Waldfläche abzuschätzen, 

ist derzeit ein verlässliches, schnelles, günstiges und qualitatives Monitoring der Waldbereiche. Hier stellen 

operationelle Fernerkundungstechniken eine wertvolle Quelle für Wissenschaft und Politik dar, da sie 

flächendeckende Aufnahmen der Vegetationsbedeckung liefern. Veränderungen in der Waldstruktur und in 

der Dynamik bestimmter Waldflächen können zudem durch wiederholte Abschätzung biophysikalischer 

Eigenschaften gemessen werden. Im Gegensatz zu einer diskreten Darstellung der Landbedeckung verändern 

sich biophysikalische Variablen kontinuierlich in Zeit und Raum und haben dadurch das Potential, frühe 

Veränderungen in Ökosystemen darzustellen. 

In diesem Zusammenhang ist der Blattflächenindex (leaf area index, LAI) eine der Schlüsselvariablen. Er 

charakterisiert die Struktur von Waldbeständen und spiegelt deren phänologische Veränderung wider. Auch 

die Unterschiede in den Assimilationsraten verschiedener Spezies und Biomasseveränderungen werden in 

der Blattfläche sichtbar. Operationelle fernerkundliche Standardprodukte zum LAI, wie z.B. das MODIS LAI 

Produkt, können so zu einem konstanten und wiederholten Monitoring von Waldflächen beitragen. 

Eine Voraussetzung für den Gebrauch operationeller Satellitenprodukte ist jedoch die Überprüfung ihrer 

Genauigkeit durch eine sogenannte Validierung. Validierung ist in diesem Zusammenhang definiert als ein 

Prozess, in dem anhand unabhängiger (Feld-)Daten die Qualität von Fernerkundungsprodukten überprüft wird. 

Bis zum heutigen Zeitpunkt sind Validierungsflächen für das MODIS LAI Produkt in tropischen Regenwäldern 

unterrepräsentiert. Besonders in Afrika steht bislang keine einzige derartige Validierungsfläche zur Verfügung 

Dementsprechend beschäftigt sich diese Dissertation mit der Validierung des MODIS LAI Produkts für zwei 

Testflächen in Ostafrika, die im Rahmen des BIOTA-Ost-Projekts zur Verfügung standen. Feldmessungen 

zum Blattflächenindex wurden im Budongo Forest (Uganda) und im Kakamega Forest (Kenia) durchgeführt 

und mithilfe von hochaufgelösten Fernerkundungsdaten (ASTER, SPOT-4) auf die räumliche Auflösung des 

MODIS LAI Produkts skaliert. Schließlich wurde das MODIS LAI Produkt auf Basis dieser Datengrundlage 

validiert und auf seine räumliche Genauigkeit und zeitliche Konsistenz hin überprüft.

Im ersten Schritt wurden die existierenden Methoden zur in situ Messung von LAI im Hinblick auf ihre 

Anwendbarkeit für tropische Regenwälder begutachtet. Zusätzlich wurde ein repräsentatives und gültiges 

Aufnahmeschema für die Feldmessungen erarbeitet. Dieses basiert im Wesentlichen auf Empfehlungen der 

Netzwerke VALERI und CEOS-LPV. LAI-2000 PCA und digitale hemisphärische Bilder wurden schließlich 

als geeignet erachtet, um LAI im Feld abzuschätzen. Beide Methoden erlauben eine relative schnelle Aufnahme 

des Blattflächenindex für große Flächen. Allerdings sind die Messbedingungen für beide Instrumente in den 

Tropen nicht ideal. So stellt zum Beispiel die Tatsache, dass indirekte Einstrahlung kaum und meist nicht für 

ausreichend lange Zeiträume vorherrscht, ein Problem für die LAI-2000 PCA Messungen dar. Aus diesem 

Grund wurden die Messergebnisse beider Instrumente im Hinblick auf verschiedene Fehlerquellen genau 

analysiert. Für LAI-2000 PCA Messungen, die unter direkten Strahlungsbedingungen aufgenommen wurden, 

wurde eine Korrekturmethode entwickelt. Andere Fehlerquellen, die sowohl LAI-2000 PCA Messungen und 

hemisphärische Fotos beeinflussten, konnten nicht korrigiert werden. Allerdings wurde ihr Einfluss auf die 

Messungen zumindest indirekt durch die Feststellung der Messgenauigkeit quantifiziert. 

Für die Hochskalierung der in situ Messungen auf Basis der hochauflösenden ASTER- und SPOT-4- 

Daten erwies sich die Theil-Sen Regression als geeignet. Sie zeigte sich als robuste Methode, selbst wenn 

Messungenauigkeiten in Feld- und Fernerkundungsdaten vorliegen. Für Budongo Forest, wo eine bessere 

Datenqualität der in situ Messungen erzielt werden konnte als für Kakamega Forest, ergab eine genaue 

Datenanalyse, dass verschiedene Regressionsmodelle für frühe und mittlere (d.h. gestörte/degradierte) 

Waldstadien etabliert werden mussten. Interessanterweise waren die mittleren LAI-Werte für mittlere und späte 

(d.h. ungestörte) Waldstadien fast gleich, allerdings führten strukturelle Unterschiede im Kronenbereich zu 

signifikanten Unterschieden in der Oberflächenreflexion. Während das Theil-Sen Regressionsmodell basierend 

auf dem Vegetationsindex „Simple Ratio“ die besten Ergebnisse für frühe und mittlere Waldstadien erzielte 

(R2 =0,94), konnte der LAI für späte Waldstadien besser durch ein Texturmaß (GLCM Varianz von ASTER, 

Band 4) modelliert werden (R2 =0,71). Dementsprechend konnte eine hochauflösende LAI-Karte mit einer 

relativen Genauigkeit von 9% für Budongo Forest erstellt werden. Dies stellt nur einen leichten Qualitätsverlust 

gegenüber der Messgenauigkeit der Feldmessungen dar. Für Kakamega Forest führte die mindere Datenqualität 

der in situ Messungen konsequenterweise zu einer niedrigeren Genauigkeit der hochauflösenden LAI-Karte. 

Ein Regressionsmodell basierend auf dem Vegetationsindex „Reduced Simple Ratio“ wurde hier verwendet, um 

LAI für das gesamte Studiengebiet abzuschätzen. Dieses Modell erreichte ein R2 von 0,53 mit einer relativen 

Genauigkeit von 16% (RMSE=0,8). 

Basierend auf den hochaufgelösten LAI-Karten wurde das MODIS LAI Produkt schließlich für die beiden 

Studiengebiete validiert. Die räumliche Validierung für Budongo Forest zeigte, dass das MODIS LAI Produkt 

die hochskalierten Felddaten mit einer Genauigkeit von 0,53 repräsentiert. Dies entspricht einer relativen 

Genauigkeit von 9% und ist somit identisch mit der relativen Genauigkeit der hochaufgelösten LAI-Karten. 

Für Kakamega Forest zeigte die Validierung des MODIS LAI Produkts dagegen eine vergleichsweise geringe 

Genauigkeit von 1,5 (relative Genauigkeit von 25%). Dies ist auf die mindere Felddatenqualität für dieses 

Studiengebiet zurückzuführen, die durch das Hochskalieren entsprechend auch eine mindere Qualität der 

hochaufgelösten LAI-Karte nach sich zieht. 

iv

v 

Die Untersuchungen der zeitlichen Konsistenz des MODIS LAI Produkts basierte auf der Analyse von Zeitserien 

der Jahre 2000-2005. Die Ergebnisse zeigten, dass die Daten verlässlich und stabil sind, allerdings nur, wenn 

Pixel mit niedriger Qualität zeitlich interpoliert wurden. Die beobachtete saisonale Variabilität im LAI (0,4 für 

mittlere und späte Waldstadien) entspricht Feldmessungen anderer Studien für vergleichbare halbimmergrüne 

Regenwälder. Maximale LAI-Werte wurden am Ende der Regenzeit registriert, minimale LAI-Werte in der 

Trockenzeit. Somit darf angenommen werden, dass das MODIS LAI Produkt die saisonale LAI-Variabilität 

korrekt abbildet. 

Diese Dissertation zeigt, dass das MODIS LAI Produkt den Blattflächenindex der beiden ostafrikanischen 

Studiengebiete mit einer Genauigkeit repräsentiert, die innerhalb der Messfehler der Feldmessungen liegt. 

Saisonale Veränderungen werden außerdem korrekt wiedergegeben, wenn die Qualitätsinformation des MODIS 

LAI Produkts zur Maskierung und zeitlichen Interpolation der entsprechenden Pixel herangezogen wird. 

Obwohl strukturelle Veränderungen in Regenwäldern mit Hilfe des MODIS LAI Produkts nur beobachtet werden 

können, wenn sie auch die absoluten LAI-Werte beeinflussen (signifikante Unterschiede im in situ gemessenen 

LAI der mittleren und späten Waldstadien konnten nicht ermittelt werden), zeigen die Ergebnisse dieser 

Dissertation dennoch, dass das MODIS LAI Produkt einen Beitrag zur Abschätzung der Reaktion tropischer 

Regenwälder auf klimatische Veränderungen und anthropogene Störungen liefern kann. Die Ergebnisse können 

z.B. dazu beitragen, die Vorhersagegenauigkeit biophysikalischer Prozessmodelle zu erhöhen und als zusätzliche 

Daten in die Modellkalibrierung einfließen.

Acknowledgements 

Many people have supported me over the last few years during the completion of this thesis. Although the list of 

individuals I wish to thank extends beyond the limit of this page, I would like to show gratitude to the following 

individuals. 

First of all my thanks go to my supervisor Prof. Dr. Cyrus Samimi at the University of Erlangen. He not 

only introduced me to the field of remote sensing during my studies, but also supported this work from the 

very beginning with helpful advice and critical discussion. I am also grateful to Prof. Dr. Roland Baumhauer 

(University of Würzburg) for his role as co-reviewer. 

This thesis was written during my time at DFD-DLR and at the Chair for Remote Sensing at the University of 

Würzburg. My sincere gratitude goes to Prof. Dr. Stefan Dech, who made it possible for me to work in both 

places and to benefit from a stimulating atmosphere. Despite his manifold responsibilities he was always ready 

to lend an ear to problems and to make helpful comments. 

Writing a thesis while being involved with other projects apart from BIOTA East has not always been easy. Yet 

this thesis has greatly benefited from the constant support and steady encouragement of Dr. Michael Schmidt. 

Thanks for never giving up on me. I am further grateful to my former team leader, Prof. Dr. Günter Strunz, who 

initially offered me the opportunity to work at DFD. His door was always open, even at stressful times. 

Carrying out fieldwork in Kakamega and Budongo Forest was one of the most exciting experiences for me. 

The people I encountered and the experiences I have had have continued to influence me, both personally 

and academically. All this would not have been possible without the framework of the BIOTA East Africa 

project funded by BMBF. Yet a network only thrives thanks to the individuals who support it and thus I thank 

my colleagues, especially those from the subproject E02. The subproject leader Prof. Dr. Gertrud Schaab, and 

PhD students Tobias Lung and Nick Mitchell have helped with valuable discussions and with sharing data and 

knowledge concerning the two test sites. 

Without the assistance of Jaqueline Kennedy Ayuka, Benson Bwibo Chituyi, Kennedy Andama, and Afeku 

Alfred fieldwork would not have been possible. I greatly appreciate that you shared your valuable knowledge 

concerning the flora of Kakamega and Budongo Forests with me. Thanks go also to Mark Broich and Susanne 

Zenzinger for their help with in situ measurements. 

Several international cooperations and contacts helped advance my work. Thanks go especially to the VALERI 

network and to Frederic Baret and Marie Weiss for practical support with field equipment and software issues, 

as well as theoretical considerations with respect to field measurements. Thanks also to CEOS-LPV and to Jeff 

Morisette, Sebastien Garrigues and Jaime Nickeson for recommendations and provision of ASTER data. 

At DFD and the University of Würzburg my thesis has benefited from various discussions with and the technical 

support of several colleagues and students: I would especially like to thank Christopher Conrad, Rolf Richter, 

Gunter Schorcht, Rene Colditz and Johannes Hetzel. I am also very grateful to Martin Bachmann, Christopher 

vi

vii 

Conrad, Wouter Dorigo, Susan Giegerich, Rosslynne Grandinger, Miriam Machwitz and Katharina Winterroth 

for proofreading and commenting on various parts of this thesis. Nadine Hauff was a great help in completing 

some of the figures, Conny Schödl with the layout. 

Finally, this dissertation would not have been possible without the support and understanding of the most 

important people in my life: my family. Especially my husband Marcus was not only patient when I spent more 

time with my computer than with him, but also provided me with help and encouragement when it was most 

needed. This thesis is dedicated to my parents who have taught me never to give up. 

Tanja Kraus München, July 2008

Table of Contents 

1 Introduction 1 

1.1 Status of research 3 

1.2 Thesis objectives 6 

2 The study areas 9 

2.1 Budongo Forest 10 

2.1.1 Physiogeographic aspects 10 

2.1.2 Vegetation characteristics 13 

2.1.3 Forest management 18 

2.2 Kakamega Forest 20 

2.2.1 Physiogeographic aspects 20 

2.2.2 Vegetation characteristics 22 

2.2.3 Forest management 25 

3 Theoretical background 27 

3.1 Definition of LAI 27 

3.2 Characteristics of LAI in tropical rain forests 28 

3.3 Ground-based measurements of LAI 29 

3.3.1 Methods 31 

3.3.2 Instruments 34 

3.3.3 Limitations of indirect optical methods 37 

3.3.4 Spatial sampling strategies 40 

3.4 Satellite-based derivation of LAI 41 

3.4.1 Empirical approaches 42 

3.4.2 Physical models 43 

3.4.3 Spatial and spectral scale of sensors 45 

3.4.4 Suitability and limitations of empirical methods 46 

3.4.5 Validation approaches 51 

4 Remote sensing data and preprocessing 55 

4.1 Available high resolution satellite data 55 

4.1.1 ASTER 55 

4.1.2 SPOT-HRVIR 56 

4.2 Preprocessing of high resolution satellite data 56 

4.2.1 Geometric correction 56 

4.2.2 Atmospheric correction 57 

4.3 The MODIS LAI product 58 

4.3.1 The MODIS instrument 59 

4.3.2 Data set characteristics 60 

viii

ix 

4.3.3 Algorithm description 61 

4.3.4 Algorithm refinements in C5 63 

4.3.5 Validation status of the MODIS LAI product 64 

5 Ground-based LAI measurements 67 

5.1 Spatial sampling strategy 67 

5.1.1 Modifications made to the CEOS-LPV/VALERI methodology 67 

5.1.2 Field campaigns 69 

5.1.3 Measurement set-up 71 

5.2 Processing, correction and error analysis 73 

5.2.1 LAI-2000 PCA 73 

5.2.2 DHP 80 

5.2.3 Comparison between LAI-2000 PCA and DHPs 85 

5.3 Results of in situ measurements 87 

5.3.1 Budongo Forest 88 

5.3.2 Kakamega Forest 93 

5.4 Conclusion 98 

6 Derivation of high resolution LAI maps 101 

6.1 Calculation of spectral vegetation indices and texture measures 101 

6.2 Establishment of transfer functions 101 

6.2.1 Quantification of measurement errors 102 

6.2.2 Regression analysis 107 

6.3 Results 109 




7 Validation of the MODIS LAI product 127 

7.1 Upscaling of the high resolution LAI e map 127 

7.2 Correction for foliage clumping 127 

7.3 Results 128 




8 Discussion and Outlook 145 

References 151 

Appendix of Figures 173 

Appendix of Tables 181 

Appendix of Equations 191 

Curriculum Vitae 192

Index of Figures 

Figure 1-1 Validation sites of CEOS LPV 5 

Figure 2-1 Map of East Africa with the study sites in Kenya and Uganda 9 

Figure 2-2 Budongo Forest 10 

Figure 2-3 Climatic characteristics of Budongo Forest 12 

Figure 2-4 Assumed movement of forest species after 10,000 B.P. across Uganda 13 

Figure 2-5 Present distribution of forest in Uganda and estimated distribution of forest before clearance 13 

Figure 2-6 Forest types in Budongo Forest in 1990 16 

Figure 2-7 Vegetation types in Budongo Forest 16 

Figure 2-8 Characteristic features of trees belonging to the upper tree layer 17 

Figure 2-9 Logging compartments in Budongo Forest 19 

Figure 2-10 Illegal logging activities in the N15 nature reserve 19 

Figure 2-11 Kakamega Forest and its associated forest areas 20 

Figure 2-12 Surroundings of Kakamega Forest 21 

Figure 2-13 Climatic characteristics of Kakamega Forest 22 

Figure 2-14 Vegetation in Kakamega Forest 23 

Figure 2-15 Illegal activities in Kakamega Forest 26 

Figure 3-1 Vertical LAI variation in a semi-evergreen tropical rain forest in Panama 29 

Figure 3-2 LI-COR LAI-2000 PCA 35 

Figure 3-3 Illustration of the different viewing angles of the LAI-2000 PCA 36 

Figure 3-4 Digital camera with fish-eye lens 36 

Figure 3-5 Hemispherical photograph taken in Budongo Forest, Nature Reserve on 11 October, 2005 36 

Figure 3-6 Effect of foliage clustering on gap fraction 38 

Figure 3-7 Sampling schemes 41 

Figure 3-8 Reflectance from one to six layers of cotton leaves 42 

Figure 3-9 Simulation of vegetation leaf canopy as onedimensional turbid medium 44 

Figure 3-10 Simulation of a 3D vegetation canopy with a Monte Carlo Ray Tracing model 45 

Figure 3-11 Relationship of NDVI versus SR derived from ASTER data 48 

Figure 3-12 Structural differences in primary rain forest and selectively logged forest 51 

Figure 3-13 Schematic illustration of validation strategy 52 

Figure 3-14 VALERI validation scheme of medium resolution satellite data 53 

Figure 4-1 MODIS tiles for processing levels 2G, 3 and 4 60 

Figure 5-1 Location of ESUs in Kakamega Forest 69 

Figure 5-2 Location of ESUs in Budongo Forest 70 

Figure 5-3 Modified sampling scheme 71 

Figure 5-4 Transects on ESUs in Kakamega Forest and Budongo Forest 71 

Figure 5-5 Schematic illustration of measurement set-up with three LAI-2000 PCA devices 72 

Figure 5-6 Measurement set-up of LAI-2000 PCA 73 

Figure 5-7 LAI e (LAI2000) calculated from rings 1-4 versus LAI e (LAI2000) calculated from ring 4 only 74 

Figure 5-8 Test measurements with LAI-2000 PCA 75 

x

xi 

Figure 5-9 Hemispherical photograph 75 

Figure 5-10 LAI e (LAI2000) calculated from continuous B1 readings on ESU 7 in Budongo Forest 77 

Figure 5-11 LAI e (LAI2000) derived from continuous B1 measurements on ESUs 16, 18 and 30 78 

Figure 5-12 Scatter plot of LAI e (LAI2000, cen) and θ sun and correction for e θsun 79 

Figure 5-13 Optical centre of Nikon Coolpix 4300 camera used in Budongo Forest 80 

Figure 5-14 Illustration of DHP area used for LAI derivation 81 

Figure 5-15 Subset of hemispherical photograph as original and classified image 81 

Figure 5-16 Variation of the G-function with the average leaf inclination angle θ 83 

Figure 5-17 Equidistant projection 83 

Figure 5-18 Illustration of calculated projection function for NIKON coolpix 4300 and FC-E8 fisheye 84 

Figure 5-19 LAI e (DHP) derived from classification of gap fraction by two different operators 85 

Figure 5-20 LAI e (DHP) plotted against θ sun 85 

Figure 5-21 Illustration of field of view of LAI-2000 PCA with 45° view cap and DHP 86 

Figure 5-22 Boxplots showing light variables (P 0 and τ 0 ) 87 

Figure 5-23 LAI e (LAI2000) plotted against LAI e (DHP) 87 

Figure 5-24 Thinned upper canopy and illegal logging activities on ESU 28 in Budongo Forest 88 

Figure 5-25 Comparison of percentage of forest stages of Budongo Forest and of sampled ESUs 89 

Figure 5-26 Frequency distribution of LAI e (LAI2000) on sample level 89 

Figure 5-27 ESU 21 and ESU 15 in Budongo Forest 89 

Figure 5-28 Frequency distributions of LAI e (LAI2000) , LAI e (DHP) and LAI true (DHP) on ESU level 90 

Figure 5-29 DHP of understorey on ESU 4 and ESU 28 91 

Figure 5-30 Mean LAI per forest stage in Budongo Forest 91 

Figure 5-31 ESU 19 in Kakamega Forest 93 

Figure 5-32 Comparison of percentage of forest stages of Kakamega Forest and sampled ESUs 94 

Figure 5-33 Frequency distribution of LAI e (LAI2000) on sample level in Kakamega Forest 95 

Figure 5-34 ESU 25 and ESU 27 in Kakamega Forest 95 

Figure 5-35 Frequency distribution of LAI e (LAI2000) , LAI e (DHP) and LAI (DHP) on ESU level 96 

Figure 6-1 Illustration of bias, precision and accuracy 103 

Figure 6-2 Relation between c v and LAI e (2000) for all ESUs in Budongo Forest 105 

Figure 6-3 Derivation of reference values for regression analysis 108 

Figure 6-4 Bivariate plot of LAI e (DHP) with ρ red derived from ASTER 110 

Figure 6-5 Bivariate plot of LAI e (DHP) and SR derived from ASTER data 110 

Figure 6-6 Bivariate plots of LAI e (DHP) and SR (ASTER) and LAI (DHP) and NDVI c (SPOT) 111 

Figure 6-7 Bivariate plots of LAI e (DHP) and NDVI (ASTER) and both variables in log transformation 112 

Figure 6-8 Bivariate plot of LAI e (DHP) and variance (band 4, 9x9 kernel) derived from ASTER 113 

Figure 6-9 Relationship between observed LAI e (DHP) and mean LAI e (DHP) values over associated groups 

and observed surface reflectances and mean values over associated groups in the red 

and NIR spectral bands of ASTER, taking into account observation errors in field LAI 114 

Figure 6-10 Relationship between observed surface reflectance and mean values over associated groups 

in the red and NIR spectral bands of ASTER and observed LAI e (DHP) and mean LAI e (DHP) 

values over associated groups, taking into account observation errors in ASTER data 115

Figure 6-11 Relationship between observed LAI e (DHP) and mean LAI e (DHP) values over associated groups 

and observed surface reflectance and mean values over associated groups in the red 

and NIR spectral bands of ASTER, taking into account observation errors in both 

field LAI and ASTER data 116 

Figure 6-12 Bivariate plots of log transformed LAI e (DHP) and NDVI and LAI e (DHP) and SR 116 

Figure 6-13 Bivariate plot of LAI e (DHP) and SR 117 

Figure 6-14 Relation between LAI e estimated from ASTER data and in situ measured LAI 118 

Figure 6-15 High resolution LAI e map of Budongo Forest estimated from ASTER data 119 

Figure 6-16 Focus areas of Budongo Forest displayed in Figure 6-15. 120 

Figure 6-17 Bivariate plot of LAI e (LAI2000) and RSR derived from SPOT 121 

Figure 6-18 Bivariate plot of LAI e (LAI2000) and RSR derived from SPOT for Kakamega Forest 122 

Figure 6-19 Bivariate plot of LAI e (DHP) and homogeneity (band 4, 7x7 kernel) derived from SPOT 122 

Figure 6-20 High resolution LAI e map of Kakamega Forest modelled from SPOT data. 123 

Figure 6-21 Focus areas of Kakamega Forest displayed in Figure 6-22 124 

Figure 7-1 Quality information on cloud cover etc. of the MODIS data sets JD 273-353 2005 129 

Figure 7-2 Percentage of invalid (i.e. cloud contaminated) pixels for JD 273 to 353 in 2005 129 

Figure 7-3 Type 3 of MOD12Q1 data sets 2001-2004 130 

Figure 7-4 Quality information on the applied MODIS LAI algorithm for JD 289 and 329 131 

Figure 7-5 (Effective) LAI maps for the Budongo Forest test site 132 

Figure 7-6 Frequency distribution of (effective) LAI values derived from maps of Figure 7-6 133 

Figure 7-7 Regression between ASTER LAI e (1000 m) and C5 MODIS LAI product 134 

Figure 7-8 Quality information on the applied MODIS LAI algorithm (C4 data) for JD 329 135 

Figure 7-9 Relative frequencies of C5 MODIS LAI algorithm usage for the years 2000-2005 136 

Figure 7-10 Mean MODIS LAI retrieved for the centre of the Budongo Forest test site 137 

Figure 7-11 Annual MODIS LAI trajectories for the different forest stages of Budongo Forest 138 

Figure 7-12 Quality information of the MOD15A2 data sets JD 281-361 2004 138 

Figure 7-13 Percentage of invalid (i.e. cloud contaminated) pixels for JD 281 to 361 in 2004 139 

Figure 7-14 Type 3 of MODIS land cover of 2004 and MODIS LAI algorithm applied to data set JD 281 139 

Figure 7-15 (Effective) LAI maps for the Kakamega Forest test site 140 

Figure 7-16 Frequency distribution of (effective) LAI values 141 

Figure 7-17 Regression between SPOT LAI e (1000 m) and the C5 MODIS LAI product 

and SPOT LAI (1000 m) and the MODIS LAI product 142 

Figure 7-18 Annual MODIS LAI trajectory for the northern part of Kakamega Forest 143 

xii

xiii 

Index of Tables 

Table 2-1 Formations of tropical moist forests 14 

Table 3-1 Review of LAI values reported for tropical rain forests in the literature 30 

Table 3-2 Nominal angular coverage of the different rings of LAI-2000 PCA 35 

Table 3-3 cos θ and W i values for LAI-2000 PCA rings 1-5 and rings 1-4 35 

Table 3-4 Optical remote sensing data types, mainly used LAI derivation methods 47 

Table 3-5 Overview of most important SVIs with respect to LAI derivation for tropical rain forests 49 

Table 4-1 Spectral and spatial properties of the ASTER VNIR and SWIR instruments 55 

Table 4-2 Spectral and spatial properties of SPOT-4 HRVIR 56 

Table 4-3 Reference data for geometric correction 57 

Table 4-4 Acquisition and preprocessing information for the high resolution satellite data 58 

Table 4-5 Spectral and spatial properties of the first seven bands of MODIS 59 

Table 4-6 MOD15A2 data set characteristics 61 

Table 4-7 Biome classes in MOD12 land cover product 62 

Table 4-8 Canopy structural attributes of the six biomes used in the C4 LAI algorithm 63 

Table 4-9 Overview on validation studies of the MODIS LAI product (C4) 66 

Table 5-1 Methodology proposed by VALERI for in situ sampling and modifications made 68 

Table 5-2 Statistical measures for light intensity recorded by rings 1-4 of LAI-2000 PCA 74 

Table 5-3 Descriptive statistics of continuous LAI-2000 PCA measurements 76 

Table 5-4 Allocation of ESUs to the respective radiation regimes and necessary processing steps 76 

Table 5-5 Descriptive statistics for LAI e (LAI2000) derived from continuous B1 measurements 78 

Table 5-6 Mean and standard deviation of LAI e (DHP) per ESU processed by two different operators 84 

Table 5-7 Spearman Rho correlation matrix between τ 0 (LAI-2000 PCA) and P 0 (DHP) 86 

Table 5-8 Forest stage, logging status, NFA zone, comp. no. and LAI measures of 30 ESUs 

in Budongo Forest 92 

Table 5-9 Differences in LAI sampling between the first and the second field campaign 93 

Table 5-10 Forest stage, logging status, NFA zone, compartment number and LAI measures 

of the 30 ESUs in Kakamega Forest 97 

Table 5-11 Comparison between the properties of LAI-2000 PCA and DHP 99 

Table 6-1 Second-order texture measures used in this thesis 102 

Table 6-2 Descriptive statistics of LAI e (LAI2000) per ESU derived from B1 measurements 104 

Table 6-3 Descriptive statistics for LAI e (DHP) derived from DHP time series 106 

Table 6-4 Relative precision α for red, NIR and SWIR spectral bands 107 

Table 6-5 Theil-Sen and OLS regression models based on LAI e (DHP) and SVIs for early 

and intermediate forest stages in Budongo Forest 111 

Table 6-6 Theil-Sen and linear OLS regression models based on LAI variables and 

texture measures for late forest stages in Budongo Forest 113 

Table 6-7 Relative bias, precision and accuracies of relationships between LAI e (DHP) 

and surface reflectance derived from ASTER data 117 

Table 6-8 Theil-Sen and linear OLS regression models based on LAI e (DHP) and SVIs 117

Table 6-9 Theil-Sen and linear OLS regression models based on LAI e (LAI2000) and SVIs 

for early and intermediate forest stages in Kakamega Forest 122 

Table 6-10 Theil-Sen and linear OLS regression models based on LAI variables and texture 

measures for late forest stages in Kakamega Forest 122 

Table 7-1 Comparison of λ derived from DHP with λ given in Chen et al. (2005) 128 

Table 7-2 Mean (and standard deviation) of (effective) LAI derived from ASTER 

and MODIS for different forest stages 134 

Table 7-3 Relative frequency of C4 and C5 LAI algorithm usage for the Budongo Forest test site 135 

Table 7-4 Mean (and standard deviation) of (effective) LAI derived from SPOT 

and MODIS for the northern part of Kakamega Forest 142 

xiv

xv 

Glossary 

6S Second Simulation of Satellite Signal in the Solar Spectrum 

AATSR Advanced Along Track Scanning Radiometer 

ALA Average Leaf Inclination Angle 

ANN Artificial Neural Network 

ASTER Advanced Spaceborne Thermal Emission and Reflection Radiometer 

AVHRR Advanced Very High Resolution Radiometer 

BRF Bi-directional Reflectance Factors 

C4 Collection 4 (version of MODIS products and respective algorithms) 

C5 Collection 5 (version of MODIS products and respective algorithms) 

CAR Central African Republic 

CCD Charge Coupled Device 

CCRS Canada Centre for Remote Sensing 

CEOS Committee on Earth Observing Satellites 

CNES Centre National d’Etudes Spatiales 

COI Circle of Interest 

CYCLOPES Carbon Cycle and Change in Land Observational Products from an Ensemble of Satellites 

DAAC Distributed Active Archive Center 

DART Discrete Anisotropic Radiative Transfer 

DEM Digital Elevation Model 

DFD Deutsches Fernerkundungsdatenzentrum (German Remote Sensing Data Center) 

DHP Digital Hemispherical Photograph 

DLR Deutsches Zentrum für Luft- und Raumfahrt e. V. (German Aerospace Center) 

DRC Democratic Republic of the Congo 

EC European Commission 

EDC EROS Data Center 

EDOS EOS Data and Operations System 

EOS Earth Observing System 

ERS Earth Resources Satellite 

EROS Center for Earth Resources Observation and Science 

ESA European Space Agency 

ESU Elementary Sampling Unit 

ETM+ Enhanced Thematic Mapper Plus 

EVI Enhanced Vegetation Index 

FAO Food and Agricultural Organization 

FPAR Fraction of Photosynthetically Active Radiation absorbed by vegetation 

FRA Global Forest Resources Assessment 

GCP Ground Control Point 

GLCM Grey Level Co-Occurrence Matrix 

GPP Gross Primary Production

GPS Global Positioning System 

GSFC Goddard Space Flight Center 

HDF Hierarchical Data Format 

IDL Interactive Data Language 

IFOV Instantaneous Field Of View 

INRA Institut National de la Recherche Agronomique 

IMF Institut für Methodik der Fernerkundung (Remote Sensing Technology Institute) 

IPCC Intergovernmental Panel of Climate Change 

ITCZ Intertropical Convergence Zone 

JD Julian Day 

KFS Kenya Forest Service 

KWS Kenyan Wildlife Service 

LAI Leaf Area Index 

LAI-2000 PCA LAI-2000 Plant Canopy Analyzer 

LP DAAC Land Processes Distributed Active Archive Center 

LPV Land Product Validation 

LUT Look-Up-Table 

MERIS Medium Resolution Image Spectrometer Instrument 

MODAPS MODIS Adaptive Processing System 

MODIS Moderate Resolution Imaging Spectroradiometer 

MODTRAN Moderate Resolution Atmospheric Transmission 

MSR Modified Simple Ratio 

NASA National Aeronautics and Space Administration 

NDMI Normalized Difference Moisture Index 

NDVI Normalized Difference Vegetation Index 

NDVI c 

Corrected Normalized Difference Vegetation Index 

NFA National Forestry Authority of Uganda 

NIR Near Infrared (part of the electromagnetic spectrum between 0.78 and 1.40 µm) 

NOAA National Oceanic and Atmospheric Administration 

NSIDC National Snow and Ice Data Center 

OASIS Optimising access to SPOT infrastructure for Science 

OLS Ordinary Least Squares 

PAI Plant Area Index 

PAR Photosysthetically Active Radiation 

POLDER Polarization and Directionality of the Earth’s Reflectances 

(view imaging radiometer developed by CNES) 

PROSPECT Leaf Optical Properties Spectra 

REDD Reducing Emissions from Deforestation and Degradation 

RMA Reduced Major Axis 

RMSE Root Mean Square Error 

RSR Reduced Simple Ratio 

SAIL Scattering by Arbitrarily Inclined Leaves 

xvi

xvii 

SDS Scientific Data Set (MODIS) 

SLA Specific Leaf Area 

SPOT Satellite Pour l’Observation de la Terre 

SR Simple Ratio 

SRTM Shuttle Radar Topography Mission- 

SVI Spectral Vegetation Index 

SWIR Shortwave Infrared (part of the electromagnetic spectrum between 1.4 and 3.0 µm) 

TiSeG Time Series Generator 

TIR Thernal Infrared (part of the electromagnetic spectrum between 7.0 and 15.0 µm) 

TM Thematic Mapper (sensor carried on Landsat 3, 4, and 5) 

UNECSO United Nations Educational, Scientific and Cultural Organization, Organisation 

UNFCCC United Nations Framework Convention of Climate Change 

USGS United States Geological Survey 

VALERI Validation of Land European Remote Sensing Instruments 

VGT VEGETATION (sensor carried on SPOT-4) 

VIS Visible (part of the electromagnetic spectrum between 0.38 and 0.78 µm) 

WAI Woody Area Index 

WGCV Working Group on Calibration and Validation 

ZEF Zentrum für Entwicklungsforschung

xviii

1 

Introduction 

Tropical rain forests are the most important habitat 

type for biodiversity conservation worldwide (Myers 

et al. 2000). Although they only cover about 7% of 

the global land surface (Hansen & DeFries 2004), 

they shelter a large variety of life. At least 44% of 

the world’s vascular plants and 35% of terrestrial 

vertebrate species are endemic to 25 global 

biodiversity hotspots, 15 of which are tropical rain 

forests (Brooks et al. 2002, Myers et al. 2000). 

Additionally, tropical rain forests are playing an 

important role in the world’s carbon cycle. Vast 

amounts of carbon are stored in their vegetation 

and soils and are processed in photosynthesis 

and respiration processes (Clark 2007). Although 

the net-carbon balance of mature forests is still 

the subject of various scientific discussions, there 

is no doubt that carbon is released if degradation 

and deforestation processes take place, thus adding 

to the concentration of atmospheric greenhouse gases 

(Gullison et al. 2007). 

In the context of current global changes pan-tropical 

deforestation has already destroyed between 8 and 

12 million km² (i.e. 35-50%) of the original primary 

forest cover (Wright & Muller-Landau 2006). 

According to the latest global Forest Resources 

Assessment (FRA) of the Food and Agricultural 

Organization (FAO) the annual net loss in tropical 

forest area during the years 1990-2005 was highest 

in Africa with approximately 43,000 km²/a (FAO 

2006). Whereas in Europe net forest cover is actually 

increasing, growing human population numbers in 

Africa and a continuing high amount of people living 

in rural areas are predicted to maintain high levels of 

net deforestation on the African continent through 

2030 (Wright & Muller-Landau 2006). Additionally, 

an estimated two-thirds of the remaining forest cover 

in the tropics show clear signs of human disturbance 

and degradation (FAO 2006). Habitat loss, habitat 

modification and fragmentation lead to altered 

vegetation structure and species composition and, 

as a result, disturbed natural processes (such as seed 

dispersal or pollination), lower species abundance 

and extinction (Cuarón 2000). Further, the possible 

carbon sink capacity is reduced and additional 

greenhouse gases are emitted. Gullison et al. (2007) 

state in this context that deforestation and degradation 

processes contributed to almost 20% of anthropogenic 

greenhouse gas emissions during the 1990s. 

Rising rates of atmospheric CO 2 and temperature 

will in turn affect various ecosystem processes in 

tropical rain forests, for example, photosynthetic 

rates, net carbon balance, or carbon sequestration. 

In turn, declining productivity or even forest die-off 

could enforce regional and global climate change 

as precipitation patterns, surface temperature 

and carbon uptake are modified (Clark 2007, 

Jin & Zhang 2002). Several studies have also 

revealed latitudinal or altitudinal shifts of species 

ranges due to changing climatic conditions in the 

last 30 years (e.g. Parmesan & Yohe 2003, Root 

et al. 2003). According to the latest report of the 

Intergovernmental Panel on Climate Change (IPCC) 

increasing global mean temperatures are likely to 

aggravate this effect with the risk of ecosystem 

disruption and species extinction (Schneider et al. 

2007). 

1

2 

Within this context, the prospects of jointly addressing 

concerns about climate change, biodiversity loss and 

poverty by Reducing Emissions from Deforestation 

and Degradation (REDD) have attracted growing 

attention from the international environment and 

development communities (Gullison et al. 2007). 

Especially tropical countries that are severely 

affected by deforestation (e.g. Costa Rica and Papua 

New Guinea), but often lack the financial power to 

stop ongoing processes, have submitted proposals 

to the United Nations Framework Convention on 

Climate Change (UNFCCC) on how to share the 

responsibility for these processes among nations. 

The objective is to reduce emissions by the 

implementation of trading mechanisms, as 

compensation for protection of existing forests and 

funding for afforestation and reforestation. While 

the political implementation of the so-called REDD 

mechanism is still subject to various discussions, 

there are also methodological challenges that need 

to be addressed. They mainly relate to the definition 

of a baseline to which future (and past) deforestation 

rates can be compared, as well as to the establishment 

of reliable monitoring systems for both deforestation 

and degradation (Ernsting & Rughani 2007). 

In this context, whenever surveying, monitoring and 

modelling issues are to be addressed on a regional to 

global scale, operational remote sensing techniques 

form a valuable data source for the scientific (and 

political) community since such techniques permit 

repetitive and synoptic observations of vegetation 

cover. 

Subtle changes in forest structure and dynamics 

can, for instance, be monitored through repeated 

measurement of remotely derived biophysical 

attributes. In contrast to discrete representations 

of land cover and land cover change, biophysical 

variables alter continuously over space and time. They 

may thereby reveal starting ecosystem modifications 

that do not necessarily lead to a land cover conversion 

in the observed time frame (Lambin 1999). 

One of the key biophysical variables in this context 

is the leaf area index (LAI), which refers to the onesided 

foliage area per unit ground area. It characterises 

canopy structure, quantifies the size of canopyatmosphere 

interface and accounts for differences in 

phenological development, assimilation and biomass 

growth among plant species (Weiss et al. 2004, 

Wilson et al. 2007). As forest degradation also affects 

phenological cycles in semi-evergreen rain forests 

(Lambin 1999), operational standard products of LAI 

derived from satellite data are here a valuable data 

source. 

One of the most widely used operational standard 

LAI products is derived from optical remote sensing 

data acquired by the Moderate Resolution Imaging 

Spectroradiometer (MODIS) on board the Terra and 

Aqua platforms. The MODIS LAI product provides 

global information on LAI at 1 km resolution based 

on a compositing period of 8 days (Knyazikhin et 

al. 1999). Modelled relationships between spectral 

reflectances of vegetation canopies and their 

structural characteristics are stored in a look-uptable 

(LUT) that is used to achieve inversion of the 

radiative transfer problem. If the main algorithm fails, 

a backup method based on the normalized difference 

vegetation index (NDVI) is used for LAI retrieval 

(Myneni et al. 2002). The product is distributed free of 

charge together with additional per pixel information 

on product quality by the National Aeronautics and 

Space Administration (NASA). 

Yet a key issue in working with operational satellite 

products is product accuracy. Though the MODIS 

LAI algorithm is constantly refined, various errors 

still persist. Comparison of the MODIS LAI product 

with comparable products derived from the POLDER, 

MERIS, and VGT sensors (for explanation see 

Glossary) revealed significant discrepancies (Baret et

1 Introduction 

al. 2006). An important step towards the evaluation 

of satellite product accuracy - and in consequence 

of the scientific results based on these products - is 

therefore the validation of algorithm outputs. 

1.1 

Status of research 

According to the Working Group on Calibration and 

Validation (WGCV) within the Committee on Earth 

Observing Satellites (CEOS), validation is defined 

as “the process of assessing by independent means 

the quality of the data products derived from the 

system outputs” (Justice et al. 2000, see also WGCV 

2007). Validation refers in this context to the detailed 

examination of satellite data products with respect to 

their accuracy in depicting actual surface conditions. 

Based on analytical comparison to independent 

reference data, products must be analyzed for 

different ecosystems, varying atmospheric conditions 

and temporal consistency. 

Independent reference data can only be gathered 

through the sampling of accurate and representative 

field measurements for a certain test site. These in situ 

methods for LAI assessment have been developed 

for more than 70 years. Early studies focused on 

the direct determination of LAI to calculate growth 

rates of crops and grasslands (Rozhnyatovsky 1954, 

Warren-Wilson 1959, Watson 1947), but soon novel 

methods were developed that also allowed semidirect 

and indirect LAI estimation for more complex 

vegetation types such as forests (e.g. Anderson 

1964, Ogawa et al. 1961, Rogers & Hinckley 1979). 

As direct and semi-direct methods are usually very 

elaborate in terms of time and labour, the focus has 

further shifted towards the development of new and 

faster indirect methods (Chen & Black 1991, Lang & 

Yuequin 1986, Nilson 1971, Welles 1990, Welles & 

Norman 1991). Today, direct and indirect approaches 

have been thoroughly tested for crops, grasslands 

as well as temperate and boreal forest ecosystems. 

Discussions on error sources have been based on 

theoretical and practical considerations (Bréda 2003, 

Fassnacht et al. 1994, Ferment et al. 2003, Weiss et 

al. 2004, among others). 

In tropical rain forests however the in situ assessment 

of LAI remains challenging due to various constraints. 

Difficulties with site accessibility, complexity of forest 

structure and canopy height need to be taken into 

account, as well as the fact that many conventional 

direct field methods, that were developed to estimate 

LAI in forest stands of higher latitudes, do not work 

in evergreen broadleaf forests owing to the high 

amount of foliage, species richness and the lack of 

seasonal litter fall (e.g. Bréda 2003, Dufrêne & Bréda 

1995). Optical methods that have been evolving 

since 1990 and are widely applied today, in turn, 

show constraints related to illumination conditions 

under which the measurements should be performed. 

Although some studies have been conducted with 

optical devices in rain forest ecosystems (e.g. Aragão 

et al. 2005 in Eastern Amazonia, De Wasseige et al. 

2003 in the Central African Republic [CAR], and 

Wirth et al. 2001 in Panama), the effects of non-ideal 

measurement conditions are not yet well understood. 

Parallel to the improvement of in situ based LAI 

estimation the derivation of LAI based on remote 

sensing data has also been subject to significant 

progress over the past 30 years. Especially through 

further development of the physical and mathematical 

foundation, techniques have evolved from simple 

empirical approaches to complex physically-based 

canopy reflectance models (Liang 2004). 

Empirical methods either rely on regression analyses 

or neural networks based on field measurements and 

canopy reflectance data. Recent examples for LAI 

estimation with optical remote sensing data can be 

found in case studies by Kalácska et al. (2004) for 

tropical rain forests (with Landsat ETM+), Fassnacht 

et al. (1997) and Jensen & Binford (2004) for 

3

4 

temperate forests (both with Landsat TM), or Chen 

& Cihlar (1996) and Cohen et al. (2003) for boreal 

forests (with Landsat TM and ETM+ respectively). 

Fernandes et al. (2003) produced Canada-wide LAI 

maps based on VGT data with a semi-empirical 

method, taking into account field data, vegetation 

indices and land use information. For global 

applications and operational processing empirical 

methods are, however, not suitable as the established 

relationships are very time and site specific. 

Constraints are thus put on the transferability of 

empirical approaches. However empirical approaches 

are still of importance when it comes to the upscaling 

of field data for validation purposes. 

By contrast, physical models, as e.g. applied in the 

MODIS LAI algorithm, use an inversion of canopy 

radiation models to estimate biophysical variables. 

Radiative transfer models describe the interaction 

between sunlight and vegetation elements within 

the canopy on a theoretical basis and are thus ideal 

for global applications. They can be run for given 

patterns of soil, canopy structure, view-illumination 

conditions etc. The major challenge lies here in 

solving the inverse problem, i.e. the derivation of 

biophysical variables (such as LAI) given canopy 

reflectance, as due to various error sources and a 

large number of possible combinations of canopy 

properties the inversion of a physical model does not 

have a unique solution (Knyazikhin et al. 1998a). 

Apart from the MODIS LAI algorithm, physical 

models have also been applied to VGT data in the 

European Commission (EC) funded CYCLOPES 

programme (Baret et al. 2007) as well as to combined 

VGT, AATSR and MERIS data in the European 

Space Agency’s GLOBCARBON project (Plummer 

et al. 2005). 

In order to validate the above-mentioned LAI 

products, CEOS established a subgroup on Land 

Product Validation (LPV) in 2000. The objective is 

to define standard guidelines and common protocols 

for validating satellite products characterising 

the land surface. A special effort has further been 

put into international collaboration in validation 

and intercomparison of land biophysical products 

(Morisette et al. 2006a). Several international 

programmes and other individual initiatives 

contributed validation exercises for LAI products 

derived from different medium spatial resolution 

sensors, e.g., BigFoot (funded by NASA’s Terrestrial 

Ecology Program), Validation of Land European 

Remote Sensing Instruments (VALERI, mainly 

supported by the French Centre National d’Etudes 

Spatiales, CNES, and the Institut National de la 

Recherche Agronomique, INRA), or the Canada 

Centre for Remote Sensing (CCRS). 

An overview of the CEOS-LPV test sites is given 

in Figure 1-1 (cf. Morisette et al. 2006a). Although 

further validation sites of other initiatives and 

networks do exist, the geographical site distribution 

is quite representative. Validation sites are mainly 

situated in Europe and North America with most 

studies being conducted in agricultural (e.g., Cohen 

et al. 2003, Pisek & Chen 2007, Tan et al. 2005) or 

forest areas (e.g., Abuelgasim et al. 2006, Cohen et 

al. 2003, Pisek & Chen 2007, Tian et al. 2002b, Wang 

et al. 2004). Only a few studies have so far dealt 

with test sites in Africa, and all of them validated 

the MODIS LAI product for savannah ecosystems 

(Privette et al. 2002, Tian et al. 2002a) in Southern 

Africa (not included in Figure 1-1). 

The studies mentioned above have contributed to the 

detection of anomalies in previous versions of the 

MODIS LAI product and quality assessment of LAI 

retrieval over different biomes. However there is a 

clear lack of adequate validation data over broadleaf 

evergreen forests (Baret et al. 2006), especially 

in Africa. So far, three validation efforts dealing 

with MODIS LAI of tropical rain forests have been 

published, all exhibiting major or minor difficulties 

with in situ LAI assessment and geolocation,


upscaling of in situ measurements or algorithm 

performance due to cloud cover. Whereas two of 

the studies were dealing with the same tropical rain 

forest site in Brazil (Aragão et al. 2006 and Cohen 

et al. 2006), the third was part of a larger nationwide 

assessment in Australia (Hill et al. 2006). Among 

the reported problems was a general overestimation 

for tropical forests and a general weak correlation 

between in situ data and the MODIS LAI product. 

It must be mentioned, however, that all three studies 

were dealing with a previous version of the MODIS 

LAI product. Improvements are supposed to be 

integrated into the current version. 

Although further field measurements were made in 

a tropical rain forest in Counami, French Guiana, in 

the context of VALERI (Baret & Rosello 2007), no 

further MODIS LAI validation has been published 

for this site. 

Figure 1-1 

An important part of validation is the establishment 

of standardised sampling schemes. Morisette et 

al. (2006a) give a good overview of the validation 

methodologies used by different LPV groups. 

Differences lie especially in the instruments used for 

field measurements and subsequent LAI calculation, 

sampling schemes, site extents and upscaling 

approaches. In order to establish common standards 

and achieve comparability of validation exercises, 

VALERI has focused on developing effective standard 

methodologies for field measurements and scaling to 

higher resolution remotely sensed imagery (Baret et 

al. submitted). High spatial resolution maps of LAI 

are generated from ancillary satellite data and used 

for the validation of moderate-resolution biophysical 

products, such as MODIS (Morisette et al. 2006a). 

So far most validation exercises have focused on the 

spatial domain, i.e., upscaling of in situ measurements 

Validation sites of CEOS-LPV. Classes represent bare soil, water bodies, deciduous broadleaf forest, evergreen 

needleleaf forest, evergreen broadleaf forest, crops, and grassland (Morisette et al. 2006a). 

5

6 

and comparing the results with MODIS LAI maps in 

order to assess algorithm correctness (e.g. Abuelgasim 

et al. 2006, Cohen et al. 2003, Pisek & Chen 2007, 

Tan et al. 2005, Tian et al. 2002a, Tian et al. 2002b, 

Wang et al. 2004). Only a few publications deal with 

the temporal dimension of the MODIS LAI product in 

terms of the correct retrieval of phenological patterns 

(Fensholt et al. 2004, Huemmrich et al. 2005, Kang 

et al. 2003, Wang et al. 2005). 

There are several reasons for focusing this thesis 

on East African rain forests. First of all, these 

ecosystems belong to a globally outstanding 

ecoregion with respect to their biological diversity. 

Although endemism rates are low, both flora and 

fauna are characterised by high levels of species 

richness (Burgess et al. 2004). However, the former 

mosaic of tropical forests mixed with savanna 

woodlands has been significantly modified. Severe 

deforestation has taken place during the last century. 

Most of the remaining forest areas are now protected, 

but nevertheless the diverse flora is imperilled by 

human disturbance and possible future climatic 

impact. Second, the close link to the VALERI project 

encouraged the establishment of rain forest study 

sites in Africa for the mentioned reasons. 

With respect to the above-described current state of 

research, the objectives of this thesis will now be 

outlined. 

1.2 

Thesis objectives 

As documented in the previous chapter, the MODIS 

LAI product holds a large amount of information on 

forest structural and phenological properties. It can 

thus contribute to monitoring and modelling purposes 

in the context of current scientific and political 

discussions. Nevertheless – and especially if financial 

support to developing countries is involved – it can 

only be applied if the product is reliable and accurate 

for different geographical regions and land surfaces. 

To date there is a clear lack of validation sites for 

the MODIS LAI product in tropical rain forests. 

Especially for African rain forests no information 

on the quality of the MODIS LAI product is thus far 

available. Further, the current version of the MODIS 

LAI product has not been validated yet for tropical 

rain forests. 

The main objective of this thesis is therefore the 

validation of the MODIS LAI product for rain 

forest ecosystems in East Africa based on ground 

measurements of LAI. For that reason adequate in 

situ sampling strategies are necessary. As explained 

earlier, the available optical measurement techniques 

have not been thoroughly tested in tropical rain 

forests. Therefore a detailed comparison of available 

and suitable instruments as well as their advantages 

and constraints is required. In addition, an in-depth 

analysis of the results is planned in order to reveal 

errors introduced by non-ideal sampling conditions. 

With respect to spatial sampling strategies the 

recommendations proposed by the CEOS-LPV and 

VALERI networks will be reviewed with respect to 

their suitability for test sites in tropical rain forests. 

If necessary, the strategies will be adapted and 

optimized. 

Direct upscaling from in situ measurements to MODIS 

data bears multiple uncertainties because a) field 

samplings of LAI representing an adequate area would 

be very time intensive and b) smaller sampling sizes 

are unable to cover the entire heterogeneity within 

one MODIS pixel. Additionally geometry problems 

can lead to errors in the direct upscaling of in situ 

measurements. Therefore the intention is to integrate 

high resolution satellite data in order to accurately 

scale the ground-based measurements to MODIS 

resolution. If possible, measurement precision of both 

field and satellite data will be included to develop 

robust regression models that lead to the production 

of high resolution LAI maps with known accuracy.


In the last step, the MODIS LAI product will be 

compared with this high resolution reference data in 

a detailed analysis. Here the analysis of additional 

information on the quality of the MODIS LAI 

product will shed light on the performance of the 

LAI algorithm for tropical rain forests. Owing to the 

observation of seasonal variability of LAI due to the 

presence of some deciduous species in the two study 

sites, the temporal consistency of the MODIS LAI 

product will also be investigated. The generation and 

analysis of time series will help to reveal product 

smoothness over time and address the feasibility 

of using the MODIS LAI product as a proxy for 

structural differences in forest stages. 

In summary, the following research objectives will be 

addressed within this thesis 

1. Sound review of in situ methods of LAI 

assessment together with a judgement on 

transferability and applicability of these 

methods in tropical rain forests. 

2. Establishment of a representative and valid 

sampling scheme (based on CEOS-LPV 

and VALERI recommendations) for in situ 

measurements of LAI on the test sites. 

3. Performance of in situ LAI measurements and 

analysis of gathered field data (also with respect 

to different forest stages). 

4. Upscaling of field measurements based on high 

resolution satellite data and derivation of high 

resolution LAI maps for the test sites. 

5. Validation of the MODIS LAI product based on 

derived high resolution LAI maps. 

6. Check of temporal consistency of the of the 

MODIS LAI product. 

This research was carried out under the framework 

of the international and interdisciplinary research 

network BIOTA AFRICA, funded by the German 

Ministry of Education and Research (BMBF). As part 

of the regional initiative BIOTA East Africa, which 

focuses on biodiversity changes in Kenyan and 

Ugandan rain forests with different degradation and 

disturbance grades, this study offered the opportunity 

to get access to two of these forests for detailed 

field and validation campaigns. Close cooperation 

with the BIOTA West and South Africa projects 

led to comparable validation exercises in Namibia 

and Burkina Faso and will serve as comparison 

for the methodological aspects of this study in 

later publications. Additionally, field data and high 

resolution LAI maps will be made available within 

the VALERI and LPV networks for the validation of 

biophysical products derived from other sensors. 

7

2 

Figure 2-1 

The study areas 

Map of East Africa with the study sites in Kenya and Uganda. 

Research in the BIOTA East Africa project focuses 

on three tropical rain forests, namely Mabira and 

Budongo Forests in Uganda and Kakamega Forest 

in Kenya. Each study site shows different grades 

of anthropogenic interference, with Kakamega and 

Budongo Forests including areas that are considered 

relatively undisturbed (Mitchell & Schaab, in print). 

By contrast, not a single part of Mabira Forest 

can be identified as unlogged (Mitchell, personal 

communication). To cover as many different forest 

types with fieldwork as possible (with respect to 

stand age, logging history and canopy structure) 

field campaigns were thus conducted in Kakamega 

and Budongo Forests (cf. Figure 2-1). Both study 

sites will be described in the following text with 

respect to their physiogeographic aspects, vegetation 

characteristics and forest management. 

9

10 

2.1 

2.1.1 

Budongo Forest 

Physiogeographic aspects 

Budongo Forest is one of Uganda’s largest remaining 

rain forests, ranking third in species diversity 

nationwide (Howard et al. 1997, Reynolds 2005). It is 

situated in the northwest of the country in the Hoima 

and Masindi districts, close to the eastern shoreline of 

Lake Albert (cf. Figure 2-1). Including all remaining 

forest fragments, its geographic location ranges 

approximately from N 169,500 to N 218,800 and E 

317,100 to E 369,100 (Projection: UTM, Zone 36 N, 

Datum: WGS 84). 

Figure 2-2 

Figure 2-2 shows a subset of a Landsat ETM+ image 

acquired on 17 February, 2000. Budongo Forest lies 

in the centre with the forested areas appearing in 

bright green colours. Lake Albert is clearly visible 

on the western side, with the Albertine Rift stretching 

along its shoreline from southwest to northeast. As 

the scene was acquired at the end of the dry season 

and severe fires had occurred in the woodlands and 

grasslands of Bugungu Game Reserve, burnt areas 

are perceptible and represented by purple colours. 

According to Howard et al. (1996) the reserve covers 

an area of 793 km² and comprises four smaller 

blocks, namely Budongo itself (364 km²), Siba in 

Landsat ETM+ scene acquired on 17 February, 2000, showing Budongo Forest (band combination 5-4-3, Projection: 

UTM 36 N, WGS84). The protected areas of Budongo Forest Reserve (including the main forest blocks), Bugungu 

and Karuma Game Reserves as well as Murchinson Falls National Park are outlined in yellow (based on data of the 

National Forestry Authority, NFA, Uganda).

2 Study areas 

the southwest (66 km²), Kitigo in the northwest (95 

km²) and Kanyo-Pabidi in the northeast (268 km²). 

An estimated 47% of the reserve comprises partly 

wooded grassland communities (Howard 1991). They 

are mainly found in Kanyo-Pabidi and Kitigo and are 

indicated by light green and reddish tones that are 

clearly distinguishable from the lush green colours of 

the forested areas. 

The forest itself lies approximately between 870 m 

and 1170 m a.s.l. with the whole area gently sloping 

down towards the escarpment of the Albertine Rift 

in the northwest. It is separated by several partly 

ephemeral rivers draining either through Weiga, 

Waisoke, Sonso, or Bubwe Rivers towards Lake 

Albert. 

Budongo Forest is surrounded by populated areas 

in the south, southwest and southeast. Subsistence 

farming is widely spread and indicated by the smallscale 

mosaic of lighter and darker reddish tones. In 

the very south of Budongo Forest large sugarcane 

fields are visible in pink (harvested) and light green 

(standing crops); they belong to the second leading 

sugar producer in Uganda, Kinyara Sugar Works 

Limited. In the north and northeast, the forest 

reserve is contiguous with Bugungu Game Reserve 

and Murchinson Falls National Park. Towards 

the northeast Karuma Game Reserve extends the 

protected area. 

Geology and Soils 

The largest part of Budongo Forest lies on an 

Archean gneissic-granulitic basement complex 

that makes up large parts of northern, eastern 

and northwestern Uganda. It was created during 

several orogenic processes along the margin of 

the Congo and Tanzania Cratons thus creating a 

polyphase mobile belt (Schlüter 1997). Rocks in the 

Budongo region were affected by Watian (2.88 Ga, 

11 

granulite facies) and Aruan (2.25 Ga, amphibolite 

facies) tectono-metamorphic events and consist of 

highly metamorphic granulites, paragneisses and 

orthogneisses together with various schists, quarzites 

and marbles (Schlüter 1997). 

The southeastern part of Budongo Forest is occupied 

by the low metamorphic sedimentary rocks of the 

Bunyoro Series (see MacDonald 1966). The series 

is of late Precambrian age and rests unconformably 

on the crystalline basement complex, probably 

representing a shallow remnant of a formerly more 

extensive sedimentary formation (Bjørlykke 1973, 

MacDonald 1966). The sequence is described 

as “sandstones and pellitic sediments” with 

“conglomerates and pebbly mudstones in the lower 

parts” of glacial origin (Bjørlykke 1973). 

To the west, Budongo Forest stretches towards the 

above-mentioned escarpment of the Albertine Rift 

that is situated only about 10 kilometres away. It is 

the result of tectonic activities that started during 

the Miocene and continue today. The uplift of the 

rift walls led to a modification of regional drainage 

patterns, as the area of Uganda and western Kenya 

was draining towards the west into the Atlantic in 

the late Miocene-Pliocene, but later changed towards 

the north through the Nile River, thus influencing 

regional and interrift climates (Schlüter 1997). 

Eggeling (1947) states that formerly existing lateritic 

layers, having their origin in the peneplanation 

following the uplift of the rift shoulders, are mainly 

eroded today and only remain on a few hills 

surrounding the forest. The soils in the Budongo 

region are heavily weathered and mainly consist 

of ferralitic clay soils (Sheil 1999). According to 

Eggeling (1947) they vary from heavy loam or sandy 

clay to very sandy loams. Immediately below the 

humus horizon this soil is reported to be less clayey 

and sandier than at greater depth. Soils in the Budongo 

region are moderately fertile (Howard 1991).

12 

Climate 

As in most parts of Uganda, the climate of Budongo 

Forest is governed by the country’s latitudinal 

position, the resulting movements of the Intertropical 

Convergence Zone (ITCZ), and altitudinal and 

topographical characteristics. In consequence 

Budongo Forest’s precipitation pattern is bimodal 

with more or less pronounced dry seasons. 

Figure 2-3a shows the mean monthly rainfall as 

recorded at Sonso Camp between 1993 and 2004 

(source: Budongo Forest Project, courtesy of Geoffrey 

Muhanguzi). Though the observed time frame is 

short and climate stations with records over several 

decades do exist outside the forest (for instance at 

Busingiro or Nyabyeya, cited in Reynolds 2005), data 

from Sonso is preferred for the illustration below, as 

precipitation can vary significantly between the forest 

and surrounding areas. Unfortunately, data from 2004 

to 2007 was not available. 

The two rainfall peaks between March to May and 

September to November are clearly visible as are 

the dry seasons, which are experienced from mid- 

December to mid-February and less pronounced 

between June and July (see also Reynolds 2005). 

Figure 2-3 

However, it has to be noted that following an El-Niño- 

Southern-Oscillation event, December 1997 was 

exceptionally wet with the precipitation summing up 

to 520 mm. The mean value for that month as shown 

in Figure 2-3a was thus strongly influenced (mean 

monthly rainfall for December drops from 84 mm 

to 40 mm when excluding data from 1997). Yearly 

rainfall at Sonso varied between 1,240 mm and 

2,187 mm in the observed period, with a mean of 

1,642 mm (±232 mm). 

Temperature records from the Sonso site (same 

source) only exist for an even shorter period due to 

inconsistent data sets from April 1999 on. Figure 2-3b 

shows the mean monthly maximum and minimum 

temperatures from 1993 to 1999. As expected, mean 

maximum temperatures are highest during the dry 

season from December to February with more than 

30°C on average. The short dry season in June/July 

has in turn no effect on mean maximum temperatures. 

Mean minimum temperatures show little variance 

throughout the year, with slightly higher values 

during the rainy seasons. Mean monthly values 

range from 14.3–15.7°C minimum and 24.3–31.0°C 

maximum temperatures with an annual mean of 

21.1°C (±1.1°C). 

Climatic characteristics of Budongo Forest. a) Mean monthly rainfall at Sonso Camp from 1993-2004 and b) mean 

monthly temperatures at Sonso Camp from 1993-1999 (source: Budongo Forest Project).

2 Study areas 

2.1.2 

Vegetation characteristics 

Biogeographic position 

Apart from recent anthropogenic influences, climatic 

fluctuations since the Upper Quaternary have been 

a major driver for the distribution of forest species 

in East Africa (Hamilton 1981). Based on pollen 

diagrams and the analysis of floristic disjunctions, 

assumptions can be made about the former 

distribution of forested areas as well as changes in 

their extent during the late Pleistocene and Holocene. 

Especially water availability greatly influenced 

rain forest distribution. Whereas rain forests were 

restricted to a small number of refuges (e.g., in the 

eastern Democratic Republic of the Congo, DRC) 

during a very dry period before 12,000 B.P., they 

greatly increased during the following warmer and 

wetter stages (Hamilton 1981, Reynolds 2005). 

Figure 2-4 shows the assumed movement of forest 

species after 10,000 B.P. across the country with 

the former refuges in DRC and a probable minor 

Figure 2-4 

Assumed movement of forest species after 

10,000 B.P. across Uganda (Hamilton 1982, 

modified). 

13 

one in the western Lake Victoria region as starting 

points. It has to be pointed out that the extent of the 

former forest cover is still a question of scientific 

discussion. Whereas generic similarities as well as 

common species of coastal forests in East Africa and 

the forested regions in the DRC are evident, there 

is no proof yet whether both areas were linked by a 

vast continuous forest cover, as assumed by Schmidt 

(1992), that would have included both Budongo 

and Kakamega Forests. Alternatively, small isolated 

forests that expanded and maybe even connected 

during wetter periods and shrank again during dry 

conditions could have contributed to seed dispersal 

of forest species towards the East African coastline 

(Hamilton 1981, Lind & Morrison 1974). 

Lowland forest cover in East Africa probably reached 

its maximum in a relatively warm and wet period 

between 7,000 and 5,000 B.P. (Kendall 1969 and 

Reynolds 2005, derived from pollen analysis). After 

that time trees characteristic for semideciduous 

forests became more abundant, whereas pollen from 

trees found in both evergreen and semideciduous 

Figure 2-5 

Present distribution of forest and estimated 

distribution of forest before clearance (Hamilton 

1974, modified).

14 

forests decreased in numbers according to a study 

from Pilkington Bay, Lake Victoria (Kendall 1969). 

Today, one of the major factors affecting forest cover 

and species diversity is anthropogenic influence. 

According to Hamilton (1974) anthropogenic 

forest destruction began already around 1,000 B.P., 

but human pressure on remaining forest patches 

tremendously increased within the last 100 years. 

As a result, tropical forests in Uganda have suffered 

massive loss. Figure 2-5 shows the forest distribution 

in 1974 (dark green) in relation to the assumed former 

limits of forest cover based on distribution of forest 

remnants, precipitation and altitude (light green). 

According to Andrua (2003) Uganda’s tropical forests 

decreased from 12.7% of the land area in 1900 to 

3.7% in 2002. This corresponds to 8,847 km² of rain 

forest cover in the country. An alarming 32% of these 

forest areas are degraded. 

Vegetation classification and floristics 

The distinction of forest formations in tropical Africa 

is usually based on floristic elements as well as 

environmental factors such as altitude and moisture. 

With respect to floristic elements, White’s descriptive 

memoir on the UNESCO vegetation map of Africa 

(White 1983), classifies Budongo Forest as belonging 

to the Lake Victoria Regional Mosaic floristic region. 

This transition zone is influenced by the phytochoria 

of the Guineo-Congolian, Sudanian, Zambezian, 

Somaila-Masai and Afromontane floristic regions. 

In consequence only very few endemic species and 

probably no endemic genera are being recorded for 

this zone (White 1983). In Budongo Forest however, 

elements of the Guineo-Congolian phytochorion 

prevail. 

Whitmore (2003) further distinguishes five types 

of rain forest formations according to elevation and 

climate (cf. Table 2-1). As Budongo Forest is situated 

in a seasonally dry climate with less than 100 mm 

Table 2-1 

Formations of tropical moist forests 

(modified according to Whitmore 2003). 

Climate Elevation a.s.l. Forest formation 

Seasonally dry Semi-evergreen 

rain forest 

Perhumid < 1200 m Lowland evergreen 


Perhumid 1200-1500 m Lower montane 


Perhumid 1500-3000 m Upper montane 


Perhumid > 3000 m Subalpine forest 

to tree line 

of rainfall during some months, it is categorized as 

semi-evergreen rain forest. Langdale-Brown et al. 

(1964) mention, however, that the transition between 

evergreen and semi-evergreen rain forests is fluid and 

in consequence somewhat arbitrary. The latter shows 

a higher amount of leaf-shedding trees for a slightly 

longer period due to climatic conditions. It should be 

noted that in semi-evergreen rain forests only some 

species shed their leaves, and not necessarily all at 

the same time (White 1983), resulting in slightly 

lower LAI values during the dry seasons. 

With respect to floristics Budongo Forest is one of 

the best described forests in Uganda (Howard et al. 

1996). Since 1930 many researchers have conducted 

vegetation studies of different scopes. Among the most 

important publications are the often cited description 

of vegetation types from Eggeling (1947), Plumptre’s 

studies on the effects of selective logging on species 

distribution (Plumptre 1995, and 1996) and some 

more recent vegetation surveys that can partly be 

seen as a continuation of Eggeling’s work. A critical 

review of Eggeling’s hypotheses on succession was 

conducted by Sheil (e.g. 1995, 1998, and 1999). Apart 

from that, inventories were carried out on behalf of the 

Uganda Forest Department in 1986 (based on a threeday 

reconnaissance and aerial imagery, published

2 Study areas 

by Howard 1991) and, more thorough, in 1993 and 

1994 (based on 42 days of fieldwork and sampling 

1,053 km of transects throughout the forest, published 

in the Budongo Forest Reserve Biodiversity Report 

by Howard et al. 1996). As a result of the latter, 

465 tree and shrub species are now known to exist 

there, 93 of which are restricted-range species that 

only occur in five other forests in Uganda. In the 

following, the data published in the above-mentioned 

sources is summarized. 

According to Eggeling (1947) four forest types can 

be distinguished in Budongo Forest, three of which 

follow an ecological succession series. Apart from 

swamp forest, that forms an edaphic climax stadium 

and occupied only 2% of the total forest area at the 

time of Eggeling’s survey, mainly colonising forest, 

mixed forest and a so-called ironwood forest can be 

distinguished. Colonising forest is the youngest stage, 

occurring mainly at the forest edge and covering 6% 

of the forested land. It is described as moderately 

species rich and comprises Maesopsis eminii stands 

on deeper and better soils and a species mixture of 

Olea welwitschii, Sapium ellipticum and Phyllanthus 

discoideus on shallow poor soils. Most of these 

species are light loving and therefore replaced very 

quickly by a mixed forest with high species richness, 

including several mahogany and sapotaceous species 

(Sheil 1999). Eggeling (1947) estimated that mixed 

forest covered the largest part of the forest reserve, 

approximately 60% of the area. The most abundant 

species include Khaya anthoteca, Entandrophragma 

spp. and Cynometra alexandri. Mixed forest species 

regenerate in gaps, but gradually and over a long time 

period only the more shade-tolerant trees survive 

and ironwood forest develops, with Cynometra 

alexandri being the dominant species in the canopy 

and Lasiodiscus mildbraedii in the understorey. This 

is also supported by Synott (personal communication 

in Reynolds 2005), who discovered that Cynometra 

spp. are less often eaten by rodents and grow more 

successfully in shady conditions. According to 

15 

Eggeling (1947), ironwood forest covered around 

32% of Budongo Forest at the time of his survey. 

Whereas the above description of colonising forest 

has found wide acceptance, Eggeling’s assumptions 

of later succession stages have prompted controversial 

discussion (Langdale-Brown et al. 1964, Lieth & 

Werger 1989, Sheil 1999). Especially his theory 

about late-successional species decline, resulting in 

the dominance of Cynometra alexandri in the upper 

tree layer, has been a subject of ongoing discussion. 

Whereas Langdale-Brown et al. (1964) agree in 

most parts with Eggeling’s description, they also 

mention that Cynometra forest might only represent 

the climax stage on poor soils, relying on evidence 

from observations in Semliki and Bugoma Forests. In 

their opinion mixed forest can also develop to Celtisor 

Chrysophyllum-dominated types. This theory is 

in line with Sheil (1999), who evaluated Eggeling’s 

successional interpretation and found inconsistencies 

of the data with time series collected on Eggeling’s 

original plots for the following 60 years. Sheil (1999) 

proposes a possible change of successional processes 

as the reason. 

It is indisputable that since Eggeling’s study the forest 

and its vegetation composition has been influenced 

strongly by anthropogenic processes. Many parts of 

the forest have been disturbed by former or present 

legal or illegal logging activities, thus annulling 

any theory of climax vegetation (though there are 

protected areas that are still pretty well intact). A very 

detailed study of Plumptre (1996) based on aerial 

imagery of 1951 and 1990, forest inventory data 

and fieldwork showed a decrease in Cynometra and 

Cynometra-mixed forest of more than 205 km² during 

the above-mentioned time span. Mixed forest in turn 

increased more than 350 km² also at the expense 

of colonizing forest. The state of forest types in 

Budongo Forest based on Plumptre’s work is shown 

in Figure 2-6. Cynometra-mixed forest is pictured in 

Figure 2-7a.

16 

Figure 2-6 

Forest types in Budongo Forest in 1990 (Plumptre 1996). 

Outside the forest reserve natural vegetation cover 

almost disappeared due to uninhibited exploitation of 

natural resources by the local population. The protected 

regions in the northwest towards Kanyo-Pabidi and 

along the escarpment line in the north and northeast, 

however, still support moist Combretum-Savanna 

(Langdale-Brown et al. 1964) with a grass layer up 

to 1.5 to 2 m thick and characterized by Hyparrhenia 

rufa (cf. Figure 2-7b). The light to moderate tree 

cover consists mainly of deciduous trees of the 

Combretaceae family with abundant Combretum 

molle, Terminalia glaucescens and Albizia zygia 3 to 

12 m in height. Patches of dry Combretum savannas 

also exist with varying combinations of Combretum- 

Terminalia-Loudetia and Combretum-Hyparrhenia 

types (Langdale-Brown et al. 1964). 

Figure 2-7 

Vegetation types in Budongo Forest. 

a) Cynometra-mixed and b) moist Combretum- 

Savanna close to Kanyo-Pabidi.

2 Study areas 

Vegetation structure 

Structural differences in rain forests are mainly 

associated with different succession or disturbance 

stages. According to FAO definitions (FAO 2006), 

primary forests are forests that have never been 

logged and have developed under natural processes 

and following natural disturbances. They are thus 

the end stage of natural succession processes. 

As already mentioned, primary forest can hardly 

be found in Budongo Forest. Solely areas that 

were under protection since the 1930s and only 

experienced selective illegal logging activities can 

still be regarded as near primary forest. Most other 

parts are rather characterized as secondary forests 

that according to FAO (2006) have been logged and 

recovered or have lost, through human disturbance, 

the structure, function, species composition or 

productivity normally associated with primary 

forests. Primary and secondary rain forests differ 

significantly in vegetation composition and structure. 

In order to describe the latter, a strata or canopy layer 

concept is usually applied. Though this is usually 

a simplification of reality, it may be a useful aid to 

description or analysis (Whitmore 2003). 

For Budongo Forest Eggeling (1947) proposed a three 

layer concept for undisturbed parts which is probably 

applicable to many other lowland and medium 

Figure 2-8 

Characteristic features of trees belonging to the upper tree layer. a) Umbrella shaped crowns and b) buttresses. 

17 

altitude forests in East Africa. In the upper layer, 

trees are approximately 21-36 m high, have umbrella 

shaped crowns and are often heavily buttressed (cf. 

Figure 2-8). Some trees emerge through this stratum, 

growing to heights above 40 m. The second layer 

is formed by individuals ca. 11 m to 12 m high. 

They usually have oblong crowns in lateral contact. 

Together with smaller trees up to 11 m high they form 

a closed canopy. Due to these dense strata, the ground 

is intensively shaded and understorey vegetation in 

the form of shrub and herb cover is hardly present 

(Schulz & Wagner 2002). Whereas this concept is 

still applicable to only slightly disturbed parts of 

Budongo Forest, it has to be altered for those areas 

where more or less intense logging took place in the 

20th century. 

In secondary forest around Sonso Camp, for example, 

where large amounts of timber had been selectively 

logged until the 1970s, no distinct canopy stories 

are present today. The upper canopy is more even as 

emergent trees are hardly present. Greater gaps are 

still perceptible and as a result of the higher amount 

of light reaching the forest floor a dense cover of 

young shrubs and trees is growing in the understorey, 

frequently comprising pioneer species. Climbers 

(lianas, woody and herbaceous climbers) are more 

abundant in this forest type than in the near primary 

forest areas (Schulz & Wagner 2002). Vascular

18 

epiphytes are also well represented throughout 

mature forest stages, especially in the upper canopy 

(Eggeling 1947). 

In addition to vertical differences in forest structure, 

horizontal variability is also present. Apart from a 

small-scale mosaic of natural growth cycles driven 

by gap dynamics in the undisturbed forest areas, 

human interference in terms of timber extraction 

patterns had a great influence on horizontal structural 

variability in Budongo Forest. Plumptre (1996) 

found major differences between areas that had been 

logged since 1941 and areas that have been protected 

for more than 80 years. The latter are characterized 

by a significantly denser canopy, more trees in the 

lower canopy, fewer lianas and less light reaching the 

forest floor. In order to understand spatial patterns of 

forest structure, the forest management and logging 

history over the last 70 years will be described in the 

following chapter. 

2.1.3 

Forest management 

According to Howard (1991) Budongo Forest, 

being famous for its precious mahogany trees, has 

been the largest and most valuable timber forest 

in Uganda. The first sawmill was established in 

1926 (Eggeling 1947). Before that time only minor 

quantities of timber had been cut by pit sawyers, 

mainly in the southern part of the forest. After 1926 

more extensive exploitation began by concession 

holders. Between 1932 and 1939 Budongo Forest was 

gazetted as a Forest Reserve by the British Colonial 

Administration (Howard 1991). At the same time, 

the first of several detailed 10-year working plans 

was elaborated in order to control exploitation. The 

forest was divided into different compartments with 

individual logging schemes. Additionally, a nature 

reserve of approximately 100 ha was established 

in the southwestern part of the forest (N15), where 

vegetation was to remain untouched. 

The working plans prescribed a rotating system of 

selective logging with species-dependent rotating 

times of 40-, 60- and 80-years and minimum girths 

(Reynolds 2005). Mechanical logging activities 

increased until the 1960s and decreased in the 1970s, 

with the most valuable timber species being the 

mahoganies Khaya anthoteca, Entandrophragma 

angolense, E. cylindricum and E. utile as well as the 

ironwood Cynometra alexandri that was used for 

flooring, amongst others (Eggeling 1947, Plumptre 

1996, Reynolds 2005). 

Efforts were also made to plant mahogany trees 

to increase the stock inside the forest, but without 

success (Reynolds 2005). Plumptre (1996) reports 

arboricide treatment in the 1950s to 70s for tree 

species with little or no market value, such as Ficus 

spp, Celtis spp. and Lasiodiscus spp. A mixture of 

2,4,5-T and 2,4-D together with diesel was applied (a 

similar mixture became famous during the Vietnam 

war under the name Agent Orange) and was supposed 

to prevent these species from competing with the 

mahoganies. As consequence, structural changes 

and changes in vegetation types can still be observed 

in Budongo Forest after 60 years of commercial 

and illegal timber extraction as well as arboricide 

treatment (Plumptre 1996). 

Today the forest is managed by the Uganda Forest 

Department through the Masindi District Forest 

Office with two local stations at Nyakafunjo and 

Biiso. The current Management Plan (1997-2007) 

foresees the expansion of strict nature reserves, 

where no human activity is allowed to take place, 

surrounded by buffer zones (Figure 2-9). It should be 

noted however, that with respect to logging history, 

only compartments N15, W17, W31-34 and KP11-13 

can be considered as unlogged for the past 80 years 

(Plumptre 1996). Though the creation of protected 

zones is definitely a step in the right direction, human 

interference in the forest remains. Even parts that 

should stay totally undisturbed, such as the nature

2 Study areas 

Figure 2-9 

Logging compartments in Budongo Forest and forest management (made available by NFA through BIOTA-E02). 

reserve in compartment N 15, showed clear signs 

of illegal logging during the time of the fieldwork 

(Figure 2-10). This is also confirmed by Reynolds 

(2005) for other parts of the forest. 

Apart from logging, further anthropogenic threats 

to Budongo Forest include charcoal burning, illegal 

encroachment (crop planting on forest meadows), 

firewood collection and illegal grazing activities 

(Aryal 2002, Reynolds 2005). Further information on 

the forest compartments can be found in Table A-1 

(cf. Appendix of Tables). 

Figure 2-10 

19 

Illegal logging activities in the nature reserve N15.

20 

2.2 

2.2.1 

Kakamega Forest 

Physiogeographic aspects 

Kakamega Forest is situated in Kenya’s Western 

Province between N 41,236 and N 15,984 and 

E 696,777 and E 717,761 E (Projection: UTM, 36 N, 

WGS 84) at an altitude of about 1,460-1,760 m a.s.l. 

Figure 2-11 shows a subset of a Landsat ETM+ scene 

that was acquired on 10 January, 2003. Forested areas 

are shown in dark green colours, whereas bushland, 

tea cultivation zones and grasslands appear in brighter 

green tones. It is obvious that Kakamega Forest and 

its associated forest areas are much more fragmented 

than Budongo Forest. 

Figure 2-11 

Within its official boundaries the forest covers an area 

of 236 km². However only 123 km² comprise natural 

forest (Mitchell et al. 2006). The rest is composed 

of bushland, grassland and plantations. Kakamega 

Forest is furthermore surrounded by forest patches 

of various sizes, altogether adding up to 257 km² of 

natural forest cover. Apart from the South and North 

Nandi Forests with which Kakamega Forest was once 

connected via South Nandi, the largest remaining 

forest patch is Kisere Forest in the north with an 

approximate area of 4 km² (Mitchell et al. 2006). 

Two major rivers run through the forest in a westerly 

direction, the Isiukhu in the north and the Yala in the 

south, both draining towards Lake Victoria. 

Landsat ETM+ scene acquired on 10 January 2003, showing Kakamega Forest and its associated forest areas 

(band combination 5-4-3, Projection: UTM 36 N, WGS8). The forest coverage as marked in topographic maps 

1:50,000 published in 1970 is outlined in yellow (based on aerial photography from 1967).

2 Study areas 

Kakamega Forest is surrounded by densely populated 

areas (Blackett 1994). Agricultural areas around the 

forest are shown in Figure 2-11 in pink and bluish 

colours and consist of small fields where sugarcane, 

maize and beans are grown, amongst other crops (cf. 

Figure 2-12a). In the south of Kakamega Forest tea is 

cultivated along the forest edges (Figure 2-12b). 

Figure 2-12 

Geology and Soils 

Surroundings of Kakamega Forest. 

a) Agricultural land between Buyangu and Kisere 

(photo: C. Brueckl), b) tea zone at the forest 

edge close to Isecheno Forest Station. 

The rocks beneath the Kakamega Forest region 

belong to the Nyanzian and the overlying Kavirondian 

systems (Cahen et al. 1984, Schlüter 1997). They 

were formed in the late Archean, approximately 3.0 

to 2.5 Ga years ago. The Nyanzian System comprises 

21 

predominantly volcanic rocks of basic to acidic 

composition as well as poorly sorted sediments 

occurring as greenstone belts within the Basement 

Complex of western Kenya (Hester & Boberg 2006). 

The Kavirondian system unconformably overlays the 

Nyanzian system and consists of sediments derived 

from the erosion of supracrustal material, following 

the formation of late granites (Schlüter 1997). 

Kavirondian sediments occur in an east-west oriented 

syncline between the western Nandi slopes and 

Yala Town in the southwest of Kakamega Forest. 

The northern and southeastern flanks of this basin 

structure are framed by palaeoproterozoic granitic 

intrusions (Schlüter 1997). The Kavirondian system 

can be subdivided in four lithographic formations – 

from bottom to top called Shivakala (conglomerates 

of granitic and basaltic, andesitic and rhyolitic 

volcanic sources), Igukhu (greywacke), Mroda 

(cross-stratified sandstones and minor shales and 

greywackes) and Mudaa (shales). 

As in Budongo the soils of Kakamega Forest are 

mainly deeply weathered ferralitic soils that are poor 

to moderate in their nutrient content, as occurs very 

often under topical rain forest vegetation (Musila et al. 

2004, Schultka 1974). To a lesser extent there are also 

lixisols and cambisols, in some places phaeozems can 

also be found (Musila et al. 2004). Schultka (1974) 

reported dark reddish to dark brownish-red crumbly, 

sandy mudstones and sandy, loamy mudstones, 

mainly occurring on granite. In lower horizons he 

even found laterite crusts and compressed layers. 

Climate 

Analogous to Budongo Forest the climate of Western 

Kenya is influenced by its latitudinal position and 

the resulting seasonal ITCZ movements. Kakamega 

Forest also experiences a bimodal rainfall pattern 

with a long dry season from December to February

22 

and a short, less pronounced dry spell in June and 

July. Most precipitation falls in the rainy seasons 

between March and May and August to October. The 

average monthly precipitation recorded by the former 

Forest Department (now Kenya Forest Service, KFS) 

at Isecheno Forest Station between 1982 and 2006 

is shown in Figure 2-13a. Mean yearly rainfall is 

1,972 mm (±336 mm). 

Temperature recorded by a meteorological station 

set up by the BIOTA project in the northern part of 

the forest close to the office of the Kenya Wildlife 

Service (KWS) showed daytime maximum values 

up to 32°C in February and a minimum nighttime 

temperature of 9°C recorded in October (Steinbrecher 

et al. 2004). As no full year of data records was 

available, temperature data as cited in Mutangah 

(1996) is displayed in Figure 2-13b. Minimum mean 

monthly temperatures are again recorded during the 

dry seasons. 

Figure 2-13 

2.2.2 

Vegetation characteristics 

Biogeographic position 

The main aspects concerning the history of forest 

development in East Africa were described in 

Chapter 2.2.2. It is again emphasized that according 

to Hamilton (1982) forest species diversity in this 

region decreases with increasing distance to the 

Guineo-Congolian Quaternary refuge. Althof (2005) 

mentions, that Kakamega Forest probably developed 

in the postglacial period after 10,000 B.P., an era 

favourable to forest expansion. Accordingly, this 

relatively young age together with the distance to the 

Guineo-Congolian centre of endemism results in low 

species numbers and low endemism in Kakamega 

Forest. Fischer (personal communication) states that 

only about 400 different vascular plant species are 

known from Kakamega, which is equivalent to some 

mixed deciduous forests in Germany. 

Kakamega Forest is generally referred to as the 

easternmost remnant of the former Guineo-Congolian 

rain forest belt (Kowkaro 1988). On the other hand 

Lind & Morrison (1974) claim, “botanical evidence 

suggests that there have been at least two periods 

Climatic characteristics of Kakamega Forest. a) Mean monthly rainfall at Isecheno Forest Station from 1982-2006 

(source: Forest Department) and b) mean monthly temperatures at Kakamega Agricultural Station from 1980-1992 

(source: Mutangah 1996).

2 Study areas 

of connection between the coastal rain forests [of 

Kenya and Tanzania] and the Zaire rain forest” (p. 

203). This would imply that some still existing forest 

patches along the coastline of East Africa could also 

be regarded as relicts with strong Guineo-Congolian 

influence. Althof (2005), however, argues that the 

influence of other floristic elements in these forests 

is higher, whereas in Kakamega Forest species of 

Guineo-Congolian origin prevail. 

On a regional scale Mitchell (2004) analysed the 

exploitation and disturbance history of Kakamega 

Forest and its surrounding forest patches based on 

published data, oral records, and place name evidence. 

According to his study, the forest was still linked to 

South Nandi Forest in 1959, which in turn formed an 

23 

entity with North Nandi Forest (cf. Figure 2-11). No 

evidence has been found that Kakamega Forest had 

ever been fully joined with the forest patches (often 

referred to as “fragments”) in the north and west. 

Vegetation classification and floristics 

White (1983) also classifies Kakamega Forest as 

belonging to the floristic region of the Lake Victoria 

Regional Mosaic. Kakamega Forest, however, 

harbours a much higher number of afromontane 

species than Budongo. These might have survived 

arid conditions in close-by mountainous regions in 

Kenya (e.g., around Mount Elgon, as reported by 

Hamilton 1982). According to Althof (2005) about 

Figure 2-14 Vegetation in Kakamega Forest. a) Dracena fragrans, b) Deinbollia kilimandscharica-Markhamia lutea alliance in 

Kisere Forest, c) Harungana madagascariensis-Desmodium adscendens alliance close to Buyangu Camp and 

d) forest glade close to Isecheno Nature Reserve.

24 

33% of the woody species recorded in Kakamega 

Forest are afromontane floristic elements. Another 

26% are so-called transition species and 41% are 

of Guineo-Congolian origin. Of the latter, several 

species such as Aningeria altissima, Cordia milennii, 

Entandrophragma angolsense, Maesopsis eminii 

and Moodora myristica reach their easternmost limit 

here. Just as Budongo Forest, Kakamega Forest is 

classified as semi-evergreen rain forest following 

Whitmore’s forest formations (see Table 2-1). 

Compared to Budongo Forest, relatively few studies 

on floristics variation have been conducted in 

Kakamega Forest. Though Beentje (1990), Mutangah 

(1996), and Fashing & Mwangi Gathua (2004) 

analysed the tree species composition on certain 

plots in Kakamega Forest and Schultka (1974) 

studied forest regeneration in different surrounding 

grasslands and shrublands, Althof (2005) was the first 

to describe plant communities in more detail. Her 

study was based on 200 phytosociological relevés that 

were distributed throughout the forest and resulted in 

the differentiation of two vegetation alliances (I and 

II) with thirteen communities and subcommunities. 

Though a detailed and spatially explicit vegetation 

map of the whole forest is still missing, her results 

will be briefly described below. 

Alliance I mainly refers to more or less disturbed 

mature forest sites. The dominant tree species in this 

alliance are Antiaris toxicaria, which is of guineocongolian 

origin, and Diospyros abyssinica, an 

afromontane element. In the lower canopy layers 

Bequaertiodendron oblanceolatum, Trilepisium 

madagascariense as well as Trichilia emetica, 

Heinsenia diervilleoides, Cassipourea ruwensorensis, 

Morus mesozygia and Chrysophyllum albidum are 

frequent. Dracena fragrans is highly abundant in 

the shrub layer (cf. Figure 2-14a). Lianas also occur 

under the dense canopy cover. 

Alliance I can be further split up into two major 

groups, the Deinbollia kilimandscharica-Markhamia 

lutea (cf. Figure 2-14b) and the Celtis mildbraedii- 

Craibia brownii alliances. Whereas the first group 

mainly occurs in the northern parts of Kakamega 

(Kisere Forest and Buyangu Nature Reserve) and 

comprises different stages of middle-aged, old 

secondary and near primary forests, the second one 

mainly occurs in the middle and southern forest parts 

and reflects a combination of different anthropogenic 

treatment and natural variation. 

Alliance II indicates young initial forest stages 

that occur after clear felling when bushland and 

subsequent young secondary forest develop. Two 

species are characteristic for this alliance: Harungana 

madagascariensis as dominant tree and Desmodium 

adscendens in the herb layer (cf. Figure 2-14c). 

For both major alliances, different communities and 

subcommunities can be distinguished. As they are, 

however, less important for this thesis, the reader is 

referred to Althof (2005) for further information. 

Within its official forest boundary, Kakamega 

Forest also contains various glades where grassland 

is present (cf. Figure 2-14d). Whether their origin 

is natural or anthropogenic remains unclear. 

Mitchell (2004) mentions that some glades might 

possibly be remnants of former unforested areas. 

Due to their shallow soils forest tree species 

could not establish. Furthermore, human induced 

fire and grazing animals might have kept the 

glades open. However, Mitchell (2004) reports 

that forest succession is taking place, resulting 

in the shrinking of most glades with some 

even disappearing altogether, a process that 

has been especially marked in the northern 

part of the forest since the establishment there 

of the Kakamega National Reserve in 1986. 

Althof (2005) further mentions several plantations 

that were mainly established in the southern

2 Study areas 

parts of the forest primarily consisting of Maesopsis 

eminii, Eucalyptus saligna, Pinus patula or Cupressus 

lusitanica. 

Vegetation structure 

Concerning vertical forest structure Althof (2005) 

describes the less disturbed forest stages as having 

canopy heights of 20 m to 35 m in the upper tree 

layer with some emergents reaching up to 40 m. A 

more or less developed second layer exists with tree 

heights of 10 m to 12 m. Understorey vegetation, 

mainly comprising of Dracena fragrans, is dense in 

most parts. Younger forest stages of the Harungana 

madagascariensis-Desmodium adscendens alliance 

are in turn characterized by a single tree layer up 

to 12 m high with more or less dense herb cover. 

Understorey vegetation in the form of larger shrubs 

is hardly present. 

Just like in Budongo Forest, horizontal variability 

of forest structure is mainly driven by human 

exploitation of the rain forest ecosystem. Valuable 

timber species once abundant in Kakamega, such as 

the mahogany Entandrophragma angolense, are now 

very rare and only occur in restricted areas, though 

seedlings and saplings were recorded in other areas 

by Althof (2005). Thus it is not astonishing that a true 

primary forest no longer exists in Kakamega Forest 

(Althof 2005). Today the forest is rather a mixture 

of different disturbance stages, such as near primary 

forest, secondary forest and clearings mixed with tea 

and timber plantations. The different forest stages are 

however not spatially grouped in compartments as in 

Bodongo Forest, but rather irregularly distributed. 

Based on Mitchell (2004), Kisere Forest, as well as the 

Yala and Isecheno Nature Reserves can be identified 

as the most undisturbed parts of the Kakamega Forest 

area. Buyangu Nature Reserve in the northern part 

of Kakamega Forest has in turn undergone extensive 

25 

exploitation. Natural succession can now be studied 

on areas that are now protected from fire and grazing, 

which formerly maintained them as grassland. 

2.2.3 

Forest management 

According to Mitchell (2004) Kakamega Forest was 

gazetted as a Forest Reserve in 1933 after the former 

Forest Department took over the management in 

1931. Commercial timber extraction started during 

the 1930s, triggered by the need of the then existing 

gold mines (Mitchell 2004). The number of logged 

trees continually rose until the 1980s, supported by 

numerous sawmills (Bleher et al. 2006, Mitchell 

2004), and had a major influence on forest structural 

properties and forest species composition. Clear 

felling of indigenous forest took place until the mid- 

1970s when it was partly replaced with fast growing 

softwood or tea plantations, especially in the southern 

forest parts. In the north, more selective logging took 

place. Olea capensis had been identified as the best 

timber tree, but other species were also cut, such as 

Zanthoxylum gillettii, Aningeria altissima, Funtumia 

africana, Prunus africana and Cordia africana, 

after 1948 also Celtis mildbraedii amongst others 

(Mitchell 2004). 

Until the nature reserves of Yala, Isecheno and 

Kisere (cf. Figure 2-11) were established in 1967, 

local people were allowed to use the whole forest 

for their own purposes (collection of firewood, poles 

for construction material, cattle grazing). In 1986 

Kisere Forest and the northern part of Kakamega 

Forest, Buyangu, were officially handed over to 

the management of the KWS. Since the 1980s the 

conversion of indigenous forest to plantation, the 

cutting of indigenous trees, as well as cultivation 

have been banned from all parts of the forest. Cattle 

grazing and the collection of firewood and grass are 

still allowed under permit outside the nature reserves 

(Mitchell 2004). Though legislation is clear, KFS and

26 

KWS do not enforce these laws with equal strictness. 

Whereas the northern part of the forest managed by 

KWS shows less anthropogenic disturbance today, the 

southern part under KFS direction still experiences 

high levels of illegal logging activity as well as 

frequent charcoal burning, firewood collection (cf. 

Figure 2-15) and cattle grazing (Bleher et al. 2006, 

Mitchell & Schaab, in print). 

Figure 2-15 

Illegal activities in Kakamega Forest. a) Firewood 

collection at Isecheno and b) charcoal burning 

close to Shandarema.

3 Theoretical background 

The following chapter provides the theoretical 

background for the MODIS LAI product validation. 

LAI estimation from both field and satellite data 

will be described with special emphasis on the 

derivation for tropical rain forest canopies. Starting 

with the definition of LAI and an overview of LAI 

characteristics in tropical rain forests, existing in 

situ methodologies, instruments, constraints and 

sampling strategies are to be reviewed. The state 

of research in satellite-based LAI derivation will 

be further discussed with respect to physical and 

empirical approaches, as well as constraints related to 

sensor resolution. Focus will then be put on spectral 

vegetation indices (SVIs) and texture measures as 

they form the basis for the upscaling of field data in 

this thesis. Last but not least validation approaches 

for satellite-based LAI products will be appraised. 

3.1 

Definition of LAI 

27 

LAI is a dimensionless variable that was first 

mentioned by Watson (1947) following earlier 

studies of foliage amount estimation. According to 

his definition, it refers to the total one-sided area of 

photosynthetic tissue per unit ground surface area. 

Whereas this definition is adequate for broad leaves, 

problems occur with needle leaves as here the onesided 

area is not clearly defined (Jonckheere et al. 

2004). Therefore several authors (for instance, Ross 

1981, Smith et al. 1991) proposed a projected leaf 

area to take into account asymmetrical forms of 

certain foliage types. Yet if the projection angle is 

not clearly specified, the chosen projection does not 

necessarily result in the highest possible LAI values 

of the examined vegetation. Consequently Myneni 

et al. (1997) defined LAI as the maximum projected 

leaf area per unit ground surface area. 

Whereas the projected leaf area is more relevant 

to radiometric assessments of LAI, where the 

parameter is estimated indirectly from solar radiation 

interception, Chen & Black 1992 in agreement with 

Lang (1991) suggested using half the total green leaf 

area per unit ground surface as definition of LAI. 

The basis for their theoretical reasoning was that the 

hemisurface leaf area is a simple definition that also 

allows for further calculation of projected leaf area 

with a few assumptions (Hyer & Goetz 2004). This 

adheres to the original concept of Watson (1947), as 

at least for broad-leaved trees the one-sided leaf area 

is equivalent to the hemisurface leaf area. 

This last definition for LAI is now widely used in 

literature (Casa & Jones 2005, Eriksson et al. 2006, 

Leblanc et al. 2005, Nilson & Kuusk 2005, Weiss 

et al. 2004, among others). The latter also described 

optical methods of in situ LAI estimation as well as 

the MODIS LAI algorithm, which are based on this 

definition (Privette et al. 2002). Consequently it will 

be the one used throughout this study. However, it is

28 

important to note that different definitions result in 

significant discrepancies of calculated LAI values 

and for comparison purposes the respective definition 

should be checked carefully. 

3.2 

Characteristics of LAI 

in tropical rain forests 

As mentioned in Chapter 2, the leaf area of tropical 

rain forest stands varies according to species 

composition, phenological stage (in the case of 

semi-evergreen rain forests with a certain percentage 

of deciduous species), disturbance and climatic 

conditions. Additionally local changes in LAI can be 

observed due to forest dynamics (Welles 1990). 

Compared to higher latitudes, field measurements of 

LAI in tropical environments are rare. Doubtlessly the 

reason for that is the higher complexity of vegetation 

structure and species diversity in rain forests that 

limits the transferability of well-established in situ 

methods of, for instance, agricultural lands or boreal 

forests. In the following a literature review will shed 

light on available information on LAI of tropical rain 

forests. 

One of the most extensive and global LAI data sets 

was gathered by Scurlock et al. (2001) and is further 

summarized by Asner et al. (2003). Consisting of 

almost 1000 LAI measurements from nearly 400 

unique and globally distributed field sites, it is based 

on published in situ data sampled between 1932 and 

2000. However, LAI was estimated with different 

methods and no information on measurement 

precision is available. Unfortunately the database 

does not comprise a single test site for East Africa 

and only 6% of all valid observations refer to tropical 

evergreen broadleaf forest sites. For these sites the 

mean reported LAI is 4.78 with a standard deviation 

of 1.70 and minimum and maximum values of 1.48 

and 8.0 respectively. Interestingly, Asner et al. (2003) 

mention that an analysis of the collected LAI values 

over time revealed a decline in mean measured LAI 

values over the decades. This is most noticeable in 

the 1990s and is probably related to the methods 

predominantly used. Before 1990 mainly direct 

and semi-direct methods were applied, whereas for 

optical approaches that were developed after 1990 

underestimations are frequently reported (cf. Chapter 

3.3.1). 

One of the first studies on LAI of tropical rain forests 

is the one by Ogawa et al. (1961). It was performed 

using destructive measurements and reports LAI 

values between 7 and 11 for evergreen tropical rain 

forests in Thailand. Slightly lower LAI values are 

reported by Alexandre (1981), who reviewed several 

studies, mostly from tropical rain forests in South 

America and Asia. He comes to the conclusion that 

a maximum annual LAI of 8.2 seems to be a reliable 

value for tropical rain forests in general. 

By contrast, more recent studies show the abovementioned 

tendency for lower LAI values. Aragão 

et al. (2005) measured LAI in evergreen primary 

and secondary rain forests in Eastern Amazonia 

with optical devices. Their very detailed statistical 

analysis reveals a spatial variability of LAI at 

different landscape units (individual measurements to 

plot level). For the plot level, which is probably the 

most comparable to other studies, they recorded LAI 

values between 3.06 (±0.31) for secondary forest 

and 5.85 (±1.19) for primary rain forest. Comparable 

values were retrieved by Wirth et al. (2001), who 

examined the vertical and horizontal variability of 

canopy structures and seasonal variability of LAI 

in a semi-evergreen tropical rain forest in Panama. 

Although the mean decrease in LAI from wet to dry 

season was low with 5.5% from 5.41 (±1.27) to 5.11 

(±0.68), Wirth et al. (2001) report considerable smallscale 

heterogeneities in horizontal LAI distribution. 

The results of Wirth et al. (2001) concerning the 

vertical variation of LAI are displayed in Figure 3-1.


Summing up to a mean total LAI of 6.04, 

measurements along five vertical profiles revealed 

that more than 50% of canopy leaf area is found 

in the uppermost 5 m due to the clustering of tree 

crowns in the upper canopy. Below 30 m the canopy 

is dominated by almost leafless trunks until a second 

maximum of leaf area arises at the height of trees 

in the second tree layer or understorey (Wirth et 

al. 2001). In another study from Central America 

Kalácska et al. (2004) estimated in situ LAI values 

of 4.80 (±0.82) for intermediate and 6.90 (±1.96) for 

late forest stages in a tropical moist forest in Costa 

Rica. 

The mean LAI values given by Wirth et al. (2001) 

correspond closely to one of the very few examples 

for LAI measurements in African rain forests. De 

Wasseige et al. (2003) investigated seasonal changes 

in leaf area of a nearly undisturbed semi-evergreen 

tropical rain forest in the CAR, which is comparable 

to Budongo and Kakamega Forests in terms of 

vegetation type, average yearly rainfall and length of 

dry season. At the end of the rainy season an average 

LAI of 5.47 was measured at plot level. This value 

changed to 5.13 on the same site at the end of the 

dry season. Slightly higher values were derived by 

Laumonier et al. (1994) for an evergreen tropical rain 

forest in Cameroon. A mean value of 5.6 is reported 

here, with minimum and maximum values of 3 and 

7 at sample level. So far, no published LAI values 

could be found for East African rain forests. 

Table 3-1 summarizes the above-mentioned 

publications. The examples show that the variability 

of in situ LAI for tropical rain forests can be relatively 

high. Whereas a minimum value of 3.06 (mean LAI 

at plot level) is reported by Aragão et al. (2005) for a 

secondary evergreen rain forest in Panama, Ogawa et 

al. (1961) derived maximum LAI values of 11 with 

destructive measurements in an evergreen tropical 

rainforest in Thailand. These values do not only reflect 

differences in forest formation and structure due to 

Figure 3-1 

29 

natural environmental parameters or anthropogenic 

influences (mainly represented by different forest 

stages); the variation of the results might also be partly 

attributed to the instrumentation and methodologies 

used (e.g. the destructive measurements employed 

by Ogawa et al. [1961] versus studies using LAI- 

2000 PCA). For a better understanding of these 

discrepancies, the most important methods for in situ 

leaf area index determination will be described in the 

following. 

3.3 

Vertical LAI variation in a semi-evergreen tropical 

rain forest in Panama (Wirth et al. 2001). 

Ground-based measurements 

of LAI 

Performing ground-based measurements of LAI is 

quite challenging in tropical rain forests. Besides 

problems in transferability or applicability of methods 

to tropical rain forests, in situ sampling is complicated 

by species diversity and stand complexity. Further, in 

situ data should be representative for the study sites, 

which can be very difficult in areas where the spatial 

distribution of sampling sites is restricted by the 

accessibility of remote forest regions. Therefore an 

in-depth understanding of LAI, its definition and its 

characteristics in tropical rain forest stands, as well as 

the knowledge about state-of-the-art methodologies 

of in situ LAI estimation is crucial. After an overview 

on the available field methods, the instruments used 

in this study will be described and limitations and

30 

5.85 (±1.19) for primary rain 

forest (plot level) 

3.06 (±0.31) for secondary rain 


Alto Tapajós, Eastern 

Amazonia, Brazil 

(2°24’S/4°01’1S; 

55°30’W/54°29’W) 

Table 3-1 

Evergreen rain forest Primary and secondary 


LAI-2000 PCA One-sided leaf area 

per unit ground area 

Aragão et al. (2005) 


6.90 (±1.96) for primary rain 


3.20 (±0.82) for secondary rain 

Los Innocentes, Costa 

Rica 

(11°1’N, 85°30’W) 

Evergreen rain forest Primary and secondary 


17°00’E/17°30’E) 

(wrong definition for 

LAI-2000 PCA) 

LAI-2000 PCA One-sided leaf area 


Kalácska et al. (2004) 

(plot level, seasonal changes) 

5.13 to 5.47 

Ngotto Forest, Central 

African Republic 

(3°40’N/4°10’N; 

Semi-evergreen 


specified) 

based on radiative 

measurements 

Primary forest LAI-2000 PCA Total leaf area per unit 

ground area 

De Wasseige et al. 

(2003) 

4.78 (±1.70) for evergreen rain 

forest 

3.92 (±2.53) for semi-evergreen 


5.11 to 5.41 

(seasonal changes) 

Barro Colorado Island, 

Panama 

(9°10’N, ° W not 

Semi-evergreen rain 


Semi-evergreen 


Primary forest Inversion of light 

interception model 

Not mentioned Wirth et al. (2001) 

Worldwide (no East 

African rain forest 

included) 

Evergreen rain forest 

n/a Various methods 

applied 

Not mentioned Asner et al. (2003)/ 

Scurlock et al. (2001) 

(5°15’N/ 52°55’W) 

4.51 to 5.44 Paracou, French 

Guiana 

Evergreen rain forest Primary forest LAI-2000 PCA One-sided leaf area 


Review of LAI values reported for tropical rain forests in the literature. 

Ferment et al. (2001) 

5.6 Campo Nature 

Reserve, South 

Cameroon 

Evergreen rain forest Not mentioned, primary 

forest assumed 

LAI-2000 PCA Not mentioned Laumonier et al. (1994) 

Maximum value of 8.2 Worldwide Evergreen rain forest n/a Leaf litter fall method Total leaf surface area 


Alexandre (1981) 

7 to 11 Thailand Evergreen rain forest Not mentioned, primary 

forest assumed 

Destructive 

measurements 

One-sided leaf area 


Ogawa et al. (1961) 

In situ LAI Study area Forest formation Forest stage Methodology/ 

instruments 

LAI definition Source


caveats of in situ LAI assessment will be reviewed. 

The subchapter closes with the description of 

adequate ground sampling strategies. 

3.3.1 

Methods 

Methods for LAI determination have been developed 

since the early 1930s. Most of the literature since 

published in this context is related to methodology 

and constraints, validation of methods and 

instruments, the development of new instrumentation 

and the comparison of results with reference to leaf 

area measurements both in stands and in laboratory 

conditions. 

Ross (1981) gives a good overview of the historical 

methodological development and their theoretical 

background until 1980. He separates the methods 

into direct and indirect approaches, as well as several 

others that have lost importance since the 1980s and 

will not be mentioned here. More recently Bréda 

(2003) published a useful review of methods, giving 

further detail to newer optical instruments and 

current controversies such as error analysis, crosscalibration, 

sampling strategy, spatial validation or 

scaling problems. Jonckheere et al. (2004) discuss 

different methods for in situ LAI determination with 

special regard to indirect measurements. The most 

important methods for tropical rain forest stands will 

be described in the following. 

Direct and semi-direct measurements 

Generally, direct methods measure the leaves 

themselves. In consequence they are the most 

accurate method of retrieving LAI for a certain 

sample, yet the direct approach is usually very time 

consuming and labour intensive and is therefore only 

feasible for small areas. According to Ross (1981) 

the earliest method of leaf area determination was 

31 

drawing the contours of a leaf on a piece of paper 

and relating it to a reference grid. Though this is very 

accurate (Rozhnyatovsky [1954] employing a very 

similar method found an error in the order of 1%), 

it is not very efficient and practically inapplicable to 

small compound and crimpy leaves. Today opticallybased 

automatic area measurement systems exist, 

such as the LI-3000 or the LI-3100 Area Meters (LI- 

COR, Lincoln, Nebraska). Single leaves are scanned 

(attached or unattached) and the respective leaf area 

is automatically calculated. 

However, when scaling direct LAI measurements 

up to a whole stand, errors can be introduced that 

might lead to a large over- or underestimation of the 

parameter. This also puts the above-mentioned early 

work of Ogawa et al. (1961) into perspective (cf. 

Table 3-1). 

In general, direct methods are not compatible with the 

long-term monitoring of spatial and temporal patterns 

of LAI that can be done with optical devices. Yet they 

are still of relevance when it comes to the validation 

of indirect measurement methods (Jonckheere et al. 

2004). 

Semi-direct methods are based on direct 

measurements of other vegetation parameters that 

are closely related to LAI, e.g. sapwood area. Here, 

measurement errors mainly result from the empirical 

relationship established between LAI and the variable 

taken into account, as well as the applied upscaling 

approach. A common example is the collection of 

leaf litter fall in temperate deciduous forests during 

the autumn exfoliation period (e.g. Dufrêne & Bréda 

1995, Eriksson et al. 2005). A litter trap with a known 

surface area is emptied on a regular basis to avoid 

weight loss through decomposition. The collected 

leaves are usually weighed and converted to leaf 

area by determining the specific leaf area (leaf area/ 

leaf mass) for sub-samples (Eriksson et al. 2005). 

The method is believed to give reliable results and is

32 

therefore widely used as reference for other methods 

of LAI determination and their validation. However, 

it should be noted that the LAI assessment using litter 

traps is also based on an estimation technique, as it 

depends on the specific leaf area (SLA) relationship 

which is species dependent and also varies according 

to tree vertical position. Eriksson et al. (2005) e.g. 

found that the top leaves of beech and oak stands 

had a higher SLA than lower leaves. If an SLA 

relationship is only based on low leaves, the resulting 

LAI from litter trap measurements is likely to be 

overestimated. 

In tropical rain forests the use of leaf litter traps is 

rather difficult. The correlation of seasonal litter fall 

shows a correlation to the average life span of leaves 

or climatic variability rather than to real leaf area. 

Scurlock et al. (2001) e.g. mention leaf turnover times 

for tropical rain forests of up to 6 years. However, 

Kalácska et al. (2005) used leaf litter data from a 

tropical dry forest to determine LAI. As they also 

mention that especially in late succession stages not 

all trees on the sample plots shed their entire leaves, 

the LAI determination with litter traps is questionable 

in this case. Moreover, as SLA is species dependent, 

representative relationships with LAI are not easy to 

establish in a species rich environment. 

LAI can also be retrieved by semi-direct allometric 

methods, which describe the relation between leaf area 

and an independent variable (Gower et al. 1999). These 

independent variables are commonly characterized by 

easy-to-measure physical dimensions, as e.g. stem 

diameter at breast height (DBH). Species- or standspecific 

relationships are then applied on the basis of 

detailed destructive measurements of sub-samples. 

Good correlation coefficients have been found 

between leaf area and sapwood area (Kaufmann & 

Troendle 1981, O’Hara & Valappil 1995, Rogers & 

Hinckley 1979, Vertessy et al. 1995), stem basal area 

(Bartelink 1997) and DBH (e.g. Le Dantec et al. 2000) 

for different species or vegetation types. 

The disadvantage is that allometric relationships 

are very stand-, species- and time-specific. Neither 

phenological change over the year nor LAI recovery 

after a natural or human-induced canopy opening can 

be described (Bréda 2003). In addition, considerable 

fieldwork is required to establish a stable relationship 

for a whole tree stand, especially when it consists of 

a high number of different species. Therefore it is 

hardly – if at all – applicable to tropical rain forests. 

Indirect methods 

In recent years indirect measurements, in particular, 

have grown in importance as they allow rapid LAI 

assessment of large areas. A good overview of stateof-the-art 

methodologies is given by Weiss et al. 

(2004) and Jonckheere et al. (2004). Indirect methods 

usually refer to optical measurements of radiation 

with photosensitive instruments. The transformation 

into LAI is then performed either by a calibrated 

relationship or an inversion model that takes statistic 

and probabilistic models of canopy geometry into 

account (see e.g. Bréda 2003, Eriksson et al. 2005, 

Frazer et al. 2001, Jonckheere et al. 2004, Leblanc et 

al. 2005, Ross 1981, Welles 1990, Welles & Norman 

1991, Zhang et al. 2005, among others). While these 

methods have the advantage of convenience and 

low labour costs, many difficulties arise from the 

complexity of radiation transfer in forest canopies. 

Major problems include unknown foliage angle 

distribution, improper model assumptions and 

the contribution of woody material to radiation 

interception. The theoretical background of indirect 

optical measurements will be described briefly in the 

following. 

As it penetrates the canopy, incident solar radiation 

is intercepted by plant organs. Consequently light 

attenuation corresponds to the vertical depth of the 

canopy. Though this process can be quite variable 

in space and time, mean horizontal flux densities

l that 3, that ERIKSSON takes statistic et al. and 2005, probabilistic FRAZER models et models al. 2001, of of canopy JONCKHEERE geometry et into 

al. 2004, 

If we consider G ( , ) 

constant along the path and LAI bein 


constant along the path and LAI being de 

981, RIKSSON IKSSON WELLES et et al. al. 1990, 2005, 3 Theoretical WELLES FRAZER background & et et NORMAN al. al. 2001, 1991, JONCKHEERE ZHANG et et et al. al. al. 2005, 2004, 

among Z 

Z 

33 

WELLES have ELLES the 1990, advantage WELLES of & convenience & NORMAN THEORETICAL 1991, and 1991, low ZHANG labour et BACKGROUND 

et costs, al. al. 2005, many among 

difficulties LAI u dz , 

LAI 

f the the radiation advantage transfer of of 

7 u dz , 

0 THEORETICAL BACKGROUND 

convenience in forest canopies. and low labour Major labour costs, problems costs, many include difficulties 

unknown 0 

ndirect methods usually refer to optical measurements of radiation 

iation tion proper transfer model in in assumptions forest canopies. and Major the Major contribution problems include of include woody unknown 

material to 

z 

ransformation into (2004) usually LAI and decrease is JONCKHEERE then performed resulting et in al. either (2004). a downward by a Indirect calibrated pattern methods usually where where refer z is z is the to the optical total total canopy measurements canopy height of and radiation ε z , , it follows aft 

heoretical r er model assumptions background assumptions of and indirect the the contribution optical measurements of of woody material will material be described to to where z is the total canopy height and cos , it follows after L 

takes statistic and with of light probabilistic photosensitive reduction models (Saeki instruments. of 1975). canopy The The geometry transformation relationship into into it follows LAI is after then Lang performed et al (1985) either that by the a calibrated gap cosfraction 

 

tical cal background of of indirect optical measurements will be be described 

ON et al. 2005, 

7 

described by 

relationship between FRAZER light et or al. reduction an 2001, inversion JONCKHEERE and model canopy that et geometry takes al. 2004, statistic can and described is described probabilistic by by models THEORETICAL of canopy BACKGROUND 

geometry into 

LES incident 1990, solar WELLES radiation account be & characterized NORMAN (see is intercepted e.g. 1991, by BRÉDA statistical ZHANG by 2003, plant models et organs. ERIKSSON al. 2005, that Consequently among describe et al. 2005, light FRAZER et al. 2001, LAI 

GLAI 

, JONCKHEERE et al. 2004, 

(2004) and JONCKHEERE et al. (2004). Indirect methods usually refer to optical cos measurements nt e ent vertical solar radiation depth radiation the of is the is probability intercepted canopy. Though of by by radiation plant this organs. process interception Consequently can be (or quite non- light variable light P ( , 

) 

 

of radiation 

advantage of convenience LEBLANC et and al. low 2005, labour ROSS costs, 1981, many WELLES difficulties 1990, WELLES in & NORMAN o Ge 

, 

1991, cos 

. 

P 

ZHANG 

et al. 2005, among 

o ( , 

) 

e . 

(3.3) 

with photosensitive instruments. The transformation into LAI is then performed either by a calibrated 

ntal tical cal transfer depth flux depth densities of in of the forest the others). interception) canopy. usually canopies. While Though decrease Though within Major these this resulting the methods problems process canopy. in can have a can Usually include downward be be the quite advantage these unknown variable pattern models of in of in convenience light and low labour costs, many difficulties 

relationship or an inversion model that takes statistic and probabilistic models of canopy geometry into 

lux x odel e densities relationship densities assumptions usually arise are between decrease and based decrease from the light the on resulting contribution resulting complexity simplified reduction in in a a downward of assumptions and of downward woody radiation canopy pattern material pattern transfer concerning geometry of of to light in light can forest Lang be canopies. (1986) showed Major that problems the gap include fraction unknown is equivalent 

account (see e.g. BRÉDA 2003, ERIKSSON et al. 2005, FRAZER et al. 2001, JONCKHEERE et al. 2004, 

lationship background tionship odels that between describe of foliage the indirect light radiative light the angle optical reduction probability reduction and distribution, measurements spatial and of canopy characteristics radiation improper canopy will geometry be interception model described of canopy assumptions can (or be be non- to and radiation the contribution transmittance of τ woody through material the canopy, to 

LEBLANC et al. 2005, ROSS 1981, WELLES 1990, WELLES & NORMAN 1991, ZHANG et al. 2005, among 

ls y. that Usually describe these radiation structural the models the probability are interception. properties, based of on of radiation simplified as radiation The e.g. theoretical leaf interception assumptions interception area density, background (or concerning (or leaf non- 

of indirect the if the latter optical is measured measurements at wavelengths will be described where the 

others). While these methods have the advantage of convenience and low labour costs, many difficulties 

istics allyually these of these canopy models briefly angle structural are are based and in based the properties, foliage on following. on simplified distribution as e.g. assumptions leaf among area density, concerning others. leaf The angle the the and assumption of opaque leaves (i.e. non-reflecting and 

olar radiation is arise intercepted from the by complexity plant organs. of Consequently radiation transfer light in forest canopies. Major problems include unknown 

thers. f of canopy The structural key key variable properties, variable properties, in in those as those as e.g. models leaf area is is gap density, gap density, fraction leaf angle P angle 

o ( , 

and ) 

and , 

which 

depth of the canopy. As it Though penetrates this the process canopy, can incident be quite solar variable radiation in is intercepted non-transmitting) by plant is valid. 

foliage angle distribution, improper model assumptions and the contribution organs. of woody Consequently material light to 

ky . The vs. 

key 

foliage key variable seen which in 

at in point describes those 0 

models 

in models the zenith proportion is is gap 

gap and 

fraction of fraction azimuth sky vs. P oP 

( ofoliage 

( 

 

, , 

 

directions 

) , 

) , which seen which 

ensities usually radiation attenuation decrease resulting interception. corresponds in a downward The to the theoretical vertical pattern depth background of light of the canopy. of 

from 

indirect Though optical this measurements process can be will quite be variable described in 

. foliage 

ERE foliage et 

seen 

al. seen 2004). 

at at point at point 0 0 in in zenith and azimuth directions from 

ship between space light and reduction time, mean and horizontal canopy geometry flux densities can be usually decrease A good resulting overview in of a downward other theoretical pattern of models light of 

briefly in the following. 

t et al. al. 2004). 

THEORETICAL beneath the canopy (Jonckheere et al. 2004). 

different at describe statistical the reduction 

BACKGROUND 

probability (SAEKI 

models exist of radiation 1975). 

(each characterizing interception The relationship 

a (or different non- between light canopy reduction geometry and and canopy their relation geometry to gap can fraction be 

As it penetrates the canopy, incident solar radiation statistical is intercepted 

nt ent ts), these statistical one statistical models of the models are models characterized 

best based known exist on (each and simplified (each by 

most characterizing statistical characterizing frequently assumptions models 

applied a a different concerning different that describe 

in the statistical 

estimation the the is probability given by by 

of Nilson plant 

radiation (1971), organs. 

interception who, Consequently in addition light 

(or non- to the 

er to optical measurements attenuation of of 

Though a large corresponds radiation 

number of different statistical models 

e ne e nopy e.g. of of the MONSI the structural best known & known interception) 

SAEKI properties, and 1953, most as within 

THEORETICAL to the vertical 

MUSSCHE e.g. frequently leaf the area canopy. 

et applied al. applied density, Usually 

BACKGROUND depth of the canopy. 

2001, in in NILSON leaf the the angle these estimation models 

1999, and WEISS of of are Poisson Though 

based on model, this process 

simplified describes can 

assumptions binomial be quite variable 

concerning models in 

the (the 

then performed space either et 

exist and by 

(each time, a calibrated 

characterizing mean horizontal a different statistical 

canopy e MONSI ONSI key variable & is & SAEKI divided SAEKI radiative in 1953, into 1953, those N MUSSCHE and 

statistically models spatial et is et al. characteristics THEORETICAL flux densities 

independent gap al. 2001, fraction NILSON layers, P of 

o ( canopy 

BACKGROUND usually decrease 

1999, , 1999, ) 

in , which WEISS structural WEISS leaves et et properties, stand is resulting divided as e.g. into in a 

leaf area a downward finite density, number pattern 

leaf of angle statistically of light 

and 

irect ilistic methods models of usually reduction canopy refer geometry to are 

distribution (SAEKI optical into 

of canopy 1975). measurements 

elements), The relationship of radiation 

one of the between best light 

iage endently y is is divided seen divided at of into point each into foliage N other 0 statistically in zenith distribution 

(NILSON independent independent and among 

1999). azimuth If layers, others. 

N layers, 

directions 

in , in The 

the which key 

number leaves from leaves variable 

of are overlaps are in independent reduction 

those models layers, and 

is gap a canopy positive geometry 

fraction binomial can 

P o ( , 

) 

, model be 

which for 

et 04). sformation al. Indirect 2001, into JONCKHEERE methods characterized LAI is usually then et refer performed al. 2004, to optical either measurements by a calibrated of radiation 

known THEORETICAL and most by BACKGROUND 

frequently statistical applied models in that the estimation describe the regular probability dispersion of radiation of foliage, interception a negative (or binomial non- 

. son The MAN kes tly ntly 2004). of statistic distribution. of transformation 1991, each other ZHANG and other describes The (NILSON gap the fraction 1999). proportion If is If N then N 

of the 

, 

sky , the probability the vs. number foliage of for of seen overlaps zero overlaps at overlaps point 0 in zenith and azimuth directions from 

interception) probabilistic et into al. LAI 2005, 

of LAI is the within models is among then 

Poisson the of performed 

model canopy. canopy 

(see Usually geometry either 

e.g. Monsi these by into a calibrated 

& models Saeki are model based on for simplified clumped assumptions dispersion concerning of foliage) the and 

ually stribution. and istribution. l al. that et (1985) low refer al. takes 2005, labour to be The statistic optical described FRAZER gap beneath gap fraction as the canopy is is then (JONCKHEERE the the probability et for for al. zero 2004). zero overlaps 

tatistical models radiative costs, measurements and many et 

1953, exist Mussche (each and probabilistic al. difficulties spatial 2001, of 

characterizing et al. characteristics JONCKHEERE radiation models of canopy 

2001, Nilson a different 1999, of et canopy al. geometry 2004, 

statistical Weiss structural into 

et al. properties, Markov models as e.g. (layers leaf area are density, dependent, leaf i.e. angle probability and 

S es. 85) LAI 985) RIKSSON 1990, Major be be is described WELLES then described et problems performed al. as as 

f the best known foliage Though & 2005, NORMAN include FRAZER 

2004). and most It distribution a either 

assumes large unknown 1991, by et 

frequently number a al. ZHANG 

that among calibrated 2001, 

applied the of canopy others. et JONCKHEERE different al. 2005, 

in the is The estimation divided statistical among key et 

into variable al. 

of models 2004, 

N in of exist those transmission (each models characterizing differs is gap according fraction a different P to o ( contact 

, ) 

statistical , which or non- 

vantage e probabilistic WELLES contribution of 1990, convenience models 

SI & SAEKI 1953, describes distribution WELLES of woody of and canopy & material 

statistically MUSSCHE the 

NORMAN of low 

independent proportion et canopy geometry labour to 1991, 

al. 2001, elements), costs, 

layers, of into ZHANG 

NILSON sky 

many 

in vs. one 1999, which foliage 

difficulties et of al. the 2005, 

WEISS leaves seen best among 

et at are known point (3.1) contact and 0 in most zenith in frequently adjacent and azimuth layers). applied in The the directions Poisson estimation model from of is 

ptical ransfer e AZER the advantage measurements et in al. forest 2001, (3.1) 

ough divided the into canopy N statistically LAI of canopies. convenience JONCKHEERE 

dispersed is will the be 

in direction randomly independent Poisson Major described and et problems model low al. 2004, labour 

( , ) 

, and u is layers, independently (see include 

the leaf in e.g. costs, 

area which MONSI unknown many 

density of leaves each & difficulties 

beneath the canopy (JONCKHEERE et al. 2004). SAEKI 

(defined are other 1953, MUSSCHE et al. 2001, NILSON 1999, WEISS et 

as the simplest of the above-mentioned models, as no 

& el iation NORMAN assumptions transfer 1991, 

of 

photosynthetic 

he the each canopy other in in direction (NILSON direction al. in and 

(Nilson 2004). forest ZHANG the 

( 1999). tissue per 

( , It canopies. contribution et 

, ) 

assumes al. 2005, Major 

unit 

, 

) u , If u 

canopy 

is is N the the that 

among of 

leaf leaf 

, the problems woody 

the volume), 

area number canopy material 

number density is include 

of of 

and G 

(defined divided to unknown 

overlaps (defined overlaps into 

( , ) 

is 

as can as 

the 

the the N statistically additional independent parameters layers, (as e.g. in layer which thickness) leaves are are 

per ckground nience model and of assumptions low indirect Though mean 


ynthetic 

d ution. 

synthetic 

by plant The 

direction 

tissue organs. gap dispersed labour optical costs, a and large 

per 

be fraction 

( , 

per 

described Consequently 

) . 

unit is randomly many measurements the number contribution difficulties of different 

canopy then by the 

volume), 

the light and 

volume), 

probability Poisson independently will of be woody described statistical 

and distribution. 

G 

for 

G ( ( 

zero of 

, , ) 

) 

is 

overlaps each material models 

is 

The other to exist (each characterizing a different statistical 

the the mean 

gap (NILSON 1999). If N , the number of overlaps 

mean 

needed. For further reading see also Weiss et al. 

etical canopies. background Major distribution 

gh be this described process as can problems of 

fraction be indirect of 

quite described include canopy optical 

is then variable the by probability the unknown elements), measurements one of 

in Poisson for distribution. will the be best described known and most frequently applied in the estimation of 

zero overlaps The and gap fraction (2004). is then the probability for zero overlaps 

irection ection and the ( ( , contribution , ) 

. 

) . LAI is the 

nt sulting r radiation along in the a is downward path intercepted and can and according can of Poisson 

LAI pattern according woody 

by being plant to of Lang defined to material model (see 

light organs. LANG et al. as (1985) et to e.g. MONSI & SAEKI 1953, MUSSCHE et al. 2001, NILSON 1999, WEISS et 

Consequently al. (1985) be described be light described as as 

irect 

dent ction g ng th the 

optical 

of the solar path the and path canopy. radiation canopy and 

measurements al. 2004). It assumes 

and LAI Though is geometry being intercepted being this defined 

will be that 

defined can process by as 

described the canopy is divided into N statistically independent layers, in which leaves are 

be as plant can be organs. quite Consequently variable 

(3.1) in light Based on the above-described models, measurements 

dispersed randomly ( 

uG 

, dand 

) independently of each other (NILSON 1999). If N , the number of overlaps 

0 

rtical sities y of depth usually radiation of decrease the Pinterception 

o ( canopy. , 

resulting 

) eThough 

(or in non- a this downward , process pattern can be quite of light variable (3.1) in (3.2) of gap fraction (or transmittance) at ground (3.1) level can 

anopy in direction can ( be , described 

) 

tercepted flux ip simplified between densities by assumptions plant light usually organs. reduction decrease concerning 

, u 

Consequently 

is the by 

resulting and 

leaf the 

the canopy 

area Poisson 

in light a 

density distribution. 

downward geometry 

(defined 

can pattern 

as 

be 

the The gap fraction is then the probability for zero overlaps 

(3.2) of (3.2) light be used to derive LAI. Several detailed reviews, as 

and 

hetic tissue per 

where can 

unit canopy 

according is the 

volume), 

path to length LANG 

and 

through et al. (1985) 

G ( , 


) is 

canopy be described 

the mean 

in direction as ( , ) 

, u is the leaf area density (defined as the 

y. elationship describe as Though e.g. leaf the this area between probability process where density, light can ε leaf is of be the reduction angle quite radiation path and variable length and interception in canopy through geometry the (or canopy non- can in be e.g. Jonckheere et al. (2004) or Bréda (2003), describe 

z 

on rease ight ese ls ( and , ) 

, it follows after LANG et al (1985) that the gap fraction is 

z. 

total one-sided leaf area of photosynthetic tissue per unit canopy volume), and G ( , ) 

odels that models resulting is describe 

is the mean 

zgap 

are cos fraction in based direction a the downward on probability P osimplified 

( ( 

, 

u) 

Gpattern 

, u which , is 

of assumptions dthe 

) radiation of leaf light area concerning interception density (defined the (or as non- these respective sensors together with their advantages 

0 

d nd , it , it follows after LANG et et al al (1985) that the the gap fraction is is 

ith sually py t reduction structural these cos and cos azimuth models 

properties, and projection 

Po 

( , 

) 

e 

the canopy total are directions based one-sided as of e.g. geometry unit on leaf simplified leaf 

from leaf area area 

, 

can area density, in assumptions be of the photosynthetic direction leaf angle concerning ( and , tissue 

) . 

(3.1) 

the and problems. Instruments for transmittance 

e path and LAI being defined as 

key obability of canopy variable of structural in radiation where per those unit properties, models interception canopy is volume), as gap e.g. (or fraction leaf non- and area P Go 

density, ( , 

) 

, is which leaf the angle mean and 

If we consider 

is the path 

G ( , 

length 

) 

constant 

through 

along 



canopy 

path 

in 

and 

direction measurements 

LAI being 

( , ) 

defined 

, u is the include 

as 

leaf area e.g. density the Sunfleck (defined or AccuPAR as the 

e ased s. seen The on at key simplified point variable total 0 projection in assumptions zenith one-sided in those of unit and models concerning leaf leaf azimuth area area is of in gap the the photosynthetic directions fraction direction P o ( from , 

tissue ) 

. , which per 

ch characterizing a different (3.2) (3.3) unit Ceptometers canopy volume), (Decagon and Devices G ( , ) 

is Inc., the mean Pullman, 

Z statistical 

s. perties, 

04). foliage as seen e.g. leaf at point area 0 density, in 

LAI 

(3.3) 

frequently applied in the u dz estimation , 

zenith leaf angle and and 

projection of unit leaf azimuth directions from WA, USA), the Demon device (CSIRO, Canberra, 

of area in the direction ( , ) 

. 

(3.2) 

those et al. models 2004). is If gap we consider fraction 0 P Go 

( , 

) 

, constant which along the path ε and Australia), or the LAI-2000 Plant Canopy Analyzer 

HE istical et zal. 

models 2001, NILSON exist (each 1999, characterizing WEISS et a different statistical 

in zenith , it follows and If LAI we azimuth after being consider LANG defined G directions et ( al as , (1985) ) 

constant from that the along gap the fraction path is and LAI (LI-COR being defined Inc., as NE, USA). A study from Dufrêne & 

rent independent e best cosstatistical 

known 

layers, and models most in which exist frequently (each leaves applied characterizing are in the estimation a different zof 

statistical 

where z is the total canopy height and , it follows Bréda after (1995) LANG shows et al (1985) good agreement that the gap between fraction Demon is 

Z 

one 9). & SAEKI If of N the 1953, 

best , the known MUSSCHE number and et most of al. overlaps 2001, frequently NILSON applied 1999, in WEISS the estimation coset 


LAI 

xist . ided then MONSI (each into the & probability N characterizing SAEKI statistically described u dz , (3.2) and LAI-2000 Plant Canopy Analyzer (PCA). (3.2) With 

1953, for independent MUSSCHE zero a by 0 different overlaps et layers, statistical al. 2001, in which NILSON leaves 1999, are WEISS et 

respect to this thesis the main disadvantage of the 

each py d most is other divided frequently (NILSON into N applied 1999). statistically in If the N independent estimation LAI 

, the number of layers, of in overlaps which 

 

(3.3) leaves are 

G 

, 

cos 

ion. ently MUSSCHE The of each gap et al. other fraction where P2001, 

( , 

) 

 

z 

o (NILSON is NILSON z is then ethe 

1999). the total 1999, probability canopy If WEISS . N height 

et 

for , the zero and number overlaps of overlaps , it follows after LANG et al (1985) that the gap fraction (3.3) is 

cos 

istically distribution. described independent as The gap layers, fraction in which is then (3.1) leaves the probability are for zero overlaps 

described by 

1985) ON 1999). be described If N as 

, the number of overlaps 

u 

ction 

is the 

is 

leaf 

then 

area 


density 

probability 

(defined 

for zero 

as LAI the 

G 

, 

 

overlaps 

cos 

(3.1) 

py volume), and PoG 

( 

( 

, 

, 

) 

 

 

) 

eis 

the mean . (3.3) 

(3.1)

quadrants of 0°, 45°, 90°, 180°, or 270° (cf. Figure 3-2). Due to a 

34 

wavelengths between 0.32 µm and 0.49 µm is recorded (LI-COR 1992 

Measurements are usually taken above and below the canopy to coll 

sky radiation (no direct radiation present). Transmittance can subseq 

Demon and the AccuPAR Ceptometer is that data 

I trans ( ) 

must be collected over several hours over the course 9 ( ) , 

(3.4) 

I o ( ) 

of the day to capture an adequate range of sun angles. 

Gap fraction can further be calculated directly from where where I trans ( ) is the intensity of the transmitted light as measured 

fish eye photographs (e.g. Jonckheere et al. 2005, intensity of the hemispheric light source (i.e. the sun) as measured 

Leblanc et al. 2005, Zhang et al. 2005), which are function of zenith angle . 

independent of illumination conditions (Baret 2005, 

personal communication). 

3.3.2 Instruments 

In order to determine LAI in the field, LAI-2000 PCA 

and digital hemispherical photographs were used in 

this thesis. Both are standard methods applied in the 

CEOS, VALERI and BigFoot validation efforts and 

will thus be described in more detail. 

The standard method to 

of the above-mentioned Poisson model. In wavelengths where 

radiation, leaves are assumed to be non-reflecting and non-transmitti 

LAI-2000 Plant Canopy Analyzer 

fraction and is further averaged over azimuth, Equation (3.3) become 

LAI 

The LAI-2000 Plant Canopy Analyzer (LI-COR, 

G 

cos 

o e 

Lincoln, Nebraska, cf. Figure 3-2) records light 

intensity with an optical sensor (LAI-2050) consisting 

of five concentric silicon detectors with fields of view 

between 0° and 74.1°. The nominal coverage of each 

detector is shown in Table 3-2. Incoming radiation is 

projected through a fisheye lens onto the photoelectric 

sensor rings, where it is integrated over the complete 

azimuth range and stored in the associated control 

unit (LAI-2070). In order to avoid unwanted effects, 

such as operator shadow, the 360° azimuthal field of 

view may be restricted with view caps into quadrants 

of 0°, 45°, 90°, 180°, or 270° (cf. Figure 3-2). Due 

to an optical filter, only radiation with wavelengths 

between 0.32 µm and 0.49 µm is recorded (LI-COR 

1992). 

Measurements are usually taken above and below the 

canopy to collect above and below canopy diffuse sky 

radiation (no direct radiation present). Transmittance 

τ can subsequently be calculated for each ring as 

 

, 

which can also be written as 

G ( ) LAI ln 

( ) cos 

. 

For flat leaves with a random distribution of foliage and leaf azimuth 

/ 2 

G ( ) sind 

0. 

5 . 

0 

Inserting Equation 3.6 in 3.7 leads to 

/ 2 

LAI 2 lnocos 

sin 

d 

0 

 


is the intensity of the transmitted 

where I trans ( ) is the intensity of the transmitted light as as measured below the canopy and I o ( ) is the 

intensity of the hemispheric light source (i.e. the sun) as measured above the canopy, both varying as a 

function of zenith angle . 

Table 3-2: Nom 

rings of LAI-2000 

LAI-2000 PCA 

1 

2 

3 

4 

5 

Figure 3-2: LI-CO 

The standard method to calculate LAI is based on inversion 

of the above-mentioned Poisson model. In wavelengths where the LAI-2000 PCA sensor records 

radiation, leaves are assumed to be non-reflecting and non-transmitting. As ( ) is thus equivalent to gap 

fraction and is further averaged over azimuth, Equation (3.3) becomes 

LAI 

G 

cos 

o e 

. 

WELLES & NORMAN (1991) offer a numerical solution for Equation 3 

who introduced the measurement of light transmittance in different s 

transmittance data for the five different rings: 

5 

LAI 2 ln 

i cos Wi 

, 

i1 

where i stands for rings 1-5 and Wi are normalized weighting factors a 

with the different rings. Whereas cos values are constant for each 

viewing angle, weighting factors can vary according to the number of 

3-3 shows the W i factors in the event that all five rings are employe 

used in this thesis, cf. Chapter 5.2.1). The W i factors differ slightly fo 

 

, (3.5) 


G ( ) LAI ln 

( ) cos 

. (3.6) 

For flat leaves with a random distribution of foliage and leaf azimuth angles, MILLER (1967) showed that 

/ 2 

G ( ) sind 

0. 

5 . (3.7) 

0 


/ 2 

LAI 2 lnocos 

sin 

d 

0 

 

foliage angle distribution, improper model assumptions and the contribution of woody material to 


radiation interception. The theoretical background of indirect optical measurements will be described 

where 

briefly 

I 

in the following. 

trans ( ) is the intensity of the transmitted light as measured below the canopy and I o ( ) is the 

intensity As it of penetrates the hemispheric the canopy, light incident source (i.e. solar the radiation sun) as is measured intercepted above by the plant canopy, organs. both Consequently varying as a light 

function attenuation of zenith corresponds angle . to the vertical depth of the canopy. Though this process can be quite variable in 

space and time, mean horizontal flux densities usually decrease resulting in a downward pattern of light 

reduction (SAEKI 1975). The relationship between light reduction and canopy geometry can be 

characterized by statistical models that describe the probability of radiation interception (or noninterception) 

within the canopy. Usually these models are based on simplified assumptions concerning is the 

radiative and spatial characteristics of canopy structural the intensity properties, of the hemispheric as e.g. leaf area light density, source (i.e. leaf the angle and 

foliage distribution among others. The key sun) variable as measured in those above models the canopy, is gap fraction both varying P o ( , 

as ) 

a , which 

describes the proportion of sky vs. foliage seen function Table at point of 3-2: zenith 0 in Nominal zenith angle angular . and azimuth coverage directions of the different from 

beneath the canopy (JONCKHEERE et al. 2004). rings of LAI-2000 PCA (LI-COR 1992) 

The standard 

LAI-2000 

method 

PCA ring 

to calculate 

Aangular 

LAI is 

coverage 

based on 

Though a large number of different statistical models exist (each characterizing a different statistical [°] 

inversion of the above-mentioned Poisson model. In 

distribution of canopy elements), one of the best known and 1 most frequently applied 0.0°-12.3° in the estimation of 

The standard wavelengths method where to the calculate LAI-2000 LAI PCA is based sensor on records inversion 

LAI is the Poisson model (see e.g. MONSI & SAEKI 1953, MUSSCHE 2 et al. 2001, NILSON 16.7°-28.6° 1999, WEISS et 

of the above-mentioned Poisson model. In wavelengths radiation, leaves where are the assumed LAI-2000 to PCA be non-reflecting sensor records 

al. 2004). It assumes that the canopy is divided into N statistically 3 independent layers, 32.4°-43.4° in which leaves are 

radiation, leaves are assumed to be non-reflecting and non-transmitting. As ( ) is thus equivalent to gap 

dispersed randomly and independently of each other (NILSON 4 1999). If N , the 47.3°-58.1° number of overlaps 


can be described by the Poisson distribution. The gap fraction 5 is then the probability 62.3°-74.1° for zero overlaps 

and can according LAI 

G 

to LANG et al. (1985) be described as 

cos 

Figure 3-2: LI-COR LAI-2000 PCA (LI-COR 2007) 

o e 

 

( 

uG, 

d 

) 

0 

Po 

( , 

) 

e , (3.1) 

where is the path length through the canopy in direction ( , ) 

, u is the leaf area density (defined as the 

total one-sided leaf area of photosynthetic tissue per unit canopy volume), and G ( , ) 

is the mean 

projection of unit leaf area in the direction ( , ) 

. 


constant along the path and LAI being defined as 

Z 

LAI u dz , (3.2) 

0 

z 

where z is the total canopy height and , it follows after LANG et al (1985) that the gap fraction is 

cos 

described by 

LAI 

G, 

 

cos 

P 

 

o ( , 

) 

e . (3.3) 

. (3.8) 

WELLES & NORMAN (1991) offer a numerical solution for Equation 3.8. It is based on LANG et al. (1985), 

who introduced the measurement of light transmittance in different sun angles, and includes the measured 


5 

LAI 2 ln 

i cos Wi 

, (3.9) 

i1 

where i stands for rings 1-5 and Wi are normalized weighting factors associated to the relative area covered 

with the different rings. Whereas cos values are constant for each ring as they depend on the zenithal 

viewing angle, weighting factors can vary according to the number of rings used for LAI calculation. Table 

3-3 shows the W i factors in the event that all five rings are employed, as well as the use of four rings (as 

used in this thesis, cf. Chapter 5.2.1). The W i factors differ slightly for both cases, but they always sum up 

 

, (3.5) 


G ( ) LAI ln 

( ) cos 

. (3.6) 

For flat leaves with a random distribution of foliage and leaf azimuth angles, MILLER (1967) showed that 

/ 2 

G ( ) sind 

0. 

5 . (3.7) 

0 


/ 2 

LAI 2 lnocos 

sin 

d 

0 

 

9 

where I trans ( ) is the intensity of the transmitted light as measured 

intensity of the hemispheric light source (i.e. the sun) as measured a 


Table 3-2: Nominal angular coverage of the different 

rings of LAI-2000 PCA (LI-COR 1992) 

LAI-2000 PCA ring Aangular coverage [°] 

1 0.0°-12.3° 

2 16.7°-28.6° 

3 32.4°-43.4° 

4 47.3°-58.1° 

5 62.3°-74.1° 

Figure 3-2: LI-COR LAI-2000 PCA (LI-COR 2007) 


of the above-mentioned Poisson is thus model. equivalent In wavelengths to where t 

radiation, gap fraction leaves and are is assumed further averaged to be non-reflecting over azimuth, and non-transmittin 

fraction Equation and (3.3) is further becomes averaged over azimuth, Equation (3.3) becomes 

LAI 

G 

cos 

o e 

. (3.8) 

WELLES & NORMAN (1991) offer a numerical solution for Equation 3.8. It is based on LANG et al. (1985), 

who introduced the measurement of light transmittance in different sun angles, and includes the measured 


5 

LAI 2 ln 

i cos Wi 

, (3.9) 

i1 






 

, 


G ( ) LAI ln 

( ) cos 

. 

For flat leaves with a random distribution of foliage and leaf azimuth a 

/ 2 

G ( ) sind 

0. 

5 . 

0 


/ 2 

LAI 2 lnocos 

sin 

d 

0 

 

9 






LAI-2000 PCA 

1 

2 

3 

4 

5 



of the above-mentioned Poisson model. In wavelengths where t 

radiation, leaves are assumed to be non-reflecting and non-transmittin 


(3.5) 

LAI 

G 

cos 

o e 


. 


who introduced the measurement of light transmittance in different su 


5 

LAI 2 ln 

i cos Wi 

, 

i1 




3-3 shows the W i factors in the event that all five rings are employed 

used in this thesis, cf. Chapter 5.2.1). The W i factors differ slightly for 

 

, 


G ( ) LAI ln 

( ) cos 

. 


/ 2 

G ( ) sind 

0. 

5 . 

0 


/ 2 

LAI 2 lnocos 

sin 

d 

0 

 

9 






LAI-2000 PCA 

1 

2 

3 

4 

5 






LAI 

G 

cos 

o e 

(3.6) 

For flat leaves with a random distribution of foliage 

and leaf azimuth angles, Miller (1967) showed that 

. 




5 

LAI 2 ln 

i cos Wi 

, 

i1 






 

, 


G ( ) LAI ln 

( ) cos 

. 


/ 2 

G ( ) sind 

0. 

5 . 

0 


/ 2 

LAI 2 lnocos 

sin 

d 

0 

 

9 






LAI-2000 PCA 

1 

2 

3 

4 

5 






LAI 

G 

cos 

o e 

(3.7) 


. 




5 

LAI 2 ln 

i cos Wi 

, 

i1 






 

, 


G ( ) LAI ln 

( ) cos 

. 


/ 2 

G ( ) sind 

0. 

5 . 

0 


/ 2 

LAI 2 lnocos 

sin 

d 

0 

 

9 


intensity of the hemispheric light source (i.e. the sun) Table as measured 3-2: Noma 



LAI-2000 PCA 

1 

2 

3 

4 

5 






LAI 

G 

cos 

o e 

. 

(3.8) 

WELLES Welles & Norman NORMAN (1991) offer offer a numerical a numerical solution 

for Equation 3 

who for Equation introduced 3.8. the It is measurement based on Lang of et light al. (1985), transmittance who in different su 

transmittance introduced the data measurement for the five of different light transmittance 

rings: 

in different sun angles, and includes the measured 

5 

transmittance LAI 2 lndata 

i cos for the 

Wi 

five , different rings: 

i1 






 

, 


G ( ) LAI ln 

( ) cos 

. 


/ 2 

G ( ) sind 

0. 

5 . 

0 


/ 2 

LAI 2 lnocos 

sin 

d 

0 

 

9 

where I trans ( ) is the intensity of the transmitted light as 

intensity of the hemispheric light source Table (i.e. the 3-2: sun) Nom as 



LAI-2000 PCA 

1 

2 

3 

4 

5 


The standard m 

of the above-mentioned Poisson model. In wavelength 

radiation, leaves are assumed to be non-reflecting and non 

fraction and is further averaged over azimuth, Equation (3. 

LAI 

G 

cos 

o e 

. 




5 

LAI 2 ln 

i cos Wi 

, 

(3.9) 

i1 

where i i stands for rings 1-5 1-5 and and Wi are normalized weighting factors a 





 

, 


G ( ) LAI ln 

( ) cos 

. 

For flat leaves with a random distribution of foliage and lea 

/ 2 

G ( ) sind 

0. 

5 . 

0 


/ 2 

LAI 2 lnocos 

sin 

d 

0 

 

Table 

rings o 

LAI 

Figure 

. 

WELLES & NORMAN (1991) offer a numerical solution for 

who introduced the measurement of light transmittance in 


5 

LAI 2 ln 

i cos Wi 

, are normalized 

i1 

weighting factors associated to the relative area 

where i stands for rings 1-5 and Wi are normalized weightin 

with the different rings. Whereas cos values are constan 

viewing angle, weighting factors can vary according to the n 

3-3 shows the W i factors in the event that all five rings ar 

used in this thesis, cf. Chapter 5.2.1). The W i factors differ



35 

/ 2 

LAI 2 lnocos 

sin 

d 

0 

covered with the different rings. Whereas cos θ 

values are constant for each ring as they depend on 

the zenithal viewing angle, weighting factors can 

vary according to the number of rings used for LAI 

calculation. Table 3-3 shows the 

 

LAI 2 ln o cos 

sin 

d 

0 

. (3.8) 

WELLES & NORMAN (1991) offer a numerical use solution of four for rings Equation (as used 3.8. in It is this based thesis, on cf. LANG Chapter et al. (1985), 

who introduced the measurement of light transmittance 5.2.1). The in different sun angles, and includes the measured 


5 

LAI 2 ln 

i cos Wi 

, factors in the 

(3.9) 

i1 

event that all five rings are employed, as well as the 






. 

WELLES & NORMAN (1991) offer a numerical solution for Equation 3.8. It is 

who introduced the measurement of light transmittance in different sun angle 


5 

LAI 2 ln 

i cos Wi 

, factors differ slightly for both cases, 

i1 

but they always sum up to unity. LAI is therefore the 

where i stands weighted for rings sum 1-5 of and measured Wi are normalized canopy transmittance, 

weighting factors associate 

with the different where the rings. highest Whereas weight cos is put values on the are outmost constant ring for as each ring as 

viewing angle, it covers weighting the largest factors part can of vary the according canopy (as to illustrated the number of rings us 

3-3 shows the in Figure W i factors 3-3). in the event that all five rings are employed, as we 

used in this thesis, cf. Chapter 5.2.1). The W i factors differ slightly for both c 

Over the last 15 years the LAI-2000 PCA has gained 

wide acceptance. Though several constraints remain 

(cf. Chapter 3.3.3), the device has been used in a 

large number of studies in forested environments. 

Examples for applications in boreal forests can be 

found in Chen et al. (1997), Fassnacht et al. (1994), 

Stenberg et al. (1994). Cutini et al. (1998), Dufrêne & 

Bréda (1995), and Eriksson et al. (2005). Strachan & 

McCaughey (1996) used the LAI-2000 PCA for LAI 

determination in deciduous forests and De Wasseige 

et al. (2003) and Kalácska et al. (2005) in tropical 

rain forests. 

Figure 3-2 

Table 3-2 

LI-COR LAI-2000 PCA (LI-COR 2007). 

Nominal angular coverage of the different rings 

of LAI-2000 PCA (LI-COR 1992). 

LAI-2000 PCA ring Angular coverage [°] 

1 0.0°-12.3° 

2 16.7°-28.6° 

3 32.4°-43.4° 

4 47.3°-58.1° 

5 62.3°-74.1° 

Table 3-3 cos θ and W values for LAI-2000 PCA rings 1-5 

i 

and rings 1-4 respectively (LI-COR 1992). 

Ring cos θ W i (rings 1-5) W i (rings 1-4) 

1 0.993 0.034 0.034 

2 0.921 0.104 0.103 

3 0.788 0.160 0.158 

4 0.602 0.218 0.706 

5 0.375 0.484 n/a 

 

 

Digital hemispherical photography 

Gap fraction (and in consequence LAI) can also 

be derived from hemispherical photography. 

Hemispherical photographs have been used since the 

1960s to study canopy structural information, with 

Anderson (1964) being the first to compute light 

penetration through the canopy with that technique. 

Hemispherical pictures have been analysed with 

regard to LAI by various authors (e.g. Anderson 

1981, Bonhomme et al. 1974, Jonckheere et al. 2005, 

Leblanc et al. 2005, Van Gardingen et al. 1999, Zhang 

et al. 2005 among others), with studies on forest LAI 

obtained by Leblanc et al. 2005, Jonckheere et al. 

2005 (boreal forests), Neumann et al. 1989, Wang & 

Miller 1987 (temperate deciduous forests), and Frazer 

et al. 2001 (temperate mixed forests). 

Pictures are taken through a fisheye lens (cf. Figure 

3-4) from below the canopy or, in the case of low

36 

View angle (θ) 7° 23° 38° 53° 68° 

Figure 3-3 

Illustration of the different viewing angles of the LAI-2000 PCA. 

canopies or understorey vegetation, placed above 

looking downward. The hemispherical view results 

in an angular field of sight of –90 to 90° in zenith and 

360° in azimuth directions (cf. Figure 3-5). Levelling 

guarantees that the camera is oriented towards 

the zenith, allowing for a detailed analysis of the 

zenithal and azimuthal variation of gap fraction. LAI 

is subsequently derived from inversion of Equation 

3.3, e.g. applying a LUT approach and certain 

assumptions concerning leaf angle distribution 

(further details will be given in Chapter 5.2.2). 

In contrast to the LAI-2000 PCA, hemispherical 

photographs provide not only a single gap fraction 

value for each image, but also a graphic record of 

site-, species- and age-specific differences in canopy 

architecture (Jonckheere et al. 2004). 

The development of digital cameras eliminated 

potential error sources resulting from the development 

and scanning of films needed for analogue techniques. 

Digital cameras nowadays have a high spatial 

resolution (decreasing the problem of mixed pixels) 

and a better radiometric image quality than analogue 

cameras. Critical steps in acquisition and analysis of 

digital hemispherical photographs (DHP), however, 

remain the selection of an adequate exposure time and 

shutter speed (see Zhang et al. 2005), as well as the 

right threshold to distinguish vegetation from gaps. 

As the latter is somewhat arbitrary and subjective 

it is a potential source of error for gap fraction and 

subsequent LAI retrieval. 

Figure 3-4 Digital camera with fish-eye lens. 

Figure 3-5 Hemispherical photograph taken in Budongo 

Forest, Nature Reserve on 11 October, 2005.


Several software packages have been developed for 

the processing of DHPs, such as HemiView (Delta-T 

Devices Ltd., Cambridge, UK), Gap Light Analyzer 

(GLA, Frazer et al. 1999), WinSCANOPY (Régent 

Instruments Inc., Canada) and CAN-EYE (Baret 

2004a). 

3.3.3 

Limitations of indirect 

optical methods 

The assessment of leaf area index using indirect 

optical methods offers the possibility of fast and 

non-destructive sampling over large areas compared 

to direct or semi-direct methods. However, certain 

assumptions and simplifications are made concerning 

the structural and radiative properties of plant 

canopies that are sometimes unrealistic. As a result, 

errors associated with the respective LAI estimation 

have been recorded; these have been described and 

analysed by various authors (e.g. Cutini et al. 1998, 

Hyer & Goetz 2004 and Inoue et al. 2004 for LAI- 

2000 PCA; Frazer et al. 2001 and Zhang et al. 2005 

for hemispherical photography, Jonckheere et al. 

2005, and Mussche et al. 2001 for both). 

Usually validation of indirect optical methods is 

performed by comparison to direct or semi-direct 

approaches (e.g. Cutini et al. 1998, Leblanc & 

Chen 2001, Mussche et al. 2001). Although a high 

correlation usually indicates a good sensitivity of 

optical instruments, a general underestimation of 

LAI with respect to direct and semi-direct methods 

cannot be disregarded. This underestimation seems 

to be higher for coniferous stands than for broadleaf 

vegetation types (Nilson 1999), but can still result in 

an underestimation of “true LAI” of up to 30% for 

the latter (Cutini et al. 1998) under certain conditions. 

The main limitations to LAI assessment with indirect 

optical methods will therefore be described in the 

following section together with appropriate correction 

methods. 

Simplification of leaf optical properties 

37 

When LAI is calculated from transmittance 

measurements (e.g. from LAI-2000 PCA data), one 

of the main assumptions used to derive Equation 3.5 

is that foliage elements are only absorbing incoming 

radiation, i.e. they are non-transmitting and nonreflecting. 

Transmission recorded below the canopy 

is equivalent to gap fraction only if these conditions 

are fulfilled. If the conditions are not met, radiometric 

measurements will be biased and lead to an 

overestimation of below canopy light transmittance. 

In consequence LAI will be underestimated (Hyer & 

Goetz 2004). Although the LAI-2000 PCA records 

radiation in wavelengths, where foliage absorption 

reaches 95% and more, a small amount is also 

transmitted and reflected. Chen (2001) estimated the 

bias due to the ignorance of first and second order 

scattering effects to be around 8%. 

Several studies recommended the exclusion of certain 

rings from calculation in order to retrieve better 

results. Leblanc & Chen (2001) e.g. found that LAI 

measured with LAI-2000 PCA under direct sunlight 

conditions was underestimated by 20% compared 

to a canopy radiative transfer model. To correct for 

that, they developed an empirical correction of LAI 

as a function of sun zenith angle θ . They state that 

sun 

LAI calculated only from ring 4 proved to be more 

consistent. Planchais & Pontailler (1999), however, 

found in a both theoretical and experimental analysis 

that the underestimation of LAI must rather result 

from an inappropriate use of the Poisson model. 

Clumping effects 

Foliage in plant canopies – and especially forest 

canopies – is usually not distributed randomly, either 

horizontally or vertically as assumed by the Poisson 

model. In fact, leaves are grouped along branches 

and branches are associated with boles resulting in a

38 

clumped distribution of canopy elements at various LAI is therefore the true variable measured with 

e 

scales (Hyer & Goetz 2004). This clumped and thus optical devices if no correction for clumping is 

non-random spatial distribution has consequences for 

light transmittance through the canopy. Clumping 

applied. 

results in a higher canopy transmittance than Various studies have dealt with the correction for non- 

13 predicted by random models (such as the Poisson randomness and the THEORETICAL derivation of BACKGROUND 

λ, especially when 

model, which assumes random distribution) and indirect optical methods were applied to coniferous 

thus in an underestimation of LAI (Black et al. 1991, canopies. According to Walter et al. (2003) they can 

Chen & Cihlar 1995a, Fassnacht et al. 1994, among be grouped into optical-spectral methods (e.g. Chen 

others). Though clumping is higher in conifer forests et al. 1997, Fassnacht et al. 1994, Kuusk et al. 2002, 

compared 13 to broadleaf forests as conifer needles are Smith et al. 1993 or THEORETICAL Stenberg et al. 1994) BACKGROUND and methods 

additionally grouped closely in shoots (see Figure incorporating stand dimension analyses (Kucharik et 

3-6 for illustration), it also affects measurements in 

broadleaf forests. 

al. 1999, Nilson 1999). 

Chen & Cihlar (1995a) developed a procedure to 

Figure 3-6: Effect of foliage clustering on gap fraction according to NILSON (1999). Same amount of 

needles 

According 

is dispersed 

to Nilson 

(a) 

(1971) 

randomly 

leaf 

and 

area 

b) 

index 

clustered. 

can be derive a clumping index based on gap size and gap 

derived from gap fraction in the presence of clumping fraction analyses. The underlying theory is that within 

According through an to adjustment NILSON (1971) of Equation leaf area 3.3, index that includes can be derived a random from gap canopy fraction the in probability the presence of of the clumping occurrence 

through the clumping an adjustment index λ, of Equation 3.3, that includes the clumping of large gaps index can , be derived from the distribution of 

LAI 

gap sizes. The iterative removal of large gaps from 

Figure 3-6: Effect 

G, 

 

cos of 

 

foliage P , 

 

clustering on gap fraction according to NILSON (1999). Same amount of 

o e 

(3.10) the total gap fraction is equivalent to the (3.10) design of 

needles is dispersed (a) randomly and b) clustered. 

an imaginary plant canopy with gaps inserted at 

with < 1 in the case of aggregated canopies, > 1 for regular canopies and 1 in the case of true random 

with According λ < 1 in to the NILSON case of (1971) aggregated leaf area canopies, index can > 1 be for derived random from gap and fraction having the in the same presence total gap of clumping fraction as 

leaf spatial distribution (BLACK et al. 1991). It is clear that if there is no account for clumping in non- 

regular through canopies an adjustment and 1 in of the Equation case of 3.3, true that random includes leaf the the clumping real canopy. index , From this new canopy the canopy 

random canopies, the LAI derived from optical measurements will be under- or overestimated. To 

spatial distribution (Black et al. 1991). It is clear that element area index is calculated and compared to the 

LAI 

account for that, 

G, 

BLACK et al. (1991) introduced the term “effective leaf area index”, which is defined as 

if there is 

no account cos 

P , 

 


clumping in non-random 

o e 

original canopy to derive the clumping index. (3.10) Gap 

LAIe= canopies, LAI the LAI derived from optical measurements size distribution may be estimated with the (3.11) TRAC 

with < 1 in the case of aggregated canopies, > 1 for regular canopies and 1 in the case of true random 

LAIe 

will be is therefore under- or the overestimated. true variable To measured account with for that, optical devices instrument if no (Tracing correction Radiation for clumping and Architecture is applied. of 

leaf spatial distribution (BLACK et al. 1991). It is clear that if there is no account for clumping in non- 

Black et al. (1991) introduced the term “effective leaf Canopies, 3rd Wave Engineering). 

Various random studies canopies, have the dealt LAI with derived the correction from optical for measurements non-randomness will and be the under- derivation or overestimated. of , especially To 

area index”, which is defined as 

when account indirect for that, optical BLACK methods et al. (1991) were applied introduced to coniferous the term “effective canopies. leaf According area index”, to WALTER which is defined et al. (2003) as 

A simpler approach was developed by Lang & Xiang 

they can be grouped into optical-spectral methods (e.g. CHEN et al. 1997, FASSNACHT et al. 1994, KUUSK 

LAIe= LAI . 

(3.11) (1986). They showed that in the case of large-scale (3.11) 

et al. 2002, SMITH et al. 1993 or STENBERG et al. 1994) and methods incorporating stand dimension 

analyses LAIe is therefore (KUCHARIK the et true al. variable 1999, NILSON measured 1999). with optical devices if no correction for clumping is applied. 

CHEN Various & studies CIHLAR have (1995a) dealt developed with the a correction procedure for to non-randomness derive a clumping and index the based derivation on gap of size , especially and gap 

fraction when indirect analyses. optical The underlying methods were theory applied is that to within coniferous a random canopies. canopy According the probability to WALTER of the et occurrence al. (2003) 

of they large can gaps be grouped can be derived into optical-spectral from the distribution methods (e.g. of gap CHEN sizes. et The al. 1997, iterative FASSNACHT removal of et al. large 1994, gaps KUUSK from 

the et al. total 2002, gap SMITH fraction et is al. equivalent 1993 or STENBERG to the design et al. of 1994) an imaginary and methods plant incorporating canopy with stand gaps inserted dimension at 

random analyses and (KUCHARIK having the et al. same 1999, total NILSON gap fraction 1999). as the real canopy. From this new canopy the canopy 

element CHEN & area CIHLAR index (1995a) is calculated developed and compared a procedure to the to derive original a clumping canopy to index derive based the clumping on gap size index. and Gap gap 

size distribution may be estimated with the TRAC instrument (Tracing Radiation and Architecture of 

Figure fraction 3-6analyses. 

Effect of The foliage underlying clustering on theory gap fraction is that according within a to random Nilson (1999). canopy Same the amount probability of needles of is the dispersed occurrence 

Canopies, of large gaps 3rd (a) can Wave randomly be Engineering). derived and b) clustered. from the distribution of gap sizes. The iterative removal of large gaps from 

A the simpler total gap approach fraction was is developed equivalent by to LANG the design & XIANG of (1986). an imaginary They showed plant canopy that in the with case gaps of inserted large-scale at 

clumping random and (i.e. having at plant the level) same a total procedure gap fraction of spatial as the logarithmic real canopy. averaging From of this transmittance new canopy can the give canopy an 

approximation element area index of true is calculated LAI. Assuming and compared the Poisson to the model original – and canopy its condition to derive the of randomly clumping index. distributed Gap 

foliage size distribution – is more applicable may be estimated locally to with smaller the canopy TRAC sectors instrument (as e.g. (Tracing derived Radiation from DHPs), and Architecture Po is calculated of 

Canopies, 3rd Wave Engineering). 

for each sector as well as its logarithm. Instead of averaging Po over whole azimuth ranges (as e.g.


clumping (i.e. at plant level) a procedure of spatial 

logarithmic averaging of transmittance can give an 

approximation of true LAI. Assuming the Poisson 

model – and its condition of randomly distributed 

foliage – is more applicable locally to smaller canopy 

sectors (as e.g. derived from DHPs), P is calculated 

O 

for each sector as well as its logarithm. Instead of 

averaging P over whole azimuth ranges (as e.g. 

O 

determined with the LAI-2000 PCA) its logarithm 

derived for small sectors is averaged, approaching 

true LAI in the presence of clumping. This method 

can easily be used to derive the clumping index from 

hemispherical photographs and is implemented e.g. 

in the CAN-EYE software. 

Chen & Cihlar (1995) and Law et al. (2001) noticed 

that it is more difficult to estimate clumping (and 

therefore true LAI) for high and dense canopies 

due to darkness and multiple scattering inside the 

canopy. According to Russell et al. (1989) it has been 

estimated that without any degree of leaf grouping, a 

tree could not sustain an LAI greater than 6.0 because 

of self-shading. 

Radiation interception by non-foliage elements 

Another issue associated with LAI estimation from 

light transmittance measurements is absorption by 

non-photosynthetic canopy elements. Stems and 

branches prevent light from reaching the optical 

device (cf. Figure 3-5) and as most instruments 

cannot discriminate between the effects caused by 

foliage or woody elements (cf. Cutini et al. 1998, 

Hyer & Goetz 2004, Kucharik et al. 1998, among 

others) they estimate Plant Area Index (PAI) instead 

of LAI. In order to derive LAI from PAI another term 

is introduced here: the Woody Area Index (WAI), 

where 

PAI = LAI + WAI. 

(3.12) 

39 

Some studies (e.g. Cutini et al. 1998, Dufrêne & 

Bréda 1995, Kalácska et al. 2005) used leafless times 

during the year to determine WAI with LAI-2000 

PCA or DHP. However, care has to be taken when 

using this approach, as the contribution of stems and 

branches to LAI when leaf area is at its maximum 

is far less than the WAI determined during leafless 

periods. Alternatively Chapman (2007) developed an 

interesting method with hemispherical photography 

in the near-infrared to better distinguish foliage and 

woody elements. However, it is hard to estimate 

the amount of leaves that is obscured by stems and 

branches, so if the latter are subtracted from PAI, this 

leads to an underestimation of real LAI. 

Kucharik et al. (1998) demonstrated with 

measurements in the visible and near-infrared 

wavelengths of a Multiband Vegetation Imager 

(MVI) that for different boreal forest types in Canada 

the contribution of branch area to PAI did not exceed 

10% and did therefore not significantly bias estimated 

LAI. However, they suggest that account should be 

taken of stem area, as stems comprised 30 to 50% 

of the total woody area in their study. Stems may in 

most cases not, or only partially, be shed by leaves. In 

the same context Smolander & Stenberg (1996) state 

that the contribution of branches and stems to PAI is 

a function of foliage amount. 

Apart from the above-mentioned, further errors exist 

when LAI is estimated from measurements with 

optical devices. These include instrument-inherent 

errors, such as instrument sensitivity or accuracy 

of the analogue-to-digital signal converter for the 

LAI-2000 PCA. With regard to DHPs, possible lens 

distortion has to be taken into account. According 

to Jonckheere et al. (2005) even small amounts of 

uncorrected angular distortion might cause substantial 

error in the measurement of gap area and distribution. 

As every lens invariably exhibits a small deviation 

from the theoretical projection, a lens correction 

should be applied.

40 

Apart from that, measurement errors may also 

result from the inappropriate use of the devices. 

According to Welles & Norman (1991) LAI- 

2000 PCA measurements, for example, should 

be made under diffuse light conditions, i.e. under 

homogeneously overcast skies or close to sunrise/ 

sunset (Chason et al. 1991, Fassnacht et al. 1994). In 

this case the assumption that leaves are exclusively 

absorbing radiation is closest to reality. In spite 

of the optical filter, sunlit leaves in the canopy are 

expected to transmit and reflect light, which adds 

to the recorded radiation below the canopy. This is 

supported by De Wasseige et al. (2003) and Leblanc 

& Chen (2001), who found an influence of θ on sun 

the retrieved transmission values and consequently 

LAI. Additionally, if measurements are taken under 

direct sunlight and the optics of the LAI-2000 PCA 

are not carefully shaded, the ring containing the sun 

measures incoming radiation at θ , φ rather than 

sun sun 

over the entire azimuth range (Hyer & Goetz 2004). 

In addition, heterogeneous skies should be avoided 

with the LAI-2000 PCA instrument (Hyer & Goetz 

2004). If above and below canopy measurements are 

not taken at the same spot (as it is e.g. the case in 

forest canopies, where the reference sensor is located 

in large gaps or outside the forest), heterogeneous 

skies or moving clouds introduce a bias in the 

transmission measurements, as I is not retrieved 

0 

from the same sky sector. 

For hemispherical photographs taken with analogue 

cameras, diffuse illumination conditions were also 

recommended (Jonckheere et al. 2004). With the 

use of digital cameras however, the problem of 

distinguishing sunlit leaves from small underexposed 

canopy gaps becomes less important. Better 

radiometric image quality and a spatial resolution 

comparable to that of an analogue camera allows 

image acquisition even under direct sunlight 

conditions. However, care has still to be taken with 

respect to exposure. Zhang et al (2005) showed that 

automatic exposure is not reliable for the calculation 

of gap fraction and resulted in an overestimation of 

gap fraction (i.e. underestimation of LAI). 

3.3.4 

Spatial sampling strategies 

The set-up of field measurements and the spatial 

sampling strategy are key issues when performing 

ground-based measurements. The main factors in 

devising a sampling scheme include the size of the 

complete test site, the size and spacing of individual 

sampling units, the sampling pattern, number 

of measurements to be made, distance between 

measurements and measurement height. 

The size of the complete test site mainly depends on 

the study objectives. Whereas for the characterization 

of different forest stands test sites of several km 

border length might be necessary, the description 

of LAI variability on small-scale agricultural fields 

could be accomplished within a much smaller area. 

Test sites should be flat to avoid terrain effects in 

satellite data and should be relatively homogeneous 

at medium resolution scale. 

On the test sites, smaller sampling units are usually 

established where the actual ground measurements 

occur. According to McCoy (2005) random sampling, 

stratified random sampling, systematic sampling and 

clustered sampling approaches can be applied to 

identify appropriate locations for the sampling units. 

If ground measurements are to be related to satellite 

imagery, Justice & Towneshend (1981) suggest a 

minimum area of 

a = p (1+2l), (3.13) 

with a being the minimum area of a sampling unit, p 

the pixel dimension and l the location error in pixels. 

Especially when geolocation is difficult and l is large, 

a homogeneous stand geometry is desirable. This


can be evaluated in advance by using meaningful 

reference data and recent satellite imagery. 

To cover the whole variability of the measured 

variable, it is further important to set up a large 

enough number of sampling units on the test site, 

which should also be equally spread. This can e.g. 

be accomplished by stratifying the test site into 

landscape units, based on vegetation, terrain and soil 

features (Aragão et al. 2005). With respect to the 

later validation of remote sensing data, care should 

be taken that sampling units can be accurately linked 

to satellite pixels. Scale effects should be taken into 

account (cf. Chapters 3.4.2 and 3.4.4). To avoid the 

effect of mixed pixels, sampling units should be 

located far enough from landscape boundaries. 

With respect to the sampling scheme, Garrigues 

et al. (2002) compared different patterns with 

geostatistical methods (cf. Figure 3-7). According 

to them there is no significant difference in the 

spatial representativeness between the random 

square and cross sampling. Apart from these two, 

transect sampling is very common in forests (e.g. De 

Wasseige et al. 2003, Erikson et al. 2005, Le Dantec 

et al. 2000). Other studies try to get an area-wide 

coverage of the sampling units (Kalácska et al. 2004, 

Stenberg et al. 2004). 

The number of individual measurements per ESU 

mainly depends on ESU size and canopy height. 

Ideally single measurements represent well LAI 

variability on a given ESU but are at the same time 

spatially independent to meet all conditions for later 

statistical analysis and to estimate mean LAI according 

to a given precision (De Wasseige et al. 2003, 

Ferment et al. 2001, and Weiss et al. 2004). Therefore 

an adequate sampling distance must be selected; 

this will depend mainly on the instrument’s field of 

view, but also on the vegetation type. Especially in 

dense vegetation, the area actually represented by a 

single measurement can be much smaller than the 

41 

theoretical one. For transect measurements varying 

sampling distances can be found in the literature. 

Whereas generally distances of 2 to 30 m are 

reported (e.g. Dufrêne & Bréda 1995, Eriksson et al. 

2005, Mussche et al. 2001), many studies use a 10 

m lag (e.g. Aragão et al. 2005, Chen et al. 1997, Le 

Dantec et al. 2000). Very few studies however really 

tested whether individual measurements are spatially 

autocorrelated. 

Figure 3-7 

3.4 

a) Random square and b) cross sampling 

schemes (Garrigues et al. 2002). 

Satellite-based derivation of LAI 

The scientific community discovered the potential 

of remote sensing to monitor vegetation properties 

in the mid-seventies with the availability of Landsat 

data. The earliest attempts to derive LAI were 

purely empirical and mainly focused on agricultural 

landscapes. Kanemasu (1974) was among the first 

to correlate in situ measured LAI of soybean and 

sorghum with band ratios derived from Landsat 

MSS data. Comparable studies followed, all based 

on band ratios or spectral vegetation indices (e.g. 

Pollock & Kanemasu 1979, Kanemasu et al. 1977). 

After promising results were received for various 

crop types, LAI estimation of forested areas started in 

the mid-eighties. Peterson et al. (1987) were the first 

to assess LAI of coniferous forests in the U.S. with 

Landsat TM data. Other empirical studies dealing 

with forest LAI followed (e.g. Curran et al. 1992, 

Herwitz et al. 1990, Nemani et al. 1993, and Spanner 

et al. 1994, with Landsat TM; Spanner et al. 1990b 

with AVHRR).

42 

Concurrently the development of canopy reflectance 

models led to physical approaches of LAI estimation. 

Several studies at the beginning of the eighties 

showed that LAI could be estimated with sufficient 

accuracy from in situ measured canopy reflectance 

data and model inversion if ancillary data on leaf 

optical properties as well as soil reflectance are 

provided (Goel & Strebel 1983, Goel & Thompson 

1984a, Goel & Thompson 1984b). However, due to 

various constraints these techniques were transferred 

to data acquired by spaceborne sensors only in the 

mid-nineties (Jacquemoud et al. 1995, Kuusk 1995). 

Both methods, empirical and physical approaches of 

LAI derivation from remote sensing data, have been 

applied and further developed in various studies since 

the 1990s. They will consequently be described in the 

following together with their applicability to different 

sensor types. The subchapter closes with a discussion 

of the limitations of remotely-sensed LAI derivation 

with respect to tropical rainforest ecosystems. 

3.4.1 

Empirical approaches 

Empirical approaches of LAI derivation rely on 

relationships between in situ and satellite reflectance 

data or SVIs calculated from the latter. A prerequisite 

is the assumption that variations in surface reflectance 

(or SVI) are due to variations in LAI only. Figure 

3-8 shows the variation of surface reflectance with 

increasing LAI. SVIs make use of these distinct 

characteristics in the NIR and SWIR wavelengths: 

typically NIR and SWIR reflectances increase with 

increasing LAI whereas reflectance in the visible 

wavelengths (especially in the red part of the 

spectrum) remains low due to chlorophyll absorption. 

SVIs ideally maximize the sensitivity to biophysical 

variables and normalize both external (e.g. sun and 

viewing geometry, atmospheric disturbance) and 

internal effects (e.g. topography, soil variation) at 

the same time. The SVIs used most often include the 

NDVI, the SR or SWIR corrected versions of both 

(cf. Chapter 3.4.3) 

Once a relationship between in situ measured LAI and 

an SVI is found, an empirical model is established 

through parametric or non-parametric regression 

analyses or artificial neural networks (ANN). Most 

often, the relationship between spectral data and LAI 

is modelled by simple or multiple regression analysis 

(Cohen et al. 2003). Here, both linear and non-linear 

models have been used to predict LAI from SVIs. 

Whereas Heiksanen (2006), Schlerf et al. (2005) 

and Stenberg et al. (2004) applied linear regression 

to estimate LAI from SVIs for temperate and boreal 

forests, Chen & Cihlar (1996), Turner et al. (1999) 

and Kalácska et al. (2004) observed non-linear 

regression to be more accurate for boreal forests in 

Canada, temperate forests in the U.S. and a moist 

tropical rainforest in Costa Rica. Multiple linear 

regressions were found to be useful for the estimation 

of LAI of Wisconsin forests and tropical rain forests 

in Brazil by Fassnacht et al. (1997) and Aragão et 

Figure 3-8 

Reflectance from one to six layers of 

cotton leaves (McCoy 2005).


al. (2005). However, no general conclusion can be 

drawn from this as certainly the choice of regression 

model depends on the used SVI (e.g. NDVI is very 

often reported to have a non-linear relationship to 

LAI, whereas simple ratio (SR) is more often used 

in linear models) and unique site characteristics. This 

will be discussed in more detail in Chapter 3.4.4. 

As atmospheric effects, viewing geometry and 

anisotropic reflectance behaviour of vegetation 

canopies are usually not taken into further account, the 

resulting regressions are mostly site and time specific. 

Therefore operational use of empirical models for 

LAI derivation is scarcely possible because in situ 

measurements are needed for every major land cover 

type to establish correct relationships. Consequently 

empirical relationships between remotely-sensed 

SVIs and in situ LAI are used most often for singledate 

and local to regional applications. The method is 

easy to implement and provides optimal results when 

extensive ground measurements were performed. 

An exception is an approach that was developed by 

Sellers et al. (1994) based on empirically derived 

vegetation type properties and NDVI data from the 

National Oceanic and Atmospheric Administration’s 

(NOAA) Advanced Very High Resolution 

Radiometer (AVHRR). Calculation of LAI is here 

based on the linear relationship of the fraction of 

photosynthetically active radiation absorbed by 

vegetation (FPAR) to NDVI after which LAI can 

be estimated from FPAR. As potential error sources 

exist in the original data due to cloud contamination, 

atmospheric constituents, and viewing geometry 

(Sellers et al. 1994), and NDVI seems to saturate for 

LAI values above 2.0 to 3.0 (Sellers et al. 1996), this 

approach must be utilized with care. Nevertheless, it 

was the first method that allowed the derivation of 

LAI data sets from operational processing (see, e.g., 

Dech et al. 1998). A similar approach, but directly 

based on NDVI values, is used by the ECOCLIMAP 

project (Masson et al. 2003). 

3.4.2 

Physical models 

43 

In contrast to empirical methods, physical models 

are based on a theoretical description of canopy 

radiative transfer. They characterize the interaction 

of incoming electromagnetic radiation with canopy 

elements and its reflection based on certain leaf and 

canopy structural properties and soil background 

(Baret 1995). Ross (1981) gives a detailed description 

of radiative transfer modelling. Further explicit 

reviews on this topic were compiled by Myneni & 

Ross (1991), and Myneni (1990a and b). 

One-dimensional models act on the presumption that 

the canopy is horizontally homogeneous and infinite, 

but variable and finite in the vertical direction. The 

canopy is assumed to contain only small and flat 

leaves that are randomly distributed in space. So are 

leaf azimuths (cf. Figure 3-9). This turbid medium 

approach allows a number of approximations and 

simplifications with respect to canopy scattering 

behaviour (Disney 2001). These models are therefore 

based on a restricted amount of input parameters and 

are computationally efficient. Based on the work of 

Suits (1972), Verhoef (1984) developed a canopy 

radiative transfer model called SAIL. This model 

was adapted and extended by several authors (e.g., 

Gastellu-Etchegorry et al. 1996a, Kuusk 1995), and 

coupled with leaf optical models, e.g., PROSPECT 

(Jacquemoud et al. 1995). Comparisons to similar 

models gave promising results (Jacquemoud et al. 

2000). Bacour et al. (2006) successfully tested the 

generation of LAI products based on MERIS data 

using a coupled SAIL+PROSPECT radiative transfer 

model within the CYCLOPES programme. According 

to Disney (2001) a major drawback of the onedimensional 

turbid-medium approach is that the size 

of scattering objects, i.e. leaves, is not considered. As 

certain properties of the observed canopy reflectance 

are directly controlled by the size and orientation 

of scattering objects (e.g. hotspot effects), different 

approaches are thus needed. Further, as lateral

44 

homogeneity at the scale of interest is presumed for 

1D models, their application leads to problems if 

input data is heterogeneous (e.g. mixed pixels in the 

case of coarser resolution satellite data). Nevertheless, 

1D models provide a valid basis for radiative transfer 

modelling in small and homogeneous canopies (e.g. 

grasslands), that can be considered as turbid media 

(Shabanov et al. 2000). 

To account also for lateral heterogeneity in vegetation 

stands (e.g. single trees in forest stands), threedimensional 

models were developed. They divide 

the canopy in a rectangular matrix of parallelipipedic 

cells (voxels). Optical properties are computed with 

optical and structural element characteristics within 

each voxel, where also multiple scattering is allowed 

(Gastellu-Etchegorry et al. 1996b). Radiation 

transport is restricted to propagate only in a finite 

number of directions (Liang 2004). A widely applied 

example of a 3D canopy radiative transfer model is 

the DART model (Gastellu-Etchegorry et al. 1996b). 

Disney (2001) gives a good overview of more 

recent developments in 3D canopy radiative transfer 

modelling. Although 3D radiative transfer models 

have been developed since the 1980s, computational 

effort restricted the operational use of such algorithms 

for more than a decade (Jacquemoud et al. 2000). In 

principle, numerically solving 1D radiative transfer 

equations is carried out for one horizontal layer after 

the other, but solving 3D equations must be carried 

out cell by cell, which makes the whole process very 

time-consuming. 

Apart from the above-mentioned physically-based 

models, other approaches were developed (Disney 

2001, Liang 2002). Li and Strahler introduced e.g. 

a geometric optical model for conifer forests (Li 

& Strahler 1985 and 1986) that consists of simple 

geometric objects (e.g. cones or spheroids). According 

to Shabanov et al. (2000) this model suffers from an 

inaccurate description of multiple scattering and the 

propagation of radiation into deeper canopy parts. 

Figure 3-9 

Simulation of vegetation leaf canopy as onedimensional 

turbid medium (Pinty et al. 2001). 

Computer simulation models, like the Monte Carlo 

Ray Tracing (cf. Figure 3-10) and the radiosity 

method (Disney 2001, Liang 2002, Shanbanov et al. 

2000), are on the other hand computationally very 

extensive. 

A critical step in deriving LAI from canopy radiative 

transfer models is model inversion. It allows the 

retrieval of biophysical variables from remotelysensed 

measurements, but is only possible if a unique 

solution of the inverse problem exists (Knyazikhin 

et al. 1998). Model uncertainties and uncertainties 

in sensor-derived radiation measurements may, 

however, lead to a large number of possible solutions 

(Weiss et al. 2000). Different approaches exist for 

model inversion (see, e.g., Kimes et al. 2000, Liang 

2004) with varying computational demand. 

A very efficient method of canopy radiative transfer 

inversion is the LUT method, which is also used in 

the MODIS LAI algorithm. As the information on 

remotely-measured mean canopy-leaving radiance 

averaged over the three-dimensional radiation field 

is not sufficient to derive LAI, the range of variables 

determining the plant radiative regime has to be 

specified. Based on a global biome map, certain 

structural parameters of individual plants and the 

canopy as a whole, as well as optical properties of 

vegetation and soil, are predefined within certain 

ranges for each biome (Knyazikhin et al. 1998). 

Directional reflectance is computed in advance based 

on a 3D model and the results are stored in a LUT.


Spectral reflectances derived from the MODIS sensor 

are then compared to these model entries in order to 

find LAI (Myneni et al. 2002). 

Other inversion approaches include numerical 

optimization methods (e.g. Goel & Thompson 1984a 

and b, Jacquemoud et al. 1995, Liang 2002) and ANN 

(e.g. Atzberger et al. 2003, Baret et al. 1995, Weiss & 

Baret, 1999). 

3.4.3 

Spatial and spectral scale of sensors 

LAI and other biophysical variables have been 

derived from various remote-sensing data sources 

over the past 30 years. Usually the choice of spatial 

and spectral resolution of the input data depends on 

the study objectives and on sensor availability. 

With respect to forest studies, research on localto-regional 

scales is mainly aimed at an improved 

assessment of forest structure and contributions 

to forest management and monitoring (Rautiainen 

2005). Due to data availability most studies since the 

1980s have focused on high resolution multispectral 

Figure 3-10 

Simulation of a 3D vegetation canopy with 

a Monte Carlo Ray Tracing model (Disney 2001). 

45 

satellite data, e.g., sensors aboard the Landsat, SPOT 

and IRS satellites (e.g. Berterreche et al. 2005, Brown 

et al. 2000, Chen & Cihlar 1996, Cohen et al. 2003, 

Fassnacht et al. 1994, Rautiainen 2005, Spanner et al. 

1990a, 1994, Stenberg et al. 1994, Turner et al. 1999, 

White et al. 1997, among many others). 

In recent years airborne hyperspectral and very high 

resolution multispectral data has become available. 

Lee et al. (2004) compared, for instance, the power 

of hyperspectral AVIRIS data with simulated 

multispectral ETM+ data for predicting LAI of 

needleleaf and broadleaf forests in North America. 

According to their results, the use of narrow bands 

has no clear advantage over the use of broadband 

data. Although regression models using AVIRIS data 

performed better, they suspected an overfit of the 

applied models. Schlerf et al. (2005) however found 

a better correlation between LAI and SVIs derived 

from hyperspectral HyMap data than simulated 

multispectral data for a needleleaf forest in Germany. 

Whereas the studies dealing with high resolution 

multispectral data found empirical regressions with 

SVIs to be best suited for LAI derivation, studies 

using very high resolution data as IKONOS or 

QuickBird, reported texture measures to be better 

related to LAI (Colombo et al. 2003, Leboeuf et 

al. 2007, Peddle & Johnston 2000, Seed & King 

2003). This is clearly a function of spatial scale: the 

IFOV of very high spatial resolution sensors does 

not integrate over the reflectance caused by clusters 

of ground elements, e.g. a group of trees, but only 

represents parts of these elements (e.g. parts of a 

shadowed crown). Therefore very high resolution 

pictures have a much higher internal variance caused 

by stand structure and shadow fractions. This, in turn, 

is related to the complexity of forest stands, which 

increases with stand age and thus LAI. 

Though the major part of studies at very high and 

high spatial resolution used empirical approaches

46 

based on in situ data for LAI derivation, attempts 

to infer LAI from the inversion of canopy radiative 

transfer models were also made. Gascon et al. (2004) 

used for example a LUT inversion of a 3D canopy 

radiative transfer model to derive LAI of a mixed 

forest in France from SPOT-HRV and IKONOS data. 

Whereas the performance with SPOT data was good 

(see also Rautiainen 2005), problems occurred due 

to the high spatial resolution of IKONOS data. Here 

the inversion procedure had to be applied based on 

segmentation of homogeneous areas rather than on a 

pixel-per-pixel basis. 

In contrast to the above-mentioned studies that 

required a comparatively high spatial resolution for 

their applications, a lower spatial resolution (300 m 

to 7 km) is sufficient for research on continental to 

global scales. Biophysical variables derived from 

these sensors (e.g. MODIS, MERIS, POLDER, 

VEGETATION, AVHRR) mainly serve as input 

for climate or ecosystem modelling, for example 

trying to quantify net primary production and carbon 

sequestration under different climate scenarios. 

Though empirical approaches do exist at these 

relatively coarse spatial scales (e.g. Chen et al. 

2002 for Canada, Sellers et al. 1996), the majority 

of models is physically based. This is mainly due to 

the fact that at the spatial scale of medium to coarse 

resolution sensors, observed pixels always represent 

a mixture of different land cover types. Empirical 

relationships are on the other hand very sensitive 

to the land surface and viewing geometry, so that 

universal relationships are hard – if not impossible – 

to establish. 

Physical models do not require ground data or a priori 

knowledge about canopies. Weiss & Baret (1999) 

consequently suggest the estimation of biophysical 

variables based on the modelling of radiative transfer 

process in the canopy for these sensors. Recent 

examples are manifold and comprise operational and 

non-operational models presented by e.g. by Myneni 

et al. (2002), Bacour et al. (2006), Bicheron & Leroy 

(1999), Deng et al. (2006) and Bicheron et al. (1998). 

Table 3-4 gives an overview of the above-mentioned 

studies. The summary shows that empirical 

approaches are mainly applied to very high and high 

spatial resolution multispectral imagery. However, 

as empirical models are limited by vegetation 

type and related optical properties, sun-sensor 

geometry, background reflectance and the quality 

of atmospheric correction, they are not suitable for 

operational processing. Consequently for medium to 

coarse spatial resolution data mainly physical models 

are applied. 

With respect to spectral resolution, the above 

mentioned results Schlerf et al. (2005) offer 

an interesting perspective especially for future 

hyperspectral satellite missions, such as EnMAP. 

With a spatial resolution of 30 m this sensor seems 

to be well suited for LAI derivation with empirical 

and physically-based models. It is acknowledged that 

LAI can also be derived from passive microwave 

sensors, which have certain advantages especially 

over tropical regions (e.g. independency from cloud 

cover). Yet as radar data was not available for this 

thesis, this kind of remote sensing data is not further 

taken into account. 

3.4.4 

Suitability and limitations 

of empirical methods 

As mentioned in the previous chapter the major 

part of studies dealing with regional applications 

focused on empirical methods with high resolution 

multispectral data. As multispectral data will be used 

in this thesis for the upscaling of in situ measurements 

based on empirical relationships, a special focus of 

this subchapter will be on SVIs and the influence of 

structural characteristics on the establishment of such 

a model. As only few studies have analysed data for


tropical rain forests, publications on temperate and 

boreal forests will also be taken into account. 

Sensitivity of SVIs 

As mentioned above, empirical models based on SVIs 

derived from broadband multispectral sensors have 

shown to be useful for LAI assessment in various 

Table 3-4 

Data type Spatial 

resolution 

Multispectral Very high 

(0.60 m-4 m) 

Multispectral High 

(10 m-30 m) 

Multispectral Medium to 

coarse 

(300 m-7 km) 

Hyperspectral Very high 

to high 

(5 m) 

Optical remote sensing data types, mainly used LAI derivation methods, applications and recent examples. 

LAI derivation method Applications Recent examples 

Mainly empirical models, 

best regressions models 

reported for texture 

measures 

Mainly empirical models 

established with in situ 

data and SVIs, 

some physically-based 

approaches with canopy 

radiative transfer models 

Mainly physical 

modelling, inversion of 

radiative transfer models, 

some empirical or semiempirical 

approaches 

Site and time specific 

applications, single 

date (forest inventories, 

forest structure, forest 

monitoring) 

Site and time specific 

applications, mostly 

single date (forest 

inventories, forest 

structure, forest 

monitoring) 

Global derivation 

of biophysical 

variables as input 

for process models 

(evapotranspiration, 

crop yield, primary 

productivity, changes in 

vegetation structure), 

high temporal resolution 

Mainly empirical models Site and time specific 

applications, single 

date (forest inventories, 

forest structure, forest 

monitoring) 

CASI: Wulder et al. (1998) 

IKONOS: Colombo et al. (2003) 

QuickBird: Leboeuf et al. (2007) 

Aster: Eckert (2006), Heiskanen (2006) 

Landsat 5/7: Aragão et al. (2005), 

Berterreche et al. (2005), 

Brown et al. (2000), Butson & 

Fernandes (2004), Cohen et al. 

(2003), Kalácska et al. (2004), 

among many others 

SPOT-4/5: Gascon et al. (2004), 

Rautiainen (2005) 

MERIS: Bacour et al. (2006) 

MODIS: Myneni et al. (2002) 

POLDER: Bicheron et al. (1998) 

SPOT-VGT: Chen et al. (2002), 

Deng et al. (2006) 

AVIRIS: Lee et al. (2004) 

HyMap: Schlerf et al. (2005) 

47 

studies (see Jensen & Binford 2004, and Liang 2004 

for general overviews). However, the major part 

of these studies focused on natural and agricultural 

vegetation types in semi-arid, temperate and boreal 

regions. Especially for biomes with little above ground 

biomass, SVIs showed a considerable sensitivity to 

LAI variation (e.g. Cohen et al. 2003, Lu et al. 2004, 

among others). Problems occur more frequently when 

higher LAI values are to be analysed.

48 

Gobron et al. (1997) investigated the estimation of 

LAI from visible and near-infrared satellite data on 

a theoretical basis. According to their findings the 

main constraint of LAI estimation from reflectance 

data is that the radiation field emerging at the top 

of the canopy does not always remain sensitive to 

increasing LAI values. Especially when the canopy 

is optically not thin enough to allow a significant 

illumination of the underlying soil, the radiation 

field becomes independent from further increases in 

foliage area. If this happens, LAI can consequently 

not directly be estimated from reflectance data at 

particular wavelengths without further input. This 

problem mainly occurs in dense canopies with high 

LAI values, as e.g. forests. Consequently saturation 

of SVIs at higher LAI values is reported for various 

studies. Especially the classic NDVI shows saturation 

for high LAI values as its dynamic range is stretched 

in favour of low-biomass conditions. 

open and very dense canopies respectively. These 

findings are in line with Horler & Ahern (1986), who 

reported that SWIR bands contain more information 

on forest structure in Western Canadian conifer and 

hardwood forests than others. 

SWIR data is also included in the normalized 

difference moisture index (NDMI), which is basically 

an NDVI including SWIR information instead of 

red reflectance (sometimes also called normalized 

difference water index or infrared index). Hardisky et 

al. (1983) reported that NDMI was highly correlated 

to canopy water content and tracked changes in plant 

biomass better than the NDVI. 

Another SVI incorporating SWIR information was 

introduced by Nemani et al. (1993), who applied a 

SWIR correction factor to the NDVI in order to 

improve the relation to LAI. This index is called the 

corrected NDVI (NDVI ). According to them the 

c 

In contrary, SR is more sensitive to LAI variation in SWIR correction acts here as a scalar for canopy 

high biomass vegetation as indicated in Figure 3-11 

(assuming that high LAI corresponds to high NDVI 

closure. 

values) and shows a higher linearity to biophysical In this context, one of the rare studies in tropical rain 

variables (Chen 1996). The same is true for the forest environment was undertaken by Kalácska et 

modified simple ratio (MSR), as it is a function of SR al. (2004). They tested the relationship between LAI 

(cf. Table 3-5). Yet MSR increases more slowly with and different SVIs for a tropical moist forest in Costa 

NDVI. 

Rica. Although MSR and SR were especially sensitive 

24 at higher LAI values (RTHEORETICAL BACKGROUND 

2 =0.77 and 0.80 respectively), 

Apart from SVIs that are solely THEORETICAL based on information BACKGROUND 

In contrary, SR is more sensitive to LAI variation in high biomass vegetation as indicated in Figure 3-11 

in the red and NIR wavelengths, many studies 

LAI variation in (assuming high biomass that vegetation high LAI corresponds as indicated in to Figure high NDVI 3-11 values) and shows a higher linearity to biophysical 

dealing with forest environments have also included 

ds to high NDVI variables values) (CHEN and shows 1996). a higher The same linearity is true to for biophysical the modified simple ratio (MSR), as it is a function of SR (cf. 

SWIR information in their analyses. In order to 

true for the modified Table simple 3-5). Yet ratio MSR (MSR), increases as it is more a function slowly with of SR NDVI. (cf. 

create nation-wide LAI maps for Canada, Chen et al 

slowly with NDVI. 

(2002) Apart e.g. from applied SVIs a SWIR that are modification solely based of on the information SR in the red and NIR wavelengths, many studies 

sed on information called dealing the in reduced the with red forest simple and environments NIR ratio wavelengths, (RSR). have According also many included studies to SWIR information in their analyses. In order to create 

ve also included them, SWIR nation-wide RSR information improves LAI LAI maps in their derivation for analyses. Canada, for In mixed CHEN order cover et to al create (2002) e.g. applied a SWIR modification of the SR called 

CHEN et al (2002) types the and e.g. reduced suppresses applied simple a SWIR background ratio modification (RSR). information, According of the as to SR e.g. them, called RSR improves LAI derivation for mixed cover types 

ording to them, soil, RSR and as the suppresses improves SWIR LAI is background most derivation sensitive information, for to mixed the amount cover as e.g. types soil, as the SWIR is most sensitive to the amount of 

ation, as e.g. of soil, vegetation as the SWIR containing is most liquid sensitive water. water. to swir the max amount and swir of min (cf. Table3-5) are directly SR derived from the study 

swir max and swirarea 

min (cf. and Table3-5) represent are SWIR are directly reflectance derived of from very the open the study and Figure very 3-11 dense Relationship canopies of respectively. NDVI versus SR These derived findings from are 

of very open study and in very area line dense and with represent canopies HORLER respectively. SWIR & AHERN reflectance (1986), These of findings who very reported are that SWIR ASTER bands data for contain the study more area of information Budongo Forest. on 

986), who reported forest that structure SWIR in bands Western contain Canadian more conifer information and hardwood on forests than others. 

conifer and hardwood forests than others. 

SWIR data is also included in the normalized difference moisture index (NDMI), which is basically an 

normalized difference NDVI moisture including index SWIR (NDMI), information which is instead basically of an red reflectance (sometimes also called normalized 

on instead of red difference reflectance water (sometimes index or infrared also called index). normalized HARDISKY et al. (1983) reported that NDMI was highly 

index). HARDISKY correlated et al. to (1983) canopy reported water content that NDMI and tracked was highly changes in plant biomass better than the NDVI. 

NDVI


Kalácska et al. (2004) report a saturation of these 

SVIs above LAI of 4.7. In their case this affects all 

late forest stages, which have LAI values of 4.9- 

8.9. For boreal forests however, sensitive relations 

between SVIs and LAI were found at much higher 

LAI values (see Schlerf et al. 2005 for an overview). 

This could possibly be attributed to the higher 

clumping of coniferous forests and thus a higher 

sensitivity of SVIs to high LAI values (Stenberg 

et al. 2004). Curran et al. (1992) e.g. showed that 

NDVI and LAI were positively correlated for North 

American pine forests with LAI values between 1.84 

and 9.24 (R2 =0.35-0.86 for different dates). Other 

studies producing similar results were e.g. undertaken 

by Spanner (1990a and b) for various conifers (LAI 

of 1-16). 

However, for broadleaf forest stands in Wisconsin 

(LAI of 4.4-8.4) Fassnacht et al. (1997) found that 

after stratification, regressions between LAI and SVIs 

Table 3-5 

Overview of most important SVIs with respect to LAI derivation for tropical rain forests. 

49 

with R2 between 0.60 and 0.98 could be achieved. 

A study on mixed forest types (aspen, pine and 

spruce) by Brown et al. (2000) showed that RSR had 

a linear relationship to LAI up to around 8. Here, no 

saturation occurred. The respective R2 (0.55) further 

increased if only data from coniferous stands were 

analysed (R2 =0.66). Brown et al. (2000) attribute the 

good performance of RSR to the greater sensitivity of 

the SWIR band to LAI. 

Generally it must be noted that SVIs have only a 

limited robustness, especially when transferred 

to other sites. They are sensitive to atmospheric 

conditions and viewing geometry (anisotropic 

behaviour of vegetation). For tropical rain forest 

environments a further problem can be the saturation 

of indices. 

25 

SVI 25 Definition 

THEORETICAL BACKGROUND 

THEORETICAL 

Reference 

BACKGROUND 

25 

25 SVI 

Simple 25 SVI Simple ratio ratio (SR) (SR) 

Simple SVI ratio (SR) 

SVI 

SVI Simple ratio (SR) 

Normalized Normalized Simple Normalized ratio difference (SR) difference vegetation vegetation index (NDVI) index 

Simple 

Normalized (NDVI) ratio (SR) 

(NDVI) difference vegetation index 

(NDVI) Corrected Normalized 

Corrected NDVI 

difference (NDVIc) 

vegetation index 

Corrected NDVI (NDVI ) 

Corrected (NDVI) 

Normalized difference vegetation index 

Normalized c 

(NDVI) NDVI difference (NDVIc) vegetation index 

(NDVI) Corrected NDVI (NDVIc) 

Modified Corrected Modified simple NDVI ratio (NDVIc) (MSR) 

Corrected Modified simple NDVI ratio (NDVIc) (MSR) 

Modified Modified simple simple ratio (MSR) ratio (MSR) 

Modified simple ratio (MSR) 

Modified simple ratio (MSR) 

Reduced simple ratio (RSR) 

Reduced simple ratio (RSR) 

Reduced Reduced simple ratio (RSR) 

Normalized Reduced simple simple ratio (RSR) 

difference ratio (RSR) 

Reduced Normalized simple difference ratio (RSR) moisture index 

Normalized (NDMI) difference moisture index 

(NDMI) Normalized difference moisture index 

(NDMI) 

Normalized difference moisture index 

Normalized Normalized (NDMI) difference difference moisture index moisture (NDMI) index 

(NDMI) 

Definition 

THEORETICAL Reference 

BACKGROUND 

THEORETICAL BACKGROUND 

Definition 

THEORETICAL Reference BIRTH & MCVEY Birth & McVey (1968) 

NIR 

BIRTH & MCVEY BACKGROUND (1968) 

NIR 

Definition 

BIRTH 

 

Reference & MCVEY (1968) (1968) 

Definition NIR 

red 

Reference 

Definition 

 

NIR 

Reference BIRTH & MCVEY (1968) 

NIR 

red 

ROUSE et al. (1974) 

NIR 

BIRTH red 

ROUSE & et MCVEY al. (1974) 

red 

Rouse et (1968) 

NIR 

al. 

 

 

BIRTH & MCVEY 

NIR 

 

NIR 

ROUSE et 

red 

(1974) al. (1974) 

(1968) 

NIR red red 

red red 

 

NIR red 

 

 

NIR 

red 

SWIR NEMANI ROUSE et red 

et al. (1993) 

SWIR min 

 

 

 

1* 

 

 

 

 

NIR 

red 

SWIR NEMANI et 

al. 

al. 

(1974) 

red 

 

(1993) 

NIR 

SWIR min ROUSE et 

NIR 

red 

 

1* 

 

SWIR max 

SWIR min 

 

 

Nemani al. (1974) 

red 

et al. 

NIR 

NIR 

 

red SWIR max 

 

NIR 

 

NIR red 

red 

SWIR ROUSE NEMANI et et al. al. (1974) 

red 

(1993) 

SWIR min 

NIR 

red 

SWIR min 

 

1* 

 

 

 

(1993) 

NIR 

red 

 

SWIR max 

SWIR min 

 

NIR 

red 

SWIR NEMANI et al. (1993) 

NIR 

red 

SWIR min CHEN (1996) 

NIR 

 

 

1 

 

1* 

 

 

NIR 

red 


SWIR min 

NIR 

red 

CHEN (1996) 

 

1* 

SWIR max 

SWIR min 

 

 

NIR 

red 


SWIR min 

NIR NIR 

red 

red1 

 

 

1* SWIR max 

SWIR min 

 

NIR CHEN (1996) 

NIR 

red SWIR max 

SWIR min 

Chen 

 

red 

NIR 1 

CHEN (1996) 

NIR 

NIR 

 

11 

red 1 

CHEN (1996) (1996) 

NIR 

 

NIR 

red 

1 red 

1 

red 

NIR 

red 

BROWN 

NIR 

SWIR et al. (2000) 

NIR 

SWIR min 

 

 

 

1 

1* 

 

 

 

BROWN 

NIR 


 

SWIR min 

red 

 

red 

 

 

1 NIR 

1* 

1 

SWIR max 

 

SWIR min 

 

BROWN 

NIR 

SWIR 

red SWIR max 

et al. (2000) 

red 

SWIR 

min 

 

SWIR min 

 

1* 

 

 

 

red 

 

 

 

red NIR 

 

HARDISKY et al. (1983) 

NIR 

SWIR max 

BROWN 

NIR 


SWIR 

SWIR 

min 

 

min 

 

 

1* 

SWIR 


SWIR 

 

 

BROWN 

NIR 

SWIR et Brown al. (2000) et al. 

SWIR min 

 

 

red 

NIR 

NIR 

 

1* 

 

 

NIR 

SWIR max 

 

SWIR min 

BROWN (2000) 

NIR 


SWIR min 

 


red 

SWIR 

1* 

 

SWIR 

SWIR max 

SWIR min 

 

 

 

NIR 

NIR 

red 

 

HARDISKY 

 

et al. (1983) 

SWIR 

SWIR 

SWIR max 

SWIR min 

NIR 

 


SWIR 

NIR 

NIR 

HARDISKY 

SWIR 

Hardisky et al. (1983) et al. 

SWIR 

NIR 

SWIR 

NIR 

(1983) 

SWIR 

Generally it must be noted that SVIs have only a limited robustness, especially when transferred to other 

Generally sites. They it are must sensitive be noted to that atmospheric SVIs have conditions only a limited and robustness, viewing geometry especially (anisotropic when transferred behaviour to other 

sites. They are sensitive to atmospheric conditions and viewing geometry (anisotropic behaviour of 

sites. Generally 

vegetation). They it 

vegetation). For are must 

For tropical sensitive be noted 

tropical rain to that 

rain forest atmospheric SVIs have 

forest environments conditions only a limited 

environments a further and robustness, 

further problem viewing 

problem can geometry especially 

can be the saturation (anisotropic when transferred 

saturation of indices. behaviour to other 

Generally it must be noted that SVIs have only a limited robustness, especially when transferred of 

indices. to other 

Generally 

vegetation). sites. They it must 

For are tropical sensitive be noted 

rain to that 

forest atmospheric SVIs have 

environments conditions only a limited 

a further and robustness, 

problem viewing can geometry especially 

be the saturation (anisotropic when transferred 

of indices. behaviour to other of 

sites. They are sensitive to atmospheric conditions and viewing geometry (anisotropic behaviour of 

sites. vegetation). They For are tropical sensitive rain to forest atmospheric environments conditions a further and problem viewing can geometry be the saturation (anisotropic of indices. behaviour of 

vegetation). For tropical rain forest environments a further problem can be the saturation of indices. 

vegetation). For tropical rain forest environments a further problem can be the saturation of indices. 

Influence of structural characteristics 

Influence of structural characteristics

50 

Influence of structural characteristics 

Schlerf et al. (2005) state that in coniferous 

forest plantations the relationship between LAI 

and vegetation indices can be disturbed as other 

biophysical stand characteristics may influence the 

reflectance signal. Especially differences in canopy 

structure and stand density can affect the reflectance 

signal. Canopy structure, which is dependent on 

canopy depth, leaf area index, leaf orientation and leaf 

distribution, becomes more complex as the vegetation 

amount increases, so that incoming photons are 

trapped within the canopy causing a decrease in the 

intensity of reflected radiation. 

The same can be true for the highly complex 

structure of tropical primary rain forest ecosystems. 

Figure 3-12 shows the differences in vertical and 

horizontal canopy structure between primary forest 

stands and disturbed secondary forests. Emergent 

trees are missing in the upper canopy of selectively 

logged forests causing the upper canopy boundary to 

be smoother. In contrast, the complexity of canopy 

structure increases with stand age in undisturbed 

forested areas. As illustrated, the primary forest site 

has a very rough upper canopy causing mutual shading 

and thus reducing the reflectance signal, even though 

the leaf area might be comparable between both sites 

(selectively logged forests tend to have a denser 

ground cover, compare Chapter 2.1.2 for Budongo 

Forest). Especially if satellite data is acquired at low 

sun elevation angles, the rough canopy surface of 

primary forest stands can trap a large proportion of 

solar irradiation. The positive relationship between 

SVIs and LAI might thus be disturbed. 

Several studies have therefore used structural 

information to improve the relation between SVIs and 

LAI. According to Seed & King (2003) the shadow 

fraction remains sensitive to increases in LAI even 

after saturation of SVIs. Aragão et al. (2005) applied 

for example spectral unmixing to extract the shadow 

fraction for tropical rain forest in Brazil from Landsat 

ETM+ imagery. Together with SVIs it improved 

the derivation of LAI through the application of a 

multiple linear regression model. 

Apart from spectral unmixing, the shadow component 

of vegetation stands can also be derived indirectly 

from statistical measures such as variance, standard 

deviation, and mean brightness over a predefined 

region (see e.g. Leboeuf et al. 2007). Here, image 

texture is analysed, i.e. the spatial variation in surface 

reflectance. As the spatial variability of pixel values is 

related to the spatial distribution of forest vegetation 

elements, forest structure is accordingly represented 

by texture (Wulder et al. 1998). Texture measures 

are usually statistics that are calculated for the 

central pixel within a moving window of fixed size. 

The so-called first-order texture measures calculate 

a representative statistical value for the respective 

central cell of that window and are based on the 

original image values. Second-order measures in 

turn consider the relationship between neighbouring 

windows (Hall-Beyer 2007, Wulder et al. 1998). 

Therefore a grey level co-occurrence matrix (GLCM) 

has to be calculated, which according to Wulder et al. 

(1998), p. 66, and after Haralick et al. (1973) can be 

defined as a “matrix of relative frequencies for which 

two neighbouring pixels that are separated by a userdefined 

distance and angle can occur in the image.” 

Statistical measures are then applied to the GLCM. 

These can include homogeneity, contrast and entropy, 

which characterize the specific textural characteristics 

of the image, or diversity, mean, variation and 

standard deviation that relate to the complexity of 

grey level transitions (Wulder et al. 1998). 

Texture indices were successfully applied to a study 

site in northern Italy by Colombo et al. (2003). They 

found that the inclusion of texture indices strongly 

increased the linear relationship between NDVI 

derived from IKONOS data and LAI (increase in 

R2 from 0.19 to 0.73, mean LAI values from 3 to 6).


Also Wulder et al. (1998) found that image texture 

provides a quantification of structural variability of 

mixed forest stands. If texture information was added, 

the prediction of LAI from NDVI was significantly 

improved. The inclusion of texture measures, 

however, requires that the image from which texture 

is calculated has an adequate spatial resolution to 

represent canopy objects. Often, very high resolution 

imagery such as IKONOS or QuickBird is used 

(Colombo et al. 2003, Leboeuf et al. 2007). An 

alternative are airborne sensors (Seed & King 2003, 

Wulder et al. 1998). 

3.4.5 

Figure 3-12 

Validation approaches 

When focusing on the validation of biophysical 

properties derived from satellite data, in situ 

measurements form an independent reference data set. 

As described in Chapter 3.3.4 the sampling scheme 

is of major importance. When dealing with satellitebased 

LAI at very high to high spatial resolution, in 

situ sampling units can be scaled correspondingly 

(cf. Equation 3.13) and field data can be compared 

directly to satellite-derived LAI. 

However, when biophysical variables derived from 

medium resolution sensors as MODIS are to be 

validated, an upscaling of in situ measurements to 

high spatial resolution satellite data is needed as an 

intermediate step (cf. Figure 3-13). Here a direct 

comparison between in situ data and the MODIS 

output is not feasible due to possible geolocation 

Structural differences in primary rain forest and selectively logged forest (NIES 2007, modified). 

51 

errors and surface heterogeneity at MODIS resolution 

(Liang et al. 2002). Upscaling can be achieved by 

the methods mentioned in Chapters 3.4.1 and 3.4.2, 

i.e. either empirical approaches or physically-based 

models. The latter is, however, not recommended by 

VALERI (Baret et al. submitted), as the majority of 

operational algorithms for estimating LAI employ 

a radiative transfer model. This means that circular 

dependencies with the LAI product to be validated 

might occur. 

For the validation purpose in this thesis, empirical 

methods are sufficient as only single time steps are 

needed for validation. Additionally, in contrast to 

physical models that are limited by saturation of nadir 

spectral measurements unless a priori information is 

applied (Knyazikhin et al. 1998a), empirical models 

are only constrained by the sampling distribution. 

In consequence, ecophysiological limits on LAI are 

incorporated to make predictions past the saturation 

point of reflectance data (Fernandes et al. 2005). 

Within the framework of CEOS-LPV, LAI product 

validation has been performed by various academic 

groups and research networks, as summarized later 

in Table 4-16). Although each of the groups applied 

slightly different methodologies for validation, e.g. in 

terms of instruments chosen for in situ LAI estimation 

or in terms of the upscaling procedure used, there are 

common guidelines on how to validate medium scale 

resolution satellite data based on field measurements. 

They will be described in the following and is based 

mainly on the synthesis publications of Yang et al.

52 

(2006), Morisette et al. (2006a), and Baret et al. 

(submitted). 

First of all, the spatial distribution of predominant 

land cover types on the test site is to be assessed. LAI 

measurements must be taken in several homogeneous 

patches within each land cover type in order to sample 

the natural range of LAI. In situ measurements 

should be as accurate as possible and the precision 

of the measurements should be known. Usually high 

resolution broadband satellite data, e.g. derived from 

SPOT, ASTER or ETM+, is used for the generation 

of high spatial resolution LAI maps. If available, also 

very high resolution optical, hyperspectral, or radar 

data could be used. Based on site-specific relationships 

between high resolution satellite data and ground 

measurements, a so-called transfer function needs to 

be established. This transfer function links both data 

sets based on empirical methods and ideally results 

in high resolution LAI maps covering the whole test 

site. 

The next step is the aggregation of the high resolution 

LAI map to 1 km MODIS resolution. Usually 

the fine resolution LAI map is averaged and then 

Figure 3-13 

Schematic illustration of validation strategy (Yang et al. 2006, modified). 

aggregated. According to Yang et al. (2006) a pixelby-pixel 

comparison is not recommended due to e.g. 

geolocation uncertainties and pixel-shift errors caused 

by the point spread function. Further the MODIS LAI 

algorithm is not designed to retrieve a deterministic 

value, but instead generates a mean LAI value from 

all possible solutions within a specified level of input 

satellite data and model uncertainties. Comparison 

should thus be rather performed at multipixel (patch) 

scale, where LAI is statistically stable. 

According to Morisette et al. (2006a) global validation 

of the MODIS LAI product is the final stage of 

validation exercises. The prerequisite is, however, the 

availability of data coming from various validation 

sites that represent the global variability and range 

of LAI and canopy types. Based on these global 

validation efforts but also on information coming 

from individual validation initiatives, an algorithm 

refinement is also envisaged. Within CEOS-LPV, 

the VALERI network, especially, has put some 

effort into the development of adequate sampling 

strategies for medium resolution satellite biophysical 

product validation (Baret et al. submitted). VALERI 

recommends test sites of at least 3 km x 3 km, with


larger sizes being desirable for coarser resolution 

sensors as e.g. POLDER (Baret et al. submitted). 

Local ground measurements are performed in each 

of the nine 1 km2 subsquares (cf. Figure 3-14) in 

so-called elementary sampling units (ESUs). ESUs 

ideally correspond approximately to the pixel size 

of a high spatial resolution satellite image, which 

can be later used for upscaling. Three to five equally 

spread ESUs should at least be sampled per 1 km2 . 

They should be located far enough from landscape 

boundaries and represent all cover types of the entire 

test site. 

The sampling strategy proposed by VALERI was 

followed in this thesis as far as possible. However, 

for several reasons a suite of modifications had to be 

applied to the VALERI scheme when transferring it 

to the tropical rain forest environment. These will be 

discussed in Chapter 5.1.1. 

Figure 3-14 

VALERI validation scheme of medium resolution 

satellite data (Baret et al. submitted). 

53

4 

Remote sensing data 

and preprocessing 

In the tropics the main constraint on the availability 

of optical satellite imagery is cloud cover. Although 

radar data acquisition is independent of clouds, 

it was not available for the study sites. Hence a 

suite of optical high resolution satellite imagery 

was analysed. Recent Landsat ETM+ scenes made 

available to the BIOTA East Africa project were 

utilized for visual assessments prior to fieldwork. 

However, as the sensor’s scan line corrector failed 

in 2003, Landsat data was not taken into further 

account for the derivation of LAI. Instead the 

OASIS (Optimising Access to Spot Infrastructure for 

Science) programme financed by the EC provided 

two SPOT-4 scenes for Budongo and Kakamega 

Table 4-1 

55 

Forests. The cooperation with CEOS-LPV led to the 

further possibility of analysing level 2 ASTER data 

for Budongo Forest (courtesy of Jaime Nickeson and 

Jeff Morisette from Goddard Space Flight Center, 

GSFC). In the following the available data sets for 

this study will be briefly described with respect to the 

sensor characteristics and preprocessing steps. 

4.1 

4.1.1 

Available high resolution 

satellite data 

ASTER 

The Advanced Spaceborne Thermal Emission and 

Reflection radiometer (ASTER) was built in a 

joint effort between NASA and Japan’s Ministry of 

Economy, Trade and Industry and is orbiting aboard 

the Earth Observing System (EOS) platform Terra. 

Launched in December 1999, Terra is in a sunsynchronous 

near-polar orbit with a local equator 

crossing time of 10:30 am. The ASTER instrument 

consists of three different subsystems (VNIR, SWIR 

and TIR) and has a swath width of 60 km. Table 4-1 

shows ASTER’s spatial, spectral and radiometric 

characteristics for the two subsystems used in this 

thesis. VNIR consists of two telescopes, one with a 

nadir-viewing three-spectral band detector (visible to 

Spectral and spatial properties of the ASTER VNIR and SWIR instruments (Abrams & Hook 2002). 

Band/System Wavelength [µm] Spatial resolution at nadir [m] Quantization [bit] 

(1) VNIR (green) 0.52 - 0.60 15 8 

(2) VNIR (red) 0.63 - 0.69 15 8 

(3a) VNIR nadir looking 0.76 - 0.86 15 8 

(3b) VNIR backward 0.76 - 0.86 15 8 

(4) SWIR 1.60 - 1.70 30 8 

(5) SWIR 2.145 - 2.185 30 8 

(6) SWIR 2.185 - 2.225 30 8 

(7) SWIR 2.235 - 2.285 30 8 

(8) SWIR 2.295 - 2.365 30 8 

(9) SWIR 2.360 - 2.430 30 8

56 

near-infrared wavelengths) and the other backward 

looking in the near-infrared range. All VNIR 

telescopes can rotate 24° across track. The SWIR 

instrument consists of a nadir-looking telescope that 

operates in six spectral bands. With the help of a 

pointing mirror, imagery can be acquired 8.55° across 

track. ASTER data is only acquired on demand. For 

more detailed information please refer to Abrams & 

Hook (2002). 

Compared to the other high resolution satellite data 

used in this study, ASTER has the highest spectral 

resolution with nine bands in VIS to SWIR and the 

highest spatial resolution in the VIS/NIR with 15 m. 

Unfortunately ASTER data was only available for 

Budongo Forest exactly one year after fieldwork. 

However, the fact that imagery was also acquired 

at the end of the rainy season justifies the analysis 

of this data set. Acquisition details for the respective 

scenes are shown in Table 4-4. 

4.1.2 

SPOT-HRVIR 

Launched in March 1998, the fourth member of 

the Satellite Pour l’Observation de la Terre (SPOT) 

family is one of the three SPOT satellites that are still 

in operation. It is in a sun-synchronous near-polar 

orbit at an altitude of 822 km with an inclination of 

98° and a repetition time of 26 days for the same 

ground track. The local equator crossing time on 

a descending node is around 10:30 am. Designed 

and developed by CNES, SPOT 4 carries a linear 

pushbroom radiometer, the HRVIR instrument. 

Table 4-2 

It features a pointable optics and offers off nadir 

viewing up to 27° (Henry & Meygret 2001). The 

spatial, spectral and radiometric properties of SPOT- 

4 are introduced in Table 4-2. Table 4-4 provides the 

acquisition details for the study sites. 

Preprocessing of high resolution 

satellite data 

Band Wavelength [µm] Spatial resolution at nadir [m] Quantization [bit] 

(1) green 0.50 - 0.59 20 8 

(2) red 0.61 - 0.68 20 8 

(3) NIR 0.78 - 0.89 20 8 

(4) SWIR 1.58 - 1.75 20 8 

4.2 

The preprocessing of satellite data aims to take into 

account the spatial and spectral comparability of 

different data sources. Geometric correction is the 

prerequisite for a precise co-registration of multiple 

satellite scenes. Although atmospheric correction is 

not necessarily needed to apply empirical models, 

Turner et al. (1999) found a significant improvement 

in the relationship between in situ measured LAI and 

SVIs after conversion of DNs to surface reflectance 

for three sites in temperate zones including a broadleaf 

forest. Effects of topographic corrections were in 

turn less evident. Consequently in this thesis the 

preprocessing steps of high resolution satellite scenes 

comprised geometric and atmospheric correction. 

4.2.1 

Spectral and spatial properties of SPOT-4 HRVIR (Henry & Meygret 2001). 

Geometric correction 

According to Richards & Jia (2006) there are many 

sources for geometric distortion, both systematic 

and non-systematic. While systematic distortions are 

mainly caused by the sensor’s forward movement 

(scan skew), its scanning system (mirror velocity) or 

the earth curvature (panoramic effect), nonsystematic

4.2.1 Geometric correction 

According to RICHARDS & JIA (2006) there are many sources for geometric distortion, both systematic 

4 Remote sensing data and preprocessing 

and non-systematic. While systematic distortions are mainly caused by the sensor’s forward movement 

(scan skew), its scanning system (mirror velocity) or the earth curvature (panoramic effect), nonsystematic 

distortions can be related to variations in altitude, attitude and velocity of the satellite platform and terrain- 

distortions can be related to variations in altitude, 4.2.2 Atmospheric correction 

induced effects. Systematic distortion is usually corrected for level 1 imagery by the data provider, but 

attitude and velocity of the satellite platform and 

non-systematic distortions still need to be rectified by the user. In this thesis topographic map sheets 

terrain-induced effects. Systematic distortion is Radiance reaching the sensor is usually not only 

1:50,000 of both study sites (made available by the BIOTA East project) were used as reference data for 

usually corrected for level 1 imagery by the data influenced by the properties of the reflecting ground 

geometric correction (cf. Table 4-3). 

provider, but non-systematic distortions still need to surface in the sensor’s instantaneous field of view 

be rectified by the user. In this thesis topographic map 

Table 4-3: Reference data for geometric correction 

(IFOV), but also by relief induced topographic 

sheets 1:50,000 of both study sites (made available illumination effects as well as atmospheric scattering 

by the BIOTA REFERENCE East project) were used COVERAGE as reference SCALE and absorption. YEAR Especially SOURCE for multitemporal 

data for 

Topographic 

geometric correction 

Map Sheets 

(cf. Table 4-3). 

Budongo 

30/III, 30/IV, 39/I, 39/II, 38/IV, 39/III Forest 

applications and 

East 


Africa 

derivation 

1:50 000 


(Uganda), 

quantities 

Series 


1:50,000 1966 

physical models, image Y732 calibration (D.O.S. 426), and Edition the I correction 

Each scene Topographic was georeferenced Map Sheets with 15-20 Kakamega evenly 1:50,000 of atmospheric East 

1970 effects Africa is 1:50 crucial. 000 (Kenya) Many Series different 

102/2 and 102/4 

Forest 

Y731 (D.O.S. 423), Edition 6 

distributed ground control points (GCPs) and radiative transfer codes for atmospheric correction 

Each assigned scene to was UTM georeferenced projection (Zone with 15-20 36N), evenly WGS 84 distributed exist, ground including control the points widely (GCPs) applied and 6S assigned (Second 

to datum. UTM Table projection 4-5 indicates (Zone 36N), the respective WGS 84 root datum. mean Table Simulation 4-5 indicates of the respective Satellite root Signal mean in square the Solar 

error square (RMSE), error (RMSE), which is which calculated is calculated as as 

Spectrum; Vermote et al. 1997) code or MODTRAN 

(Berk et al. 1998). In this thesis an atmospheric 

RMSE 

n 1 2 2 

XRi 

YRi 

n i1 

(4.1) correction with the IDL code of ATCOR-3 (4.1) was 

performed (Richter 1998, 2007). 

with n n being the the number of GCPs, of GCPs, and and XR and XR YR and the residual at GCP i in X and Y direction. An RMSE 

of YR less the than residual one pixel at GCP was i in achieved X and for Y direction. all scenes. An It should ATCOR-3 be noted, uses however, an atmospheric that GCPs database were mainly compiled 

located RMSE at of roads less than and intersections one pixel was outside achieved the forest for all as distinct from features the MODTRAN-4 are hard to identify code to in represent the forested a wide 

areas. scenes. The It should clearing be of noted, the Sonso however, Camp that Site GCPs in were Budongo range Forest of is atmospheric the only exception conditions because based buildings on various 

were mainly already located indicated at roads on and the intersections topographic outside maps of the 1966. combinations It should be mentioned of aerosol that types, the water RMSE vapour indicates content, 

the forest error as close distinct to features the selected are hard GCPs to and identify does in not the necessarily ground serve elevation, as an absolute solar zenith quality angles measure and inside visibility 

the forested forests. areas. Nevertheless The clearing detailed of the comparisons Sonso Camp of Site the different (Richter images 2007). and Based their on good an atmosphere agreement in type overlay selected 

suggest in Budongo that the Forest remaining is the internal only exception variation of because the scenes by is low. the user, an iterative process modifies properties 

buildings were already indicated on the topographic of the selected standard aerosol type to match scene 

maps of 1966. It should be mentioned that the RMSE conditions. The so-called dark dense vegetation/dark 

4.2.2 indicates Atmospheric the error close to correction the selected GCPs and does dense object approach uses dark scene areas of low 

not necessarily serve as an absolute quality measure reflectance to derive the scene specific optical depth. 

Radiance reaching the sensor is usually not only influenced by the properties of the reflecting ground 

inside the forests. Nevertheless detailed comparisons For details of this approach see Song et al. (2001). 

surface in the sensor’s instantaneous field of view (IFOV), but also by relief induced topographic 

of the different images and their good agreement in 

illumination effects as well as atmospheric scattering and absorption. Especially for multitemporal 

applications 

overlay suggest 

and 

that 



derivation 

remaining 


internal 

quantities 

variation 


For 

physical 

the calculation 

models, 


image 

surface 

calibration 

reflectance 

and 

ρ, 


sensor 

correction 

of the scenes 


is 

atmospheric 

low. 

effects is crucial. Many different 

specific 

radiative 

calibration 

transfer 

files 

codes 

and 


a digital 

atmospheric 

elevation 

correction exist, including the widely applied 6S (Second Simulation of the Satellite Signal in the Solar 

Table 4-3 

Reference data for geometric correction. 

Reference Coverage Scale Year Source 

Topographic Map Sheets 

30/III, 30/IV, 39/I, 39/II, 38/IV, 39/III 

Topographic Map Sheets 

102/2 and 102/4 

Budongo Forest 1:50,000 1966 East Africa 1:50 000 (Uganda), 

Series Y732 (D.O.S. 426), Edition I 

Kakamega Forest 1:50,000 1970 East Africa 1:50 000 (Kenya) 

Series Y731 (D.O.S. 423), Edition 6 

57

58 

model (DEM) that allows the correction of radiation 

scattered into the satellite’s IFOV from adjacent 

terrain (in this case a combined SRTM/ERS DEM 

from the W42 database at DLR-DFD) are needed 

as additional input. ρ can then be approximated 

iteratively according to Equation A.1 in the Annex. 

For further information on ATCOR and atmospheric 

correction theory please refer to Jensen (2007) and 

Richter (2007). 

All SPOT-4 scenes were corrected with ATCOR-3. 

Since ASTER data had only been available as ondemand 

level 2 surface reflectance product (USGS 

2006) and already contained atmospherically 

corrected data for both the VNIR and SWIR bands, 

ATCOR was not applied. The ASTER SWIR data 

had been further corrected automatically for the 

crosstalk phenomenon based on statistically derived 

radiometric sensitivity coefficients (USGS 2006). 

Acquisition and preprocessing details of the high 

resolution satellite scenes that were used in this thesis 

are summarized in Table 4-4. 

Table 4-4 

4.3 

Acquisition and preprocessing information for the high resolution satellite data. 

Study site Sensor Acqu. 

date 

Budongo Forest ASTER 15 Oct. 

2006 

Budongo Forest SPOT-4 

HRVIR 

Budongo Forest Landsat 

ETM+ 

Kakamega Forest SPOT-4 

HRVIR 

Kakamega Forest Landsat 

ETM+ 

01 Dec. 

2005 

23 May 

2000 

14 Dec. 

2004 

10 Jan. 

2003 

Solar 

azim. [°] 

Solar 

elev. [°] 

The MODIS LAI product 

In the following the MODIS LAI product 

(MOD15A2) will be described briefly with respect to 

instrument and data set characteristics, the algorithm 

used to derive LAI and the current validation status. 

It should be noted that both MODIS collection 4 (C4) 

and collection 5 (C5) were available for this thesis, as 

the former were processed within the BIOTA EAST 

AFRICA project. However, as C5 is the current 

version of the algorithm, validation will focus on 

these products. Because the following description 

only covers information relevant for validation, the 

interested reader is referred to Knyazikhin et al. 

(1998a, b) and Myneni et al. (1997) for further details 

concerning the theoretical basis of the algorithm. 

Tian et al. (2000) and Zhang et al. (2000) describe 

the algorithm prototyping, Panferov et al. (2001), 

Shabanov et al. (2003), Tian et al. (2002a), Wang 

et al. (2003a), and Shabanov et al. (2005) deal with 

the evaluation of the algorithm physics. Myneni et 

al. (2002), Tan et al. (2006), Wang et al. (2001), and 

Yang et al. (2006) provide product diagnostics. 

Tilt 

ang. [°] 

RMSE 

x [m] 

RMSE 

y [m] 

RMSE MODTRAN 

total [m] atmosphere 

118.20 68.51 n/a 6.95 9.23 11.55 Atm. correction 

already applied 

142.50 60.20 4.2 (L) 6.52 6.39 9.14 Tropical rural 

52.81 56.69 n/a 7.86 8.03 11.24 Tropical rural 

136.30 56.80 29.0 (R) 10.52 7.72 13.05 Tropical rural 

53.77 55.61 n/a 11.53 10.38 15.52 Tropical rural


4.3.1 

The MODIS instrument 

The MODIS sensor is a key instrument aboard 

the Terra and Aqua platforms (former EOS-AM 

and EOS-PM). The satellites were launched on 

18 December, 1999 and 4 May, 2002 respectively 

in a sun-synchronous, near-polar orbit at an altitude 

of 705 km. While Terra is on a descending node with 

a mean equator crossing time of 10:30 a.m. (see also 

Chapter 4.1.1 on ASTER data), Aqua operates on 

an ascending node with an equatorial overpass at 

1:30 p.m. Because afternoon cloud convection in 

tropical regions hinders ground observations with 

optical data, only data acquired by MODIS-Terra 

were used for this thesis. 

MODIS is a cross-track scanning spectroradiometer. 

It acquires data in 36 spectral bands with bands in 

the visible, near-, shortwave- and thermal infrared 

portions of the electromagnetic spectrum ranging 

from 0.4 to 14.54 µm (cf. Table 4-5 for bands relevant 

for MOD15A2, see Guenther et al. (2002) for all 

bands) in different spatial resolutions (250 m to 1000 

m). With a cross-track swath dimension of 2,330 km 

and 10 km along-track at nadir, global coverage can 

be achieved daily at latitudes higher than 30° and 

every two days for lower latitudes. Because pixel 

size increases with swath angle, scan lines overlap 

Table 4-5 

59 

towards the scene edges (known as bow-tie effect, 

see Masouka et al. [1998] for illustration). 

According to Wolfe et al. (2002) the mean geolocation 

error of MODIS data is quantified as 27 m alongtrack 

and 25 m along-scan. For further information 

on the MODIS instrument please refer to Barnes et 

al. (1998) and Guenther et al. (2002), as well as the 

special issues in Remote Sensing of the Environment 

(2002, Volume 83, Issues 1-2) and IEEE Transactions 

on Geoscience and Remote Sensing (1998, Vol. 36, 

Issue 4). 

Following reception, data are sent to the EOS Data 

and Operations System (EDOS) at the GSFC, where 

level 1A, level 1B, geolocation and cloud mask 

products as well as higher-level MODIS land and 

atmosphere products are generated by the MODIS 

Adaptive Processing System (MODAPS). The 

products are supplied to the user by three Distributed 

Active Archive Centers (DAACs) at GSFC, EROS 

Data Center (EDC) and the National Snow and Ice 

Data Center (NSIDC). Land Products are available 

through the Land Processes DAAC at EDC, also 

including data relevant to this research. Figure 4-1 

shows the global distribution of MODIS tiles with the 

tile h08v21, marked in red, covering the study sites. 

Spectral and spatial properties of the first seven bands of MODIS (relevant for MOD15A2). 

For further information on bands 8-36 (1000 m resolution) please refer to NASA (2007a). 

Band Wavelength [µm] IFOV at nadir [M] Quantization [bit] Key Applications 

1 0.620 – 0.670 250 12 Chlorophyll, land cover transformations 

2 0.841 – 0.876 250 12 Cloud amount, vegetation land cover transformation 

3 0.459 – 0.479 500 12 Soil/vegetation differences 

4 0.545 – 0.565 500 12 Green vegetation 

5 1.230 – 1.250 500 12 Leaf/canopy differences 

6 1.628 – 1.652 500 12 Snow/cloud differences 

7 2.105 – 2.155 500 12 Cloud properties, land properties

60 

Figure 4-1 

4.3.2 

MODIS tiles for processing levels 2G, 3 and 4. Study area marked in red (Wolfe 2002, modified). 

Data set characteristics 

The MODIS LAI/FPAR product (MOD15A2) is 

processed at 1 km spatial resolution and composited 

over 8 days based on the maximum FPAR value. Daily 

surface reflectance and cloud states (MOD09) as well 

as a biome map derived from the MOD12 land cover 

product and ancillary information serve as input to 

the MODIS LAI algorithm (cf. Figure 4-2). As LAI is 

a modelled output, MOD15A2 is regarded as level 4 

product. Further descriptions of the algorithm will be 

provided in Chapter 4.3.3. 

Similar to most other MODIS data products, 

MOD15A2 is stored in the EOS-HDF format 

and provided in the Sinusoidal Projection (cf. 

Figure 4-1). Table 4-6 shows the scientific data 

sets (SDS) and data specifications. Relevant 

for this thesis are only the SDS LAI_1km and 

LAIStDev_1km as well as the quality control data 

set FPAR_LAI_QC, for which the specifications 

are given in the Appendix in Table A-2. The key 

indicator for the retrieval quality is the SCF_QC 

bitfield. In order to provide a complete overview, the 

second quality control SDS, FparExtra_QC can be 

found in Table A-3 as well as the values assigned to 

non-vegetated surfaces of the SDS FPAR_1km and 

LAI_1km (cf. Table A-4). 

All available C4 MOD15A2 data sets from 2000 

to 2006 were processed within the BIOTA project 

to derive phonological characteristics for the study 

sites. For C5, at the time of writing only data sets 

covering the years 2000 through 2005 were available 

to analyse temporal consistency. Preprocessing was 

accomplished with TiSeG (Colditz et al. 2008). C5 

data sets of Julian Days 281 of 2004 (Kakamega 

Forest) and 329 of 2005 (Budongo Forest) served 

as input for validation with in situ data (coinciding 

with the field campaigns at Budongo and Kakamega 

Forest, respectively).


Figure 4-2 

Table 4-6 

Schematic processing chain of the MODIS LAI product (Knyazikhin et al. 1998, modified). 

MOD15A2 data set characteristics (USGS 2006b). 

SDS Units Data type (bit) Fill values Valid range Scale factors 

FPAR_1km % 8-bit unsigned integer 249-255 0 - 100 0.01 

LAI_1km m 2 /m 2 8-bit unsigned integer 249-255 0 - 100 0.10 

FPARLAI_QC Class flag 8-bit unsigned integer 255 0 - 254 n/a 

FPARExtra_QC Class flag 8-bit unsigned integer 255 0 - 254 n/a 

FPARStdDev_1km Percent 8-bit unsigned integer 248-255 0 - 100 0.01 

LAIStdDev_1km m 2 plant/m 2 ground 8-bit unsigned integer 248-255 0 - 100 0.10 

4.3.3 

Algorithm description 

As briefly addressed in Chapter 3.4.2, the physical 

modelling of LAI given surface reflectance is a 

complex 3D inversion problem. Radiation leaving 

the canopy of a vegetation stand depends not only 

on the LAI, sun and viewing geometries, but also 

on other canopy parameters, such as canopy closure 

and understorey vegetation. Therefore the socalled 

“space of canopy realization” is introduced 

(Knyazikhin et al. 1998b). In C5 it is based on an 

eight biome map, which describes canopy structural 

types of global vegetated land cover. However, as this 

61 

map is not available online so far, the classes of the 

old six biome map are summarized in Table 4-7 (land 

cover type 3 in the MOD12Q1 product). According 

to NASA (2007b) the only difference is that for C5 

the classes “broadleaf forest” and “needleleaf forest” 

are split up into deciduous and evergreen types. 

Each biome represents structural patterns of trees 

and/or other canopy types as well as other structural 

and spectral attributes relevant to radiative transport 

theory (cf. Table 4-8). For each variable the range 

of variation is limited within the respective biome 

(Knyazikhin et al. 1998b).

vegetated land cover. However, as this map is not available online so far, the classes of the old six biome 

map are summarized in Table 4-7 (land cover type 3 in the MOD12Q1 product). According to NASA 

62 

Table 4-7: Biome classes in (2007b) the only difference is that for C5 the classes “broadleaf forest” 

MOD12 land cover product as and “needleleaf forest” are split up into deciduous and evergreen types. 

underlying the MOD15A2 

The algorithm relation (USGS between 2006b). measured Also 

Each biome represents structural patterns of trees and/or other canopy 

surface reflectance 

refer to Table 4-10. 

types as well as other structural and spectral attributes relevant to 

and canopy structure is based on a 3D radiative 

radiative transport theory (cf. Table 4-8). For each variable the range of 

transfer model (compare Chapter 3.4.2). Bi-directional 

variation is limited within the respective biome (KNYAZIKHIN et al. 

reflectance factors (BRF) are modelled for the above- 

1998b). 

mentioned suite of canopy structures and soil patterns 

covering a range of expected natural The conditions relation between (Tian measured surface reflectance and canopy structure 

et al. 2000). The transfer model further is based fulfils on the a 3D law radiative transfer model (compare Chapter 3.4.2). Bi- 

of energy conservation, i.e. incident directional radiation equals reflectance factors (BRF) are modelled for the aboveabsorbed, 

transmitted and reflected mentioned radiation suite by the of canopy structures and soil patterns covering a range 

canopy (Knyazikhin et al. 1999). of The expected results natural are conditions (TIAN et al. 2000). The transfer model 

stored in a LUT, which is the interface further to the fulfils MODIS the law of energy conservation, i.e. incident radiation 

LAI algorithm. The algorithm compares equals observed absorbed, and transmitted and reflected radiation by the canopy 

modelled BRFs, taking into account (KNYAZIKHIN the uncertainties et al. 1999). The results are stored in a LUT, which is the 

in both (Myneni et al. 2002). Only interface those solutions to the MODIS LAI algorithm. The algorithm compares 

are accepted, where the magnitude observed of residuals and in modelled the BRFs, taking into account the uncertainties in 

comparison both (MYNENI does et not al. exceed 2002). uncertainties Only those in solutions modelled are accepted, where the magnitude of residuals in the 

and comparison observed does BRFs, not i.e. exceed uncertainties in modelled and observed BRFs, i.e. 

 

 

2 

obs 

mdl 

1 BRF ( , L, 

, , 

) ( , , , , 

) 

 

LC BRF L LC 

1 , (4.2) 

N 

 

 

 

 

obs 

mdl 

where BRF denotes observed and BRF modelled BRFs for N spectral bands with their 

uncertainties . 

and 1998b). In C5 it is based on an eight biome map, which describes canopy structural types of global 






algorithm (USGS 2006b). Also 




CLASS LAI/FPAR radiative transport theory (cf. Table 4-8). For each variable the range of 


0 water 

1998b). 

1 grasses/cereal crops 

2 shrubs 

The relation between measured surface reflectance and canopy structure 

is based on a 3D radiative transfer model (compare Chapter 3.4.2). Bi- 

3 broadleaf crops 

directional reflectance factors (BRF) are modelled for the above- 

4 savanna 

mentioned suite of canopy structures and soil patterns covering a range 

5 broadleaf forest 

of expected natural conditions (TIAN et al. 2000). The transfer model 

6 needleleaf forest further fulfils the law of energy conservation, i.e. incident radiation 

7 unvegetated 

equals absorbed, transmitted and reflected radiation by the canopy 

8 urban 

(KNYAZIKHIN et al. 1999). The results are stored in a LUT, which is the 

interface to the MODIS LAI algorithm. The algorithm compares 

254 unclassified 

observed and modelled BRFs, taking into account the uncertainties in 

both (MYNENI et al. 2002). Only those solutions are accepted, where the magnitude of residuals in the 

comparison does not exceed uncertainties in modelled and observed BRFs, i.e. 

 

 

2 

obs 

mdl 

1 BRF ( , L, 

, , ) ( , , , , ) 

 

LC BRF L 

LC 

1 , (4.2) 

N 

 

 

 

 

obs 

mdl 

where 

BRF denotes observed and BRF modelled BRFs for N spectral bands with their 

indicate the sun and view directions and LC stands for the land cover type. 

uncertainties . Although the use of biome dependent canopy attributes reduces the possible number of solutions of the 

inverse problem/for the inversion, there are still multiple possible results because observed BRFs can 

relate to several combinations of different LAI values and soil background. Therefore, mean LAI values 

averaged over all acceptable solutions are computed as output of the algorithm. Whereas only average LAI 

values for each pixel were provided by the C4 algorithm, a new SDS in C5 specifies also the standard 

deviation corresponding to the acceptable solution indicated (cf. Table 4-6). 

If the main algorithm fails to produce a valid solution for a certain pixel, a backup algorithm is applied. It 

estimates LAI based on biome dependent regressions as a non-linear function of NDVI. For low and 

medium NDVI values SHABANOV et al. (2005) report consistency of the LAI-NDVI curve with main 

 

and 1998b). In C5 it is based on an eight biome map, which describes canopy structural types of global 











CLASS LAI/FPAR 


0 water 

1998b). 



2 shrubs 


3 broadleaf crops directional reflectance factors (BRF) are modelled for the above- 

4 savanna 


5 broadleaf forest of expected natural conditions (TIAN et al. 2000). The transfer model 

further fulfils the law of energy conservation, i.e. incident radiation 

6 needleleaf forest 


7 unvegetated 


8 urban 


254 unclassified observed and modelled BRFs, taking into account the uncertainties in 



 

 

2 

obs 

mdl 

1 BRF ( , L, 

, , ) ( , , , , ) 

 

LC BRF L 

LC 

1 , (4.2) 

N 

 

 

 

 

obs 

mdl 

where BRF denotes denotes observed and and BRF modelled BRFs for N spectral bands with their 

uncertainties . indicate the sun and view directions and LC stands for the land cover type. 

Although the use of biome dependent canopy attributes reduces the possible number of solutions of the 









 

and vegetation. Therefore the so-called “space of canopy realization” is introduced (KNYAZIKHIN et al. 

1998b). In C5 it is based on an eight biome map, which describes canopy structural types of global 










CLASS LAI/FPAR radiative transport theory (cf. Table 4-8). For each variable the range of 

0 water 


1 grasses/cereal crops 1998b). 

2 shrubs 


3 broadleaf crops is based on a 3D radiative transfer model (compare Chapter 3.4.2). Bi- 

4 savanna 

directional reflectance factors (BRF) are modelled for the abovementioned 

suite of canopy structures and soil patterns covering a range 





7 unvegetated 


8 urban 


254 unclassified interface to the MODIS LAI algorithm. The algorithm compares 




 

 

2 

obs 

mdl 

1 BRF ( , L, 

, , ) ( , , , , ) 

 

LC BRF L 

LC 

1 , (4.2) 

N 

 

 

 

 

modelled 

obs 

mdl 

where BRF BRFs denotes for N spectral observed 

indicate bands and with the 

BRF 

sun their and 

modelled uncertainties view directions 

BRFs for 

and 

N 

LC 

spectral 

stands for 

bands 

the land 

with 

cover 

their 

type. 

uncertainties 

Although the 

. 

use of biome dependent canopy attributes reduces the possible number of solutions of the 









 

and egetation. Therefore the so-called “space of canopy realization” is introduced (KNYAZIKHIN et al. 


egetated land cover. However, as this map is not available online so far, the classes of the old six biome 

ap are summarized in Table 4-7 (land cover type 3 in the MOD12Q1 product). According to NASA 

able 4-7: Biome classes in (2007b) the only difference is that for C5 the classes “broadleaf forest” 

OD12 land cover product as and “needleleaf forest” are split up into deciduous and evergreen types. 

nderlying the MOD15A2 

lgorithm (USGS 2006b). Also 


efer to Table 4-10. 





0 water 

1998b). 



2 shrubs 



directional reflectance factors (BRF) are modelled for the above- 

4 savanna 







7 unvegetated 


8 urban 




oth (MYNENI et al. 2002). Only those solutions are accepted, where the magnitude of residuals in the 

omparison does not exceed uncertainties in modelled and observed BRFs, i.e. 

 

 

2 

obs 

mdl 

1 BRF ( , L, 

, , ) ( , , , , ) 

 

LC BRF L 

LC 

1 , (4.2) 

N 

 

 

 

 

obs 

mdl 


ncertainties . indicate the sun and view directions and LC stands for the land cover type. 

Although the use of biome dependent canopy attributes reduces the possible number of solutions of the 









 

and vegetation. Therefore the so-called “space of canopy realization” is introduced (KNYAZIKHIN et al. 




able 4-7: Biome classes in (2007b) the only difference is that for C5 the classes “broadleaf forest” 

OD12 land cover product as and “needleleaf forest” are split up into deciduous and evergreen types. 

nderlying the MOD15A2 

lgorithm (USGS 2006b). Also 


efer to Table 4-10. 



LASS LAI/FPAR 


water 

1998b). 

grasses/cereal crops 


shrubs 


broadleaf crops 

directional reflectance factors (BRF) are modelled for the abovesavanna 


broadleaf forest 



needleleaf forest 


unvegetated 


urban 






 

 

2 

obs 

mdl 

1 BRF ( , L, 

, , ) ( , , , , ) 

 

LC BRF L 

LC 

1 , (4.2) 

N 

 

 

 

 

obs 

mdl 


uncertainties . and indicate the sun and view directions and LC stands for the land cover type. 

lthough the use of biome dependent canopy attributes reduces the possible number of solutions of the 

nverse problem/for the inversion, there are still multiple possible results because observed BRFs can 

elate to several combinations of different LAI values and soil background. Therefore, mean LAI values 

veraged over all acceptable solutions are computed as output of the algorithm. Whereas only average LAI 

alues for each pixel were provided by the C4 algorithm, a new SDS in C5 specifies also the standard 

eviation corresponding to the acceptable solution indicated (cf. Table 4-6). 

f the main algorithm fails to produce a valid solution for a certain pixel, a backup algorithm is applied. It 

stimates LAI based on biome dependent regressions as a non-linear function of NDVI. For low and 

edium NDVI values SHABANOV et al. (2005) report consistency of the LAI-NDVI curve with main 

 

and Table 4-7 Biome classes in MOD12 land cover product as 

underlying the MOD15A2 algorithm 

(USGS 2006b). Also refer to Table 4-10. 

DN Biome class 


0 Water 

0 water 

1 Grasses/cereal crops 

2 Shrubs 


3 Broadleaf crops 

2 shrubs 

4 Savanna 


5 Broadleaf forest 

4 savanna 

6 Needleleaf forest 


7 Unvegetated 


8 Urban 

7 unvegetated 

254 Unclassified 

8 urban 


(4.2) 

low and medium NDVI values Shabanov et al. 

(2005) report consistency of the LAI-NDVI curve 

indicate the sun and view directions directions and with LC stands main algorithm for the land retrievals. cover type. However for NDVI 

Although the and use LC of biome stands dependent for the land canopy cover type. attributes Although reduces the the possible values greater number than of 0.82 solutions a constant of the LAI value of 6.1 

inverse problem/for use of biome the inversion, dependent there canopy are still attributes multiple reduces possible results is computed, because leading observed to retrieval BRFs anomalies can especially 

relate to several the combinations possible number of different of solutions LAI values of the and inverse soil background. for broadleaf Therefore, forests. mean LAI values 

averaged over problem/for all acceptable the solutions inversion, are computed there are as still output multiple of the algorithm. Whereas only average LAI 

values for each possible pixel results were provided because observed by the C4 BRFs algorithm, can relate a new to SDS Even in C5 though specifies the radiative also the transfer standard model is designed 

deviation corresponding several combinations to the acceptable of different solution LAI values indicated and (cf. soil Table to 4-6). incorporate up to seven MODIS bands, currently 

background. Therefore, mean LAI values averaged only surface reflectances of bands 1 (red) and 2 


over all acceptable solutions are computed as output (NIR) are incorporated (Shabanov 2001). In previous 


of the algorithm. Whereas only average LAI values versions of the MOD15A2 product (C1 to C3) 


for each pixel were provided by the C4 algorithm, a AVHRR land cover maps and modelled BRF values 

new SDS in C5 specifies also the standard deviation based on SeaWiFS data were processed (Cohen et al. 

corresponding to the acceptable solution indicated 2006). 


If the main algorithm fails to produce a valid 

solution for a certain pixel, a backup algorithm is 

applied. It estimates LAI based on biome dependent 

regressions as a non-linear function of NDVI. For 

Although some product anomalies have been resolved, 

other problems did still exist for C4 MODIS LAI. 

An analysis by Shabanov et al. (2005) indicated, for 

instance, a retrieval anomaly for broadleaf forests, for 

which a low portion of best quality main algorithm


retrievals are observed over broadleaf forests in 

northern America and Amazonia. Apparently there 

were still some inconsistencies between modelled and 

observed surface reflectances in the C4 algorithm. At 

the time of writing there are no available publications 

which analyse or address possible improvements of 

these issues with the C5 algorithm (cf. Chapter 4.3.4). 

Further details on the theoretical background of the 

MODIS LAI algorithm are provided by Knyazikhin 

et al. (1998b), Knyazikhin et al. (1999), Myneni et al. 

(1997), Myneni et al. (2002), Tian et al. (2000), and 

Wang et al. (2001). 

4.3.4 

Algorithm refinements in C5 

Beginning March 2005 the C5 version of MODIS 

level 1 science algorithms was released. Reprocessing 

of MODIS/Terra land products started in July 2006 

and was almost completed at the time of writing. 

Table 4-8 

63 

Apart from an improved atmospheric correction of 

surface reflectance products that serve as input to the 

MODIS LAI algorithm, major refinements include 

the replacement of six biome map with the new eight 

biome map (NASA 2007b, cf. Chapter 4.3.3). LUT 

values were further defined and refined for all eight 

biomes. A new stochastic radiative transfer model 

based on the work of Shabanov et al. (2007) was 

incorporated, which improved the representation of 

canopy structure and spatial heterogeneity intrinsic 

to woody biomes. Further, biome-dependent 

uncertainties were introduced (cf. Equation 4.2) 

as thresholds of allowable discrepancies between 

modelled and observed BRFs. For woody vegetation 

the uncertainties equal 30% for red and 15% for NIR 

reflectances. 

As a result of the above-mentioned changes, higher 

retrieval rates and improved consistencies with field 

measurements were observed for savannas, broadleaf 

Canopy structural attributes of the six biomes used in the C4 LAI algorithm (Knyazikhin et al. 1998). 

Grasses and 

cereal crops 

Shrubs Broadleaf 

crops 

Savannas Broadleaf 

forests 

Horizontal heterogeneity No Yes Variable Yes Yes Yes 

Needleleaf 

forests 

Ground cover 100% 20-60% 10-100% 20-40% > 70% > 70% 

Vertical heterogeneity 

(leaf optics and LAD) 

No No No Yes Yes Yes 

Stems/trunks No No Green stems Yes Yes Yes 

Understorey No No No Grasses Yes Yes 

Foliage dispersion Minimal 

clumping 

Random Regular Minimal 

clumping 

Clumped Severe 

clumping 

Crown shadowing No Not mutual No No Yes, mutual Yes, mutual 

Brightness of canopy 

ground 

Medium Bright Dark Medium Dark Dark

64 

and needleleaf forests (NASA 2007b). Improvements 

are also obvious for the study sites of this thesis and 

will be discussed in Chapter 7. 

4.3.5 

Validation status of the MODIS LAI 

product 

In order to evaluate the performance of the MODIS 

LAI algorithm with respect to various circumstances, 

i.e. for different biomes, seasons and atmospheric 

conditions, a comprehensive validation is needed. 

Generally and depending on the efforts already made 

for the validation of a certain satellite product, CEOS 

distinguishes three validation stages (Morisette et al. 

2006b): 

Stage 1 Product accuracy has been estimated using a 

small number of independent measurements 

obtained from selected locations and time 

periods and ground-truth/field program 

effort. 

Stage 2 Product accuracy has been assessed over 

a widely distributed set of locations and 

time periods via several ground-truth and 

validation efforts. 

Stage 3 Product accuracy has been assessed, and 

uncertainties in the product are wellestablished 

via independent measurements 

made in a systematic and statistically robust 

way that represent global conditions. 

The MODIS LAI product is currently classified as 

stage 1, indicating that more ground truth data and 

validation efforts are needed. As already mentioned 

in Chapter 1 validation activities have mainly been 

completed within the CEOS-LPV context (including 

the VALERI and BigFoot networks). Table 4-9 lists 

published studies concerned with the analysis of the 

C4 MODIS LAI product, the applied in situ methods, 

upscaling approaches and validation results. Further 

validation work, e.g. by Privette et al. (2002), Cohen 

et al. (2003), Kang et al. (2003), Scholes et al. (2004) 

or Tian et al. (2002a and b) was not taken into 

account as those studies employed previous versions 

of the MODIS LAI product (C1 and C3). Studies on 

C5 data have not been published so far. 

Generally, sites with low LAI values are reported to 

be mapped with satisfying accuracy by the MODIS 

LAI product (Cohen et al. 2006, Hill et al. 2006, 

Pandya et al. 2006, Tan et al. 2005). For woody 

biomes a weak correlation between in situ LAI and 

MODIS LAI is reported (Abuelgasim et al. 2006, 

Aragão et al. 2005) as well as an overestimation of 

MODIS LAI (Abuelgasim et al. 2006, Aragão et al. 

2005, Cohen et al. 2006, Hill et al. 2006). The studies 

mentioned in Table 4-9 also confirm a dominance 

of backup algorithm retrievals in C4 for forest sites 

(Aragão et al. 2005, Cohen et al. 2006). 

With respect to validation of C4 MODIS LAI for 

tropical rain forests three studies have been published 

so far (Aragão et al. 2005, Cohen et al. 2006 and Hill 

et al. 2006). However the methodology applied in Hill 

et al. (2006) to a tropical rain forest site in Australia 

involved the direct comparison of field measurements 

and MODIS LAI without intermediate upscaling 

steps. This methodology is seen as problematic 

by Tan et al. (2005) and Yang et al. (2006) among 

others, because of scale mismatch. Hill et al. (2006) 

also mention that the observed variations between 

MODIS LAI and in situ measurements are likely to 

be due to the exclusion of understorey vegetation, 

which was not included in field measurements. 

Cohen et al. (2006) took LAI-2000 PCA measurements 

as reference for validation of the C4 MODIS LAI 

product for a BigFoot primary rain forest site in 

the Tapajós National Forest in Brazil. Understorey 

vegetation was also not included in their study. 

Similar to Hill et al. (2006), this can partly explain


the reported overestimation of MODIS LAI. Cohen 

et al. (2006) further acknowledge problems with 

geolocation, which restricted the available test sites 

to only 10 of 100 plots for analysis. This limitation 

has been partly attributed to the small plot size of 

25 x 25 m, which had to be registered accurately to 

the respective Landsat ETM+ pixel to avoid errors in 

regression analysis. 

Aragão et al. (2005) also validated the C4 MODIS 

LAI product for the Tapajós region, but used larger 

plot sizes with a nested sampling design of three 

50 x 50 m subplots within an area of 270 x 270 m. 

Field measurements were again conducted with 

LAI-2000 PCA, thus excluding understorey 

vegetation. Aragão et al. (2005) also report an 

overestimation of MODIS LAI compared to a 

fine resolution LAI map with an RMSE of 1.54. 

Apparently no difference was made between MODIS 

LAI data derived from main and backup algorithms. 

To the author’s knowledge, no studies have been 

published on the validation of the C5 MODIS LAI 

product based on in situ data. 

Although there are some common approaches in 

the scientific community (cf. Table 4-16) especially 

with respect to the applied upscaling approaches, the 

description of validation results is fairly different. It 

ranged from purely descriptive quality assessments 

(Hill et al. 2006) to quantitative estimates of accuracy, 

precision and uncertainty (Tan et al. 2005). 

65

66 

comparison between aggregated fine resolution map and 

MODIS data 

Table 4-9 

Wang et al. 

(2004) 

Needleleaf forest 

(Finland) 

LAI-2000 PCA Empirical transfer function, univariate regression (RSR 

derived from Landsat ETM+ and in situ data), patch 

C4 MODIS LAI is accurate within a range of 0.5, 

precision: 48% 

average LAI values derived from fine resolution maps 

factor of 2 

measurement errors), comparison: MODIS LAI and 

Tan et al. 

(2005) 

Grasses & cereal crops 

(France) 

LAI-2000 PCA Empirical transfer function (SR derived from Landsat 

ETM+ and in situ data, including observation & 

C4 MODIS LAI is accurate within a range of 0.3 

with a precision of 20% and an uncertainty of 25%, 

biome misclassification decreases the accuracy by 

MODIS LAI and average LAI values derived from fine 

resolution maps 

Pandya et al. 

(2006) 

Grasses & cereal crops 

(India) 

LAI-2000 PCA Empirical transfer functions, non-linear regressions (NDVI 

derived from LISS-III and in situ data), comparison of 

MODIS C4 data were positively correlated to LISS- 

III derived LAI maps (r = 0.62-0.73), RMSE better 

than with C3 data (0.55-0.82) 

in tropical rain forest. Misclassifications in biome 

map strongly affect quality of MODIS LAI. 

Hill et al. 

(2006) 

Broadleaf forest, 

savanna (Australia) 

LAI-2000 PCA, 

hemispherical 

photography 

Direct comparison between in situ and MODIS LAI C4 MODIS LAI shows good correspondence with 

field measurements in savanna, overestimates LAI 

Fensholt et al. 

(2004) 

Savanna 

(Senegal) 

LAI-2000 PCA Direct comparison between in situ and MODIS LAI 

(homogeneity of sites evaluated with Landsat ETM+) 

Seasonal dynamics of in situ LAI was captured well 

by MODIS LAI (C4), but overestimation of MODIS 

LAI by approximately 2-15% 

and South America) 

> 50% for evergreen broadleaf forests 

Overview on validation studies of the MODIS LAI product (C4). 

Cohen et al. 

(2006) 

Grasses & cereal crops, 

needle-leaf forest, 

broadleaf forest (North 

LAI-2000 PCA, 

destructive sampling, 

allometry 

Empirical transfer functions (Landsat ETM+ and in situ 

data), comparison of MODIS LAI and average LAI values 

derived from fine resolution maps 

Overprediction of C4 MODIS LAI in forested biome 

sites with low LAIs correctly mapped (especially 

arid and semiarid regions), backup algorithm used 

MODIS LAI with respect to field LAI of 1-2 

Aragão et al. 

(2005) 

Broadleaf forest (Brazil) LAI-2000 PCA Empirical transfer functions (multivariate regression with 

terrain and spectral variables derived from Landsat ETM+) 

Weak correlation between C4 MODIS LAI and fine 

resolution LAI maps (predominance of backup 

algorithm over region), general overprediction of 

resolution maps, both aggregated to 3 km resolution 

forests (bias = -2.2 to 2.4) 

Abuelgasim 

et al. (2006) 

Needleleaf forest, and 

broadleaf forest, 

(Canada) 

LAI-2000 PCA, 

TRAC, hemispherical 

photography 

Empirical transfer functions, linear regression (RSR 

derived from Landsat ETM+ and in situ data), comparison 

of MODIS LAI and average LAI values derived from fine 

C4 MODIS LAI showed weak correlation to mixed 

forest stands (R2 =-0.06 to 0.34), overestimation 

of MODIS LAI for mixed and broadleaf dominated 

Reference Biome Field instrumentation Upscaling approach Results

5 Ground-based 

LAI measurements 

As indicated earlier, methods of in situ LAI sampling 

need an adaptation to the tropical rain forest 

environment. Based on the theoretical background 

given in Chapter 3, the establishment of an appropriate 

spatial sampling strategy will be described in the 

following. Subsequently, processing and analysis of 

field data will be presented together with a correction 

method that was developed for LAI-2000 PCA data 

acquired under non-diffuse illumination conditions. 

Finally the results of the LAI field sampling will 

be described, forming an independent basis for the 

validation of MODIS LAI data. 

Lessons learnt during the first field campaign in 

Kakamega Forest in October/November 2004 led 

to improvements in the sampling strategy and an 

elaborate correction of in situ data. Thus data of 

higher quality could be derived from the second 

field campaign that took place in Budongo Forest in 

October 2005. Data presented in Chapters 5.1 and 

5.2 originate from this study site only. Results from 

Kakamega Forest will, however, also be presented 

for comparison purposes in Chapter 5.3.2. 

5.1 

5.1.1 

Spatial sampling strategy 

Modifications made to the CEOS- 

LPV/VALERI methodology 

67 

The validation strategy of CEOS-LPV/VALERI was 

introduced in Chapter 3.4.5. These recommendations 

had to be modified in this thesis with respect to the 

complex vegetation structure in tropical rain forests, 

the area sampled with individual in situ measurements, 

as well as considerations of geolocation accuracy and 

upscaling. The main modifications related to the size 

of the whole test site, the size of individual sampling 

units and the spatial sampling scheme in terms of 

pattern and distance. 

Table 5-1 gives an overview of the recommendations 

made by VALERI on the above-mentioned issues and 

necessary modifications. First of all, the recommended 

size of the test site – 3 x 3 km – is too small to cover 

all forest stages and thus the variability in structure 

and LAI (compare logging compartments in Budongo 

Forest in Figure 2-9). The size of the test sites was 

therefore increased to cover the whole forest. 

Second, the VALERI scheme is based on sampling 

of ESUs. These correspond to the size of one pixel of 

the high resolution satellite image that is later used 

for upscaling of field measurements. In tropical rain 

forests this leads to difficulties for two reasons: 

a) Geolocation with GPS is difficult as 

the satellite signal is obstructed by the 

dense canopy. Often the reception of a 

sufficient number of satellites to locate the 

respective position with an acceptable error 

(i.e. 5-10 m) cannot be accomplished. 

Thus it can not be certain that the mean 

field LAI of an ESU can be assigned to the 

correct reflectance values retrieved from 

the corresponding pixel of high resolution 

satellite data (cf. Equation 3.13).

68 

b) To provide a statistically sound basis for 

later analysis, single LAI measurements 

should be spatially independent. At the same 

time a sufficient number of measurements 

per ESU should be taken (cf. Chapter 

3.3.4). This prerequisite cannot be fulfilled 

in tropical rain forests if the ESU size 

corresponds to pixel sizes of high resolution 

satellite imagery (i.e. 15 to 30 m) as the 

theoretical field of view of the instruments 

used is more than 30 m based on the settings 

applied in this thesis (cf. Figure A-1 and 

Equations A.2 and A.3 in the Appendix). On 

the other hand, in reality canopy elements 

(especially in dense canopies) block the 

sensor’s field of view at a smaller distance 

than the theoretical one. For example, in 

a seasonal moist tropical forest in French 

Guiana, Ferment et al. (2001) found LAI 

measurements derived from LAI-2000 PCA 

and DHP to be independent of each other at 

a distance of 10 m (at a significance level of 

5%). According to their results, a minimum 

distance between two measurements of 

10 m was defined, decreasing to 5 m in 

lower canopy types of young forest stages. 

In consequence of a) and b), and taking account 

of Equation 3.13, ESU size was increased to 200 

x 200 m. Also the spatial sampling pattern had to 

be modified. Instead of following the sampling in 

square or cross patterns recommended by Garrigues 

et al. (2002), a transect approach was followed as 

in De Wasseige et al. (2003), Leblanc et al. (2002), 

Le Dantec et al. (2000), Mussche et al. (2001), 

and Strachan & McCaughey (1996). This had the 

advantage that in both study sites existing transects 

established for scientific purposes by either the 

BIOTA EAST AFRICA project or the Budongo 

Conservation Field Station could be used as the test 

sites were simply not otherwise accessible due to 

dense understorey vegetation on many ESUs. As the 

different sampling patterns tested by Garrigues et 

al. (2002) gave similar results with respect to their 

representativeness of ESUs, the impact of the transect 

sampling is considered to be marginal. 

Table 5-1 Methodology proposed by VALERI for in situ sampling and modifications made with respect to the rain forest 

environment. 

Parameter VALERI recommendation Modification Reason for modification 

Size of test 3 x 3 km Size of forest Forests not easy to access (larger area 

site 

needed), focal point on sampling of 

representative amount of ESUs for each 

forest type 

Size of ESU Pixel size of high resolution 200 x 200 m GPS location difficult (larger ESU size 

satellite data (e.g. 15 x 15 m for 

minimizes error), individual measurement 

ASTER data) 

covers area larger in high canopies 

Sampling 

pattern 

Sampling 

distance 

Square or cross pattern 3 transects á 200 m Established transects could be used in 

Budongo and Kakamega Forest 

Not specified, 12 measurements 

per ESU recommended 

(cf. Figure 3-7) 

Either 20 or 40 per transect 

(distance 10 m/5 m inside/ 

outside forest respectively) 

Distance between individual measurements 

chosen according to canopy height to 

guarantee spatial independence

5 Ground-based LAI measurements 

5.1.2 

Field campaigns 

The field measurements took place from 25 October to 

29 November, 2004 in Kakamega Forest (main forest 

block and Kisere Fragment) and from 5 October to 

3 November, 2005 in Budongo Forest (Budongo and 

Kanyo-Pabidi forest blocks). In situ measurements of 

LAI were conducted in both study sites with DHP and 

LAI-2000 PCA. 30 ESUs were selected in each of the 

forests based on a stratified random approach. Land 

cover classifications (Lung 2004, cf. Figure A-2) 

and information on forest management and logging 

regimes in different compartments (cf. Chapter 2.1.3, 

available for Budongo only) were used to establish 

Figure 5-1 

69 

ESUs in different forest stages. Here the concept of 

Kalácska et al. 2004 was adopted, who instead of only 

distinguishing between primary and secondary forest 

areas, focused in their analyses on early, intermediate 

(disturbed) and late (undisturbed) forest stages. 

Early forest stages are thus referring to very young 

secondary (i.e. colonizing) forests, intermediate 

forest stages are selectively logged secondary forests 

and late forest stages signify natural (or near natural 

in case of Kakamega) primary forest areas. 

Whereas the first two correspond to early succession 

stages and disturbed secondary forest, the latter 

resembles nearly undisturbed primary forest. ESUs 

Location of ESUs in Kakamega Forest. a) Buyango Nature Reserve, Kambiri, Shanderema, Lugusi, 

b) Kisere fragment, c) Isecheno and Yala and d) Ikuywa (background: Landsat-ETM+ scene, 10 January, 2003).

70 

were always located more than 100 m away from 

the forest border to avoid mixed pixels in the high 

resolution satellite imagery. Yet the main restriction 

for ESU distribution in both forests was the 

accessibility to test sites. 

Figure 5-1 illustrates the spatial distribution of the 

30 ESUs on the test site in Kenya. The major part of 

ESUs was situated in the northern part of Kakamega 

Forest (21 ESUs in Buyangu Nature Reserve and the 

areas of Kambiri, Shanderema, Lugusi and the Kisere 

fragment, cf. Figure 5-1a and b). In addition, efforts 

were made to sample Isecheno and Yala (6 ESUs, 

Figure 5-1c), and Ikuywa (1 ESU, Figure 5-1d). 

Figure 5-2 

In Budongo Forest, all 30 ESUs were located in the 

forest blocks of Budongo and Kanyo-Pabidi (cf. 

Figure 5-2a-d). Whereas the major part was situated 

within a 3 km radius around Sonso camp site (Figure 

5-2a, 17 ESUs), some remote areas in Kanyo-Pabidi 

(Figure 5-2b, 4 ESUs), Busingiro (Figure 5-2c, 

3 ESUs) and Kidwera (Figure 5-2d, 5 ESUs) were 

sampled as well. 

Recent satellite imagery (Landsat 7 ETM+ satellite 

scenes acquired on 17 February 2000 for Budongo 

Forest and 10 January 2003 for Kakamega Forest, 

available within the BIOTA project) was analysed 

to determine relative homogeneity or heterogeneity 

Location of ESUs in Budongo Forest, a) around Sonso Camp site, b) in Kanyo-Pabidi, c) Busingiro and d) Kidwera 

(background: Landsat ETM+scene, 17 February, 2000).


of surface reflectance in the proposed area. Here 

especially brightness values of ETM bands 3, 4 and 

5 were taken into account. As larger changes in the 

forest canopy could have occurred in between the 

acquisition of the older satellite data and the time 

of fieldwork, visual assessments were accomplished 

in the field before sampling the respective ESU. 

Proposed ESUs were only accepted if the site and its 

surroundings proved to be homogeneous in structure 

(at ESU scale) and the topography was flat in order to 

minimize errors due to terrain effects. 

Within each ESU measurements were taken along 

three transects (Figure 5-3). In Budongo Forest 

transects established for chimpanzee research could 

be used in all forest stages except the early ones. In 

Kakamega Forest existing transects established by 

BIOTA scientists could be used on 14 of the 30 ESUs. 

On the others new transects had to be established. 

Examples from both test sites are displayed in 

Figure 5-4. 

In order to determine the correct distance between the 

individual measurements, two tapes of 100 m length 

each were laid out on the transects. GPS readings 

were taken with a non-differential GPS (GARMIN 

GPS III) at the beginning and end of the transects. 

Care was taken to only include readings with dilution 

of precision < 5 (i.e. satellites well distributed over 

the sky) and error of precision < 10 m, so that most 

measurements should be accurate to within 20 m. If 

valid readings could not be retrieved for a certain 

point, coordinates were calculated based on readings 

taken at other locations along the respective transect. 

5.1.3 

Measurement set-up 

LAI-2000 Plant Canopy Analyzer 

To calculate transmission (cf. Equation 3.4), reference 

readings were collected by a so-called “A sensor” 

Figure 5-3 

Figure 5-4 

71 

Modified sampling scheme (crosses 

represent single measurements; note that the 

distances in between varied according to canopy 

height, cf. chapter 5.1.1). 

Transects on ESUs in a) Kakamega Forest 

(ESU 2) and b) Budongo Forest (ESU 9). 

that was placed outside the forest or in forest glades 

as close as possible to the respective ESU (Figure 

5-5). The sensor was levelled and the unit set to 

remote logging mode with a sampling frequency of 

60 seconds. A shading construction behind the sensor 

(outside its field of view) assured that no direct 

radiation could hit the lens and cause errors due to 

reflections. As ideal illumination conditions for 

LAI-2000 PCA measurements, i.e. diffuse radiation 

only, are seldom present in the tropics and sunset 

and sunrise are too short for sampling a whole ESU, 

measurements were also made during the day (cf. De

72 

Wasseige et al. 2003). The influence of changing sky 

conditions on the measurements was recorded with 

a stationary LAI-2000 PCA unit set up at the ESU 

center (B1 sensor, cf. Figure 5-6a). The device took 

continuous measurements with the same frequency as 

the A sensor. This later helped to analyse the influence 

of θ and errors introduced by the spatial distance 

sun 

between the ESU and the A device due to moving 

cloud cover. Following the recommendations of 

Leblanc & Chen (2001) for the use of LAI-2000 PCA 

under sunny conditions, the sensors were always 

oriented towards north, i.e. away from the sun. The 

third device was used for readings along transects 

(B2 sensor). Prior to fieldwork, the instruments 

were intercalibrated under diffuse light conditions 

(LI-COR 1992, Weiss 2002). Intercalibration was 

repeated several times during the field campaigns. 

As LAI is highly non-linear to gap fraction (cf. 

Equation 3.3), but each reading represents a linear 

average of radiation recorded in the azimuthal view 

range, gaps in the canopy will be overweighted if 

the sensor sees large gaps and dense foliage at the 

same time. To minimize this error and the subsequent 

underestimation of LAI, all three sensors were 

equipped with a view cap with a 45° azimuth opening. 

Measurement of ground cover with the LAI-2000 PCA 

was avoided, as the instrument also responds non- 

Figure 5-5 

linearly to light interception by vegetation close to 

the sensor (Hyer & Goetz 2004). 

With the LAI-2000 PCA a minimum distance to 

foliage has to be assured that varies according to 

stand architecture, leaf size and to the view cap 

employed (LI-COR 1992). Consequently for many 

studies in forest environments LAI was sampled 

above the understorey vegetation (Mussche et al. 

2001, Stenberg et al. 2004). De Wasseige et al. 

(2003) further recommends a minimum distance 

to understorey foliage of 2 m when measuring 

in tropical rain forests. Consequently the optical 

sensor was levelled at 80 cm above ground inside 

the forest, thus disregarding the possible influence 

of understorey vegetation below that height (cf. 

Figure 5-6b). In the wooded grasslands outside the 

forest the sensor was held as close to ground level 

as possible. Measurements were repeated 4 times at 

each location. 

Digital hemispherical photography 

DHPs were acquired on the same spots and at the 

same heights as LAI-2000 PCA measurements in 

order to be able to compare both methods. They 

were obtained with a Nikon Coolpix 4300 camera 

Schematic illustration of measurement set-up with three LAI-2000 PCA devices (shaded area represents forest).


in Budongo Forest and a Nikon Coolpix 4500 (both 

Nikon Corporation, Tokyo, Japan) in Kakamega 

Forest (courtesy of Frédéric Baret, INRA, France and 

Jörg Szyarzynski, formerly ZEF, Bonn). Both have a 

charge-coupled-device (CCD) array with 4.1 million 

pixels (2272 x 1704). The cameras were equipped 

with a Nikon FC-E8 fisheye converter (183° field of 

view) with equidistant projection. Images were stored 

in JPEG format in fine resolution mode for storage 

reasons (compression factor 1:4). Frazer et al. (2001) 

found no significant influence of this compression 

on mean stand estimates of LAI compared to 

uncompressed TIFF images. 

The camera was mounted on a tripod, levelled at 

each location and oriented (cf. Figure 3-4) so that 

magnetic north was always at the top of the picture. 

Images were taken with 1 f-stop underexposure as 

underexposure was found to give better contrasts 

between foliage and sky (Frazer et al. 2001, Hale & 

Edwards 2002, Leblanc et al. 2005, Mussche et al. 

2001). Quality and contrast were controlled according 

to Leblanc et al. (2005) using the instant viewing 

mode. To capture LAI of the understorey vegetation, 

additional DHPs were taken at the same height with 

the camera looking downwards. 

5.2 

5.2.1 

Processing, correction and 

error analysis 

LAI-2000 PCA 

Data processing and LAI calculation 

Files stored in the LAI-2070 control unit were 

downloaded and processed with the FV 2000 software 

(LI-COR 2005). Prior to LAI calculation, false 

readings (e.g. double readings) were removed. Then 

B1/B2 readings were combined with the temporally 

closest A record to calculate transmission (Equation 

3.4). LAI was calculated according to Equation 3.9 

Figure 5-6 

Measurement set-up of LAI-2000 PCA. 

a) Stationary set-up of B1 instrument on ESU 5 

in Budongo Forest, b) measurements with B2 

instrument along transect 2 on the same ESU. 

73 

based on rings 1-4 only as forest gaps, in which the A 

sensor was placed, were not always large enough in 

diameter to assure that that ring 5 was not influenced 

by vegetation. No corrections for clumping were 

made, so that the output refers to effective LAI, from 

now on called LAI . As Leblanc & Chen (2001) 

e (LAI2000) 

recommended using solely the fourth ring when 

measurements are made under sunlit conditions, this 

calculation was also performed. 

Figure 5-7 shows LAI e (LAI2000) calculated from rings 

1-4 plotted against LAI e (LAI2000) calculated from 

ring 4 only. The results indicate that despite a high 

correlation between both data sets (adjusted R2 =0.97), 

mean LAI values retrieved from ring 1-4 

e (LAI2000) 

are with 4.51 (±2.17) slightly higher than those 

calculated from ring 4 only (4.18±2.36). In order to 

test whether both samples differ significantly from 

each other, a non-parametric Wilcoxon signed-rank 

test was performed. Although measurements from 

ring 4 are attributed the highest weight of all rings 

(cf. Table 3-4) and therefore contribute significantly 

to the calculation of LAI from rings 1-4, the 

e (LAI2000) 

results show a significant difference (p < 0.001%). 

Two arguments can, however, be made to justify the 

usage of LAI based on rings 1-4: first, one of 

e (LAI2000)

74 

the underlying assumptions of Equation 3.9 is that 

gap fraction is integrated over the whole range of 

zenith angles from 0 to π/2. This range is never fully 

met by the LAI-2000 PCA (cf. Chapter 3.3.2), with 

the outmost ring being always overweighted, yet it is 

better represented by four rings instead of only one. 

Second, LAI is likely to underestimate true 

e (LAI2000) 

LAI (cf. Fassnacht et al. 1994). As LAI based 

e (LAI2000) 

on rings 1-4 has a higher mean, it is probably closer 

to true LAI. 

Figure 5-7 Scatter plot showing LAI calculated from 

e (LAI2000) 

rings 1-4 versus LAI calculated from 

e (LAI2000) 

Table 5-2 

ring 4 only (N=12,659). 

Quantification of instrument inherent error 

In order to quantify the instrument inherent error of the 

LAI-2000 PCA, an experiment was conducted under 

stable light conditions in the spectroscopy laboratory 

of the Remote Sensing Technology Institute (Institut 

highe high 

LAI. 

Figu 

LAIe LAIe LAI 

versu vers 

only 

für Methodik der Fernerkundung, IMF) at the German 

only 

Quantification Aerospace Quantification Center of (Deutsches of instrument Zentrum inherent für Luft- error und 

In Raumfahrt In order to e. quantify V., DLR). the instrument Measurements inherent were error made of the LAI-2000 

under under stable a stable calibrated light conditions lamp for six in in hours the spectroscopy (no vegetation laboratory of the R 

(Institut present). (Institut für The für Methodik analysis of der the Fernerkundung, results indicated IMF) that the at at the German Aer Ae 

für instrument für Luft- und noise Raumfahrt is negligible e. e. V., DLR). with DLR). a Measurements coefficient of were made unde und 

vegetation variation vegetation (c present). ) of 1% or below for each single ring (cf. 

v present). The analysis of the results indicated that the in ini 

coefficient Table coefficient 5-2), of with of variation c being defined as 

v (cv) of 1% or below for each single ring (cf. Tab Ta 

 

 

c v 

. v . . 

 

 

(5.1) 

and are standard deviation and and mean mean of the of measurements. 


measurements. 

Table 5-2: Statistical measures for light intensity recorded by rings 1-4 

conditions (N= 397). 

Influence Ring of illumination N conditions Minimum Maximum 

1 1 397 0.506 0.518 0.512 0.51 

In order 2 2 to analyse 397 the effect of 0.760 non-ideal ideal 0.776 0.767 0.76 

illumination 3 3 conditions 397 on LAI-2000 0.636 PCA 0.660 0.651 0.65 

4 4 397 0.185 0.198 

measurements, test measurements were conducted 0.198 

on 30 October, 2005 in Budongo Forest over the 

whole day. The three devices were put up around 

Sonso Camp. The A sensor was placed at the camp 

side for reference readings, the B1 and B2 sensors in 

two different locations in secondary forest close to 

the camp. The canopy on both sites had a relatively 

dense understorey with trees of up to 3 m height. The 

remote logging mode of all devices was set to one 

minute. Figure 5-8 indicates the problems related to 

changing sun zenith angle and moving clouds. Figure 

5-8a shows LAI values plotted against θ . 

e (LAI2000) sun 

0.193 0.19 

Statistical measures for light intensity recorded by rings 1-4 of LAI-2000 PCA under laboratory conditions (N= 397). 

Ring N Minimum Maximum μ σ c v 

1 397 0.506 0.518 0.512 0.002 > 0.01 

2 397 0.760 0.776 0.767 0.003 > 0.01 

3 397 0.636 0.660 0.651 0.005 0.01 

4 397 0.185 0.198 0.193 0.002 0.01


It is obvious that LAI e (LAI2000) shows a certain variation 

throughout the day, probably due to moving clouds 

and scattering effects. What is remarkable, however, 

is the strong decrease in LAI with higher 

e (LAI2000) 

θ , i.e. with the sun being close to the horizon. If 

sun 

LAI values at θ > 70° are ignored, mean 


LAI over the day is 7.19 (cf. Table 5-5). At 

e (LAI2000) 

θ higher than 70° LAI strongly decreases, 

sun e (LAI2000) 

with a minimum of 2.54 at 90°. Transmission values 

plotted against θ in turn show a strong increase 

sun 

at higher θ (Figure 5-8b). As transmission is 

sun 

calculated as quotient of I (θ) and I (θ) (see 

trans o 

Equation 3.4) this indicates that the LAI-2000 PCA 

is probably not sensitive enough for below canopy 

measurements under a dense vegetation canopy 

and faint light conditions (see also De Wasseige et 

al. 2003). Light variations at high zenith angles are 

obviously not captured by the B1 sensor inside the 

forest, whereas the A sensor outside is still sensitive 

(cf. Figures 5-8c and d). The restriction to a 45° field 

of view additionally contributes to this error as less 

radiation can be recorded by the device. Figure 5-9, a 

hemispherical picture taken besides the B1 sensor at 

6:45 p.m. illustrates the dense canopy and the small 

amount of light reaching the forest floor at that time. 

In order to test the dependence of LAI from 

e (LAI2000) 

θ , erroneous measurements were excluded for all 

sun 

three devices (A, B1 and B2), i.e. values recorded 

at θ >70°. Table 5-3 displays the descriptive 

sun 

statistics of the resulting data sets. Mean LAIe (LAI2000) 

values calculated from A/B1 and A/B2 sensors are 

very similar (7.19 and 7.27) as well as the standard 

deviations (0.29 and 0.27 respectively). The relatively 

large range of values (2.36 and 2.71) however 

indicates a strong variation over the day, probably 

due to errors introduced by changes in blue light 

scattering (changing θ ) and moving clouds. 

sun 

A Kolmogorov-Smirnov test was applied to test 

whether LAI is normally distributed (Sachs 

e (LAI2000) 

2004). As the resulting probability of error p was 

Figure 5-8 

Figure 5-9 

Test measurements with LAI-2000 PCA made 

on 30 October, 2005. a) LAI calculated 

e (LAI2000) 

from continuous A/B1 readings, b) transmission 

calculated from continuous A/B2 readings. The 

respective solar zenith angle is shown on the x 

axis. 

75 

Hemispherical photograph taken alongside to the 

above-mentioned B1 sensor at 6:45 p.m.

76 

in both cases > 0.05, the null hypothesis (normal 

distribution) can be accepted. In consequence the 

relation between LAI and θ was tested with 


Pearson’s correlation coefficient r. For both test sites a 

correlation significant at 0.01 level could be retrieved 

between LAI and θ , yet with the relation 


between the variables being relatively weak (r = 0.39 

and 0.35 respectively). As A, B1 and B2 readings 

are presumably influenced by both θ effects and 

sun 

errors introduced by moving clouds (i.e. changing 

illumination conditions), these error sources will be 

analysed in more detail. Additionally a correction 

scheme for both effects is proposed in the following. 

Table 5-3 

Correction of illumination effects 

Apart from the instrument inherent error, in situ LAI 

calculated with LAI-2000 PCA is mainly biased by 

non-diffuse illumination conditions. The following 

two error components influence the readings: 

a) errors introduced by changing sun zenith 

angle (and with it the amount of blue light 

scattered within the canopy), e , and 

θsun 

b) errors introduced by differences in incident 

radiation due to cloud movement and the 

spatial distance between A sensor and B 

sensor, e . cloud 

Descriptive statistics of continuous LAI-2000 PCA measurements (A/B1 and A/B2) recorded on 30 October, 2005. 

A/B1 devices A/B2 devices 

LAI recorded at starting time (8:04 a.m.) 6.90 7.48 

LAI recorded at ending time (5:16 p.m.) 6.49 7.66 

LAI recorded at minimum θ sun (14.86° at 12:40 am) 6.83 6.83 

Maximum LAI 8.04 (at 12:45 a.m.) 7.80 (at 8:52 a.m.) 

Minimum LAI 5.68 (at 2:21 p.m.) 5.09 (at 5:10 p.m.) 

Mean LAI (θ sun < 70°) 7.19 7.27 

Standard deviation (θ sun < 70°) 0.29 0.27 

N 553 539 

Table 5-4 

Allocation of ESUs to the respective radiation regimes and necessary processing steps. 

Radiation regime ESU number Necessary processing steps 

Diffuse radiation 1, 5, 14, 15, 16, 20, 24, 28 No correction necessary 

Direct radiation 2, 3, 7, 11, 18 Correction for e θsun 

Changing radiation regime 4, 6, 8, 9, 10, 12, 13, 17, 19, 21, 22, 

23, 25, 26, 27, 28, 30 

Filter for valid measurements


Figure 5-10 LAI calculated from continuous B1 readings made on ESU 7 on 10 October, 2005, together with hemispherical 

e (LAI2000) 

pictures, taken at θ =38° (10:14 a.m.) and 61° (8:41 a.m.) respectively. Note cloud presence in right, cloud absence 

sun 

in right picture. 

Under perfectly diffuse light conditions, both e θsun 

and e cloud are 0. In case of clear skies e cloud is 0, so 

that LAI e (LAI2000) is influenced by e θsun only. In case 

of changing illumination conditions, LAI e (LAI2000) 

derivation is biased by both e θsun and e cloud . 

Figure 5-10 illustrates the influence of e θsun and e cloud 

on LAI e (LAI2000) retrieved from B1 readings taken 

on 10 October, 2005 on ESU 7. The A sensor had 

been situated in the camp. This was the closest large 

enough clearing and approximately 1.6 km away. 

Measurements were made between 8:12 a.m. and 

11:20 a.m. (corresponding to θ of 67° and 22°). The 

sun 

morning hours up until 9:47 a.m. (θ = 44°) were 

sun 

influenced by passing clouds, thus both, e and e θsun cloud 

77 

biased the measurements over both A and B1 sensors. 

After 10 a.m. the skies remained clear, so that only 

e influenced the reading. 

θsun 

As the location of the B1 sensors remained stable 

during the time of measurement, the derived 

LAI is not influenced by local differences in 

e (LAI2000) 

canopy structure, unlike the B2 device. Consequently 

B1 measurements derived for each ESU were used for 

analysing and partly correcting the above-mentioned 

errors. Therefore ESUs were separated into three 

groups based on notes made during measurements 

and evidence coming from the digital hemispherical 

pictures (cf. Table 5-4):

78 

1) ESUs on which the measurements were made 

under diffuse radiation conditions, 

2) ESUs that were visited under completely clear 

skies, and 

3) ESUs where the conditions changed during the 

time of measurement. 

As expected, the majority of ESUs (17 out of 30) visited 

had experienced a changing radiation regime over 

time. As overcast skies represent ideal conditions for 

LAI-2000 PCA measurements, no further correction 

is applied to group 1. In this case, the mean LAIe (LAI2000) 

for the respective ESU was calculated directly from 

A and B2 measurements. For the other two cases 

correction schemes or filters had to be applied. 

Figure 5-11 LAI derived from continuous B1 

e (LAI2000) 

measurements on ESUs 16 (blue), 18 (red) 

and 30 (green). 

Figure 5-11 shows LAI calculated from B1 

e (LAI2000) 

readings on three ESUs for different θ . Each ESU 

sun 

represents one of the above-mentioned radiation 

regimes. Whereas LAI values on ESU 16 

e (LAI2000) 

(blue) show only little variation, the increasing eθsun with decreasing θ is obvious for ESU 18 (red). ESU 

sun 

30 (green) is a characteristic example of the strong 

influence of e on derived LAI . 

cloud e (LAI2000) 

Table 5-5 indicates that, as expected, ESU 16 has 

the smallest c v of LAI e (LAI2000) values with 1%. This 

corresponds to the instrument inherent error. Although 

c for ESU 18 also seems to be relatively small (3%), 

v 

it could be much higher at ESUs measured at smaller 

θ sun since under direct radiation conditions e θsun is 

assumed to be 0 at dawn (only diffuse radiation 

present) and increasing with decreasing θ (amount 

sun 

of direct radiation penetrating the canopy increases). 

The largest c v was retrieved for ESU 30. Here the 

variation accounts for 10% of the arithmetic mean and 

individual LAI values show a range of almost 

e (LAI2000) 

2. Certainly the large variation is also a function of 

distance between the A and B1/B2 sensors. 

To develop a correction for e , all five ESUs that 

θsun 

were measured at least partly under cloud-free 

conditions were analysed. B1 measurements taken 

under changing radiation conditions (as e.g. at 

θ > 44° on ESU 7, cf. Figure 5-10) were excluded 

sun 

from analysis. To compare LAI values from different 

ESUs LAI was centred for θ = 50° (as LAI sun e (LAI2000) 

values measured at a sun zenith angle of 50° were 

present on all ESUs) with 

Table 5-5 Descriptive statistics for LAI derived from continuous B1 measurements on the ESUs shown in Figure 5-10. 

e (LAI2000) 

ESU N Minimum Maximum μ σ c v 

16 117 8.23 8.60 8.38 0.070 0.01 

18 118 6.09 6.77 6.45 0.166 0.03 

30 130 3.08 5.07 3.98 0.394 0.10


cent. LAI e (LAI2000) = LAI e (LAI2000) -LAI e (LAI2000, 50°) , 

(5.2) 

where LAI e (LAI2000, 50°) is LAI e (LAI2000) measured at 

θ sun = 50° on the respective ESU. Consequently 

centered LAI e (LAI2000) is 0 at θ sun = 50° with negative 

values at smaller and positive values at higher sun 

zenith angles on average. Figure 5-12a displays 

centered LAI values for the five ESUs 

e (LAI2000) 

measured under direct radiation plotted against θ . sun 

Linear regression was performed subsequently with 

θ as an independent variable. The relation between 

sun 

centered LAI and θ can thus be modelled as 


cent. LAI e (LAI2000) = 0.012* θ sun - 0.532. 

(5.3) 

Presumably LAI values measured at very 

e (LAI2000) 

high sun zenith angles (i.e. close to dawn or dusk 

with mostly diffuse radiation) are not influenced by 

e θsun , so that they represent the actual LAI e (LAI2000) . In 

order to avoid sensitivity problems with the device 

(cf. Figure 5-8), centered LAI at θ = 80° 


(and not 90°) was considered to be unaffected by e . θsun 

Consequently a correction factor θ /LAI sun e (LAI2000, 80°) 

could be derived for each ESU, with LAIe (LAI2000, 80°) 

being centered LAI calculated with Equation 

e (LAI2000) 

R²=0.84 

Y=-0.532+0.012 X 

Figure 5-12 a) Scatter plot of LAIe 

(LAI2000, cen) and θsun , solid line represents linear regression and b) same data corrected for eθsun . 

79 

5.3 for θ = 80°. Figure 5-12b displays this correction 

sun 

factor applied to centered LAI , which is thus 

e (LAI2000) 

corrected for e θsun . To derive corrected LAI e (LAI2000) for 

each ESU, the corrected centered LAI must 

e (LAI2000) 

be summed up with the ESU specific LAIe (LAI2000, 50°) 

(cf. Equation 5.2). B2 measurements were corrected 

correspondingly, assuming that e had a similar 

θsun 

effect on B1 and B2 measurements taken at the same 

θ . Whereas e can thus be eliminated, variability 

sun θsun 

due to random noise (e.g. caused by sunflecks) still 

remains (cf. Figure 5-12b). 

For several reasons it is difficult to establish a 

correction for e . First of all, the changing radiation 

cloud 

regimes and proportions of direct and diffuse 

radiation influencing A and B1 sensors can probably 

not be reconstructed. Modelled curves of light 

intensity would be needed each day for the reference 

site and the ESU. Whereas this could probably be 

accomplished for the A sensor (as it was usually 

taking readings over the whole day), problems occur 

with the B1 sensor, as measurements rarely cover 

more than three hours per ESU. Effects introduced by 

moving clouds can thus not sufficiently be modelled 

and corrected for.

80 

Consequently instead of developing a correction 

for e , a filter was applied to LAI transect 

cloud e (LAI2000) 

measurements that were taken under changing sky 

conditions. B1 measurements served here as “quality 

control”. Mean LAI and standard deviation 

e (LAI2000) 

were calculated first for B1 derived LAI for each 

ESU. If the corresponding value was within μ±σ, the 

B1 value was declared as valid. LAI calculated 

e (LAI2000) 

from transect measurements were then compared to 

the time-corresponding B1 measurements. If these 

were valid, LAI calculated from B2 were 

e (LAI2000) 

included in further analysis. 

5.2.2 

DHP 

Data processing and LAI calculation 

All DHPs were processed using the CAN- 

EYE software (Version 4.2, cf. Baret 2004a). 

Compared to other available software packages 

(e.g. GapLightAnalyzer by Frazer et al. 1999, or 

HemiView Canopy Analysis Software by Delta-T 

Devices Ltd., 1999) the programme has several 

advantages, such as, for instance, a simultaneous 

analysis of up to 20 pictures acquired under the 

same canopy and illumination conditions (with 

most other software packages, pictures have to be 

analysed individually). In addition, the software is 

based on colour photography, which simplifies the 

classification process and is also more accurate than 

single- or two-value thresholds based on black and 

white pictures. Before processing, the optical centre 

of the camera system was determined. Azimuthal and 

zenithal computations are based on this, but usually 

there is slight deviation from the centre point of the 

images due to lens distortions. Not correcting this 

factor would introduce an error to LAI calculation. 

The determination of the optical centre was done 

according to the method proposed by Baret (2004b). 

The black cross in Figure 5-13 (see below) marks the 

centre point of an image (pixel coordinates 1136, 852) 

taken with the Nikon Coolpix 4300 camera. Blue dots 

mark the position of holes in an objective lid that had 

been rotated after image acquisition. The procedure 

was repeated 12 times. As the holes rotate around 

the optical centre of the system, its coordinates can 

be calculated. For the Nikon Coolpix 4300 system, 

the optical centre is slightly shifted towards to 

the upper left compared to the image centre (from 

pixel coordinates 1180, 862 to 1163, 852). Before 

processing, all images were further checked for 

quality, i.e. same illumination conditions for pictures 

processed together. 

The CAN-EYE software starts with the definition of 

a calibration parameter file that identifies the image 

dimensions, the optical centre, as well as the circle of 

interest (COI) in °, i.e. the zenithal area to be analysed 

(cf. Table A-5 for the camera system used in Budongo 

Forest). According to Frédéric Baret (personal 

communication), the COI should be restricted to 

60° since, depending on the camera system, the 

image sharpness may decrease significantly at larger 

zenith angles and mixed pixels (i.e. containing gaps 

and foliage) occur. Also Frazer et al. (2001) noticed 

strong blurring at zenith angles higher than 45° with 

1,500 

1,000 

500 

0 

0 500 1,000 1,500 2,000 

Figure 5-13 

Optical centre (red cross) of Nikon Coolpix 4300 

camera used in Budongo Forest. X and Y units 

are pixels.


a Nikon Coolpix 950 camera and FC-E8 fisheye. 

However, the loss of fine canopy details in their 

study may partly be attributed to the comparably low 

resolution of their camera system (1.92 Megapixel). 

For processing, all images are divided into small 

sectors of 2.5° zenith and 5° azimuth angular 

resolution (cf. Figure 5-14). The former are needed to 

compare derived gap fraction with the modelled one 

Figure 5-14 

Figure 5-15 

Illustration of DHP area used for LAI derivation: 

COI (60°) split up into 2.5° zenith and 5° azimuth 

sectors. 

Subset of hemispherical photograph as a) original and b) classified image. Red circles illustrate the problem of mixed 

pixels and their classification. 

81 

(cf. Equation 5.6) and estimate LAI. Azimuth sectors 

help to derive the clumping index. 

As the software cannot cope with more than 20 

pictures per ESU, processing was done on transect 

basis and compiled later with an extension of CAN- 

EYE developed by Weiss (2006). The images were 

masked if necessary (appearance of solar disk, 

illuminated stems and branches, feet of the user 

in case of downward-looking pictures, etc.). After 

a reduction to 324 colours, gaps are classified 

interactively (cf. Figure A-3 for illustration). 

A critical step in determining gap fraction from 

hemispherical photographs is the classification 

process, as illustrated in Figure 5-15. Transitions 

between sky and foliage are not always sharp (Figure 

5-15a), so that mixed pixels occur. This makes the 

selection of pixels belonging to gaps in the canopy 

somewhat arbitrary and subjective (cf. classified 

image in Figure 5-15b). The consequences of this 

operator effect will be analysed later. 

Based on the classification, gap fraction is computed 

for each sector as the relation of pixels classified as 

foliage and non-foliage. In principle and analogue 

to the LAI retrieval with LAI-2000 PCA, Miller’s 

derivation (see Equation 3.7) could be used to solve

82 

inclination angle (WEISS et al. 2004) 

(2) The second method is based on model inversion. Assuming an 

Fig. 5-16: Variation of the G-function (mean projection of unit foliage area) with the average leaf 

inclination angle (WEISS et al. 2004) a large range of random LAI and ALA combinations and th 

(2) The second method is based on model inversion. defined zenith Assuming angles an are ellipsoidal stored in leaf a LUT. inclination Possible distribution, LAI values 

Equation 3.3 and compute LAI from gap fraction weighted cost-function, where the distance of 

derived from hemispherical photographs. As already the k 

mentioned, the basic assumption with Miller’s 

formula however is that gap fraction is integrated 

from 0 to π/2. Gap fraction retrieved from DHP at 

larger zenith angles might, however, include large 

uncertainties due to blurring. Therefore an approach, 

that allows LAI retrieval from gap fraction at lower 

zenith angles only, is preferred. 

th a large range of random LAI and ALA and combinations ALA 10° and 80° the (in corresponding steps of 2°). The P LUT element close 

o values for userdefined 

zenith angles are stored in a LUT. 

chosen 

Possible element based 

LAI 

on 

values of a the weighted 

lie LUT between to cost-function, measured 0 and 10 gap (in 

where 

steps of 


0.01) 

distanc 

and ALA 10° and 80° (in steps of 2°). The 

measured fraction LUT is element 

gap computed fraction 

closest as is computed 

to the derived 

as 

gap fraction is then 

chosen based on a weighted cost-function, where no. 

zenith the angles distance of the k 

LUT ( k ) MES 2 

wi 

P0( i 

) P0 

( i 

) 

LUT ( k) 

i1 

ALA 60 

ck 

 

 

MES 

( ( )) 

30 

MOD P0 

i 

th element of the LUT to 

measured gap fraction is computed as 

no. 

zenith angles 

LUT ( k ) MES 2 

wi 

P0( i 

) P0 

( i 

) 

LUT ( k) 

i1 

ALA 60 

c 

 

(5.6) 

k 

(5.6) 

MES 

( ( )) 

30 

MOD P0 

i 

IN SITU LAI MEASUREMENTS 

17 

16 

LUT(k) MES The IN SITU CAN-EYE LAI MEASUREMENTS software offers two solutions for with P (θi ) and P (θi ) being modelled 

0 

0 

16 

Equation derived 3.3 from (Weiss hemispherical et al. 2006): photographs. As already mentioned, and measured the basic gap assumption fractions for with azimuthally Miller’s 

derived from hemispherical photographs. As already mentioned, 17 

formula however is that gap fraction is integrated from 0 to averaged the basic assumption with Miller’s 

/2. Gap LUTzenith 

( kfraction 

) sector retrieved MES i. w from is a DHP weighting 

i with P at 

0 ( i ) and P0 ( i ) being modelled and measured 

(1) formula 

larger According however 

zenith angles to is that 

might, Warren-Wilson gap fraction is 

however, include (1963) integrated 

large the from 0 to 

uncertainties function /2. Gap fraction retrieved from DHP at 

zenith due sector to blurring. i. wi is a Therefore weighting an function approach, 

that 

larger 

allows G-function zenith 

LAI 

angles 

retrieval can might, be considered from 

however, 

gap fraction as include independent at lower 

large uncertainties 

zenith angles only, 

due LUT to 

is ( blurring. 

preferred. 

k ) Therefore MES an approach, 

with P 

that allows of leaf LAI inclination retrieval at from a zenith gap angle fraction of 57.5°, at lower i.e. zenith angles only, 

Npix 0 

is preferred. 

i ( 

Nmask i ) and P 

i 0 ( i ) being modelled and measured 

The CAN-EYE software offers two solutions for Equation 3.3 

zenith 

w i 

sector 

(WEISS et Npix al. 

i. (5.7) 

wi 2006): 

is a weighting , 

function 

i 

The CAN-EYE here G=0.5 (cf. software Figure offers 5-16). two Based solutions on Warren for Equation 3.3 (WEISS Npixet 

al. 2006): 

i Nmaski 

(1) According Wilson’s work, to WARREN-WILSON Bonhomme et (1963) al. (1974) the G-function where wcan 

i be Npix considered iis 

is the number as independent of of pixels in of in zenith leaf sector i and Nm 

i 

(1) According 

inclination derived LAI at 

to WARREN-WILSON 

for a zenith young angle crops and of 57.5°, 

(1963) 

found i.e. 


a good here 

G-function 

G=0.5 (cf. 

can 

Figure 

be considered Npix i 

5-16). Based 

as 

on 

independent 

Warren Wilson’s 

of leaf 

sector i. The function i and Nmask takes is into the account number that of masked 

i some zenith directions 

work, 

inclination at a zenith 

agreement BONHOMME to direct et measurements. al. 

angle 

(1974) 

of 57.5°, 

derived 

i.e. 

Therefore LAI 

here 


G=0.5 

young 

(cf. 

crops where Figure 

and found 

5-16). Based 

a good 

on 

agreement 

Warren Wilson’s 

pixels in and Npix i. is The inormalized 

is function the number takes by of into pixels account to in direct zenith that sector i and Nm 

measurements. 

work, BONHOMME 

LAI can be calculated Therefore 

et al. (1974) 

according LAI can 

derived 

tobe 

calculated 

LAI for 

according 

young crops 

to 

and found a good agreement to direct 

i. some The function zenith directions takes into may account contain that a some large zenith directions m 

no. 

zenith angles 

measurements. Therefore LAI can be calculated according to 

LAI 

pixels amount and wof 

is 

masked i 1. normalized pixels by and is normalized by 

0. 

5 

P 57. 5 

cos 57. 

7 

o e LAI 

i 1 

0. 

5 , (5.4) 

(5.4) 

no. 

zenith angles 

P 57. 5 

cos 57. 

7 

o e , The root (5.4) 

mean w square error between modelled and measured gap 

i 1. 

(5.8) 

which is equivalent to 

i 1 

which is is equivalent to 

sectors (cf. Equation 5.6) is divided by a modelled standard deviatio 

lnPo57. 5 

The root mean square error between modelled and measured gap 

LAI 

lnP . This modelled standard deviation is derived (5.5) by fitting a second o 

0o. 

93 57. 5 

LAI 

. . (5.5) The sectors (5.5) 

0. 

93 

empirical 

root (cf. mean Equation 

standard 

square 5.6) 

deviation 

error is divided 


between by a modelled 

measured 

modelled standard deviatio 

gap fraction in a certai 

Yet as this method was verified only for agricultural and vegetation This images 

measured modelled processed types gap standard and together 

fractions the portion deviation and 

over 

belonging of all is mixed derived user-defined 

to pixels the by same fitting transect. a second Tho 

Yet 

increases Yet as as this this with 

method method zenith 

was 

angle, was verified 

this verified calculation 

only for only agricultural 

was for not zenith performed. 

vegetation types and the portion of mixed pixels 

empirical called 

sectors 

regularization standard (cf. Equation deviation term that 

5.6) of puts measured is 

constraints 

divided gap by 

on fraction a 

the retrieved in a certai ALA 

increases agricultural with vegetation zenith angle, types this and calculation the portion was of not modelled performed. 

images processed standard together deviation and of belonging the measured to the gap same transect. The 

mixed pixels increases with zenith angle, this fraction The LUT gap fraction providing the minimum value of c k and its c 

called regularization values. This modelled term that standard puts constraints deviation on is the retrieved ALA 

calculation was not performed. 

derived then considered by fitting as a solution. second This order output polynomial will be further to called LAIe 

The MES LUT gap fraction providing the minimum value of c k and its c 

σ (P (θi Quantification )), the empirical of system standard inherent deviation errors of 

0 

(2) The second method is based on model inversion. measured then considered gap fraction as solution. in a certain This zenithal output will direction be further called LAIe ( 

Assuming an ellipsoidal leaf inclination resulting 

Apart from the above-described deviation of the optical centre f 

Quantification from all images of system processed inherent together errors and 

distribution, a large range of random LAI and belonging 

distortions 

to the 

can 

same 

also introduce 

transect. The 

errors 

second 

in the 

term 

calculation 


of LAI. Fi 

Apart 

ALA combinations and the corresponding P 

between from the the image above-described radius r in pixels deviation and the of corresponding the optical centre zenit f 

Equation 5.6 is a so-called regularization term that 

0 

distortions equidistant can projection, also introduce which is errors in the calculation of LAI. Fig 

values for user-defined zenith angles are stored puts constraints on the retrieved ALA values (Weiss 

between the image radius r in pixels and the corresponding zenith 

in a LUT. Possible LAI values lie between 0 2006). The d LUT gap fraction providing the minimum 

e 

equidistant . projection, which is 

and 10 (in steps of 0.01) and ALA 10° and 80° value 90of 

cr and its corresponding LAI and ALA values 

k 

(in steps of 2°). The LUT element closest to the are then 

dconsidered In order e as solution. This output will be 

to . test the projection quality for the camera system in us 

derived gap fraction is then chosen based on a further 90 

called r LAI . 

Fig. 5-16: Variation of the G-function (mean projection 

(2004b) 

of unit 

was built e (DHP) 

foliage 

(cf. 

area) 

Figure 

with 

A-4). 


Based 

average 

on this 

leaf 

the projection fu 

inclination Fig. 5-16: Variation angle (WEISS of the et G-function al. 2004) (mean projection In The order results of unit to test in foliage Figure the projection area) 5-18 with show quality the that average for the the projection leaf camera system functions in use 

inclination angle (WEISS et al. 2004) 

(2) The second method is based on model inversion. Assuming 

(2004b) equidistant an 

was 

ellipsoidal projection. built (cf. 

leaf 

Figure It inclination can A-4). thus be Based 

distribution, assumed on this that the no projection further error fu 

(2) The second method is based on model inversion. Assuming The lens system. results an ellipsoidal in Figure leaf 5-18 inclination show that distribution, the projection functions 

a large range of random LAI and ALA combinations and the corresponding P o values for user- 

a large range of random LAI and ALA combinations equidistant and the projection. corresponding It can P thus 

o values be assumed for user- that no further error 

defined zenith angles are stored in a LUT. Possible LAI 

lens 

values 

system. 

lie between 0 and 10 (in steps of 0.01) 

and 

defined 

ALA 

zenith 

10° and 

angles 

80° 

are 

(in 

stored 

steps 

in 


a 

2°). 

LUT. 

The 

Possible 

LUT element 

LAI values 

closest 

lie between 

to the derived 

0 and 10 

gap 

(in 

fraction 

steps of 

is 

0.01) 

then


17 IN SITU LAI MEASUREMENTS 

LUT ( k ) 

MES 

with P0 ( i ) and P0 ( i ) being modelled and measured gap fractions for azimuthally averaged 

zenith sector i. wi is a weighting function 

w 

Npix 

Nmask 

i 

i 

i (5.7) 

Npix i 

where Npix i is the number of pixels in zenith sector i and Nmaski is the number of masked pixels in 

i. The function takes into account that some zenith directions may contain a large amount of masked 

pixels and is normalized by 

no. 

zenith angles 

w i 

i 1 

1. (5.8) 

The root mean square error between modelled and measured gap fractions over all user-defined zenith 

sectors (cf. Equation 5.6) is divided by a modelled standard deviation of the measured gap fraction values. 

MES 

Variation of the G-function (mean projection of unit foliage area) with the average leaf inclination ( P0angle( θ i 

(Weiss et al. 2004). 

This Figure modelled 5-16 standard deviation is derived by fitting a second order polynomial to )) , the 

empirical standard deviation of measured gap fraction in a certain zenithal direction resulting from all 

images processed together and belonging to the same transect. The second term of Equation 5.6 is a socalled 

Quantification regularization of system term that inherent puts constraints errors on the retrieved ALA values (WEISS 2006). 

The LUT gap fraction providing the minimum value of c k and its corresponding LAI and ALA values are 

Apart from the above-described deviation of the 

then considered as solution. This output will be further called LAIe (DHP). 

optical centre from the image centre, geometric 

lens Quantification distortions can of also system introduce inherent errors errors in the 

Apart 

calculation 

from 



LAI. 

above-described 

Figure 5-17 illustrates 

deviation 



linear 

the optical centre from the image centre, geometric lens 

distortions 

relation between 

can also 


introduce 

image radius 

errors 

r in 

in 

pixels 

the calculation 

and the 

of LAI. Figure 5-17 illustrates the linear relation 

between 

corresponding 

the image 

zenith 

radius 

angles 

r in 


pixels 

a hemispherical 

and the corresponding 

lens 

zenith angles for a hemispherical lens with 

equidistant 

with equidistant 

projection, 

projection, 

which 

which 

is 

is 

de 

90 

r 

. (5.9) 

(5.9) 

In order order to to test the projection quality for the camera 

system in use, an experimental design after BARET 

(2004b) system in was use, built an (cf. experimental Figure A-4). design Based after on this Baret the projection function was calculated (BARET 2004b). 

The (2004b) results was in built Figure (cf. Figure 5-18 show A-4). Based that the on this projection the functions correspond well with the theoretical 

equidistant projection projection. function was It can calculated thus be assumed (Baret 2004b). that no further error is introduced through the camera and 

lens The system. results in Figure 5-18 show that the projection 

functions correspond well with the theoretical 

equidistant projection. It can thus be assumed that 

no further error is introduced through the camera and Figure 5-17 Equidistant projection 

lens system. 

(Yamashita et al. 2004, modified). 

83

84 

Influence of operator and processing 

In order to assess the operator effect in the 

classification process, DHPs of all ESUs were 

processed twice by two different individuals. Figure 

5-19 shows the resulting LAI values per ESU 

e (DHP) 

plotted against each other. Mean LAI and 

e (DHP) 

standard deviation are very close to each other (see 

Table 5-6), with LAI calculated by operator 2 

e (DHP) 

being slightly higher (R2 =0.99). 

A more detailed analysis of outliers revealed that 

there were some misclassifications on the part of the 

less experienced operator 2. It is thus acknowledged 

that the operator has a certain influence on the 

resulting variable LAI . But in order to avoid the 

e (DHP) 

introduction of new errors, only LAI processed 

e (DHP) 

by the first operator was included in the study instead 

of taking the mean of the two results. 

angle (°) 

80 

70 

60 

50 

40 

30 

20 

10 

NIKON Coolpix 4300, INRA, 10/01/2006 

R² = 0.9987 

angle= 0.09581 * radius 

0 

0 100 200 300 400 

radius (pixels) 

500 600 700 800 

Figure 5-18 

Influence of illumination conditions 

In order to assess the influence of θ sun variation on 

LAI derived from DHP, a test was carried out in the 

centre of ESU 5 (late forest stage in Budongo Forest). 

Pictures were taken every 15 minutes from 7:45 a.m. 

to 6:00 p.m. The resulting LAI values are 

e (DHP) 

shown in Figure 5-20. LAI varied between 3.50 

e (DHP) 

(θ =55°/4.15 p.m.) and 4.35 (θ = 36°/10.30 a.m.), 

sun sun 

with a mean of 3.93 (±0.19). 

Table 5-6 Mean and standard deviation of LAIe 

(DHP) 

per ESU, processed by two different operators. 

Illustration of calculated projection function for NIKON coolpix 4300 and FC-E8 fisheye. 

μ σ 

LAI e (DHP) processed by operator 1 5.11 1.41 

LAI e (DHP) processed by operator 2 5.15 1.37


was taken per hour so that the result is less 



19 

representative and 

IN SITU LAI 

therefore not shown here. 

MEASUREMENTS 

Figure 

A more 

more 

detailed 

detailed 

analysis 

analysis 



outliers 

outliers 

revealed 

revealed 

that 

that 

there 


were 

were 

some 

some 

misclassifications 

misclassifications 

on 

on 



part 

part 





less 

less 

A more detailed analysis of outliers revealed that there were experienced some misclassifications Figure 5-20: LAIe operator 2. It on is thus the (DHP) part plotted 

acknowledged of the less against sun. Measu 

that 

Fig. 5-19: LAIe (DHP) derived from classification experienced operator 2. It is thus acknowledged that 

Fig. 5-19: LAIe (DHP) derived from classification 

of gap fraction by two different operators with experienced the operator operator 

Measurements 

has 2. a certain It is thus 

made 

influence acknowledged 

on 1 

the operator has a certain influence on 

on 


the that 

Fig. resulting 

resulting 

of 5-19: gap LAIe fraction (DHP) by derived two different from classification operators with 

of gap CAN-EYE. 

CAN-EYE. fraction by two different operators with the operator variable has LAIe a certain 

variable (DHP). influence But in order on the to resulting avoid the 

LAIe (DHP). But in order to avoid the 

CAN-EYE. 

variable Correction 

introduction LAIe (DHP). of 

of new But clumping 

errors, in order only LAIe to avoid (DHP) introduction of new errors, only 

processed the 

LAIe (DHP) processed 

by 

by 



first 

first 

operator 

operator 

was 

was 

included 

included 

in 

in 



study 

study 

instead 

instead introduction 


of In 

taking 

taking order of new 


the to 

mean 

mean correct errors, 


of only 


for the 

two 

foliage two LAIe 

results. 

results. (DHP) clumping, processed the logarithm gap fraction 

by the first operator was included in the study instead of taking (1986) the mean was of applied the two (cf. results. Chapter 3.3.3). Hence gap fraction cal 

azimuth sectors (as specified in the CAN-EYE parameter file) is t 

Influence 

Influence 



illumination 

illumination 

conditions 

conditions 

Influence of illumination conditions 

st Figu 

November, 2005 

Mea 

Correction of clumping 


R²=0.96 

In order to correct for foliage clumping, the logarithm 

In 

gap 

order 

fraction 

to correct 

averaging 

for foliage 

method 

clumping, 

by LANG 


& 

logarithm 

XIANG 

gap frac 

Y=-0.009+0.982 X 

(1986) was applied (cf. Chapter 3.3.3). Hence gap 

(1986) 

fraction 

was 

calculated 

applied 


(cf. 


Chapter 

predefined 

3.3.3). 

zenith 

Hence 

and 

gap fraction 

azimuth sectors (as specified in the CAN-EYE parameter 

azimuth 

file) 

sectors 

is taken 

(as 

into 

specified 

account. 

in 

The 

the CAN-EYE 

clumping factor 

parameter file) i 

Figure 5-19 LAI derived from classification of gap fraction Figure 5-20 LAI 

e (DHP) e (DHP) 

by two different operators with CAN-EYE. 1:1 line 

shown for comparison. 

plotted against θ . Measurements 

sun 

made on 1st o for each zenithal ring can then be derived according to WEISS et 

In order to assess the influence of sun In order 

o for 

to 

each 

assess 

zenithal 

the influence 

ring can 


then be variation on LAI derived from DHP, a test was carried out in the 

sun variation 

derived 

on 

according 

LAI derived 

to o for 

WEISS 

each 

from 

et 

zenithal 

DHP, 

al. (2006) 

ring 

a test 

as 

can then be derived according to WEISS 

sec 

was carried out in the 

centre of ESU 5 (late forest stage in Budongo Forest). Pictures ln( Po 

( were )) 

centre of ESU 5 (late forest stage in Budongo Forest). 

taken every 15 minutes from 

o Pictures sec 

sec 

were taken November, every 15 2005 minutes every 15 minutes. from 

7:45 a.m. to ln( 6:00 P ( p.m. )) 

ln( sec Po 

( )) 

o The resulting LAIe (DHP) values are shown in Figure 5-20. LAIe (DHP) 7:45 a.m. 

ln Po( ) 

to 6:00 p.m. The resulting 

varied between 

o 

o 

 

LAIe (DHP) values are shown in Figure sec 5-20. LAIe (DHP) varied between (5.10) 

sec 

3.50 (sun=55°/4.15 ln P( ) p.m.) 

ln P and 4.35 (sun 3.50 (sun=55°/4.15 p.m.) and 4.35 

= 36°/10.30 a.m.), with a mean o ( ) 

o 

of 3.93 (± 0.19). 

(sun = 36°/10.30 a.m.), with a mean of 3.93 (± 0.19). 

sec 

where ln( P o ( )) is the logarithm of gap fraction derived for e 

Pearson’s correlation coefficient reveals that there is is Pearson’s the 

Pearson’s 

logarithm correlation 

correlation sec 

of the coefficient 

coefficient 

previously reveals 

reveals 

averaged that 

that 

sec 

where ln( P gap 

o ( )) is the logarithm of gap fraction derived 

where ln( 


P oeach 

( )) 

sector 

is the 

on 

logarithm 

the DHPs, 


subsequently 

gap fraction derived fo 

Pearson’s 

no significant correlation between the two variables 

averaged 



is 

is correlation 

no 

no 

significant 

significant coefficient 

correlation 

correlation reveals 

between 

between that sec 

fraction over 

over 

azimuth 

azimuth. 

(comparable 

In contrary to 

to 

that, 

gap fraction 

ln 


the P o ( ) is the log 

sec 

averaged over azimuth. In contrary to that, ln there sec 

P ( averaged ) two is 

no variables significant at 0.05 correlation level (r = between -0.31). Pictures the 

o is the 

over 

logarithm 

azimuth. 

of the 

In contrary 

previously 

to 

averaged 

that, ln P gap o ( ) is the 

In order to assess the influence of sun variation on LAI derived from DHP, a test was carried out in the 

centre of ESU 5 (late forest stage in Budongo Forest). Pictures were taken every 15 minutes from 

7:45 a.m. to 6:00 p.m. The resulting LAIe (DHP) values are shown in Figure 5-20. LAIe (DHP) varied between 

3.50 (sun=55°/4.15 p.m.) and 4.35 (sun = 36°/10.30 a.m.), with a mean of 3.93 (± 0.19). 

at 0.05 level (r=-0.31). Pictures taken on 30 October, 

2005 in secondary forest close to the Sonso campsite 

show a comparable result. However, only one 

picture was taken per hour so that the result is less 

representative and therefore not shown here. 


Figure 5-20: LAIe (DHP) plotted against sun. 

Measurements made on 1 

In order to correct for foliage clumping, the logarithm 

gap fraction averaging method by Lang & Xiang 

(1986) was applied (cf. Chapter 3.3.3). Hence gap 

fraction calculated for the predefined zenith and 

azimuth sectors (as specified in the CAN-EYE 

parameter file) is taken into account. The clumping 

factor λ for each zenithal ring can then be derived 

0 

according to Weiss et al. (2006) as 

st taken on 30 October, 2005 in secondary 

forest close to the Sonso campsite show a 

comparable result. However, only one picture 







In order to correct for foliage clumping, the logarithm gap fraction averaging method by LANG & XIANG 

(1986) was applied (cf. Chapter 3.3.3). Hence gap fraction calculated for the predefined zenith and 

azimuth sectors (as specified in the CAN-EYE parameter file) is taken into account. The clumping factor 

o for each zenithal ring can then be derived according to WEISS et al. (2006) as 

sec 

ln( Po 

( )) 

o 

(5.10) 

(5.10) 

sec 

ln Po( ) 

sec 

where ln( P o ( )) 

where 

is the logarithm of gap fraction derived for each sector on the DHPs, subsequently 

sec 

averaged over azimuth. In contrary to that, ln P o ( ) is the logarithm of the previously averaged gap 

fraction over azimuth (comparable to gap fraction values given by LAI-2000 PCA). In the case that 

sec 

sec 

P o , 

= 0, i.e. only foliage occurring in the respective sector, Po , 

is assumed to be equal to 

sat 

P o , which stands for the gap fraction derived from a simple Poisson law using a prescribed LAI 

st two variables at 0.05 level (r = -0.31). Pictures 

taken on 30 October, 2005 in secondary 













sec 

ln( Po 

( )) 

o 

(5.10) 

sec 

ln Po( ) 

sec 

where ln( P o ( )) is the logarithm logarithm of gap fraction derived for each sector on the DHPs, subsequently 

derived for each sector on the DHPs, subsequently sec 

averaged over azimuth. In contrary to that, ln P o ( ) is the logarithm of the previously averaged gap 

averaged over azimuth. In contrary to that, 


sec 

sec 

P o , 



sat 

P o , which stands for the gap fraction derived from a simple Poisson law using a prescribed LAI 

st fraction values given over by azimuth LAI-2000 (comparable PCA). In to the gap case fraction that values given 

fraction over azimuth (comparable to gap fraction two sec fraction variables at 0.05 level (r = -0.31). Pictures 

P values , given 

over 

o = 0, i.e. by 

azimuth 

i.e. LAI-2000 

(comparable 

only foliage PCA). occurring In 

to 


gap 

in case 

fraction 

in the the respective that 

values giv 

secto 

taken sec 

sec 

on 30 October, 2005 in secondary 

P o , 

= 0, i.e. only foliage occurring in the respective P 

sat o , forest 

P o , 

close 

which 

sector, 

 

sector, = 0, sec 

to 

stands 

P 

i.e. 

the Sonso 

for only o , 


 

is 

foliage 

is 

campsite 

gap 

assumed 

occurring 

fraction 

to to be 

show 

derived 

be 

in 

equal 


a 

from 

to 

respective se 

a simple 

sat 

sat 

P o , which stands for the gap fraction derived 

to 

comparable 

from 

P o a 

, which which 

simple 

stands stands 

result. 

Poisson 

for However, 

law 

the only 

using 

gap fraction 

one 

a 

fraction 

picture 

prescribed 

derived 

LAI 

from a sim 

from a simple Poisson law using a prescribed LAI 


saturation value (LAI=10 for this study). After that 


the same LUT approach as described for LAIe (DHP) 

is applied to the modified Poisson model (Equation 

5.5). The output will be called LAI . (DHP) 


5.2.3 Comparison between LAI-2000 PCA 


and DHPs 


Analysis at gap fraction level is rarely accomplished, 


but necessary to understand differences between 


instruments (Leblanc et al. 2005). Accordingly this 


comparison was performed first, i.e. before logarithmic 

sec 

transformation of gap fraction and subsequent LAI 

ln( Po 

( )) 

o 

calculation. It should be noted that despite (5.10) an exact 

sec 

ln Po( ) 

demarcation of the respective sampling points along 

the transects, the two instruments were handled by 

sec 

where ln( P o ( )) is the logarithm of gap fraction derived for each sector on the DHPs, subsequently 

different operators, so that small displacement errors 

sec 

averaged over azimuth. In contrary to that, ln P o ( ) is the were logarithm still present. of the previously averaged gap 


sec 

sec 

P o , 



P , which stands for the gap fraction derived from a simple Poisson law using a prescribed LAI 

sat 

o 

two variables at 0.05 level (r = -0.31). Pictures 

taken on 30 October, 2005 in secondary 

85 



taken 

two 


take 

compa 

fore 

was 

com 

ta 

repres 

was 

repr

86 

As the two instruments have different fields of view 

(cf. Figure 5-21), gap fraction was extracted from 

DHP for those sectors that correspond to the field of 

view of the LAI-2000 PCA. Analogue to the latter, 

gap fraction was averaged over 45° azimuth and over 

those zenith angles that correspond to the different 

rings of the LAI-2000 PCA. As the variables differed 

significantly from normal distribution, the Spearman- 

Rho correlation coefficient r was calculated (Sachs 

s 

2004). As shown in Table 5-7, it revealed relatively 

strong relationships between transmission from 

LAI-2000 PCA data and gap fraction derived from 

DHP for the corresponding azimuth sectors (between 

0.63 and 0.75, all significant at p


sectors, LAI e values calculated from the above shown 

light variables were also compared. 

As the software code of CAN-EYE was not available, 

LAI could not be calculated for the field of 

e (DHP) 

view restricted to 45° in azimuth. Consequently 

a comparison for each sampling point would not 

be correct. Instead, mean LAI per ESU derived 

e 

from both methods was analysed with a Wilcoxon 

signed-rank test for paired data. On ESU level, 

the test indicated that LAI and LAI do 

e (LAI2000) e (DHP) 

not differ significantly from each other (Z=1.265, 

p=0.21, therefore the null hypothesis that 

LAI = LAI cannot be rejected). 


A more detailed analysis reveals that for lower 

values LAI is overestimated with respect 

e (DHP) 

to LAI . For higher values LAI is 


underestimated (cf. Figure 5-23, note that only 

LAI derived from upward-looking photographs is 

e (DHP) 

shown). This corresponds to the findings of Mussche 

et al. (2001) for a deciduous forest in Belgium where 

LAI and LAI were compared with LAI 

e (DHP) e (LAI2000) 

estimated from leaf litter fall. Here LAI was 

e (LAI2000) 

not significantly different from reference LAI for high 

Figure 5-22 Boxplots showing light variables ( P and τ ) for 

0 0 

the respective rings calculated from DHP and 

LAI2000. The box represents 50% of the values, 

whiskers include minimum and maximum values, 

solid line in box represents median. 

87 

LAI values. Lower LAI values were overestimated 

by the LAI-2000 PCA, probably because of the 

contribution of woody area to LAI . As low 

e (LAI2000) 

LAI values (i.e. < 3) in this thesis correspond to 

ESUs in very early forest stages without significant 

contribution of woody material, it is assumed that 

LAI is not overestimated in these cases. 

e (LAI2000) 

5.3 Results of in situ measurements 

The previously described preprocessing lead to 

three different LAI measures, namely LAI , 

e (LAI2000) 

LAI , and LAI . Whereas LAI excludes 

e (DHP) (DHP) e (LAI2000) 

understorey vegetation under 0.8 m and represents 

a LAI measure as used in other recent studies of 

tropical rain forests (De Wasseige et al. 2003, 

Kalácska et al. 2004, Aragao et al. 2005), LAIe (DHP) 

and LAI both include understorey vegetation. 

(DHP) 

This inclusion has been recommended by various 

authors (e.g. Wang et al. 2005, Yang et al. 2006) as 

the understorey can contribute significantly to canopy 

LAI. As a postprocessing step of DHP, a correction for 

clumping was applied to account for the non-random 

distribution of foliage elements, resulting in LAI . (DHP) 

R²=0.83 

Y=-0.224+1.079 X 

Figure 5-23 LAIe 

(LAI2000) plotted against LAI . 1:1 line 

e (DHP) 

shown for comparison.

88 

Results of the in situ measurements of all three LAI 

measures will be presented in the following. 

5.3.1 


Representativeness of sampling design 

In order to analyse the representation of different 

forest stages by the sampled ESUs, Budongo Forest 

area was classified based on NFA data on logging 

compartments (cf Figure 2-9). All areas without 

closed tree cover were assigned to the „early forest 

stage” class. The decision whether an area was 

assigned to the category intermediate (disturbed) 

or late (undisturbed), was mainly based on NFA 

information on zone classification and logging status. 

If compartments were classified as unlogged (such 

as the nature reserves) or selectively logged in 1945 

(as the buffer zones around the nature reserves) they 

were considered as late and mostly undisturbed forest 

stages. All other compartments were considered as 

intermediate and disturbed (cf. Table 5-12). 

Accordingly the three forest stages were assigned to 

the 30 ESUs, with one exception. Although ESU 28 

was situated in the nature reserve N15 which is under 

protection and thus classified as late (undisturbed), 

Figure 5-24 

ESU 28 showed clear signs of recent illegal logging 

activities (cf. Figure 5-24). Consequently it was 

assigned to the intermediate (disturbed) forest stage 

class. 

A comparison between NFA data on logging 

compartments and ESUs showed that early forest 

stages cover 21.36% of the total forest area in 

Budongo Forest and are represented by 26.67% of 

the ESUs (cf. Figure 5-25). 51.39% (ESU: 46.67%) 

comprise intermediate (disturbed) forest stages and 

27.25% (ESU: 26.67%) are undisturbed as they are put 

under protection in the nature reserves. The sampled 

ESUs are therefore regarded as representative for the 

distribution of forest stages of the whole study site. 

Spatial variability of in situ LAI 

At sample level, individual LAI e (LAI2000) ranged 

from 0.25 (on ESU 25) to 8.92 (on ESU 8), with a 

mean of 4.76 (±2.09) for a total number of 1,652 

individual measurements. The frequency distribution 

(cf. Figure 5-26) shows that there are local 

maxima between 2 to 3 and 6 to 7, indicating that 

LAI values of 4 to 5 are underrepresented. 

e (LAI2000) 

Figure 5-27 shows a comparison between an 

ESU located in an early forest stage and an ESU 

a) Thinned upper canopy and b) illegal logging activities on ESU 28 in the nature reserve N15.


representing an intermediate (disturbed) forest area. 

Whereas the first is characterized by a dense and high 

grass layer with some dispersed trees of max. 14 m 

height, the intermediate forest stage is a degraded 

primary forest site that had been selectively logged 

during the past 50 years. Succession stages in between 

these two ESUs were hardly present in Budongo 

Forest and if so, only occurred on forest edges, where 

ESUs with an adequate size could not be established. 

The frequency distribution of LAI and LAI e (DHP) (DHP) 

cannot be shown for individual measurements, as 

both are only calculated on ESU level (cf. Chapter 

Figure 5-25 

Percentage of forest stages a) of the whole study area of Budongo Forest and b) of sampled ESUs. 

Figure 5-26 Frequency distribution of LAI on sample 

e (LAI2000) 

level. 

Figure 5-27 a) ESU 21 (mean LAI of 3.1) and 

e (LAI2000) 

b) ESU 15 (mean LAI of 5.53). 

e (LAI2000) 

89 

5.2.2). Yet the frequency distribution on ESU level 

reflects the lack of intermediate LAI values for 

all three measures (cf. Figure 5-28). In situ LAI in 

Budongo Forest is thus not normally distributed. 

Whereas ESUs with LAI e (LAI2000) between 3.5 and 5 

are not present at all, LAI e (DHP) and LAI (DHP) show 

minima between 4.0 to 5.0 (LAI e (DHP) ) and 6.5 to 

8.0 (LAI (DHP) ). As expected – because this does not 

include understorey vegetation in intermediate and 

late forest stages – LAI is lowest with a mean 

e (LAI2000) 

of 5.29 (minimum of 1.56 on ESU 22 and maximum 

of 7.38 on ESU 11). LAI is slightly higher with 

e (DHP)

90 

a mean of 5.46 on ESU level (minimum of 2.53 on 

ESU 22 and maximum of 7.69 on ESU 7). This seems 

reasonable as understorey LAI ranged between 

e (DHP) 

0.18 and 0.68 with a mean of 0.36 (cf. Figure 5-29). 

LAI ranges between 5.19 and 10.47 (ESUs 22 

(DHP) 

and 6 respectively) with a mean of 8.38. In particular, 

the minimum value of 5.19 seems to be unreasonably 

high, as ESU 22 belongs to the early forest stages, i.e. 

covered mainly by grass species. 

At the forest stage level (cf. Figure 5-30), a Mann- 

Whitney U test revealed that mean LAI (all measures) 

is significantly different for early and intermediate, 

as well as early and late forest stages (p


Figure 5-29 DHP of understorey on a) ESU 4 (mean understorey LAIe 

(DHP) of 0.18) and b) ESU 28 (mean LAI of 0.53). 

e (DHP) 

Figure 5-30 

Mean LAI per forest stage in Budongo Forest. 

Upper and lower box boundaries indicate 

25th and 75th percentiles, whiskers indicate 

minimum and maximum values except outliers 

(defined as more than 3 box lengths from 25th or 75th percentile, displayed as point). Mean is 

shown as solid line. 

91

92 

30 Late (undisturbed) Logged in 1945 Protection (buffer) zone N1 6.06 5.43 8.10 

29 Late (undisturbed) Unlogged Nature reserve N15 6.13 6.02 8.88 

28 Intermediate (disturbed) Unlogged Nature reserve N15 5.39 5.68 9.82 


26 Early Logged in 1970-72 Low impact harvesting KP1 2.15 3.40 6.30 





21 Early n/a Low impact harvesting Grassland 3.16 3.69 6.70 

20 Late (undisturbed) Unlogged Recreation zone KP11 6.07 4.83 7.78 

19 Early n/a Protection (buffer) zone KP10 2.69 3.56 6.05 

18 Late (undisturbed) Unlogged Recreation zone KP11 6.00 5.36 8.16 

17 Early n/a Protection (buffer) zone KP9 3.12 3.74 7.25 

Forest stage, logging status, NFA zone, compartment number and LAI measures (LAI 

Table 5-8 e (LAI2000) , LAI e (DHP) , and 

LAI (DHP) ) of the 30 ESUs in Budongo Forest. 

16 Intermediate (disturbed) Logged in 1947-52 Site of scientific interference N3 5.96 6.08 8.47 



13 Intermediate (disturbed) Logged in 1943-44 Sawmill harvesting B5 6.32 6.69 9.98 

12 Intermediate (disturbed) Logged in 1943-44 Sawmill harvesting B5 6.08 6.35 9.13 




8 Intermediate (disturbed) Logged in 1945-47 Sawmill harvesting N2 6.06 7.07 10.14 

7 Late (undisturbed) Logged in 1945 Protection (buffer) zone N1 6.92 7.69 9.57 







ESU Forest stage Logging status NFA zone No. of compartment LAI e (LAI2000) LAI e (DHP) LAI (DHP)


5.3.2 


As previously explained, in situ measurements in 

Kakamega Forest were technically not as mature 

as the ones accomplished in Budongo Forest. 

Nevertheless they are presented for comparison. First 

of all, differences with respect to the Budongo field 

campaign will briefly be discussed. 

The field campaign in Kakamega Forest was planned 

in 2004 according to state of the art recommendations 

from other studies. First of all, two LAI-2000 PCA 

instruments are usually used for the derivation of LAI 

in forests, as explained in Chapter 5.1.3. The analysis 

of LAI-2000 PCA data sampled in Kakamega 

Forest, however, revealed that non-ideal illumination 

conditions could have a large impact on calculated LAI 

that might not be quantified or corrected without the 

help of a third device. Second, the distance between 

individual measurements was defined according to 

the theoretical field of view of the instruments (cf. 

Figure A-1 and Equations A.2 and A.3). Yet tests later 

showed that autocorrelation did not occur at smaller 

distances due to canopy density. As a higher density 

of sampling points results in better coverage of the 

sampling unit and therefore a higher precision in LAI, 

the distance between individual measurements was 

decreased to 10 m in mature and 5 m in early forest 

stages in Budongo Forest. Third, discussions within 

the VALERI network and the availability of software 

that made possible the analysis of downward-looking 

Table 5-9 

Issue Kakamega Forest Budongo Forest Reason for modification 

Number of used LAI-2000 

PCA devices 

Distance between 

individual measurements 

Differences in LAI sampling between the first and the second field campaign. 

2 3 Analyse illumination effects/ 

quantify measurement precision 

35/17.5 m 

(late/early forest stages) 

Figure 5-31 

10/5 m 

(late/early forest stages) 

Picture of ESU 19 in Kakamega Forest. 

93 

hemispherical pictures led to the acquisition of the 

latter in Budongo Forest. Here a Nikon Coolpix 4500 

with FC-E8 fisheye converter was used (cf. Table A-5 

for image parameters). Table 5-9 presents a summary 

of modifications. Results of the Kakamega field 

campaign are given in the following. 

Representativeness of sampling design 

In Kakamega Forest a detailed and spatially explicit 

logging history was not available for the classification 

of different forest stages. Instead a recent land cover 

classification generated within the BIOTA EAST 

AFRICA project by T. Lung was used (cf. Figure 

A-2, for methods refer to Lung [2004]). The land 

cover classes bushland/shrubs, secondary bushland/ 

Psidium guajava, grassland with scattered trees 

No autocorrelation at smaller distance/better 

coverage of sampling unit 

Understorey LAI Excluded Included Software for downward DHP not available 

until second field campaign

94 

and grassland were assigned to early forest stages, 

secondary forest as intermediate (disturbed) forest 

stage and near natural/old secondary forest was 

assigned to the late forest stage class. All other classes 

(plantations and water) were not considered. 

As mentioned before, different forest stages are not 

systematically distributed in logging compartments in 

Kakamega Forest. Forest stages are rather dispersed 

over the forest. Consequently ESUs were assigned to 

the forests stage covering the major part of the ESU 

with the exception of ESU 19. Though its area is 

mainly deemed to be grassland with scattered trees in 

the classification by Lung (2004) and would thus be 

affiliated to the early forest stage, it was reassigned 

to the intermediate (disturbed) forest stage. As Figure 

5-31 shows it is characterized by a closed canopy of 

up to 15 m high individuals of Psidium guajava and 

Bischoffia javanica, which justified this reassignment 

decision. 

A comparison between the proportions of the three 

forest stages on the whole test site and the forest 

stages represented by the sampled ESUs indicated 

that in Kakamega Forest late forest stages are well 

covered (cf. Figure 5-32). By contrast, early forest 

Figure 5-32 

stages were underrepresented (ESUs: 9%, whole test 

site: 36.88%), whereas intermediate forest stages 

were overrepresented with 45.03% of ESUs (but 

only 19.61% of whole test site). This might partly 

be attributed to misclassifications in the land cover 

classification provided by Lung (2004), but cannot be 

verified as an accuracy assessment is not available. 

Spatial variability of in situ LAI 

Due to the lack of a third LAI-2000 PCA device, no 

correction of e could be applied to LAI θsun e (LAI2000) 

in Kakamega Forest. In order to filter presumably 

bad values the method developed by Broich (2005) 

was used. It is based on a) the range between 

the maximum and minimum ring values of the 

LAI-2000 PCA at a certain time step and b) the 

slope between two time steps. Large ranges and 

slopes indicate quick changes in illumination 

through moving clouds and are thus excluded 

from analysis. Yet in contrast to Broich (2005), 

only measurements made in a north direction were 

included. As the impact of the filter is quite strong, 

no valid LAI values could be retrieved for 

e (LAI2000) 

ESUs 16 and 20. 

Percentage of forest stages for a) the whole study area of Kakamega Forest and b) sampled ESUs.


Figure 5-33 Frequency distribution of LAIe 

(LAI2000) 

on sample level in Kakamega Forest. 

After the elimination of presumably influenced 

LAI values individual measurements ranged 

e (LAI2000) , 

from 0.00 (ESU 6) to 9.24 (ESU 14), based on a 

total number of 367 measurements. The frequency 

distribution of LAI on sample level reflects 

e (LAI2000) 

again the underrepresentation of early forest stages 

(cf. Figure 5-34) and low to medium LAI values. As 

in Budongo Forest a local maximum occurs around 

LAI of 6. 

e (LAI2000) 

The frequency distribution on ESU level is displayed 

in Figure 5-35a) to c). Data gaps occur between 

LAI 1.5 and 3.5, LAI 3.5 and 4 and 


LAI indicating either that those values are actually 

(DHP) 

missing in Kakamega Forest or that LAI variability 

was not sufficiently sampled. Mean values on ESU 

level were lower than in Budongo Forest with mean 

LAI of 5.09, mean LAI of 3.93 and mean 


LAI of 6.30. Especially the last two measured 

(DHP) 

were remarkably lower that in Budongo Forest even 

though there were more ESUs sampled in early 

forest stages with low LAI. Once more, it has to be 

Figure 5-34 a) ESU 25 with mean LAI of 1.49 and 

e (LAI2000) 

b) ESU 27 with mean LAI 3.96. 

e (LAI2000) 

95 

emphasized that understorey vegetation was not 

taken into account in Kakamega Forest. 

A comparison on forest stage level also revealed 

a much higher variance of LAI measures than in 

Budongo Forest. A Mann-Whitney test indicated that 

only LAI differs significantly (p

96 

Frequency distribution of a) LAI 

Figure 5-35 e (LAI2000) , b) LAI e (DHP) and c) LAI (DHP) on ESU level and d) mean LAI on forest 

stage level in Kakamega Forest. For explanations of d) see Figure 5-29.


ESU Forest stage Land cover (Lung & Schaab 2004) LAI e (LAI2000) LAI e (DHP) (excl. understorey) LAI (DHP) (excl. understorey) 

1 Intermediate (disturbed) Secondary forest 3.95 3.54 5.64 

2 Late (undisturbed) Near natural and old secondary forest 5.51 3.75 6.22 






8 Early Grassland with scattered trees 1.49 2.80 6.82 








Forest stage, logging status, NFA zone, compartment number and LAI measures (LAI 

Table 5-10 e (LAI2000) , LAI e (DHP) , and LAI (DHP) ) of 

the 30 ESUs in Kakamega Forest. 

16 Late (undisturbed) Near natural and old secondary forest n/a 5.10 8.47 



19 Early Bushland/shrubs 4.31 2.82 4.65 

20 Late (undisturbed) Near natural and old secondary forest n/a 4.24 6.42 





25 Early Psidium guajava bushland 1.41 2.35 4.63 



28 Early Grassland with scattered trees 3.96 2.65 4.86 


97 

30 Late (undisturbed) Near natural and old secondary forest 5.79 4.23 6.04

98 

5.4 

Conclusion 

The preceding chapter presented the design of a valid 

sampling scheme for in situ LAI measurements in 

tropical rain forests, the analysis of field data and 

the development of correction methods as well as 

the results of these field measurements for the test 

sites Budongo and Kakamega Forest. Although 

sampling schemes recommended by CEOS-LPV 

and the VALERI network served as a basis, several 

modifications had to be made to account for problems 

associated with the tropical rain forest environment 


One of the main challenges in validating medium 

to coarse spatial resolution satellite data such as 

MODIS is to find a correct method to compare 

satellite data and in situ measurements. Whereas 

the approach employed by the CEOS-LPV and 

the VALERI networks – namely the usage of high 

resolution satellite data as an intermediate step in 

upscaling – is basically correct and applicable to 

the tropical rain forest environment, adjustments 

have to be made especially with respect to ESU 

size. Whereas in other studies ESUs equivalent to 

the size of a pixel of high resolution satellite data 

were set up, for several reasons the sampling units 

had to be increased to 200 x 200 m for this thesis. 

First of all, an accurate geolocation of the sites with 

a handheld GPS is critical. Although errors were 

minimized as far as possible, a location shift of up 

to 30 m is still theoretically possible. Second, the 

geometric correction of high resolution satellite 

data is associated with a certain error, as map sheets 

applied for geometric adjustments were more than 

35-40 years old and ground control points could 

not be set inside the forest area. Consequently 

the relationship between in situ LAI values and 

the corresponding surface reflectance data might 

still contain inaccuracies. This will be subject to 

discussion in the following chapter. The applied 

sampling pattern and sampling distance were 

based on thorough theoretical considerations and 

recommendations found in the relevant literature. 

Indirect optical instruments were used to 

derive in situ LAI. Here, for the first time three 

LAI-2000 PCA devices were employed. The 

stationary device on each ESU allowed the analysis 

and quantification of θ effects and the influence of 

sun 

changing illumination conditions (i.e. moving clouds) 

during the time of measurement. Results suggest that 

the influence of direct radiation (which is increasing 

with decreasing θ ) must not be disregarded and 

sun 

is possibly even higher in lower canopies with less 

foliage. Yet the development of a correction scheme 

led to LAI-2000 PCA results that are unbiased by 

θ . The influence of moving clouds on the data 

sun 

was minimized by a filter application. The proposed 

methodology can also be transferred to other 

vegetation types and gives a detailed insight to field 

data quality. It far exceeds the sun zenith angle 

correction scheme developed by Leblanc 

& Chen (2001), which is based on a single 

device and mean plot values only. Further it 

contradicts the recommendations made by 

De Wasseige et al. (2003), who found θ of less 

sun 

than 50° to be best suited for LAI-2000 PCA 

measurements in tropical rain forests. This might 

occur because their analysis was based on only two 

devices and three groups of data taken on the same 

transects at different times of the day. The results of 

De Wasseige et al. (2003) suggest that measurements 

taken on the same transect at θ < 50° differ least, 

sun 

which is also true for the data presented in this chapter. 

Yet if measurements are taken at low sun zenith 

angles, e is also higher (cf. Chapter 5.2.1), resulting 

θsun 

in a necessary correction for e . As shown in this 

θsun 

chapter, this correction can be achieved through the 

application of three devices. This, however, is very 

expensive and also requires accurate intercalibration. 

Further, the selection of the measurement location for 

the reference sensor is crucial, yet often hindered in 

tropical rain forests as clearings or natural gaps are not


large enough. Consequently a large distance between 

reference sensor and ESU had to be accepted as a 

trade-off, leading to a stronger influence of changing 

illumination conditions on the results and subsequent 

reduction of valid readings by filtering. 

Digital hemispherical photographs were also analysed 

for understorey LAI. Yet care has to be taken with 

respect to general interpretation, as measurement 

heights differed with forest stages and understorey 

LAI results are thus not comparable between ESUs. 

If only intermediate and late forest stages are 

considered, understorey contributed up to 14% to 

total LAI and should thus not be disregarded. 

e (DHP) 

Though the correction for non-random foliage 

distribution applied to LAI is an attempt to 

e (DHP) 

derive true LAI values that are corrected for foliage 

clumping, it must be viewed as a critical step. 

Especially LAI values of the early forest stages with 

values of up to 7.25 (plot 17 in Budongo Forest) 

indicate an overestimation of λ, as LAI seems to be 

much too high for wooded grassland. The clumping 

factors are derived from DHP processing, and lie 

around 0.50 on average for ESUs of early forest 

stages. 

As a summary, Table 5-11 lists the different 

properties of both devices. With respect to tropical 

rain forests, DHP does not only offer a cheaper 

solution, but – provided that the results are accurate 

– also the possibility of correction for clumping. 

Table 5-11 

Comparison between the properties of LAI-2000 PCA and DHP. 

System Illumination 

conditions 

LAI-2000 PCA Preferably diffuse 0-74° 

in five rings 

Digital camera with 

fisheye lens 

Zenith coverage Azimuth coverage Computer resources Cost in € 

Up to 360° 

(selectable by view caps) 

99 

DHP is also less sensitive to illumination conditions 

(direct or changing) and has a similar zenith and 

azimuth coverage as LAI-2000 PCA (theoretically 

it has a higher zenithal coverage, however zenith 

angles between 60 and 90° are usually not taken 

into consideration). Although DHP demands greater 

computer resources, there is compensation for this in 

the form of a permanent information archive of each 

study site. 

Intercomparison at the ESU level reveals that there 

are differences in the results of LAI-2000 and DHP. 

As LAI includes understorey, it should be higher 

e (DHP) 

than LAI for all ESUs. This is, however, not 

e (LAI2000) 

always the case (11 out of 30 ESUs in Budongo 

Forest, cf. Table 5-8). Classification problems 

associated with mixed pixels might be the reason for 

this, especially when DHP are acquired under high 

canopies. The distance to the highest foliage elements 

does not allow single leaves to be represented by 

one or more pixels. Pixels rather represent several 

foliage units so that misclassifications can cause 

inaccuracies in the retrieved LAI. This is underlined 

by the fact that in Budongo mean LAI and 

e (LAI2000) 

mean LAI differ for the early forest stage, but 

e (DHP) 

do not differ for the intermediate and late forest 

stages. In Kakamega Forest, where hemispherical 

pictures of the understorey were not taken, LAIe (DHP) 

is much lower than LAI for the last two stages, 

e (LAI2000) 

indicating an underestimation of LAI through DHP 

for this test site. 

Low Ca. 8.900,- 

Diffuse/direct 0-90° 360° High Ca. 800,-

100 

In order to carry out an absolute calibration of the 

applied indirect methods, supplementary direct 

or semi-direct methods (assumed to give “true” 

LAI) would be necessary. Due to the complexity of 

vegetation stands, the high species numbers and the 

lack of seasonal leaf litter fall in tropical rain forests, 

this would have exceeded the possible fieldwork 

within the scope of this thesis. Only one study 

published to date encompasses direct landscapescale 

measurements for a tropical rain forest LAI 

(Clark et al. 2008). As part of a larger project, in this 

study a modular tower was used to harvest leaves 

and branches in 55 vertical transects in old-growth 

tropical rain forest around La Selva Biological 

Station in Costa Rica. A mean landscape LAI of 

6.00 was retrieved (with individual measurements 

ranging from 1.20 to 12.94), which corresponds 

approximately to mean LAI and LAI e (LAI2000) e (DHP) 

values retrieved for this thesis from Budongo Forest. 

Further comparison to other studies using indirect 

optical measurements in tropical rain forests 

revealed that mostly higher values for effective LAI 

were retrieved for Budongo (cf. Table 3-2). Results 

reported by Aragão et al. (2005) for an evergreen 

rain forest in Brazil, De Wasseige et al. (2003) for a 

semi-deciduous forest in the Car and Laumonier et 

al. (1994) for an evergreen rain forest in Cameroon 

are slightly lower on average for late forest stages, 

yet in all three sites uncorrected LAI-2000 PCA 

measurements were used. In contrast, the LAI 

retrieved for the VALERI site in Counami (evergreen 

rain forest in French Guiana) and published by 

Rossello et al. (2007) seems to be unreasonably low 

with a mean LAI of 3.38 and a mean LAI of 

e (DHP) (DHP) 

4.92. As in the above-mentioned studies, understorey 

vegetation was not considered. Interestingly, other 

studies found a significant difference between LAI 

of late and intermediate forest stages (primary and 

secondary forests respectively), as shown in Table 

3-4. Probably the inclusion of early forest stages 

explains the comparably low LAI values of the latter. 

Summarizing, it remains hard – if not impossible 

– to judge which of the two methods applied leads 

to better results. In any case, LAI seems to 

(DHP) 

give unreasonably high results. With respect to 

LAI and LAI measurement precision is 


an important factor for further upscaling that should 

not be disregarded and will be discussed thoroughly 

in the next chapter.

6 Derivation 

of high 

resolution LAI maps 

The following chapter describes the up scaling 

of in situ LAI measurements with the help of high 

resolution satellite data introduced in Chapter 4. 

Following the recommendations of Baret el al. 

(submitted), empirical relationships between field 

data and SVIs or texture measures are established. To 

provide a solid validation basis, special focus is here 

put on the correct establishment of transfer functions 

as well as the consideration of measurement errors in 

ground and satellite data (cf. Chapter 6.2). For this 

purpose a robust regression method, namely Theil- 

Sen regression, is introduced and compared to an 

approach that accounts for measurement precision in 

input data (Tan et al. 2005). The analyses of in situ 

and satellite data revealed that different regression 

models had to be applied to early/intermediate and 

late forest stages. The resulting LAI maps for both 

study sites represent independent data sets needed for 

the validation of the MODIS LAI product (cf. Chapter 

7). At least for Budongo Forest the upscaling could 

be performed without any degradation in accuracy 

compared to field measurements. Both maps are 

presented together with their accuracies at the end of 

this chapter. 

(submitted), empirical relationships between field data and SVIs o 

provide a solid validation basis, special focus is here put on the cor 

101 

as well as the consideration of measurement errors in ground and 

purpose a robust regression method, namely Theil-Sen regressio 

approach that accounts for measurement precision in input data ( 

and 

6.1 

satellite 

Calculation 

data revealed 


that 

spectral 

different 

vegetation 

regression models had to b 

forest stages. 

indices 

The resulting 

and texture 

LAI maps 

measures 

for both study sites represen 

validation of the MODIS LAI product (cf. Chapter 7). At least for 

performed 

For both Kakamega 

without 

and 

any 

Budongo 

degradation 

Forest 

in 


accuracy 

available 

compared to 

presented 

pre-processed 

together 

SPOT-4 

with 

and 

their 

ASTER 

accuracies 

data 

at 

were 


used 

end of 

to 

this chapter. 

calculate the SVIs listed in Table 3-5. SWIR and min 

SWIR values as needed in NDVI and RSR were 

max c 

derived directly from the forest area of the respective 

For 

test 

both 

sites (for 

Kakamega 

ASTER 

and 

band 

Budongo 

4 was used). 

Forest the available pre-process 

to calculate the SVIs listed in Table 3-5. SWIRmin and SWIRmax val 

derived 

With respect 

directly 

to 

from 

image 


texture, 


second-order 

area of the respective 

statistics 

test sites (for 

With were calculated respect to based image on texture, a GLCM second-order according to statistics Table were calcu 

Table 6-1 with 6-1 with 

6.1 Calculation of spectral vegetation indices an 

P 

V 

 

i , j 

i , j 

N 1 

Vi 

, j 

i , j 0 

(6.1) 

being a normalization equation within the respective 

kernel, that 

column kernel, that j (starting divides with the 0,0) pixel by value the sum V at of row values. i and According to 

(2007) column a j horizontal (starting with offset 0,0) distance by the of sum one pixel of values. was used and diffe 

7x7, According 9x9 and to 11x11 the recommendations pixels were compared. by Hall-Beyer 

(2007) a horizontal offset distance of one pixel was 

used and different moving window sizes of 3x3, 5x5, 

7x7, 9x9 and 11x11 pixels were compared. 

For all ESUs, mean SVIs and texture measures were 

extracted as well as their standard deviation. All 

pixels whose centre lay within the 200 m x 200 m 

area of the respective ESU were taken into account. 

Additionally mean and standard deviation per ESU 

of all reflective bands were stored. 

6.2 

Establishment of transfer 

functions 

In situ data, reflectance, SVIs and texture measures 

served as input for regression analysis. The goal 

is the establishment of an empirical relationship 

that can be used to invert LAI for the whole test 

site from reflectance data (or SVIs and texture 

respectively). However, both field LAI values and

accuracy observations of In situ the data, resulting reflectance, include high measurement resolution SVIs and LAI texture errors observations maps. measures that product Knowledge influence served include is crucial about the as measurem input for derived the a mfo 

q 

product accuracy is establishment crucial of for the a meaningful of resulting an empirical high and accurate resolution relationship accuracy validation LAI information that maps. of can of the the be Knowledge resulting on used MODIS model to high inve LAI abo an 

102 

information product reflectance on model is crucial data and for (or data a SVIs meaningful uncertainties and texture and into product accurate respectively). account. errors is validation crucial and As their a for However, result, of a impact the meaning MO quan bo 

errors and information observations their impact on include on model the measurement and establishment data uncertainties errors of information an errors empirical that into influence in on account. both transfer model data the As and functio sets, derive a res dat fi 

errors in errors both accuracy data and of sets, their the field impact resulting LAI on and high the remote establishment resolution sensing errors LAI to data, and of that, maps. an their have empirical regression Knowledge impact to be assessed transfer on analys the ab 

satellite observations include measurement errors to that that, errors regression product 6.2.1 in both is Quantification analysis crucial data sets, for is a performed. field meaningful of LAI measurement and At and remote errors this accurate point applicable in sensing both validation differing data data, in the sets, have approache of event field the to tha MO be LA 

influence the derived relationships and thus lower applicable the to information in that, the event regression errors on that model input analysis data and is is data performed. not uncertainties error-free. to that, At this regression into point account. differing analysis As a is re ap 

accuracy of the resulting high resolution LAI maps. applicable errors and in their the event impact that on input the data establishment is applicable not error-free. 6.2.1 of in an the Quantification 

empirical event that transfe inpu 

Knowledge about the quality of this intermediate 6.2.1 Quantification errors To pursue in both the data quantification of sets, measurement field of LAI measurement and remote errors To errors, sensing pursue data, the quantifica have to b 

product is crucial for a meaningful and accurate To pursue 6.2.1 to the the that, terms quantification Quantification regression error, uncertainty, analysis of measurement of is measurement performed. bias, 6.2.1 precision errors, accuracy At Quantification the errors this and terms will point be error, differing defined uncer of fi 

validation of the MODIS LAI product, which accuracy also To will applicable accuracy pursue be defined will the in the be quantification first event defined as they that first are input of as frequently measurement data they To is are not misused. pursue frequently published error-free. errors, the For quantification by the this FOODY terms thesis the erro & o 

takes information on model and data uncertainties published into accuracy misused. by FOODY will For be & defined this ATKINSON thesis first as (2002) the they widely are will accuracy frequently be accepted 

xfollowed. 0 may 

will misused. be 

be 

defined 

defined Accordingly, For first 

as 

this 


as th 

d 

account. As a result, quantification of measurement definitions published by Foody & Atkinson (2002) 

x 0 may be published 6.2.1 Quantification 

defined as by the FOODY difference & ATKINSON of measurement 

between the (2002) published true value will by errors be ( FOODY followed. 

0xz ) and the & Accor ATKI meas 

errors and their impact on the establishment of an will be followed. Accordingly, an error ˆ)( 0 at 0 x. 0 ) 

x 

To 

0 may 

pursue 

be defined 

the quantification 

as the difference 

of measurement 

between x 0 may the be true 

errors, 

defined value 


as ( the 

terms 

0xz ) differe and 

er 

empirical transfer function is needed. Consequently ˆ)( 0 accuracy location 0 . xzxzx will 0 )()( 

may be defined be defined first as as the they difference are frequently between e misused. For this t 

An error is thus data base 

errors in both data sets, field LAI and remote sensing the true ˆ)( 0 value 0 . xzxzx 0 )()( 

and the measured value ˆ)( 0 

0 , . xz 

e 

0 )() 

An error is 

published 

thus data 

by FOODY 

based and refers 

& ATKINSON 

to a single 

(2002) 

measurement. error 

will 

is 

be 

well 

followed. 

In defined. this case, 

Acc 

Uncn 

data, have to be assessed as a first step. Subsequent so that 

error is well An x 0error 

may defined. be is thus defined Uncertainty data as based the in difference and turn refers is related between An to a error to single from the prediction is true regression measurement. thus value data based analysis, based ( on 0xz In ) stat and 

thi a 

to that, regression analysis is performed. At this point 

from regression error is analysis, well defined. and is Uncertainty associated to in three turn error is other related is well terms, to defined. prediction namely Uncertain bias, based pre 

differing approaches are introduced that are applicable ˆ)( 0 

0 . xzxzx 0 )()( (6.2) 

e 

from regression analysis, and is associated from to three regression other analysis, terms, namely and is 

in the event that input data is not error-free. 

An error is thus data based and refers to a single measurement. In t 

error is well defined. Uncertainty in turn is related to prediction bas 

from regression analysis, and is associated to three other terms, namel 

Table 6-1 

Second-order texture measures used in this thesis (definition according to Hall-Beyer 2007). 

Texture measure Equation 

Mean (μ i ) 

2 Variance (σ ) i 

Homogeneity 

Contrast 

Entropy 

Correlation 

2 DERIVATION OF HIGH RESOLUTION LAI M 

Table 6-1: Second-order texture measures used in this thesis (definition according to HALL-BEYER 2007) 


Mean ( i ) 

1 

, 

0, 

)( 

N 

Pi ji 

ji 

2 

Variance ( i ) 

1 

2 

, 

0, 

)( 

N 

iP ij i 

ji 

Homogeneity 

N 1 P , ji 

2 

ji 0, 

ji )(1 

Contrast 

1 

2 

, 

0, 

)( 

jiP 

N 

ji 

ji 

 

 

Entropy 

N 1 

, ji , ji )l n( 

ji 0, 

PP 

Correlation 

 

 

 

 

 

 

1 

 

 

 

0, 

 

 

22 




Mean ( i ) 

 

N 

ii 

 

ji 

P , ji 

ji ji 

For all ESUs, mean SVIs and texture measures were extracted as well as their standard deviation. 

pixels whose centre lay within the 200 m x 200 m area of the respective ESU were taken into accou 

Additionally mean and standard deviation per ESU of all reflective bands were stored. 

6.2 Establishment of transfer functions 

In situ data, reflectance, SVIs and texture measures served as input for regression analysis. The goal is 

establishment of an empirical relationship that can be used to invert LAI for the whole test site fr 

reflectance data (or SVIs and texture respectively). However, both field LAI values and sate 

observations include measurement errors that influence the derived relationships and thus lower 

accuracy of the resulting high resolution LAI maps. Knowledge about the quality of this intermed 

1 

, 

0, 

)( 

N 

Pi ji 

ji 

2 


1 

2 

, 

0, 

)( 

N 

iP ij i 

ji 

Homogeneity 

N 1 P , ji 

2 

ji 0, 

ji )(1 

Contrast 

1 

2 

, 

0, 

)( 

jiP 

N 

ji 

ji 

 

 

Entropy 

N 1 

, ji , ji )l n( 

ji 0, 

PP 

Correlation 

 

 

 

 

 

 

1 

 

 

 

0, 

 

 

22 




Mean ( i ) 

 

N 

ii 

 

ji 

P , ji 

ji ji 


pixels whose centre lay within the 200 m x 200 m area of the respective ESU were taken into acco 




establishment of an empirical relationship that can be used to invert LAI for the whole test site f 


observations include measurement errors that influence the derived relationships and thus lower 

1 

, 

0, 

)( 

N 

Pi ji 

ji 

2 


1 

2 

, 

0, 

)( 

N 

iP ij i 

ji 

Homogeneity 

N 1 P , ji 

2 

ji 0, 

ji )(1 

Contrast 

1 

2 

, 

0, 

)( 

jiP 

N 

ji 

ji 

 

 

Entropy 

N 1 

, ji , ji )l n( 

ji 0, 

PP 

Correlation 

 

 

 

 

 

 

1 

 

 

 

0, 

 

 

22 




Mean ( i ) 

 

N 

ii 

 

ji 

P , ji 

ji ji 








1 

, 

0, 

)( 

N 

Pi ji 

ji 

2 


1 

2 

, 

0, 

)( 

N 

iP ij i 

ji 

Homogeneity 

N 1 P , ji 

2 

ji 0, 

ji )(1 

Contrast 

1 

2 

, 

0, 

)( 

jiP 

N 

ji 

ji 

 

 

Entropy 

N 1 

, ji , ji )l n( 

ji 0, 

PP 

Correlation 

 

 

 

 

 

 

1 

 

 

 

0, 

 

 

22 




Mean ( i ) 

 

N 

ii 

 

ji 

P , ji 

ji ji 







1 

, 

0, 

)( 

N 

Pi ji 

ji 

2 


1 

2 

, 

0, 

)( 

N 

iP ij i 

ji 

Homogeneity 

N 1 P , ji 

2 

ji 0, 

ji )(1 

Contrast 

1 

2 

, 

0, 

)( 

jiP 

N 

ji 

ji 

 

 

Entropy 

N 1 

, ji , ji )l n( 

ji 0, 

PP 

Correlation 

 

 

 

 

 

 

1 

 

 

 

0, 

 

 

22 

2 DERIVATION OF HIGH RESOLUTION LAI MA 



Mean ( i ) 

 

N 

ii 

 

ji 

P , ji 

ji ji 






1 

, 

0, 

)( 

N 

Pi ji 

ji 

2 


1 

2 

, 

0, 

)( 

N 

iP ij i 

ji 

Homogeneity 

N 1 P , ji 

2 

ji 0, 

ji )(1 

Contrast 

1 

2 

, 

0, 

)( 

jiP 

N 

ji 

ji 

 

 

Entropy 

N 1 

, ji , ji )l n( 

ji 0, 

PP 

Correlation 

 

 

 

 

 

 

1 

 

 

 

0, 

 

 

22 

N 

ii 

 

ji 

P , ji 

ji ji 

For all ESUs, mean SVIs and texture measures were extracted as well as their standard deviation. A 

pixels whose centre lay within the 200 m x 200 m area of the respective ESU were taken into accoun 



In situ data, reflectance, SVIs and texture measures served as input for regression analysis. The goal is th

6 Derivation of high resolution LAI maps 

An error is thus data based and refers to a single 

measurement. In this case, no uncertainty exists as 

the error is well defined. Uncertainty in turn is related 

to prediction based on statistical models, e.g. derived 

from regression analysis, and is associated to three 

other terms, namely bias, precision and accuracy. 

Bias is a measure of agreement between a data set bias) and precision, i.e. accuracy = unbias + precision 

and values predicted by a statistical model. It is (cf. Figure 6-1 for illustration). As displayed in 

DERIVATION OF HIGH RESOLUTION LAI MAPS 3 

an expectation of over- or under-prediction, thus Figure 6-1b, high accuracy can only be achieved if 

Bias revealing, is a measure for instance, of agreement systematic between errors a data in set the and both values precision predicted and by unbias a statistical are high. model. It is an 

expectation measurement of process. over- It or is often under-prediction, characterized by thus the 

DERIVATION OF OF HIGH HIGH RESOLUTION LAI LAI MAPS MAPS revealing, for instance, systematic errors in 3 the 3 

measurement mean error ME, process. with It is often characterized by the mean If error an independent ME, with data set is used to assess accuracy 

Figure 6-1: Illustrations of data sets with a) large bias and high preci 

Bias Bias is a is measure a measure of of agreement between a data a data set set and and high values (as values precision in predicted the predicted case (high by with by a accuracy), statistical a in statistical situ measurements c) model. low model. bias It is It and an is used an low an for precision (low 

n 1 

ME zˆ( x i ) z( 

x i ) . 

(6.3) 

precision 

expectation of of over- over- or or under-prediction, thus thus revealing, validation 

(low 

for for instance, of 

accuracy). 

instance, the MODIS systematic product), errors errors then in in (6.3) the accuracy the 

n i 1 

measurement process. It is It often is often characterized by by the by the mean mean error If may error an error ME, be independent ME, predicted with with 

data directly set is from used the to RMSE, assess accuracy which (as in the c 

Yet ME does not take into account the the spread of of errors. Instead this is considered by precision, which 

n n 

validation is basically of the the same MODIS as in product), Equation then 4.1 (RMSE accuracy of may be predic 

1 1 

MEalso 

ME errors. ME depends 

Instead 

zˆ( xz z ˆ 

( 

on x 

this a statistical is considered 

model fitted by precision, to a certain data set. Precision is usually predicted using a 

i ) i 

i) 

z) 

( 

xz 

z i( 

( ) x 

i. 

i) 

) . 

. basically GCPs). Yet the variables same as in are Equation different, 4.1 so (RMSE that (6.3) in (6.3) 

this of GCPs). case Yet variab 

n i n 

n 1 ii 

1 

measure which also of error depends spread on a around statistical the mean model error, fitted e.g. to a the standard written it is written as deviation as of the error e with 

Yet Yet certain ME ME does data does not set. not take Precision take into into account is account usually the the predicted the spread spread of using of errors. errors. Instead this this is considered is considered by by precision, which which 

n 

n 

also also a depends measure depends 2 

eof on 

i error 

on eˆ 

a 

i statistical 

a spread statistical around model model the fitted mean fitted to error, to a certain a e.g. certain data data set. set. Precision is usually is 2usually 

usually predicted using using a a 

zizˆi measure i1 

the e standard 

of of error of error deviation spread spread , around of around the error the the mean σmean mean with error, error, e.g. e.g. the the standard RMSE deviation i 1 

of of the of the error error . error (6.4) 

e e with e 

with 

(6.5) 

n 1 

n 

n 

n 

 

i1 

e 

, 

n 1 

2 2 

where e e i is e 

e iethe 

ˆ i 

e 

ˆ 

mean error and êi is the predicted error. With 

i i 

With If the respect errors to themselves measurement are not errors known, several other issues have to be tak 

i 

i1 

i 

 

1 

emeasures 

e 

 

of precision , , can be used, such as, for (6.4) instance, important have the to standard be to taken acknowledge deviation into consideration. that of the both measurements in First (6.4) situ (6.4) 

and of all, satellite it reflectance 

n n 

n 1 

1 

(TAN et al. 2005) or c v (ARAGÃO et al. 2005, cf. Equation error is important 5.1). sources Note to in that acknowledge field precision LAI estimation is that an expectation both were in situ discussed of and in Chapter 

where where where ei is e iithe 

is the mean error and and êi is ê iithe 

the is the predicted error. 

error. If influenced satellite the If the errors errors reflectance by themselves erroneous themselves data are sensor are are not not influenced calibration, known, other by georegistration other 

errors. and atm 

the spread of errors, with a large number corresponding to small errors and vice-versa, e.g. the higher e , 

measures If the errors of of precision themselves precision can can are can be not be used, known, used, such such other as, as, measures for for instance, Possible the the standard error deviation sources deviation in of field of the the measurements 

LAI measurements 

estimation were 

the lower the precision! 

(TAN (TAN of et precision al. et al. 2005) al. 2005) can or be or c used, 

v c (ARAGÃO v 

(ARAGÃO such et as, al. et for al. 2005, al. instance, 2005, cf. cf. Equation the Equation discussed 5.1). 5.1). Note Note in that Chapter that precision 3.3.3. is an is Remote is an expectation sensing of data of 

is 

The standard last term, deviation accuracy of the (sometimes measurements also referred (Tan et to al. as uncertainty), mainly influenced is defined by as erroneous unbias (i.e. sensor the opposite calibration, 

the the spread spread of of errors, of errors, with with a large a large number corresponding to small to small errors errors and and vice-versa, e.g. e.g. the the higher higher e , e 

, 

of 2005) bias) or and c precision, (Aragão et i.e. al. accuracy 2005, cf. = Equation unbias + 5.1). precision georegistration (cf. Figure 6-1 and for atmospheric illustration). correction. As displayed in 

v 

the the lower lower the the precision! 

Figure 6-1b, high accuracy can only be achieved if both precision and unbias are high. 

The The last last term, term, accuracy (sometimes also also referred to to as to as uncertainty), as uncertainty), is defined is defined as as unbias as unbias (i.e. (i.e. the the opposite 

of of bias) of bias) and and precision, i.e. i.e. accuracy = unbias = unbias + precision + precision (cf. (cf. Figure Figure 6-1 6-1 for for illustration). As As displayed in in 

Figure Figure 6-1b, 6-1b, high high accuracy can can only only be be achieved if both if both precision and and unbias unbias are are high. high. 

Figure 6-1: Illustrations of data sets with a) large bias and high precision (low accuracy), b) low bias and 

high precision (high accuracy), c) low bias and low precision (low accuracy) and d) large bias and low 

precision (low accuracy). 

Figure 6-1 

Figure Figure If 6-1: an 6-1: Illustrations independent Illustrations (high accuracy), of data data of data set c) sets low is sets used with bias with and a) to large a) low assess large precision bias accuracy bias and (low and high accuracy) (as high in precision the precision and case d) large (low with (low bias accuracy), in accuracy), and situ low measurements precision b) b) low b) low bias (low bias accuracy). 

and used and 


high high precision validation precision (high of (high the accuracy), MODIS accuracy), c) product), c) low c) low bias bias then and and accuracy low low precision may be (low predicted (low accuracy) directly and and d) from d) large large the bias RMSE, bias and and low which low 

is 

precision (low (low accuracy). 

basically the same as in Equation 4.1 (RMSE of GCPs). Yet variables are different, so that in this case it is 

If an If written If an independent as data data set set is used is used to to assess to assess accuracy (as (as in in the in the case case with with in situ in situ measurements used used for for 

validation of of the the MODIS product), then then accuracy may may be be predicted directly from from the the RMSE, which which is is 

n 

2 

basically the the same same same zas i in 

as zˆ 

Equation in i Equation 4.1 4.1 (RMSE of of GCPs). of GCPs). Yet Yet variables are are different, so so that so that in this in this case case it is it is 

i 1 

written RMSE as as 

 

. (6.5) 

103 

where ei is the mean error and êi is the predicted error. If the e 

measures of precision can be used, such as, for instance, the st 

(TAN et al. 2005) or c v (ARAGÃO et al. 2005, cf. Equation 5.1). N 

Note that precision is an expectation of the spread of 

the errors, spread with of a errors, large with number a large corresponding number corresponding to small to small e 

the errors lower and the vice-versa, precision! e.g. the higher σ , the lower the 

e 

The 

precision! 

last term, accuracy (sometimes also referred to as uncertainty 

of bias) and precision, i.e. accuracy = unbias + precision (cf. Figu 

Figure 

The last 

6-1b, 

term, 

high 

accuracy 

accuracy 

(sometimes 

can only 

also 

be achieved 

referred 

if 

to 

both 

as 

precision a 

uncertainty), is defined as unbias (i.e. the opposite of 

Illustrations of data sets with a) large bias and high precision (low accuracy), b) low bias and high precision

104 

When finally both in situ and satellite data are 

combined, a third error source has to be taken into 

account: misregistration of both measurements. 

Spatial misregistration includes spatial location 

errors, but also errors introduced through the sensor’s 

point spread function, velocity smear and atmospheric 

scatter. Yet as pixels in a remotely sensed scene 

are spatially autocorrelated, this becomes a greater 

problem when landscapes are highly diverse in 

relation to the sensor’s spatial resolution and can be 

disregarded for the study sites in this thesis. Temporal 

misregistration occurs when satellite reflectance 

data is taken as representative for in situ conditions 

that do not match those of individual measurements 

(Curran & Hay 1986). Consequently, as in situ data 

were acquired over a period of four weeks at both 

Descriptive statistics of LAI 

Table 6-2 e (LAI2000) per ESU derived from B1 measurements after correction of e θsun and applied 

cloud filter. 

ESU N Minimum Maximum μ σ c v 

1 140 6.03 6.48 6.32 0.119 0.019 

2 164 8.09 8.55 8.34 0.090 0.011 

3 92 7.38 7.78 7.58 0.098 0.013 

4 80 7.09 7.86 7.48 0.225 0.030 

5 185 7.88 8.26 8.10 0.099 0.012 

6 99 7.66 8.21 7.89 0.146 0.019 

7 143 4.31 4.77 4.60 0.088 0.019 

8 119 5.58 6.16 5.87 0.151 0.026 

9 122 5.62 6.11 5.88 0.122 0.021 

10 92 6.16 6.81 6.49 0.192 0.030 

11 233 6.30 6.56 6.42 0.065 0.010 

12 114 5.42 6.20 5.92 0.172 0.029 

13 138 7.75 8.49 8.09 0.193 0.024 

15 58 7.89 8.58 8.29 0.230 0.028 

16 117 8.23 8.60 8.38 0.069 0.008 

17 181 2.25 3.09 2.66 0.246 0.092 

18 117 6.47 6.81 6.62 0.077 0.012 

19 174 3.45 3.99 3.77 0.117 0.031 

20 94 6.67 7.06 6.84 0.092 0.013 

21 141 1.51 2.25 1.91 0.190 0.099 

22 174 1.44 1.99 1.60 0.136 0.085 

23 125 4.51 5.19 4.87 0.197 0.040 

24 114 3.27 3.63 3.44 0.077 0.022 

25 184 2.79 3.67 3.17 0.259 0.082 

26 151 2.63 3.18 2.92 0.148 0.051 

27 103 7.94 9.17 8.57 0.353 0.041 

28 151 5.85 6.27 6.10 0.100 0.016 

29 89 7.08 8.11 7.52 0.295 0.039 

30 91 3.60 4.34 3.94 0.216 0.055


study sites and satellite data was acquired at the end 

of this period (respectively exactly one year later in 

the case of ASTER data), changing LAI conditions 

can have an effect on the resulting transfer functions. 

There was an attempt to minimize this error by taking 

measurements at the end of the rainy season when 

phenological changes are less pronounced and LAI 

is assumed to be relatively stable. In the following, 

the precision of in situ measurements and satellite 

observations will be assessed as this is a prerequisite 

for the establishment of adequate transfer functions. 

Precision of LAI e (LAI2000) 

As previously described, the precision of in situ 

data can be estimated by different measures. For 

LAI-2000 PCA data, Tan et al. (2005) used the 

standard deviation of measured LAI per ESU. In their 

study field sampling of an agricultural area with very 

homogeneous vegetation cover and low LAI values 

was undertaken, so that differences in individual 

measurements were related to noise from instrument 

set-up rather than to real heterogeneities in foliage 

amount. 

In this thesis however, LAI values are generally 

much higher and vegetation is less homogeneous, 

so that the standard deviation of LAI on each ESU 

is not only influenced by measurement precision, 

but also by local changes in forest structure and thus 

the natural variability of forest cover and foliage. 

Accordingly the standard deviation of LAIe (LAI2000) 

calculated from transect measurements (B2) contains 

both measurement errors and natural LAI variability. 

This is not the case for LAI derived from 

e (LAI2000) 

continuous B1 data. In this case, the measurement 

locality is stable and after correction of e and 

θsun 

filtering only the remaining e and instrument 

cloud 

inherent noise affect the results. Accordingly B1 

data is used to determine ESU specific LAIe (LAI2000) 

precision. Yet in lieu of the standard deviation, the 

Figure 6-2 Relation between c and LAI for all 

v e (2000) 

ESUs in Budongo Forest. 

105 

coefficient of variation was used to determine the 

measurement precision as the latter is independent of 

mean LAI per ESU (cf. Equation 5.1). Table 

e (LAI2000) 

6-2 shows the descriptive statistics of LAIe (LAI2000) 

derived from corrected B1 measurements with 

measurement precision being defined as c of the 

v 

respective ESU. 

As expected, measurement precision is lower (i.e. c v is 

higher nota bene) on ESUs measured under changing 

illumination conditions (e.g. ESU 4, 8, 10, 12). 

Noticeably it is especially low on ESUs belonging to 

the early forest stage, i.e. ESUs 17, 21, 22, 23, 25, 

and 26. Logically changing illumination conditions 

here have a higher impact on measurement precision 

due to the open – or non-existent – tree cover. Figure 

6-2 illustrates that c decreases with mean LAI (and 

v 

thus precision increases) with a mean c of 0.04 (4%). 

v 

It should be noted however that this is only true for 

LAI-2000 PCA measurements under at least partly 

changing illumination conditions. Precision is always 

highest on ESUs measured under diffuse and direct 

illumination conditions (with correction of eθsun applied to the latter).

106 

N Min Max N Mi 

Test 2 (30/10/2005) 12 

Test 1 (01/11/2005) 

2.90 

31 

3.39 

3.50 Test 

3.21 

4.35 1 (01/11/2005) 3.93 

0.140 

31 3.50 

Test 2 (30/10/2005) 12 2.90 Test 3.39 2 (30/10/2005) 3.21 12 2.90 

Precision of surface reflectance 

Precision of LAIe (DHP) 

In contrast the to proximity the precision of observed of field to LAI, true surface the precision reflectance. of surface reflectance 

Ideally field According spectrometer to this measurements m , m , …, m 1 2 N would are true serve values as ground of truth for s 

In contrast to LAI-2000 PCA measurements, difficult only to atmospherically acquire over a corrected tropical rainforest surface reflectance canopy, the derived approach of WANG 

transect data was available for DHP (equivalent this to thesis. from In this high case, spatial precision resolution values satellite derived data from at N either spectral satellite data or lite 

B2 measurements of LAI-2000 PCA). As mean the and proximity bands. of Various observed error to true sources surface can reflectance. cause deviations According to this m1, m 

standard deviation of the latter are also influenced atmospherically from these corrected values, surface leading reflectance to observations derived d , d 1 from , …, 2 high spatial res 

by vegetation heterogeneity, they could not be spectral used bands. d . The N Various latter error are treated sources as can independent cause deviations random from these true valu 

2 as a measure of precision. Instead the two tests d1, on d2, …, variables dN. The latter with finite are treated variances as independent σ , k=1, 2,…, k random N and it variables with fini 

the influence of illumination conditions described 

k k 

in Chapter 5.2.2 were analysed. As here the locality N and it is is supposed that the deviations k= 2 

k 

was not changed during photograph acquisition, it 

can be assumed that the resulting standard deviation 

is influenced by errors due to instrument set-up, 

exposure, illumination conditions and processing 

only. Again the cv of both time series was calculated. 

The results showed that both test sites had very similar 

cv of around 0.05 (cf. Table 6-3), which was taken 

as measurement precision for LAI on all ESUs. 

e (DHP) 

Although test data for early forest stages was not 

available, it is assumed that measurement precision 

for this is even higher as the smaller distance between 

foliage and lens minimizes the problem of mixed 

pixels and thus increases classification accuracy. 


In contrast to the precision of field LAI, the precision 

of surface reflectance is much harder to assess. 

Ideally field spectrometer measurements would Table 6-4: Relative precision for red, NIR and SWIR spectral bands for the high 

serve as ground truth for satellite data. As these are 

difficult to acquire over a tropical rainforest canopy, 

the approach of Wang et al. (2001) is followed in this 

thesis. In this case, precision values derived from 

either satellite data or literature are used to describe 

m d 


In contrast to the precision of field LAI, the precision of surface re 

Ideally field spectrometer measurements would serve as ground tr 

difficult to acquire over a tropical rainforest canopy, the approach o 

this thesis. In this case, precision values derived from either satellite d 

the proximity of observed to true surface reflectance. According to t 

atmospherically corrected surface reflectance derived from high sp 

spectral bands. Various error sources can cause deviations from these 

d1, d2, …, dN. The latter are treated as independent random variables 

follow Gaussian distribu 

k k 

N and it is supposed that the deviations k= 2 

Gaussian N distribution. N 

2 

2 

 

dkmThe k random variable k 

d m 

k 

2 

k1 

k1 k 

describes the proximity of observed to true values and has a chi-s 

2 

N indicates good observation quality. Dispersions 1, 2, …, 

atmospherically corrected surface reflectance and parameterized in terms of the 

(TAN et al. 2005). k is thus sensor and band specific. 

Unfortunately k cannot be assessed easily as it is influenced by the sensor-sp 

the applied atmospheric correction algorithm. According to HUANG et al. (200 

of satellite reflectance data can be derived from invariant targets, i.e. surfaces 

temporal processes such as, for example, vegetation phenology. Successive an 

of such surfaces can reveal mean, variance and thus precision, if the tar 

correction are assumed to be stable. As multi-temporal scenes were not avai 

specific relative precision had to be derived from literature in the case of SPO 

delivered as atmospherically corrected surface reflectance product, relative pre 

the data product description (cf. Table 6-4 for the bands used in SVI calculation 

Table 6-4: Relative precision for red, NIR and SWIR spectral bands fo 

m d 


In contrast to the precision of fie 

Ideally field spectrometer measur 

difficult to acquire over a tropical 

this thesis. In this case, precision v 

the proximity of observed to true 

atmospherically corrected surface 

spectral bands. Various error sourc 

d1, d2, …, dN. The latter are treated 

N and it is supposed 

 

follow that the Gaussia devia 

 

N N 

N N 

2 

2 

 

dkmk 2 

2 

d m 

k 

 

dkmk d m (6.6) 

 

k 

2 

2 

k1 

k1 

k1 

k1 k 

k 

describes the proximity of observed of observed describes to true values to the true and proximity values and of has ob 

2 

has 2 

a 

chi-square N indicates distribution, good observation where quality. 

N indicates Dispersions good observa 1, 

good observation quality. Dispersions 

atmospherically corrected surface reflectance atmospherically σ , σ , …, σ are 

1 2 N and parameterized corrected surface in terr 

precisions in the atmospherically corrected surface 

(TAN et al. 2005). k is thus sensor (TAN and band et al. specific. 2005). k is thus sensor 

reflectance and parameterized in terms of their 

Unfortunately relative values k α =σ cannot /d (Tan be assessed et al. Unfortunately 2005). easily αas is it thus is k influenced cannot be by assess the 

k k k k 

the sensor applied and band atmospheric specific. correction the algorithm. applied According atmospheric to correction HUANG 

of satellite reflectance data can be of derived satellite from reflectance invariant data targets, can i.e. be 

temporal Unfortunately processes α cannot such as, be for assessed example, temporal easily vegetation processes as it is phenology. such as, for Succ ex 

k 

of influenced such surfaces by the can sensor-specific reveal mean, calibration of variance such surfaces as and well thus can precision, reveal mea if 

correction as the applied are assumed atmospheric to be stable. correction correction As multi-temporal algorithm. are assumed scenes to be were stab 

specific According relative to Huang precision et had al. to (2006) be specific derived the relative from relative literature precision in had the to cab 

delivered precision as αatmospherically of satellite reflectance corrected delivered surface data as can reflectance atmospherically be product, correc 

k re 

the derived data product from invariant description targets, (cf. i.e. Table the surfaces 6-4 data for product that the are bands description used in (cf. SVI Tc 

not influenced by temporal processes such as, for 

example, vegetation phenology. Table Successive 6-4: Relative and precision for re 

repetitive measurements of such surfaces can reveal 

mean, variance and thus precision, if the target 

itself and atmospheric correction are assumed to be 

stable. As multi-temporal scenes were not available 

Table 6-3 Descriptive statistics for LAI derived from DHP time series taken at two test sites, each over a whole day. 

e (DHP) 

N Minimum Maximum μ σ c v 

Test 1 (01/11/2005) 31 3.50 4.35 3.93 0.182 0.046 

Test 2 (30/10/2005) 12 2.90 3.39 3.21 0.140 0.044


for this thesis, sensor specific relative precision had is define the slope LAI as of independent the relationship from between satellite X reflectance and Y, and e is the err 

to be derived from literature in the case of SPOT-4. between (and not an the easy other to way measure around), variable much (such of the as remote SVIs) and a costly 

As ASTER data was delivered as atmospherically easy sensing to literature measure reports variable the is LAI then being used modelled to predict as a the costly to 

corrected surface reflectance product, relative independent dependent variable variable, (Cohen and it et would al. 2003). be logical Alternatively to define LAI as ind 

precisions were available from the data product not the the coefficients other way for around), the regression much of model the remote could sensing be literatur 

description (cf. Table 6-4 for the bands used in SVI dependent derived with variable Y = LAI (COHEN and X et al. can 2003). subsequently Alternatively be the coeffic 

calculation). 

DERIVATION OF HIGH RESOLUTION LAI MAPS derived with by inversion Y = LAI of and Equation X can subsequently 6.7 so that be 7 derived by invers 

Sensor red 

6.2.2 Regression analysis 

NIR 

Y 0 

X e , 

SWIR 1 

Source 

(6.8) 

1% (0.15) 

Linear SPOT-4 regression HRVIR is a common 5.8% approach in 6.4% remote each 5.6% observation. 

One of the most HENRY widely & used MEYGRET regression (2001) approaches in remote 

sensing to estimate biophysical variables from remote 

squares (OLS) regression. Yet very often the underlying conditions 

sensing data, either in univariate (e.g. Tan et al. One of the most widely used regression approaches in 

one assumption is that X is measured free of error (COHEN et al 

2005) or multivariate form (e.g. Cohen et al. 2003). remote sensing research is the ordinary least squares 

6.2.2 Regression analysis 

realized as discussed earlier. The inclusion of error-related outlier 

Bivariate plots are usually used to determine whether (OLS) regression. Yet very often the underlying 

Linear regression is a common approach in remote sensing and to hence estimate the biophysical predictions variables based on from this remote model. Second, homos 

the relationship between the variables is truly linear conditions of this approach are not fulfilled. First, 

sensing data, either in univariate (e.g. TAN et al. 2005) variances or multivariate (FERNANDES form (e.g. et al. COHEN 2005). If et heteroscedasticity al. 2003). is the ca 

or if transformations have to be applied a priori. one assumption is that X is measured free of error 

Bivariate plots are usually used to determine whether the reduced relationship leading between to over the or variables under prediction is truly linear (COHEN et al. 2003). 

In the simple linear case the form of the regression (Cohen et al. 2003). In practice this can scarcely be 

or if transformations have to be applied a priori. In the simple degree linear of variance case the form attenuation of the regression is a linear model function of the co 

model is 

realized as discussed earlier. The inclusion of error- 

is 

investigation. The lower the correlation, the higher is the compress 

related outliers biases the retrieved model parameters 

Y 0 1X 

e 

(6.7) As and an hence alternative the to predictions OLS, FERNANDES based on et this al. (6.7) (2005) model. introduced the 

the Second, remote homoscedasticity sensing community. is assumed, It is based i.e. equal on the error median of ranke 

where Y is the variable to be predicted, X is the variable from which Y is predicted, 0 is the intercept, 1 

where Y is the variable to be predicted, X is the robust variances in the (Fernandes presence et of al. measurement 2005). If heteroscedasticity 

errors. A comparison to se 

is the slope of the relationship between X and Y, and e is the error. Usually a relationship is established 

variable from which Y is predicted, β is the intercept, OLS is the and case, Reduced variance Major in the Axis predicted (RMA) clearly variable showed is that OL 

0 

between an easy to measure variable (such as SVIs) and a costly to measure variable (such as LAI). The 

β is the slope of the relationship between X and Y, specification reduced leading of dependent to over or and under independent prediction variables, (Cohen leading to an 

1 

easy to measure variable is then used to predict the costly to measure variable. Although Y is the 

and e is the error. Usually a relationship is established 

15% 

et al. 

(FERNANDES 

2003). According 

et al. 2005). 

to Schlerf 

Both 

et 


al. 

RMA 

(2005) 

and 


Theil-Sen appr 

independent variable, and it would be logical to define LAI as independent from satellite reflectance (and 

between an easy to measure variable (such as SVIs) 

levels, 

degree 

but 


Theil-Sen 

variance attenuation 

regression was 

is a 

superior 

linear function 

at high error levels. 

not the other way around), much of the remote sensing literature reports the LAI being modelled as a 

and a costly to measure variable (such as LAI). The of the correlation between the variables under 

dependent variable (COHEN et al. 2003). Alternatively the coefficients for the regression model could be 

easy to measure variable is then used to predict investigation. The lower the correlation, the higher is 

derived with Y = LAI and X can subsequently be derived by inversion of Equation 6.7 so that 

the costly to measure variable. Although Y is the the compression of variances. 

independent Y variable, and it would be logical to 

0 

X e , 

 

(6.8) 

1 

with the error referring to the prediction residual for each observation. 

or if transformations have to be applied a priori. In the simple linea 

is 

Y X e 

0 1 

Table 6-4 Relative precision α for red, NIR and SWIR spectral bands for the high resolution sensors used. 

One of the most widely used regression approaches in remote sensing research is the ordinary least 

Sensor α α α Source 

red NIR SWIR squares (OLS) regression. Yet very often the underlying conditions of this approach are not fulfilled. First, 

one assumption is that X is measured free of error (COHEN et al. 2003). In practice this can scarcely be 

ASTER 1% (0.15) 

and hence the predictions based on this model. Second, homoscedasticity is assumed, i.e. equal error 

variances (FERNANDES et al. 2005). If heteroscedasticity is the case, variance in the predicted variable is 

reduced leading to over or under prediction (COHEN et al. 2003). According to SCHLERF et al. (2005) the 

degree of variance attenuation is a linear function of the correlation between the variables under 

investigation. The lower the correlation, the higher is the compression of variances. 

SPOT-4 HRVIR 5.8% 6.4% 5.6% Henry & Meygret (2001) 

As an alternative to OLS, FERNANDES et al. (2005) introduced the non-parametric Theil-Sen regression to 

the remote sensing community. It is based on the median of ranked slopes and it thus appears to be more 

robust in the presence of measurement errors. A comparison to several other regression models including 

107 

where Y is the variable to be predicted, X is the variable from whic

108 

As an alternative to OLS, Fernandes et al. (2005) of LAI values is formed including measured LAI 

introduced the non-parametric Theil-Sen regression values of all ESUs whose LAI values fall within this 

to the remote sensing community. It is based on interval. The corresponding surface reflectances are 

the median of ranked slopes and it thus appears grouped accordingly. Mean and standard deviation 

to be more robust in the presence of measurement for both groups are recorded. The process is repeated 

errors. A comparison to several other regression for case B, but here measurement precision of surface 

models including OLS and Reduced Major Axis reflectance is taken into account. Mean values for 

(RMA) clearly showed that OLS gives different LAI and surface reflectance of cases A and B are 

results depending on specification of dependent then finally combined in case C. At this point a stable 

and independent variables, leading to an average relationship is established, based on the newly formed 

difference in predicted LAI of 15% (Fernandes et data set that takes measurement and observation 

al. 2005). Both the RMA and Theil-Sen approaches 

performed similarly at low error levels, but Theil- 

Sen regression was superior at high error levels. 

Fernandes et al. (2005) conclude that the rank-based 

errors into account. 

8 Theil-Sen estimator should be used instead of OLS DERIVATION OF HIGH RESOLUTION LAI MAPS 

in situ LAI surface reflectance 

or RMA as univariate linear regression when the 

8 DERIVATION OF HIGH RESOLUTION LAI MAPS 

the sample rank-based data is Theil-Sen not error-free. estimator Accordingly should be this used type instead of OLS or RMA as univariate linear regression 

LAI LAIi 

i 

r i 


when of regression rank-based 

the sample was Theil-Sen 

data used is in not 

estimator this error-free. thesis. should The Accordingly Theil-Sen be used instead 

this type 

of OLS 

of regression 

or RMA as 

was 

univariate 

used in 

linear 

this thesis. 

regression 

The 

… 

… 

Theil-Sen 

when slope estimator the 

slope 

sample is estimator defined data is not 

is asdefined 

error-free. 

as 

Accordingly this type of regression LAI was used in this thesis. The 

j 

r j 

Theil-Sen slope 

 

estimator 

y 

is 

y 

defined as 

 

j i 

1 medians ij s ij , x j x i 1 i j n 

, 

 

x j x i 

 

y y 

 

j i 

1 medians ij s ij , x j x i 1 i j n 

, 

with the intercept 

x x 

being j calculated i as (6.9) 

case A 

(group k) 

impact of errors in in situ LAI 

(6.9) 

case B 

(group k) 

(6.9) 

impact of errors in surf. refl. 

with the intercept being calculated as 

with the median intercept y ybeing 

calculated 

median x as ... x . 

LAIi Li 

r i 

r i 

LAIi 

(6.10) 

0 

 

1... 

n 1 

1 

 

n 

j 

j 

j 

However, 0 median information y 1... 

y n on 

1median 

measurement x 1... 

x n . errors (6.10) is not explicitly included in Theil-Sen regression. (6.10) To 

mean/std. mean/std. 


evaluate 

However, 

its 

information 

performance, 

on 

results 

measurement 

derived 

errors 

for the 

is 

two 

not explicitly 

test sites 

included 

are compared 

in Theil-Sen 

to another 

regression. 

interesting 

To 

approach However, proposed information by TAN on measurement et al. (2005). errors In this is 

choose all LAIj 

that fall within 

choose all r j that fall within 

evaluate its performance, results derived for the two 

instance, 

test 

information on measurement precision is 

the sites measurement are error compared of LAIi 

to another the measurement interesting error of r i 

and corresponding r j 

and corresponding LAI 

not explicitly included in Theil-Sen regression. To 

j 

approach proposed by 

included to form a new database for regression 

TAN et al. (2005). In this instance, information on measurement precision is 

evaluate its in situ performance, LAI results surface derived reflectance for the analysis. 

included 

Basically 

to form 

information 

a new 

on the quality of in 

case database C for regression 

two test sites LAI are compared to another i 

interesting 

in situ iLAI 

surface reflectance 

i 

situ measurements and satellite (group k) 

analysis. Basically information 

observations 

on the quality 

is 


used 

in 

data set accounting for errors 

approach proposed … by Tan et al. (2005). … In this to derive new pairs in of both in input situ variables LAI and surface 

LAI LAIi 

i 

i 

situ measurements and satellite observations is used 

instance, information LAI j on measurement j 

precision reflectance data (cf. Figure 6-1). In case A it is 

… 

… 

to derive new pairs rof 

i in situ LAI 

i and surface 

is included to LAI form a new database for 

regression assumed that the true ... LAI value ... of each ESU is 

j 

j 

case A 

case B 

reflectance data (cf. Figure 6-1). In case A it is 

r 

analysis. Basically (group k) information on (group the k) quality of in 

j 

LAI j 

found 

assumed 

within 

that the 

a 

true 

certain 

LAI 

interval 

value of each 

(defined 

ESU 

by 

is 

impact of errors in in situ LAI 

impact of errors in surf. refl. 

situ measurements case A and satellite observations case B is used to measurement precision) mean/std. around mean/std. the measured LAI 

(group k) 

(group k) 

found within a certain interval (defined by 

LAIi Li 

i i 

LAIi 

derive new impact pairs of errors in of in situ in LAI situ LAI and impact surface of errors in surf. reflectance 

refl. value. Consequently a new group of LAI values is 

... 

... ... 

... measurement precision) around the measured LAI 

data (cf. Figure LAI 

 

j 6-3). L j 

j 

LAI 

i In case A it i 

is assumed i 

i 

LAI j that the 

i formed 

value. Consequently 

including measured 

a new group 

LAI values 

of LAI 

of all 

values 

ESUs 

is 

... 

... ... 

... 

Figure 6-3 Derivation of reference values for regression 

true LAI mean/std. value of mean/std. each ESU is mean/std. found within mean/std. a certain whose 

formed 

LAI 

including 

values 

measured 

fall within 

LAI 

this 

values 

interval. 

of all ESUs 

The 

LAI 

j 

j 

j 

LAI j 

analysis with measurement errors in field 

choose all LAIj 

that fall within 

choose all j that fall within 

interval (defined by measurement precision) around 

the measurement error of LAI 

the measurement error of corresponding surface reflectances are grouped 

mean/std. mean/std. i 


i 

and corresponding j 

and corresponding LAI whose LAI values measurements fall within and spectral this reflectance interval. The data 

j 

the measured choose all LAI LAI choose all that fall within accordingly. 

corresponding 

Mean 

surface 

and standard 

reflectances 

deviation 

are 


grouped 

both 

j that fall value. within Consequently a new group 

taken into account (Tan et al. 2005). 

j 

the measurement error of LAIi 

the measurement error of i 

and corresponding j case C and corresponding LAIj 

groups are recorded. The process is repeated for 

(group k) 

accordingly. Mean and standard deviation for both 

data set accounting for errors 

case B, but here measurement precision of surface 

in both case input variables C 

groups are recorded. The process is repeated for 

(group k) 

 

reflectance is taken into account. Mean values for 

data set i accounting for LAI errors i 

case B, but here measurement precision of surface 

in both input variables 

... 

... 

LAI and surface reflectance of cases A and B are 

 

j 

LAI j 

reflectance is taken into account. Mean values for 

i 

LAIi 

... ... 

then finally combined in case C. At this point a 


LAI and surface reflectance of cases A and B are 

j 

LAI j 

stable relationship is established, based on the 

... 

... ... 

... 

LAI r 

r 

LAI j


6.3 

Results 

The first step in analysing the relationship between 

SVIs and texture and LAI was a close examination of 

the reflectance obtained from different bands of high 

resolution image data. Irrespective of the satellite 

sensor (and the in situ method for LAI derivation), 

common trends could be observed. In the following, 

the Theil-Sen regression, being performed for 

LAI , LAI and LAI as well as SVIs 

e (LAI2000) e (DHP) (DHP) 

and texture measures, shed light on the relationship 

between these variables. LAI measures were declared 

as independent, SVIs and texture measures as 

dependent variables. In addition, the method proposed 

by Tan et al. (2005) was applied and compared to the 

results of Theil-Sen regression. Finally, the derived 

relationships are used to produce high resolution LAI 

maps for the two test sites with known accuracies. 

The maps are presented at the end of the respective 

chapters. 

6.3.1 


Surface reflectance 

A first analysis revealed that – as expected – surface 

reflectance was highest in the NIR (20.9-36.3%) and 

lowest in the red portion of the spectrum (2.8-8.1%). 

Interestingly, reflectance in ASTER bands was always 

slightly higher than in the respective bands of SPOT- 

HRVIR. As mentioned in Chapter 4.2.2, ASTER 

had already been acquired as pre-processed surface 

reflectance data. 

With respect to the different forest stages, a clear 

decrease in ρ , ρ , and ρ of ASTER 

red SWIR (band 4) SWIR (band 5) 

could be observed with increasing forest age (cf. 

Figure A-7). ρ is highest in intermediate stages 

NIR 

of the forest, but the difference is not significant 

with respect to the early forest stage. Results from a 

Mann-Whitney U test however suggest that there is 

109 

a significant difference between ρ of intermediate 

NIR 

and late forest stages (Z=-3.01, p

110 

SVIs 

In the next step the relationship between SVIs and 

LAI measures was assessed. Here too correlation 

analysis and bivariate plots of regressor and response 

variables served as first indicators. Although 

positive correlations (significant at p


performed best (R 2 =0.89). Generally OLS regression 

yielded higher R 2 than Theil-Sen. This can easily be 

explained since the latter is derived from the median 

of slopes between ranked data points, while OLS 

minimizes the sum of squared errors in the variable 

to be predicted. As calculation of R2 is also based on 

the sum of squared errors in relation to total sum of 

squares, OLS always leads to higher R2 . Nevertheless 

Theil-Sen regression is regarded as more robust with 

respect to outliers. 

Table 6-5 Theil-Sen and linear OLS regression models based on LAI and SVIs for early and intermediate forest stages 

e (DHP) 

in Budongo Forest. Models with the highest R2 in Theil-Sen regression are highlighted. 

Sensor Y Theil-Sen model R 2 Least squares model R 2 

SPOT-HRVIR SR 2.179+0.787 X 0.88 1.949+0.831 X 0.88 

NDVI 0.542+0.034 X 0.89 0.523+0.037 X 0.91 

NDVI c 0.163+0.065 X 0.85 0.101+0.072 X 0.90 

MSR 0.638+0.165 X 0.89 0.593+0.172 X 0.89 

RSR -0.082+0.843 X 0.89 -0.591+0.936 X 0.90 

NDMI -0.050+0.053 X 0.88 -0.096+0.060 X 0.90 

ASTER SR 2.250+0.602 X 0.94 2.407+0.583 X 0.94 

NDVI 0.518+0.031 X 0.94 0.516+0.032 X 0.94 

NDVI c 0.131+0.077 X 0.86 0.039+0.089 X 0.92 

MSR 0.636+0.130 X 0.93 0.651+0.129 X 0.94 

RSR -0.027+0.837 X 0.90 -0.672+0.911 X 0.94 

NDMI 0. 089+0.038 X 0.91 0.087+0.039 X 0.92 

Theil-Sen: R²=0.94 

Y=2.250+0.602 X 

OLS: R²=0.94 

Y=2.407+0.583 X 

Figure 6-6 Bivariate plots of a) LAIe 

(DHP) and SR derived from ASTER and b) LAI (DHP) and NDVI derived from SPOT 

c 

together with OLS (solid line) and Theil-Sen linear regressions (dashed line). 

111 

Figure 6-6a displays a bivariate plot of LAIe (DHP) 

and SR derived from ASTER with OSL (solid line) 

and Theil-Sen linear regressions (dashed line). It is 

obvious that both regression models are quite similar 

(as also indicated in Table 6-5). By contrast, a strong 

deviance between model equations (and respective 

R2 ), as, for instance, in the case of LAI and NDVI (DHP) c 

based on SPOT data (cf. Figure 6-6b and Table A-11 

in the Appendix), indicates that distinct data groups 

are present. Whereas OLS regression combines both 

groups without being influenced by the within-group 


Y=0.424+0.014 X 

OLS: R²=0.79 

Y=-0.020+0.061 X

112 

correlation, Theil-Sen regression is biased by the 

group containing more data points, in this case data 

from intermediate forest stages (cf. Equations 6.8 and 

6.9). 

Although Table 6-5 indicates that ASTER SR 

and NDVI retrieve the highest R2 in the modelled 

relationships, a closer analysis reveals that NDVI 

is not linearly related to LAI and that a non- 

e (DHP) 

2 

linear quadratic function (NDVI=0.69 LAIe (DHP) 

- 0.04 LAI + 0.34) leads to a higher R e (DHP) 2 of 0.95 

(cf. Figure 6-7a). In order to perform the robust, but 

linear Theil-Sen regression, the variables were log 

transformed (cf. Figure 6-7b. The resulting regression 

model NDVI =-0.723+0.212 X also retrieves an R log 2 

of 0.94 (OLS: NDVI =-0.736+0.218 X, R log 2 =0.95) 

with X being log transformed LAI . Other SVIs 

e (DHP) 

were also checked for non-linear relations and logtransformed 

if necessary, but the above-mentioned 

ASTER SR and NDVI were still best performing. 

OLS: R²=0.95 

Y=0.69 X²-0.04 X+0.34 

Texture 

As described previously, LAI variables and SVIs 

correlated well for the early and intermediate forest 

stages. No correlation could in turn be retrieved for 

the late forest stage. In addition, despite a significant 

difference in ρ and in ESU internal variance of 

NIR 

ASTER bands 3, 4 and 5, no significant difference 

in LAI can be detected (cf. Chapters 5.3.1, 6.3.1 

and Figure 5-24). This indicates that besides foliage 

amount there is another major influence on surface 

reflectance in old growth forest sites. As discussed 

earlier, structural differences can be the reason. 

Shadow components in particular, resulting from an 

uneven canopy with emergent trees (cf. Figure 3-12), 

may lead to changes in surface reflectance, especially 

in the NIR and SWIR. 

As structural properties are reflected in image 

texture, correlation analyses were performed to 

assess the relation between LAI and second-order 

texture variables (cf. Table 6-1) for late forest stages. 

Significant correlations to LAI variables could be 

retrieved for several texture measures from both 

ASTER and SPOT data, at p


additional correlations significant at the 0.01 level 

could be retrieved: for LAI and homogeneity 

e (LAI2000) 

(ASTER band 3, 3x3 kernel), LAI and variance 

e (DHP) 

(ASTER band 4, 9x9 kernel), as well as contrast 

(ASTER band 8, 9x9 kernel) with r of -0.88, 0.86 

s 

and -0.91 respectively. Variance of ASTER band 

4 had significant correlations with LAI in all 

e (DHP) 

kernels with r being highest for the 9x9 window 

s 

size. Contrast computed from ASTER band 8 and 

LAI e (DHP) yielded the highest r s with -0.91. 

Theil-Sen regression was subsequently applied to all 

texture measures with correlations to LAI variables 

significant at p

114 

First, the impact of measurement errors in field from newly composed groups as their estimates, bias, 

LAI on the retrieved surface reflectance values was precision and accuracy can be assessed according to 

assessed (case A). Therefore groups of field LAI, LAIi the definitions in Chapter 6.2.1. Normalization by the 

and corresponding surface reflectance were formed mean of the estimates allows the later comparison of 

14 according to 

these DERIVATION measures OF and HIGH leads RESOLUTION to relative bias, LAI MAPS precision 

and accuracy. 

LAI LAI LAI c 

(6.11) 

(6.11) 

i 

j 

i 

v 

with LAIi being LAIe (DHP) of a certain ESU and LAIj referring to LAIe (DHP) of all other ESUs. If LAIj is 

within the measurement error of LAIi, described by cv (cf. Chapter 6.2.1), LAIe (DHP) and corresponding ρred 

and ρNIR are added to the group of LAIi. The resulting groups contain 1 to 6 individual pairs (mean=3.15) 

of in situ LAIe (DHP) and surface reflectance in the red and NIR spectral bands of ASTER. 

Figure 6-9a shows the relationship between observed LAIe (DHP) and mean LAIe (DHP) of the newly formed 

groups. As expected, the correlation between observed field LAI and associated mean LAIi of groups is 

very high with R2=1, with LAI being LAI of a certain ESU and LAI i e (DHP) j 

referring to LAI of all other ESUs. If LAI is 

e (DHP) j 

within the measurement error of LAI described by c i, v 

(cf. Chapter 6.2.1), LAI and corresponding ρ e (DHP) red 

and ρ are added to the group of LAI . The resulting 

NIR i 

groups contain 1 to 6 individual pairs (µ=3.15) of in 

situ LAI and surface slope reflectance 1.00 and in offset the red 0.02. and 

e (DHP) The differences between observed ρred and ρNIR and 

associated NIR spectral mean bands ρred of and ASTER. ρNIR are in turn slightly higher as displayed in Figure 6-9b. If observed values are 

taken as true values and means from newly composed groups as their estimates, bias, precision and 

accuracy Figure 6-9a can shows be assessed the relationship according between to the definitions observed in Chapter 6.2.1. Normalization by the mean of the 

estimates LAI allows and mean the later LAIcomparison of the of newly these formed 

e (DHP) e (DHP) measures and leads to relative bias, precision and accuracy. 

groups. As expected, the correlation between 

observed field LAI and associated mean LAI of the 

i 

new groups is very high with R2 Correspondingly the proximity between observed 

ρ and ρ and mean values of associated groups 

red NIR 

is characterized by a relative bias, precision and 

accuracy of

eflectance of 17% in the red and 5% in the NIR wavelengths, which corresponds to an overall relative 

precision of 11% (TAN et al. 2005). 


In the second step (case B), the impact of observation errors in surface reflectance on the prediction of 

LAIe (DHP) was analysed. Taking into account band specific values (cf. Table 6-4), and following WANG et 

al. (2001), two observations i and j are indistinguishable if 

2 2 

i j 

(6.12) 

(6.12) 

(cf. Equation 6.7). Analogue to the previously 

described steps, another 20 groups of data are 

formed, each containing reflectance values that are 

equal within the observation precision (i.e. Equation 

6.12 holds true) and the corresponding LAIe (DHP) 

values. The resulting groups contain 1 to 5 data pairs 

(mean=2.75). The correlation between observed 

surface reflectance and mean surface reflectance of 

the associated groups is again very high with R2 =1, 

slope 1.00 and offset -0.01 for ρ and R red 2 =0.89, slope 

0.86 and offset 0.04 for ρ (cf. Figure 6-10a). 

NIR 

Figure 6-10b displays the response of LAIe (DHP) 

to variations in surface reflectance that are 

equal to within measurement precision. In situ 

LAI and mean values over associated groups 

e (DHP) 

are well correlated (R2 =0.95) with a slope of 0.99 

and an intercept of 0.05. Yet especially for higher 

LAI , field data can vary significantly (between 

e (DHP) 

5.68 and 7.07), with mean values remaining essentially 

unchanged (6.22 to 6.35). Relative bias, precision 

and accuracy are

116 

Assuming that no measurement errors are present 

in the resulting data set, OLS regression was 

performed thereafter. As expected, ASTER SR and 

log transformed NDVI again produced the highest 

R2 in the modelled relationships, with ASTER SR 

performing slightly better. Figure 6-12 displays 

bivariate plots of both SVIs and (log transformed) 

LAI derived from the previously formed groups. 

e (DHP) 

Interestingly, the results show that if measurement 

errors in both LAI and surface reflectance are 

e (DHP) 

OLS: R²=0.99 

Y=-0.04+1.01 X 

taken into account, the agreement between the two 

variables is more precise and more accurate (cf. 

Table 6-7). However, modelled relationships derived 

from the above-described method differ only slightly 

from the models derived with Theil-Sen regression 

in the previous chapter, yet yield a higher R2 . Table 

6-8 summarizes the regression models for early and 

intermediate stages derived from Theil-Sen regression 

and the method according to Tan et al. (2005). 

The models are very similar with SR performing 

slightly better than log transformed NDVI in the Tan 

approach. 

OLS: R²=1 

Y=-0.19+1.00 X 

Figure 6-11 Relationship between a) observed LAIe 

(DHP) and mean LAI values over associated groups and b) observed surface 

e (DHP) 

reflectance and mean values over associated groups in the red (circles) and NIR (asterisks) spectral bands of ASTER, 

taking into account observation errors in both field LAI and ASTER data (solid lines represent 1:1 line for comparison). 

OLS: R²=0.98 

Y=2.282+0.604 X 

OLS: R²=0.98 

Y=-0.742+0.222 X 

Figure 6-12 Bivariate plots of a) log transformed LAIe 

(DHP) and NDVI (ASTER) and b) LAI and SR (ASTER) together with OLS 

e (DHP) 

(solid line) regression derived from case C.


The SR model is also illustrated in Figure 6-13. Here 

the original data set of LAI and ASTER SR is 

e (DHP) 

displayed together with the Theil-Sen regression 

model (dashed line) and the OLS regression model 

(solid line, Tan method). As LAI calculated from 

e 

both models differs only by about 1%, the methods 

can be regarded as equal. Yet when OLS regression 

is performed on the original data set, differences in 

LAI of up to 7% can occur. Theil-Sen regression can 

e 

therefore be regarded as a robust method to derive 

statistically sound relationships between variables 

that are not free of measurement errors. Theil-Sen 

regression models will consequently be used for the 

calculation of the high resolution LAI map. 

Table 6-7 Relative bias, precision and accuracies of relationships between LAI and surface reflectance derived from 

e (DHP) 

ASTER data. Groups contain only data from early and intermediate forest stages. 

Case A Case B Case C 

ASTER reflectance [%] LAI e (DHP) [%] LAI e (DHP) [%] ASTER reflectance [%] 

ρ red ρ NIR ρ red ρ NIR 

Relative bias

SR 2. 

250 

LAIe , (6.13) 

0. 

602 

118 

as used to calculate LAIe for early/intermediate stages. LAIe of late forest stages was modelled according 

to 

GLCM variance (band 4) 4.912 

LAI e 

2. 

746 

(6.14) 

(6.14) 

As GLCM variance can be relatively high at at forest 

edges or around large forest gaps due to strong 

differences edges or around in surface large reflectance forest gaps of due adjacent to strong pixels, one further constraint was included in LAIe 

differences in surface reflectance of adjacent 

pixels, one further constraint was included in LAIe calculation: if GLCM variance exceeded the data 

space shown in Figure 6-8, i.e. it is higher than 18 

and lower than 8, Equation 6.13 was applied instead. 

Figure 6-14 displays the relationship between field 

measured LAI values and LAI derived from 

e (DHP) e 

Equations 6.13 and 6.14. R2 be – at least partially – affected. As no ESU had been 

established in these areas as they were mostly not 

accessible at the end of the rainy season, no statement 

can be made concerning the validity of LAI in the 

e 

affected areas. 

Nevertheless – as in the areas declared as late forest 

stages in Figure 6-15 – comparatively high amounts of 

Cynometra alexandri seem to be present (cf. classes 

“Cynometra” and “Cynometra mixed” in Figure 2-6). 

Therefore the relationship between LAI and SVIs 

e 

or texture measures is probably not only influenced 

by applied logging schemes, but also species 

of the relationship is 0.87 composition. On the other hand compartments W19 

with an overall RMSE of 0.47 (0.39 for early and and W20, which according to Plumptre (1996) should 

intermediate, 0.64 for late forest stages). The relative also be characterized by high Cynometra abundance, 

accuracy is 9% (7% and 11% for early/intermediate show no significant reduction in LAI compared to 

e 

and late stages respectively). 

neighbouring compartments in the south and west. 

Consequently Equation 6.14 cannot be applied to 

The resulting high resolution LA map for Budongo 

e compartments with apparently underestimated LAIe. Forest is displayed in Figure 6-15. The dashed Rather these areas will be excluded from further 

lines represent those forest compartments to which 

Equation 6.14 was applied. Due to the moving 

window size of 9x9 pixels those areas appear to be 

less detailed. Red squares in Figure 6-15 represent 

three focus areas that are displayed in Figures 6-16a, 

c and e, together with the corresponding areas in the 

ASTER image (band combination 4, 3, 2) in Figures 

6-16b, d and f. Figure 6-16a and b show the area of 

the Nature Reserve N15 in the southwest of Budongo 

Forest. The overall LAI pattern of forest and 

e 

cleared areas with wooded grassland is well covered 

(cf. Figure 6-16a). The latter has LAI values of 2.2 

e 

to 3.7. Inside the forest LAI variability is higher with 

e 

minimum and maximum values of 3.0 and 8.5 and a 

mean of 6.0 (±0.9). Figures 6-16c to f display the two 

analysis, i.e. MODIS LAI product validation. 

forest areas in the southeast and north respectively. 

OLS: R²=0.87 

Here clear patterns of areas with lower and higher 

LAI are visible. The areas with lower LAI are 

e e 

following the borders of compartments W29, W30, 

Y=0.096+1.004 X 

W41, W42, and W43. Additionally compartments 

Figure 6-14 Relation between LAI estimated from ASTER 

e 

data with equations 6.12 and 6.13 and in situ 

KP2, KP 4, KP7, S4, S5, W41, W42 and W43 seem to 

measured LAI.


Figure 6-15 High resolution LAI map of Budongo Forest estimated from ASTER data. 

e 

119

120 

Figure 6-16 Focus areas of Budongo Forest displayed in Figure 6-15. a), c) and e) represent LAI calculated from Equations 6.12 

e 

and 6.13, b), d) and f) show ASTER data for the corresponding areas (band combination 4-3-2). Dashed areas 

represent late forest stages.


6.3.2 


Data analysis for the Kakamega Forest test site 

followed the same steps as described for Budongo 

Forest. Once more it must be mentioned that especially 

in situ data quality was not comparable. First of all, 

measurement precision of LAI could not be 

e (LAI2000) 

determined due to the lack of a third LAI-2000 PCA 

device and different filter methods. Consequently 

only Theil-Sen regression was performed to model 

the relationship between LAI variables and SVIs/ 

texture variables. Second, LAI , which correlated 

e (DHP) 

best to high resolution satellite data for the Budongo 

Forest test site excluded understorey vegetation in 

Kakamega Forest. And last but not least, ASTER data, 

which gave better results in Budongo Forest for both 

SVIs and texture measures, had not been available 

for Kakamega Forest. Additionally ESUs 11, 13, and 

29 had to be excluded from analysis due to clouds 

or cloud shadows. The results of the analyses will be 

presented together with the high spatial resolution 

LAI map in the following section. 

Surface reflectance 

The overall surface reflectance pattern of the different 

forest stages was comparable to the values retrieved 

for the Budongo Forest test site, albeit with mean 

surface reflectance per forest stage being slightly 

lower in the red part of the spectrum (0.5-1%) and 

slightly higher in the NIR and SWIR bands (1-4%, 

cf. Figures 6-4 and A-7. In addition, correlation 

analysis revealed similar results as for Budongo 

Forest. LAI and LAI were again 


negatively correlated to ρ green , ρ red and ρ SWIR with r s 

ranging between -0.51 and -0.58 (p

122 

The corresponding R 2 was, however, lower with 0.65 

(OLS) and 0.63 (Theil-Sen). 

Figure 6-18 displays a bivariate plot of LAIe (LAI2000) 

and RSR derived from SPOT data, together with the 

Theil-Sen (dashed line) and OLS models (solid line) 

given in Table 6-9. The inverted regression equation 

retrieved from the Theil-Sen model was used for LAIe Figure 6-18 Bivariate plot of LAI and RSR derived 

e (LAI2000) 

from SPOT data for Kakamega Forest together 

Table 6-9 Theil-Sen and linear OLS regression models based on LAI and SVIs for early and intermediate forest stages in 

e (LAI2000) 

Kakamega Forest. 



Table 6-10 

NDVI 0.745+0.011 X 0.48 0.684+0.022 X 0.78 

NDVI c 0.306+0.038 X 0.67 0.267+0.044 X 0.72 

MSR 3.207+0.089 X 0.58 2.877+0.147 X 0.78 

RSR 2.478+0.608 X 0.72 2.002+0.678 X 0.75 

NDMI 0.651-0.023 X 0.61 0.687-0.030 X 0.65 

Theil-Sen and linear OLS regression models based on LAI variables and texture measures for late forest stages 

in Kakamega Forest. 

Sensor Kernel Band Texture measure X Theil-Sen model R 2 Least squares model R 2 

SPOT- 

HRVIR 

with OLS (solid line) and Theil-Sen linear 

regressions (dashed line). 


Y=2.478+0.608 X 

OLS (Tan): R²=0.75 

Y=2.002+0.678 X 

estimation in early and intermediate forest stages in 

Kakamega Forest. 

Texture 

With respect to the late forest stages of Kakamega 

Forest, high r were retrieved for various texture 

s 

Figure 6-19 Bivariate plot of LAI and homogeneity (band 

e (DHP) 

4, 7x7 kernel) derived from SPOT with OLS 

(solid line) and Theil-Sen linear regression 

models (dashed line). 


Y=0.447-0.047 X 

OLS (Tan): R²=0.50 

Y=0.468-0.052 X 

5x5 4 Homogeneity LAI e (DHP) 0.441-0.045 X 0.35 0.497-0.059 X 0.49 

7x7 4 Homogeneity LAI e (DHP) 0.447-0.047 X 0.36 0.468-0.052 X 0.50


measures at p

124 

Figure 6-21 Focus areas of Kakamega Forest displayed in Figure 6-22. a) and c) represent LAI calculated from Equation 6.14, 

e 

b) and d) show SPOT data of the corresponding areas (band combination 4-3-2). 

Overall this model yielded an R 2 of 0.53 with 

an accuracy of 0.8 (relative accuracy 16%). The 

resulting high resolution LAI map is displayed in 

Figure 6-20. Once more the overall LAI pattern 

e 

is well covered, with slightly higher values in the 

central part of the forest that is assumed to be less 

disturbed than the peripheral parts. Two focus areas 

are displayed in Figure 6-21. Whereas the area around 

Buyangu Hill (Figure 6-21a and b) shows reasonable 

patterns of LAI indicating low LAI values (1-5) 

e e 

in the early succession stages of Psidium guajava 

and high LAI in the relatively undisturbed forest 

e 

area of Salazar, LAI patterns in Figure 6-21c indicate 

e 

once more the model’s constraint. Very high LAI e of 

>8 is shown for areas southeast of Isecheno Forest 

Station. For large parts of these areas the classification 

of Kakamega Forest (displayed in Figure A-2) 

indicates plantation forests of Bischoffia javanica, for 

which the model is not valid. The same effect occurs 

in the south and northwest of Kakamega Forest. 

6.4 

Conclusion 

The objective of the analyses described in the 

preceding chapter was the production of high


resolution LAI maps for the two test sites with known 

accuracy based on in situ measurements and satellite 

observations. For that reason empirical transfer 

functions were established as they – in contrast to 

physical models – are only limited by the sampling 

distribution. The disadvantage is that empirical 

models cannot be transferred in time or space, 

but as they serve only as an intermediate step for 

validation of the MODIS LAI product in this thesis, 

the establishment of generally valid models was not 

necessary. Ideally high resolution LAI maps should 

be produced without loss in accuracy in relation to 

in situ data. Therefore, measurement precision of 

the input variables and their impact on the resulting 

transfer functions must be taken into account. 

Whereas precision of field measurements in Budongo 

Forest was comparable for both devices, with 4% and 

5% (LAI-2000 PCA and DHP respectively), precision 

of surface reflectance varies by up to 7%. Once more 

it must be emphasized that measurement precision 

of field LAI as discussed in Chapter 6.2.1 refers to 

relative precision, because information on true LAI 

was not available. Additionally precision of surface 

reflectance could not be directly determined, but was 

retrieved from literature. As Tan et al (2005) state, the 

estimation of LAI from spectral surface reflectance 

(or derived spectral vegetation indices) is an ill-posed 

problem. Whereas small variations in LAI result only 

in limited variations in modelled spectral surface 

reflectance (i.e. the prediction of the radiation field is 

a well-posed problem), the inverse calculation results 

in an unstable solution, i.e. small variations in input 

reflectance due to observation errors can result in a 

very low precision of LAI. 

Therefore it is necessary to derive stable relationships 

that also account for errors in input data. To evaluate 

the performance of Theil-Sen regression in the 

presence of measurement errors, the results were 

compared to a method proposed by Tan et al (2005). 

Whereas the former is very robust with respect to 

outliers, it does not include information on precision 

125 

of input data. Tan et al.’s method incorporates this 

information. However, a comparison of the results 

showed that the derived transfer functions differ only 

slightly, with discrepancies in modelled LAI of ~1%. 

Both approaches are thus assumed to give comparable 

and reliable results even in the presence of errors in 

input data. 

For Budongo Forest, correlation and regression 

analyses revealed that different transfer functions 

had to be set up for early/intermediate and late forest 

stages respectively. Whereas SVIs are sensitive to 

changes in LAI in early to intermediate forest stages 

(retrieving an R2 of 0.94), no relation could be found 

to the LAI variables from late forest stages. This is 

probably caused by a different forest structure: in 

mature undisturbed tropical rain forest emergent 

trees are present, creating a very rough canopy 

texture. The result is a higher shadow component 

than in intermediate forest stages, where emergent 

trees, in particular, have been logged. Nevertheless 

overall LAI can be comparable due to the denser 

understorey in intermediate forest stages. Due to the 

smoother canopy surface in intermediate forest sites, 

larger surface areas of crowns are exposed, causing a 

higher surface reflectance especially in the NIR and 

SWIR bands. The mutual shadowing of crowns in the 

late forest stages causes, in turn, a higher variance 

in canopy reflection. Consequently a good relation 

(R2 =0.71) between GLCM variance of ASTER band 

4 and LAI could be found for late forest stages. 

e (DHP) 

Interestingly in Budongo Forest regressions with 

the highest R2 were found for LAI irrespective 

e (DHP), 

of the forest stage. The inclusion of understorey 

(in contrary to LAI , where understorey 

e (LAI2000) 

vegetation is excluded) improves the relation between 

LAI variables and SVIs or texture variables. The 

correction for foliage clumping in LAI , however, 

(DHP) 

seems to weaken the strength of the modelled 

relationship. This has also been reported by Eklundh 

et al. (2003) where correlations between LAI and 

e

126 

SVIs were significantly stronger than between LAI 

corrected for clumping and SVIs. According to 

Chen et al. (2005), SVIs are generally proportional 

to foliage cover in the view direction. Further foliage 

clumping affects gap fraction for the same LAI and 

can thus delay the occurrence of the saturation in 

reflectance as LAI increases. Consequently LAIe has a stronger relationship to the satellite signal. In 

this context, Stenberg et al. (2004) suggest applying 

clumping corrections a posteriori to LAI estimation 

from SVIs. 

With respect to satellite sensors, ASTER data 

performed better than SPOT data in Budongo Forest. 

The reason for this remains unclear, but possible 

explanations include sensor characteristics (e.g. 

slightly different band widths), as well as applied 

atmospheric correction algorithms. Last but not least, 

forest phenology at the time of ASTER acquisition 

might be more similar to phenology at the time of 

fieldwork – even though the imagery was acquired 

one year after. 

In general, similar observations can be made for 

Kakamega Forest. Here too the exclusion of LAI data 

from late forest stages improved the relationships 

between SVIs and LAI variables. In this case RSR 

was found to be best suited for LAI estimation, 

however with a lower R2 of 0.72. In contrast to 

Budongo Forest, the best relationship was retrieved 

with LAI Apparently the fact that understorey 

e (LAI2000). 

vegetation was not covered with LAI on this test 

e (DHP) 

site had a negative impact on the resulting models. For 

late forest stages a significant relationship between 

effective LAI and a texture variables calculated from 

the SWIR band was also established, although in 

this case the relationship was comparatively weak 

(R2 =0.36). Interestingly the best correlations for 

texture variables were retrieved for a 9x9 moving 

window for ASTER (Budongo Forest) and 7x7 kernel 

for SPOT (Kakamega Forest), which correspond 

to 135 m and 140 m side length respectively. As 

individual crowns can have up to 30 m in diameter, 

this edge length apparently best captures the variation 

of surface reflectance being reflected by the canopy. 

In comparison to Broich (2005), who produced 

a high resolution LAI map for Kakamega Forest 

based on the same LAI-2000 PCA measurements 

(but including measurements in south and west 

directions, cf. Chapter 5) and Landsat ETM+ SLC-off 

data, the previously described analyses revealed no 

significant relation between SWIR reflectance of 

SPOT and LAI as shown in his study. In this thesis 

e 

better relationships were yielded by the calculated 

SVIs. The here applied SPOT-4 data has moreover 

the advantage of full spatial coverage in contrast 

to the Landsat ETM+ SLC-off data. Results from 

Budongo Forest further suggest that the estimation 

of LAI from remote sensing data might generally 

e 

be more accurate when different models are set up 

for different forest stages (though in the end only one 

model was used for Kakamega also in this thesis.). 

For both test sites, care has to be taken to integrate 

only those areas in the upscaling process, which are 

comparable to the ESUs on which the regression 

models were established. In Budongo Forest, the 

dominance of certain species (in this case Cynometra 

alexandri) may distort the modelled relationships. 

In Kakamega Forest it is quite obvious that the 

LAI in plantantions is exaggerated. Broich (2005) 

encountered similar problems in his analyses. 

The resulting high resolution LAI e map of Budongo 

Forest has a relative accuracy of 9%. The upscaling 

process thus caused only a slight – yet still acceptable 

– degradation in accuracy of the high resolution 

LAI map compared to the accuracy of in situ 

e 

measurements and satellite observations. The LAIe map of Kakamega Forest is, however, less accurate 

at 16% due to inferior in situ data quality. Map 

accuracies should be kept in mind when validating 

the medium resolution LAI product.

7 

Validation of the 

MODIS LAI product 

As described in Chapter 3.4.4 the final steps in MODIS 

LAI product validation involve the aggregation of 

the high resolution LAI maps and comparison with 

e 

the MODIS data sets to be validated. However, care 

has to be taken when using the high resolution maps 

presented in Chapter 6 as they give LAI , which is 

e 

not corrected for foliage clumping. Yet the MODIS 

LAI product is supposed to model “true” LAI. In 

order to resolve this, a clumping factor was applied 

to the high resolution maps as suggested by Stenberg 

et al. (2004). At least for intermediate and late forest 

stages this can be derived from DHP. For early forest 

stages, however λ estimated from DHPs seems to be 

too small when compared to reference data. 

In the next step the quality of input data for the 

LAI algorithm and its performance is assessed to 

find the right time steps for the final validation of 

the MODIS LAI product. Based on uncertainty 

measures introduced in chapter 6.2.1, its accuracy 

can subsequently be determined for the respective 

study sites and time steps. Here also the model 

observation precision of the MODIS LAI product 

has to be taken into account, which is reported to 

be 0.2 (Tan et al. 2005). Although a true temporal 

validation would require field measurements of LAI 

over several months, at least temporal consistency of 

the MODIS LAI product is investigated. Information 

on the applied LAI algorithm is analysed and LAI 

trajectories are derived for the different forest stages. 

7.1 

Upscaling of the high resolution 

LAI map e 

127 

To avoid resampling of the MODIS LAI product, 

Morisette et al. (2006a) suggest comparing the high 

resolution LAI map with the MODIS LAI product 

e 

in the projection of the latter. Since spatial errors are 

relative to pixel size, resampling will have less impact 

if applied to the high spatial resolution LAI maps. 

e 

The aggregation of the high spatial resolution LAIe maps should further take into account the point spread 

function of the high resolution satellite product. This 

step is however poorly addressed in current research 

and also complicated by various preprocessing 

steps and repeated resampling in product generation 

(Morisette et al. 2006a). To avoid geolocation 

uncertainties and to retrieve a statistically stable 

result, Yang et al. (2006) recommend comparison of 

both products at multipixel (patch) scale (cf. Chapter 

3.4.5). 

7.2 

Correction for foliage clumping 

As previously mentioned, the MODIS LAI algorithm 

is designed to give true LAI. On the other hand the 

regression models developed in Chapter 6 are based 

on LAI and LAI , which obviously 


correlated better to SVIs and texture measures 

derived from high resolution satellite data. Yet 

according to Stenberg et al. (2004), a correction for 

foliage clumping can also be applied to the final LAIe maps, if the clumping factor λ is known. 

Table 7-1 lists mean clumping factors derived 

from hemispherical photography for the different 

forest stages of both study sites. A comparison with 

clumping indices derived on a global scale by Chen 

et al. (2005) from multi-angular POLDER 1 data 

reveals that the clumping factor λ for intermediate 

and late forest stages is well within the expected range 

for both test sites (cf. Table 7.1). Chen et al. (2005)

128 

estimated the clumping index for evergreen broadleaf 

forests to be 0.63 on global average, with minimum 

and maximum values of 0.59 to 0.68. Mean λ derived 

from DHP for intermediate and late forest stages are 

0.67 and 0.68 respectively and thus lie on the upper 

end of that range. For early forest stages, however, 

λ derived from DHP seems to be unreasonably 

low. This was also observed in Chapter 5.4, when 

LAI of early forest stages were discussed. This 

(DHP) 

was due to relatively low λ values derived from in 

situ measurements: for Budongo and Kakamega 

Forests λ values of 0.51 and 0.52 were determined. 

The corresponding λ given by Chen et al. (2005) are 

however much higher, with 0.74 on average (0.64 to 

0.83) for herbaceous cover, and 0.73 (0.62 to 0.80) for 

deciduous open to closed shrub cover. The apparent 

exaggeration of LAI due to the low λ derived 

(DHP) 

from DHP for early forest stages probably weakened 

the performance of regression models based on this 

parameter (cf. Chapter 6.3). λ values derived for 

Kakamega Forest are also not within the range given 

by Chen et al. (2005) for early forest stages: for the 

latter consequently no correction of foliage clumping 

can be applied. 

7.3 

7.3.1 

Results 


Ideally the MODIS LAI algorithm should produce 

data that is accurate at the biome or regional level. 

Further it should respond correctly to biomelevel 

LAI trajectories associated with interannual 

climate variability and be sensitive to meaningful 

perturbations to LAI, such as from significant 

disturbances (Cohen et al. 2006). Whereas the two 

latter objectives can only be confirmed with time 

series data, the spatial validation of product accuracy 

at the regional level requires a thorough analysis of 

algorithm input and output. Since the MODIS LAI 

algorithm is based on daily surface reflectance data, 

information on cloud and land cover as well as biome 

specific LUTs (Figure 4-2) errors in these data sets, 

i.e. uncertainties in input data and uncertainties in 

the applied model, obviously influence the quality of 

the resulting LAI product. For that reason the quality 

of input data will be analysed in a first step before 

the resulting algorithm output is validated in terms 

of accuracy in comparison to the up-scaled high 

resolution LAI maps. Note that all overview figures 

e 

in this chapter refer to the area shown in Figure 

2-2, the final product comparison (cf. Figure 7-5) is 

displayed for the extent of Figure 6-16). 

Spatial validation 

For the generation of the 8-day LAI composites 

there is no information determining which surface 

reflectance data sets are used for product generation. 

Consequently acquisition date and viewing geometry 

remain unclear. As the quality flags of the input 

reflectance data are, however, passed through from 

the original data sets to the MOD15A2 product (cf. 

Table A-2), the quality of input data can be judged 

indirectly. 

Table 7-1 Comparison of λ derived from DHP for forest stages of the respective study sites with λ given in Chen et al. (2005) for 

the classes “deciduous open to closed shrub cover” and “evergreen broadleaf forest”. 

Data source λ (early stage) λ (intermediate stage) λ (late stage) 

In situ data, Budongo Forest 0.51 0.67 0.68 

In situ data, Kakamega Forest 0.52 0.64 0.65 

Chen et al. (2005) 0.73 (min. 0.62, max. 0.80) 0.63 (min. 0.59, max. 0.68)

7 Validation of the MODIS LAI product 

As for high resolution optical data, the main quality 

constraint for the MOD15A2 product is cloud cover. 

Although the product is generated as an 8-day 

composite, cloud contamination and subsequent 

errors in algorithm results are still major obstacles. 

Figure 7-1 summarizes the quality information of 

MODIS LAI products for the Budongo Forest region 

with respect to cloud cover and aerosols. Data sets 

of Julian Days (JD) 273 to 353 (30 September to 19 

December, 2005) served as input, which corresponds 

to all MODIS LAI data sets acquired during the time 

of field work as well as two time steps before and 

after. Black pixels resemble areas, where no clouds, 

cirrus clouds or aerosols could be detected at any 

time step. Red pixels are either water (Lake Albert) 

Figure 7-1 

Figure 7-2 

Quality information on cloud cover, cirrus clouds 

and aerosols of the MOD15A2 data sets 

JD 273-353 2005. 

Percentage of invalid (i.e. cloud contaminated) pixels for JD 273 to 353 in 2005. 

129 

or pixels where clouds, cirrus clouds or aerosols were 

detected at each of the 11 time steps. It is obvious 

that the forested region in the middle is more prone 

to cloud contamination than the surrounding areas. 

This is understandable, as the aerial content of water 

vapour is higher above forested regions. Local cloud 

convection and rainfall over those areas could also be 

observed during field work. 

Figure 7-2 shows the same information as Figure 7-1, 

but plotted temporally for the observed time steps. 

Obviously the data sets of JD 305, 313, 321 and 337 

had the best data quality, each with less than 40% 

invalid pixels (caused by cloud cover/aerosols). This 

value seems to be quite large, but needs to be seen 

in relation to the water pixels of Lake Albert. These 

water surfaces make up approximately 12% of the 

image extent and are invalid per definitionem. 

The second aspect influencing product quality is 

the accuracy of the applied land cover map. As the 

8-biome map used in the MODIS C5 algorithm is 

not available to the public, level 3 of the MOD12Q1 

land cover product – as used to produce C4 LAI data 

– is displayed instead in Figure 7-3. Theoretically 

the only difference between both biome maps is the 

distinction between evergreen and deciduous for the 

classes “broadleaf forest” and “needleleaf forest”. 

As shown in Figure 7-3, there is variation in forest 

area of the MOD12Q1 land cover product over

130 

time, which is evident since the MOD12Q1 product 

is produced on an annual basis. From 2001 on, the 

percentage of pixels classified as broadleaf forest 

increases from 13.0% to 16.9% in 2002 and to 

21.7% in 2003. For 2004 a slight decrease to 20.9% 

can be observed. A visual assessment revealed that 

correspondence between the forested areas in ASTER 

data and the MOD12Q1 product was best for the year 

2001. Whereas the extent of the main forest block is 

covered well in all MOD12Q1 land cover data sets, 

misclassifications of forest area are mainly visible 

to the south and west of the main forest block in the 

2001 

2003 

MODIS Land Cover (MOD12Q1, Type 3) 

Figure 7-3 

water 

savanna shrubs 

2002 

2004 

broadleaf crops broadleaf forest unvegetated 

Type 3 of MOD12Q1 data sets 2001-2004. An update is not available. 

years 2002-2004. These areas are densely populated 

and remaining forest patches hardly exist in reality. 

Consequently they should rather be mapped as 

“broadleaf crops”. Since an update of the 2004 land 

cover map has not been available, 2004 land cover 

data has been used also for LAI calculation from 

2005 onward. 

Since in the MODIS LAI algorithm biome specific 

LUT are used for the comparison of observed with 

modelled BRFs (cf. Chapter 4.3.3), misclassification 

of land cover results in the application of the wrong 

grasses/cereal crops


LUT to a specific pixel. Consequently resulting LAI 

values rely on erroneous assumptions. Whereas this 

does not affect the outcomes of this thesis, as these 

focus on the correctly classified forest area only, 

this aspect must be taken into account in subsequent 

regional applications. 

Also the applied algorithm has a major influence on 

product quality. As Shabanov et al. (2005) stated, the 

C4 LAI algorithm suffered from the dominance of 

poor-quality backup algorithm retrievals over woody 

vegetation. According to Yang et al. (2006) this was 

due to a mismatch between modelled and observed 

surface reflectances resulting in a failure of the main 

algorithm (cf. Equation 4.2). The backup algorithm, 

which is based on NDVI, is apparently consistent 

with the main algorithm at low NDVI. For high 

NDVI values (0.82-1) however, the backup algorithm 

is out of the retrieval domain of the main algorithm 

and reports single LAI values of 6.1. In C5 this issue 

seems to be solved, as consistency between simulated 

and measured MODIS surface reflectances is ensured 

(NASA 2007b). 

JD 289 (C5) 

JD 289 (C5) 

Applied LAI algorithm (MOD15A2) 

Figure 7-4 

physical perfect 

Quality information on the applied MODIS LAI algorithm (C5 data) for JD 289 and 329. 

131 

Figure 7-4 shows the algorithm performance for the 

Budongo Forest study site for two time steps: JD 289 

and 329. It is obvious that for both 8-day composites 

the derivation of LAI based on the “physical 

saturated” algorithm prevails for the forested regions. 

“Physical saturated” means that LAI is derived based 

on the physical model (i.e. the main algorithm), 

but under saturated conditions (i.e. increasing LAI 

influences less modelled surface reflectances). LAI of 

areas surrounding the forest is in turn mainly retrieved 

through the main algorithm without saturation 

(“physical perfect”). The major difference between 

both time steps is the amount of pixels calculated 

using the empirical method. Whereas the NDVIbased 

backup algorithm is used for 8% of pixels on 

JD 289 (16 October 2005), hardly any retrieval can 

be found for JD 329 (25 November 2005). Generally, 

overall data quality (no clouds, aerosols or shadow 

detected) was best on JD 329, 337, and 361. 

Finally, algorithm outputs must be validated. After 

reprojection to the Sinusoidal Projection, the high 

resolution LAI map presented in Figure 7-5a (cf. 

e 

also Chapter 6.3.4) is aggregated to 1 km spatial 

resolution (Figure 7-5b). Subsequently clumping 

JD 329 (C5) 

JD 329 (C5) 

physical saturated empirical not retrieved

132 

factors as noted in Table 7-1 are applied to the 

different forest stages (Figure 7-5c). Due to the 

above described quality analyses, the MODIS LAI 

data set of JD 329 (Figure 7-5d) is chosen for spatial 

validation as it had good data quality and was close to 

the main time of field work. As already noted, areas 

in the north and south west of Budongo Forest, where 

LAI is apparently underestimated, are excluded from 

e 

validation. In addition LAI from early forest stages 

LAIe, 15 m (ASTER) 


LAI, 1000 m (ASTER) 

LAI, 1000 m (ASTER) 

(effective) LAI 

is not taken into account, as λ derived from DHP is 

apparently underestimated. 

Frequency distributions for the forest areas of the 

four (effective) LAI products presented in Figure 

7-5 are displayed in Figure 7-6 together with the 

respective normal distributions. Whereas the original 

high resolution LAI map (Figure 7-6a) is clearly 

e 

following a normal distribution, deviation is obvious 

< 1 1-2 2-3 3-4 4-5 5-6 6-7 7-8 

8-9 >9 no data 



LAI, 1000 m (MODIS) 


Figure 7-5 (Effective) LAI maps for the Budongo Forest test site. a) High resolution LAI map derived from ASTER data, 

e 

b) same map aggregated to 1000 m spatial resolution, c) clumping correction applied to b) and d) MODIS LAI 

product of JD 329.


for the aggregated high resolution LAI e map (Figure 

7-6b). Yet the low LAI e at the forest edges are rather 

an artefact of the aggregation process. The effect is the 

same for the clumping corrected LAI shown in Figure 

7-6c as it is based on the aggregated high resolution 

LAI map. For the following analyses, and to avoid 

e 

errors introduced by different spatial resolution, 

forest areas with a distance of less than 1000 m to the 

forest border are consequently not taken into account. 

In contrast to Figures 7-6a to c, MODIS LAI data (cf. 

Figure 7-6d) shows two local maxima in its frequency 

distribution. Although minimum values and mean 

values are comparable to the original high resolution 

LAI map, there is a clear dominance of LAI values 

e 

Figure 7-6 Frequency distribution of (effective) LAI values derived from a) ASTER LAIe 

(15 m spatial resolution), b) ASTER LAIe (1000 m), c) ASTER LAI (1000 m) and d) MODIS LAI (1000 m) data sets. For each histogram the normal distribution 

is indicated. 

133 

between 5.5 and 6.0 LAI derived from ASTER is in 

e 

turn more continuous. Interestingly, the MODIS LAI 

product does not capture higher LAI values (>6) in 

the centre of the forest, but on the other hand records 

higher LAI values for the Cynometra-dominated 

areas in the southwest of Budongo. 

Table 7-2 displays mean (effective) LAI and standard 

deviation averaged for the different forest stages. 

As expected, mean LAI values per forest stage of 

e 

original and aggregated high resolution map are 

almost identical, with standard deviations being 

slightly higher for the 15 m spatial resolution map. 

In turn, MODIS LAI values underestimate ASTER

134 

derived LAI by ~3 for intermediate and late forest 

stages. Obviously MODIS LAI is closer to ASTER 

derived LAI than to LAI, which is corrected for 

e 

clumping. 

For a more detailed comparison of the different LAI 

products a patch-by-patch comparison is performed. 

Mean (effective) LAI and standard deviation are 

computed for each forest compartment included in 

the analysis (N=48). With regards to the polygons 

lying at forest edges only those areas were included 

that were more than 1000 m away from the forest 

border. Figure A-9 shows that mean LAI (ASTER) 

e 

in 15 m and 1000 m spatial resolution derived for the 

respective polygons vary slightly due to the different 

scales. Consequently only LAI products aggregated 

to 1000 m are included in the validation process. 

Table 7-2 

Figure 7-7 shows the resulting scatter plots of 

regressions between the MODIS LAI product of JD 

329 and ASTER derived LAI and LAI, respectively. 

e 

Obviously MODIS LAI is closer to LAI than LAI 

e 

(dashed lines represent the 1:1 relationship). Whereas 

the relationship between MODIS LAI and ASTER 

LAI is characterized by an accuracy of 0.53 (relative 

e 

accuracy of 9%, cf. Figure 7-7a), ASTER LAI is 

poorly represented by MODIS LAI with an RMSE of 

3.3 (relative accuracy of 60%, cf. Figure 7-7b). This 

can mainly be attributed to the large bias of the latter, 

which is -3.28 with respect to ASTER LAI. Although 

MODIS LAI also underestimates LAI , the respective 

e 

bias is much lower at -0.35 (corresponding to relative 

biases of -38% and -6% respectively). The fact that 

both regression models yield an R2 of 0.64 puts the 

significance of this measure into perspective. 

Mean (and standard deviation) of (effective) LAI derived from ASTER and MODIS for different forest stages. 

Product Early stage Intermediate stage Late stage 

LAI e , 15 m (ASTER) 4.6 (±1.5) 6.1 (±0.9) 5.9 (±0.9) 

LAI e , 1000 m (ASTER) 4.6 (±0.9) 6.0 (±0.7) 5.9 (±0.4) 

LAI, 1000 m (ASTER) n/a 8.8 (±1.5) 8.7 (±0.6) 

LAI, 1000 m (MODIS) 4.4 (±1.6) 5.8 (±0.2) 5.6 (±0.2) 

OLS: R²=0.64 

Y=-0.577+1.032 X 

OLS: R²=0.64 

Y=-0.577+0.691 X 

Figure 7-7 Scatter plots showing the regression between a) ASTER LAI (1000 m) and the C5 MODIS LAI product and 

e 

b) ASTER LAI (1000 m) and the MODIS LAI product for those forest compartments included in the analysis (N=48). 

The dashed line represents the 1:1 relationship, the solid line OLS regression.


The MODIS LAI product thus represents LAI e of 

Budongo Forest with a relative accuracy of 9%. This 

corresponds to the accuracy of the high resolution 

LAI map and is only a slight degradation in accuracy 

e 

compared to in situ measurements and satellite 

observations. 

Although a validation of the (outdated) C4 MODIS 

LAI algorithm was not performed, quality information 

on the applied algorithm as well as the final LAI 

product for JD 329 (2005) is shown for comparison in 

Figure 7-8. The main difference is a higher retrieval 

rate of the physical perfect algorithm, but a resulting 

lower LAI quality. The forested area appears to be 

less well outlined, with some unreasonably low LAI 

values resulting from the physical perfect algorithm 

QC, JD 329 (C4) 

QC, JD 329 (C4) 

Figure 7-8 

Table 7-3 



Quality information on the applied MODIS LAI algorithm (C4 data) for JD 329 and the respective MODIS LAI product. 

For legends please refer to Figures 7-4 and 7-5. 

Relative frequency of C4 and C5 LAI algorithm usage for the Budongo Forest test site. 

Collection Applied algorithm 2000 2001 2002 2003 2004 2005 2006 Average 

C4 Physical perfect 37% 36% 46% 37% 31% 47% 45% 40% 

Physical saturated 17% 17% 22% 25% 15% 21% 18% 19% 

Empirical 26% 31% 29% 32% 28% 23% 34% 29% 

Not retrieved 22% 16% 3% 6% 26% 9% 2% 12% 

C5 Physical perfect 22% 25% 26% 29% 28% 23% n/a 24% 

Physical saturated 44% 50% 53% 51% 55% 65% n/a 50% 

Empirical 14% 20% 20% 17% 16% 12% n/a 17% 

Not retrieved 20% 4% 2% 7% 0% 17% n/a 9% 

135 

within the main forest area. Retrievals from the 

physical saturated algorithm on the other hand 

capture mean LAI of the forest area quite well with 

LAI values comparable to those of the high resolution 

LAI map. e 

Temporal consistency 

In order to evaluate the temporal consistency of 

MODIS LAI data, the generated time series are 

further analysed with respect to data quality. A 

comparison between MODIS LAI products of C4 (cf. 

Figure A-10) and C5 (cf. Figure 7-9) revealed major 

differences. Note that only the forest area (without 

surroundings) was included in this analysis.

136 

2004 2005 

Julian Day 

Figure 7-9 

1 

33 

65 

97 

129 

161 

193 

225 

257 

289 

321 

353 

17 

49 

81 

113 

145 

177 

209 

241 

273 

305 

337 

0 

10 

20 

30 

40 

not retrieved 

empirical 

physical saturated 


relative frequency in [%] 

50 

60 

70 

80 

90 

Relative frequencies of C5 MODIS LAI algorithm 

usage for the years 2000-2005 over the Budongo 

Forest test site. 

100 

2000 2001 2002 2003 

Julian Day 

1 

33 

65 

97 

129 

161 

193 

225 

257 

289 

321 

353 

17 

49 

81 

113 

145 

177 

209 

241 

273 

305 

337 

1 

33 

65 

97 

129 

161 

193 

225 

257 

289 

321 

353 

17 

49 

81 

113 

145 

177 

209 

241 

273 

305 

337 

0 

10 

20 

relative frequency in [%] 

30 

40 

50 

60 

70 

80 

90 

100


Due to algorithm changes the percentage of pixels 

with the highest algorithm quality “physical perfect” 

is clearly reduced from 40% in C4 to 24% in C5 on 

average (cf. Table 7-3). In contrast the amount of 

pixels retrieved with the physical algorithm under 

supposedly saturated conditions increases from 19% 

to 50% (cf. Figure 7-4). As this algorithm seems to 

perform better over forested areas than the physical 

perfect method, this is a clear increase in quality. Also 

the fact that in C5 fewer pixels are retrieved with the 

empirical algorithm or not retrieved at all is resulting 

in higher overall quality of the product. Interestingly, 

the data sets that could not be retrieved at all differ in 

C4 and C5 with respect to JD (i.e. the data set of JD 

1 in 2001 could not be retrieved at all in C4, but in 

C5 it is available). This can only be due to processing 

problems, not input data quality. 

As expected there is also a dynamic in the quality 

data corresponding to the annual time course: for 

both C4 and C5 data the highest data quality is 

usually retrieved in the dry season (beginning and 

end of year). With respect to C5 data, in some time 

steps all pixels could be retrieved using the physical 

method (either under saturation or not). During the 

rainy seasons, the amount of pixels processed using 

the empirical method clearly increases. 

Figure 7-10 

Mean MODIS LAI retrieved for a 5x5 pixel window in the centre of the Budongo Forest test site for the year 2005. Red 

indicates original and blue interpolated data, solid lines represent mean LAI, dashed lines respective standard 

deviations. Note that slight deviations between original and interpolated LAI for a certain time step are due to 

averaging over the 5x5 window. 

137 

However, the applied algorithm has a strong effect 

on the resulting LAI and thus on data quality. Figure 

7-10 displays an annual LAI trajectory of the year 

2005 derived from MODIS data for a 5x5 pixel area 

in the centre of Budongo Forest. Red lines indicate 

the original data, with the mean displayed as a solid 

and the standard deviations displayed as dashed lines. 

Over the year, several extreme minima are visible 

for different time steps that correspond to aerosol or 

could detection over the forest areas and mostly to the 

subsequent usage of the empirical algorithm. In both 

cases unrealistically low LAI values are produced. 

It becomes quite clear that reliable LAI values can 

only be derived for very few time steps annually. 

Though seasonal variation in LAI exists in tropical 

rain forests, this biophysical parameter is more stable 

than in other ecosystems. Consequently temporal 

interpolation may be applied. For the analysis, all 

pixels with bad data quality (i.e. processed with 

the empirical algorithm or not processed at all, and 

those in which aerosols, clouds or cirrus clouds were 

detected) are excluded. To fill the data gaps, linear 

interpolation is performed with TiSeG (Colditz et 

al. 2008) as indicated in Figure 7-10 (blue line). The 

resulting LAI trajectory is much more stable, but is – 

in the worst case for this test site – only based on 7 

valid (out of 46) time steps.

138 

In order to evaluate the temporal signature of the 

MODIS LAI product, Figure 7-11 shows the resulting 

LAI curves plotted together with the respective 

standard deviations for the different forest stages 

and averaged over the years 2000-2005. Whereas 

a seasonal variation is clearly visible for the early 

forest stage with LAI maxima of up to 4.7 during the 

rainy season and minima of 2.4 in the dry season, 

LAI curves for intermediate and late forest stages 

remain more stable over the year. For both a range of 

0.4 could be observed with minima at the end of the 

dry season and maxima at the end of the rainy season. 

It is further obvious that mean LAI of intermediate 

and late forest stages are quite similar with respect to 

their large standard deviations (6.1±1.5 and 5.9±2.2), 

so that a distinction of forest stages based on LAI is 

not possible. This emphasises once more the results 

retrieved in Chapter 5.3.1. 

7.3.2 


For the spatial validation of MODIS LAI for the 

Kakamega Forest test site the quality of input data to 

the MODIS LAI algorithm was, once again, analysed 

first. Here the overview figures refer to the area 

shown in Figure 2-11, the product comparison (cf. 

Figure 7-15) is given for the subset of Figure 6-20. 

Figure 7-12 displays the quality information of 

C5 MODIS LAI of JD 281-361 (8 October to 

27 December) of the year 2004 with respect to 

cloud cover. Although the time period taken into 

consideration is almost the same for both test sites, 

the overall data quality here is slightly better than 

for Budongo Forest. This might be attributed to the 

fact that Budongo Forest receives 30% more rainfall 

on average during November and thus probably 

experiences higher cloud cover (cf. Figures 2-3 and 

2-13). Similar to Budongo Forest, the forested areas 

in Kakamega Forest are more prone to cloud cover 

than surrounding areas. 

Figure 7-11 

Figure 7-12 

Annual MODIS LAI trajectories for the different 

forest stages of Budongo Forest, averaged over 

the years 2000-2005. Light green indicates early, 

dark green late forest stages. Means are plotted 

as solid lines, standard deviations as dashed lines. 

Quality information of the MOD15A2 data sets 

JD 281-361 2004. Invalid pixels are either 

processed with the empirical algorithm or clouds/ 

aerosols are detected. 

The better data quality is also obvious in Figure 7-13. 

Only JD 297 has more than 40% invalid pixels. Time 

steps 281, 329 and 361 show the best data quality 

corresponding to the least cloud coverage. Since the 

data set of JD 281 was acquired shortly before the 

beginning of field work in Kakamega Forest at the 

end of the rainy season, it was chosen for product 

validation. At this time, LAI is assumed to be at its


yearly peak. By contrast, a slight decrease in LAI 

is supposed to correspond to the dry season, which 

begins around the end of November. 

The class “broadleaf forest” of the MOD12Q1 land 

cover map (cf. Figure 7-14a) corresponds well to the 

forested areas in the SPOT image, with some minor 

misclassifications towards the west and south of the 

main forest block. As for Budongo Forest, LAI of 

Figure 7-13 

2004 

Percentage of invalid (i.e. cloud contaminated) pixels for JD 281 to 361 in 2004 for the Kakamega Forest test site. 

MODIS Land Cover (MOD12Q1, Type 3) 

Figure 7-14 

broadleaf forest 

broadleaf crops 

savanna 

139 

those areas was mainly retrieved with the physical 

saturated algorithm (cf. Figure 7-14b). Outside the 

forest the physical perfect method prevails. 

LAI products are displayed in Figure 7-15 with the 

high resolution LA map (based on SPOT data) and 

e 

its aggregated version above and the aggregated map 

corrected for foliage clumping as well as C5 MODIS 

LAI data JD 281 (2004) below. Mean MODIS LAI 

JD 281 (C5) 

JD 281 (C5) 

Applied LAI algorithm (MOD15A2) 


grasses/cereal crops unvegetated 

empirical 


not retrieved 

a) Type 3 of MODIS land cover (MOD12Q1) of 2004 and b) MODIS LAI algorithm applied to the data set JD 281 

(2005) for the same area.

140 

effective LAI (SPOT), 20 m 


LAI (SPOT), 1000 m 

LAI (SPOT), 1000 m 

(effective) LAI 

Figure 7-15 (Effective) LAI maps for the Kakamega Forest test site. a) High resolution LAI map derived from SPOT data, b) same 

e 

map aggregated to 1000 m spatial resolution, c) clumping correction applied to b) and d) MODIS LAI product of 

JD 281 (2004). 



LAI (MODIS), 1000 m 

LAI (MODIS), 1000 m 

< 1 1-2 2-3 3-4 4-5 5-6 6-7 7-8 

8-9 >9 no data


of the forest area is slightly higher than in Budongo 

Forest. Similar to Budongo Forest, forest areas with a 

distance of less than 1000 m to the forest border are 

not considered in the validation process. 

The frequency distributions of the maps are given 

in Figure 7-16 together with their normal distributions. 

As in Budongo Forest, the original high 

resolution LAI map (Figure 7-15a) follows a normal 

e 

distribution, yet with some extremely high LAIe values on the upper end of the range. This is due 

to the aforementioned plantation areas, where the 

applied regression model is not valid (cf. Chapter 

6.3.2). These areas will be excluded from spatial 

Figure 7-16 Frequency distribution of (effective) LAI values derived from a) SPOT LAIe 

(20 m spatial resolution), b) SPOT LAIe (1000 m), c) SPOT LAI (1000 m) and d) MODIS LAI (1000 m) data sets. For each histogram the normal distribution 

is indicated. 

141 

validation. Additionally, a number of relatively low 

LAI values between 0 and 2 can be observed due to 

e 

forest glades and grasslands. The aggregated LAIe and the clumping corrected LAI maps generally show 

a similar frequency distribution. LAI derived from 

MODIS data, however, again has a relatively low 

variation of LAI values for the forested areas and is 

concentrated between 5.5 and 7. LAI values between 

1 and 4 can also be observed for more fragmented 

areas and forest glades. For the following analyses 

only the area north of Isecheno Forest Station 

is included in order to exclude plantation areas. 

Additionally a buffer area of 1000 m to the forest 

border is not taken into account for further analyses.

142 

Table 7-4 displays mean LAI (LAI e ) and standard 

deviations for the northern part of the main forest 

block. Different forest stages are not distinguished 

here, as - in contrast to Budongo Forest - no larger 

areas (compartments) belonging to the same stage are 

present. Disturbances of near-natural forest are rather 

local, but they are also dispersed over the whole 

forest, so that an upscaling of the existing 30 m land 

cover classification (cf. Figure A-2) to 1000 m spatial 

resolution would involve scaling errors. 

Once more mean LAI e values for the original and 

aggregated high resolution map are identical. Their 

standard deviations are comparable to those derived 

for Budongo Forest. Mean LAI is lower than in 

e 

Budongo as some forest glades with comparably low 

Table 7-4 

Mean (and standard deviation) of (effective) 

LAI derived from SPOT and MODIS for the 

northern part of Kakamega Forest. 

Product Mean (effective) LAI 

LAI e , 20 m (SPOT) 5.1 (±0.8) 

LAI e , 1000 m ( SPOT ) 5.1 (±0.8) 

LAI, 1000 m ( SPOT ) 8.0 (±1.3) 

LAI, 1000 m (MODIS) 6.4 (±0.3) 

OLS: R²=0.11 

Y=0.705-0.017 X 

LAI are included. In contrary to Budongo Forest, 

e 

MODIS LAI is higher than LAI . e 

In addition a patch-by-patch comparison is performed 

for Kakamega Forest. Consequently 15 polygons 

are distributed randomly over the northern part of 

Kakamega Forest and mean and standard deviation 

are computed for all (effective) LAI products at 

1000 m spatial resolution. Figure 7-17 shows the 

resulting scatter plots of regressions between the 

MODIS LAI product of JD 281 and SPOT derived 

LAI and LAI. As already indicated in Figure 7-16 

e 

and in the low standard deviation of the MODIS LAI 

product given in Table 7-4, the variability of MODIS 

LAI is very low compared to the SPOT derived 

products (dashed lines represent the 1:1 relationship). 

Consequently MODIS LAI is characterized by a 

comparatively low accuracy of 25% (RMSE of 1.5) 

with respect to SPOT LAI and an accuracy of 33% 

e 

(RMSE of 2.0) for SPOT LAI respectively. The biases 

are 1.2 and -1.6 (corresponding to relative biases of 

18% and -26% respectively). 

The MODIS LAI product thus represents neither 

LAI nor LAI of Kakamega Forest with satisfactory 

e 

accuracy. Compared to the relative accuracy of 16% 

OLS: R²=0.11 

Y=0.705-0.011 X 

Figure 7-17 Scatter plots showing the regression between a) SPOT LAI (1000 m) and the C5 MODIS LAI product and 

e 

b) SPOT LAI (1000 m) and the MODIS LAI product for those forest compartments included in the analysis (N=15). 

The dashed line represents the 1:1 relationship.


of the high resolution LAI e map presented in Chapter 

6.3.2, this is a further degradation in accuracy that 

might partly be attributed to the bad in situ data 

quality. 

Figure 7-18 shows the resulting LAI curve plotted 

together with its standard deviation for the northern 

part of Kakamega Forest and averaged over 

the years 2000-2005. Once more the LAI curve 

remains relatively stable over the year with its 

standard deviation being slightly higher than that 

of intermediate and late forest stages in Budongo 

Forest. This is obviously the result of the missing 

subdivision in forest stages for Kakamega Forest. 

However mean LAI is comparable to that retrieved 

for Budongo Forest. 

Figure 7-18 

7.4 

Annual MODIS LAI trajectory for the northern 

part of Kakamega Forest averaged over the 

years 2000-2005. Mean LAI is plotted as solid 

line, standard deviations as dashed lines. 

Conclusion 

The preceding chapter completes the validation of the 

MODIS LAI product for both Budongo and Kakamega 

Forests. Based on the analyses and correction of in situ 

data presented in Chapter 5 and the high resolution 

LAIe maps created in Chapter 6, two explicit MODIS 

LAI products could be validated: JD 329 for the year 

143 

2005 for Budongo Forest and JD 281 for 2004 for 

Kakamega Forest. The products were chosen due to 

their good input data quality in terms of cloud cover 

and algorithm performance. Obviously cloud cover 

has a major influence on the applied algorithm and 

on product quality as a consequence. This means 

that the derived product accuracies are only valid for 

the considered time steps and quality levels. LAI of 

forested areas was mostly derived with the “physical 

saturated algorithm”, which gave a good performance 

over the observed broadleaf forests. The new LUTs 

associated to the C5 LAI algorithm seem to have a 

good correspondence to observed surface reflectance, 

even though LAI retrievals over broadleaf forests are 

mostly performed over a small area in the spectral 

space (Shabanov et al. 2005). Yet it remains unclear, 

why the main algorithm is not always performed in 

the saturation domain over broadleaf forest areas. 

Temporal LAI trajectories further revealed that the 

back-up algorithm has major problems in terms of 

LAI stability. Sufficient stability can only be derived 

when pixels with deficient data quality are excluded 

and temporal interpolation is performed. 

Overall the spatial validation for the Budongo 

Forest test site revealed a product accuracy of 0.53 

corresponding to a relative accuracy of 9%, which 

is exactly the same as the relative accuracy of the 

resulting high resolution LAI map of Budongo 

e 

Forest. It can thus be concluded that MODIS LAI 

well represents (up-scaled) in situ LAI on a regional 

level. However, it must be emphasised that product 

accuracy was determined with respect to LAI and not 

e 

LAI. The latter is underestimated by MODIS LAI, 

with a bias of -3.28 with respect to ASTER LAI. 

For Kakamega Forest MODIS LAI product validation 

led to a comparatively low accuracy of 1.5 

with respect to LAI , which corresponds to a relative 

e 

accuracy of 25%. In this case MODIS LAI is higher 

than LAI derived from upscaled field data (bias of 

e 

1.2) and lower than reference LAI (bias of -1.6).

144 

However, this is mainly a result of the inferior in 

situ data quality as described in Chapter 5.3.2, the 

resulting degradation of accuracy in the upscaling 

process and the low relative accuracy of the produced 

high resolution LAI map with 16%. This problem 

e 

emphasises once more that accurate field data is 

necessary for validation. Validation results for 

Kakamega Forest must thus be handled with care. 

Interestingly mean MODIS LAI retrieved for the 

forest areas of two study sites and the respective 

validation time steps differ quite a bit. Whereas mean 

MODIS LAI for intermediate and late forest stages 

on JD 329 in Budongo Forest is 5.8 (±0.2) and 5.6 

(±0.2) respectively, mean MODIS LAI on JD 281 for 

the northern part of Kakamega Forest is 6.4 (±0.3). 

This is, however, the result of seasonal LAI variation 

rather than a true difference of leaf area between 

the two forests as the validation time steps are more 

than 1 ½ months apart. If mean LAI for JD 281 and 

329 for the respective years (2004 and 2005) are 

compared directly, differences become inessential. 

For JD 281 (2004) mean LAI of intermediate and late 

forest stages in Budongo Forest are 6.3 (±0.4) and 6.2 

(±0.4) respectively. Mean LAI for Kakamega Forest 

for JD 329 (2005) is calculated as 5.7(±0.5). This 

corresponds to seasonal climatic variations (onset 

of dry season) and suggests that even in the case of 

tropical rain forest stands - which are supposed to 

remain relatively stable over the year - the time of 

field work should not exceed a few weeks in transition 

seasons (unless seasonal variability is expressly 

to be monitored). Further it should coincide as far 

as possible with the acquisition of high resolution 

satellite data used for upscaling and acquisition of the 

validated product. 

Temporal analyses of the MODIS LAI product 

for Budongo Forest further revealed that LAI 

at the early forest stage corresponds strongly to 

interannual climate changes with a range in LAI of 

2.3 over the year. LAI of mature forest sites shows a 

comparatively low annual LAI variability with 0.4 for 

intermediate and late forest stages. This corresponds 

to the seasonal change in LAI of 0.34 found by De 

Wasseige et al. (2003) for a semi-deciduous forest in 

the CAR. 

In contrast to the validation of the C4 MODIS LAI 

product for a tropical rain forest in Brazil published 

by Aragão et al. (2006), some major improvements 

can be observed for the C5 MODIS LAI products 

analysed in this thesis. Whereas Aragão et al. (2006) 

found a predominance of the backup algorithm 

over the forested region due to a failure of the main 

algorithm over vegetation areas with high NDVI 

values, this issue can be considered as resolved for 

C5. The majority of pixels representing the forest 

areas of Budongo and Kakamega Forest are now 

generated with the “physical saturated” algorithm, 

at least under cloud free conditions. If clouds are 

present, the back-up algorithm prevails - owing to a 

failure of the main algorithm – but in this case the 

respective pixels should be excluded from further 

analysis anyway. The general overprediction of 

MODIS LAI with respect to field LAI of 1-2 found by 

Aragão et al. (2006) cannot be confirmed for the test 

sites in this thesis (note that their field measurement 

also refer to LAI ). e 

The sensitivity of the MODIS LAI product to 

structural changes must, however, be viewed 

critically. Temporal analyses mainly confirmed 

the results of monotemporal field data: although 

structural differences are present in intermediate and 

late forest stages, the corresponding LAI is relatively 

similar. Internal LAI variations in the respective 

forest stages exceed by far the differences to mean 

LAI of the other forest stages. However this does not 

necessarily imply that structural differences of forest 

stages due to degradation and human disturbances 

cannot be monitored with the MODIS LAI. However 

the perturbation of LAI must be strong enough that it 

both affects in situ and remotely measured LAI data.

8 Discussion and Outlook 

As discussed in the introduction of this thesis, 

there has been a clear lack of validation sites for 

the MODIS LAI product in tropical rain forest 

environments. Especially in Africa, no test sites 

have been available so far. Consequently, this thesis 

dealt with the validation of the MODIS LAI product 

for two test sites in East Africa, Budongo Forest in 

Uganda and Kakamega Forest in Kenya. Both sites 

were accessible within the framework of the BIOTA 

East Africa project. In the following, the main results 

of this thesis will be summarised and discussed with 

respect to their contribution to current scientific 

discussions. An outlook concerning further research 

questions is also given. 

The first part of the thesis dealt with ground-based 

measurements of LAI. The establishment of a valid 

and representative sampling scheme is an issue that has 

often not been sufficiently addressed in many studies. 

However it is of major importance for the accuracy 

and representativeness of in situ measurements, 

especially if the gathered data is subsequently used 

for large scale validation efforts together with field 

data from other studies. CEOS-LPV and VALERI 

guidelines were found to be very helpful here, yet 

an adaptation to the tropical rain forest environment 

was still required. Especially constraints related to 

geolocation errors in satellite and field data as well 

as to canopy structure in tropical rain forests were 

looked at. 

145 

The methods proposed by CEOS-LPV and VALERI 

are well suited for study sites that are characterised 

by a structurally homogeneous canopy. Internal 

variability of the vegetation here occurs at scales that 

are much smaller than the ESU area (usually defined 

by the pixel size of the high resolution satellite 

data used for upscaling). In tropical rain forests, 

however, canopy structural variability can only be 

captured at larger scales. As further constraints are 

put on the geolocation of ESUs and geocorrection of 

satellite data in this environment, ESU size had to be 

significantly enlarged. In addition, sampling distance 

was increased to guarantee a spatially independent 

field database. Interestingly, these aspects have so 

far not been addressed in CEOS-LPV and VALERI 

recommendations, although validation work thus 

far comprised studies also in temperate and boreal 

forests. 

For the standardisation of validation efforts – which 

is the goal of CEOS-LPV – appropriate sampling 

strategies should thus be proposed for different 

groups of validation sites. The outcome of this thesis 

suggests that these groups should be characterized by 

similar land cover properties (such as canopy height) 

and scale of internal vegetation variability. These, in 

turn, define appropriate ESU sizes, sampling schemes 

and the number of individual measurements. This 

standardisation is needed to assure valid sampling 

approaches for all studies and to facilitate the eventual 

comparison of validation results. As a consequence, 

“best practice” protocols for data collection and 

description should be developed. With respect to 

ESU size these could also be transferred to other in 

situ methods, only the distance between individual 

sampling points depends on the used instruments. 

Concerning in situ methods of LAI estimation, a 

review revealed that only indirect optical instruments 

were applicable for field data sampling on the two 

study sites. Other methods were either not transferable 

(e.g. leaf litter collection) or too time-consuming,

146 

with the result that a representative amount of ESUs 

could not have been sampled in the available time 

frame. Consequently, measurements were conducted 

with the LAI-2000 PCA and, for comparison, with 

DHPs. These instruments have been rarely applied 

to tropical rain forests before, so that the effect of 

non-ideal measurement conditions (e.g. no diffuse 

radiation present over longer time periods and high 

vegetation stands resulting in mixed pixels) were not 

known. A thorough analysis of the results of this thesis 

led a) to the development of a novel sampling scheme 

for LAI-2000 PCA data and b) to a quantification of 

the measurement precision of both instruments. 

With respect to LAI-2000 PCA, three devices were 

deployed to monitor the influence of changing 

illumination conditions and direct radiation on the 

measurements. For the latter, a correction scheme 

could be developed, so that LAI-2000 PCA devices 

may now also be used for large test sites under 

conditions in which indirect radiation is hardly 

present over the time period needed to conduct the 

measurements. As the influence of direct radiation 

led to an underestimation of LAI of up to 

e (LAI2000) 

10% for the ESUs measured in this thesis, this effect 

should not be disregarded. Although several authors 

have discussed this problem before (De Wasseige 

et al. 2003, Leblanc & Chen 2001), a comparable 

correction scheme had not been available prior 

to this. Based on data of the third LAI-2000 PCA 

device measurement precision for each ESU could be 

quantified (up to 10% under non-ideal conditions). 

The measurement precision of DHPs was determined 

at 5%. Whereas DHPs have the advantage of lower 

instrument cost and the availability of a permanent 

illustration of the sampling site, LAI-2000 PCA 

measurements offer continuous monitoring of light 

intensity (in the case of A and B1 sensors) and thus 

allow a more accurate determination of measurement 

precision if the recommendations for instrument setup 

from this thesis are followed. Especially under 

completely direct or diffuse illumination conditions, 

the relative measurement precision can be even be 

higher than that of DHPs (cf. Table 6-2). 

A comparison of the LAI values resulting from 

both instruments revealed the advantages and 

disadvantages of both approaches. Whereas 

understorey vegetation, which comprises up to 14% 

of site-specific LAI at the two test sites considered, 

e 

may be included in DHP, the analysis and processing 

of the photographs requires much more time than 

the analysis of LAI-2000 PCA measurements. On 

the other hand, the set-up of three LAI-2000 PCA 

devices in a tropical rainforest (with one of them 

being placed in a nearby forest gap) requires greater 

logistical effort. Although LAI values retrieved 

e 

from both instruments seem reasonable, no general 

conclusion can be made to determine which method 

leads to more accurate results. More studies like the 

one described by Clark et al. (2008) for an old-growth 

tropical rain forest around La Selva Biological Station 

in Costa Rica are needed to compare the results of 

indirect optical measurements to destructive ones. 

Although the effort of setting up modular towers and 

the accomplishment of direct measurements on more 

than 55 vertical transects can only be undertaken 

as part of a larger project network, a combination 

of indirect measurements before harvesting and a 

comparison to the results of direct measurements 

afterwards would improve the understanding of the 

relationship between LAI and LAI in tropical rain 

e 

forests. Unfortunately indirect measurements were 

not acquired in La Selva. 

A comparison of the results retrieved by Clark et 

al. (2008) and the outcomes of this thesis however 

indicate a good correspondence. Mean LAIe (LAI2000) 

and LAI values calculated for the intermediate 

e (DHP) 

and late forest stages of Budongo Forest are – at 6.14 

and 6.25 (intermediate) and 6.35 and 6.14 (late) – 

only slightly higher than mean landscape LAI of 6.00 

derived from direct measurements in the old-growth


tropical rain forest around La Selva Biological 

Station. Mean LAI and LAI values for 


Kakamega Forest were lower, at 5.13 and 4.07 for 

intermediate and 5.88 and 4.17 for late forest stages, 

respectively. This should, however, be interpreted 

as a result of inferior in situ data quality (as e.g. 

understorey vegetation was not included for LAIe (DHP) 

for the reasons discussed in Chapter 5), rather than a 

true difference in LAI of the two forests. 

e 

This comparison with the results of Clark et al. (2008) 

also puts the attempt to derive true LAI values through 

correction for foliage clumping into perspective. 

Although the clumping factors derived from DHP 

processing are in line with the findings of Chen et al. 

(2005) except for early forest stages, the clumping 

correction was applied without an attempt to correct the 

overestimation of LAI through non-foliage elements. 

In this context several authors state that, depending 

on the used instruments and applied corrections, 

LAI can be very close to true LAI (Eriksson et al. 

e 

2005, Planchais and Pontailler 1999). This might 

e.g. be attributed to the fact that the correction for 

foliage clumping and the correction for the influence 

of woody area may cancel each other out. Yet for the 

latter, no satisfactory correction approach exists so 

far. First, the sometimes applied calculation of WAI 

through DHP or LAI-2000 PCA measurements during 

leafless times in deciduous forests is not transferable 

to (semi) evergreen rain forests; second the calculation 

overestimates the real contribution of WAI to PAI as 

woody elements are at least partly covered by foliage 

during the vegetation period. Therefore, further efforts 

to develop the correct estimation of clumping factors 

and especially the quantification of WAI are needed. 

Approaches that incorporate the possibilities of nearinfrared 

photography as e.g. proposed by Kucharik et 

al. (1998) and Chapman (2007) have a large potential 

in this context and should be further advanced. 

It can be concluded that both instruments, LAI-2000 

PCA and DHP, are applicable to the tropical forest 

147 

environment if the recommendations concerning 

instrument set-up and the avoidance of certain 

radiation conditions are followed. This emphasises 

once again the need for a carefully planned field 

campaign, as severe errors in the data basis have an 

effect on all further steps in the validation process. 

This problem is currently only addressed in a very 

small number of studies (e.g. Cohen et al. 2003, 

Huang et al. 2006, Tan et al. 2005). Nevertheless, the 

results from in situ measurements described in this 

thesis provide the first quantitative ground-based 

assessment of LAI in semi-deciduous rain forests in 

East Africa. 

The second part of this thesis describes the generation 

of high resolution LAI maps based on in situ 

e 

measurements and high resolution satellite data from 

ASTER and SPOT. Acknowledging that errors are still 

present in the data sets, a robust regression method 

was sought in order to retrieve stable regression 

models. Theil-Sen regression, as introduced by 

Fernandes et al. (2005) fulfilled these needs, so that 

transfer functions could be established for both study 

sites. Once more it must be emphasised that OLS 

regression, which is widely applied in remote sensing 

studies, may lead to biased results in the presence of 

measurement errors. This aspect should also be taken 

into account when recommendations concerning the 

validation of satellite products are to be made. 

For Budongo Forest, where better quality of in situ 

data could be retrieved and more detailed ancillary 

information was available, a thorough analysis 

revealed that different transfer functions must be 

established for different forest stages. Although 

mean LAI retrieved for intermediate and late forest 

e 

stages was similar, structural differences mainly in 

the upper canopy led to significantly different surface 

reflectances. Especially for late forest stages, a more 

complex and thus very uneven upper canopy led to 

reduced surface reflectances in the NIR and SWIR 

bands compared to those for intermediate forest

148 

stages. Turner et al (1999) states in this context, that 

early overflight times in the morning promote this 

effect, since shadows are still perceptible owing to 

the large θ . sun 

Depending on the structural differences of forest 

stages, either SVIs or texture variables were better 

suited to model LAI from high resolution satellite 

e 

data. In Budongo Forest, SR performed best for early 

and intermediate forest stages with an R2 of 0.94. 

Generally SR seems to be better suited to model LAI 

of high biomass stands than NDVI, as it has a better 

dynamic range under these conditions. However, 

the variability of SVIs was comparably low due to 

high internal heterogeneity in forests; this was also 

observed by Colombo et al. (2003). In contrast to 

the latter’s study, multilinear regression did not yield 

higher R2 . For late forest stages a texture measure 

was better suited to model LAI . The GLCM variance 

e 

of ASTER band 4 retrieved an R2 of 0.71. Based on 

similar observations of structural differences in forest 

canopies, correlations between LAI and the shade 

e 

fraction were observed by Seed & King (2003) for 

temperate forests with a canopy closure of more than 

80% and by Aragão et al. (2006) for a primary rain 

forest in Brazil. Generally, the potential of texture 

measures applied to high and very high resolution 

optical satellite data should be further exploited. 

Transfer functions between LAI e (DHP) and satellite 

data led to better results in Budongo Forest for all 

forest stages. Though the derived relationships are 

only valid for the respective satellite scenes with 

their characteristic atmospheric conditions and 

phenological vegetation states, they transfer the field 

based point measurements to a high spatial resolution 

validation surface with the highest possible accuracy. 

Consequently, a high resolution LAI map could be 

e 

produced for Budongo Forest with a relative accuracy 

of 9%. This is only a slight degradation in accuracy 

relative to field measurements. 

For Kakamega Forest the previously mentioned 

inferior in situ data quality led to reduced quality of 

the high resolution LAI map compared to Budongo 

e 

Forest. For example, DHPs of understorey vegetation 

had not been taken in Kakamega Forest, so that 

transfer functions based on LAI generally 

e (LAI2000) 

performed better. Further, unlike Budongo, Kakamega 

Forest had not been managed in compartments. 

As a result, no dedicated areas of homogenous 

intermediate or late forest stages are perceptible. 

Forest stages are rather dispersed evenly over the 

forest. Consequently texture information did not lead 

to similar results as in Budongo. Rather a regression 

model based on RSR was used to estimate LAI for e 

the whole test site. Similar models have so far only 

been applied to coniferous, deciduous or mixed 

temperate forests (e.g. Chen et al. 2002, Huang et al 

2006, and Stenberg et al. 2004). Overall this model 

yielded an R2 of 0.53 with a relative accuracy of 

16% (RMSE of 0.8). Although these results are not 

as good as the ones retrieved for Budongo Forest, 

they are comparable to those of Cohen et al. (2006). 

In their study Landsat-based LAI maps were created 

for a moist tropical forest in Brazil with a correlation 

of 0.54 and an RMSE of 0.83. These maps were also 

used for C4 MODIS LAI product validation. 

The third part of this thesis involved the final 

validation of the MODIS LAI product for Budongo 

and Kakamega Forest based on the high resolution 

LAI maps. Here, algorithm input data was analyzed 

e 

first. As observed by Cohen et al. (2006) for their study 

area in Brazil, the MODIS land cover class “broadleaf 

forest” corresponded well to real land cover of the 

two test sites (although it must be noted once more 

that the new eight biome map as implemented in the 

C5 MODIS LAI algorithm had not been available 

and C4 land cover data was analyzed instead). 

The predominance of back-up algorithm retrievals 

as reported by Aragão et al. (2006) for C4 MODIS 

LAI data could not be confirmed in this thesis. At


least for the dry seasons and the validated time steps 

the physical saturated algorithm prevailed over the 

forested areas. This is probably the result of algorithm 

improvements in C5 (cf. NASA 2007b), which 

reduced the failure rate of the physical algorithm. 

If applied, the back-up algorithm still suffered from 

severe problems. In tropical regions probably cloud 

cover is one of the major issues for failure of the main 

algorithm. This might partly explain the bad data 

quality of back-up algorithm retrievals. To solve this 

problem, the operational production of composites 

over more than 8-days is suggested to ensure a cloudfree 

data basis even during the rainy season. Monthly 

composites of the C5 MODIS LAI product are, for 

instance, offered for download by the “Climate and 

Vegetation Research Group” at Boston University. 

However, only data sets covering the years 2000, 

2001, 2002 and 2007 are currently available. 

The spatial validation of the C5 MODIS LAI product 

for Budongo Forest was finally performed as a patchby-patch 

comparison. It revealed that the C5 MODIS 

LAI product represented the up-scaled in situ LAIe with an accuracy of 0.53. This corresponds to a 

relative accuracy of 9%, which is identical with the 

accuracy of the high resolution LAI maps. Those in 

e 

turn were generated with only a slight degradation 

of field data accuracy (taking into account the 

above-mentioned precision of 5% for LAI and 

e (DHP) 

assuming a negligible bias). These findings once 

more indicate an improvement of the C5 algorithm 

compared to C4. Aragão et al. (2006) reported, 

for instance, a significant overprediction of fieldmeasured 

LAI by the C4 MODIS LAI product for the 

previously mentioned Brazilian rain forest site. An 

overprediction of LAI in forested biomes was also 

reported by Cohen et al. (2006), who validated the C4 

MODIS LAI product for five boreal, temperate and 

tropical forest sites across the Western Hemisphere. 

By contrast, C5 MODIS LAI, as analysed within this 

thesis, underestimated the reference LAI slightly with 

e 

a bias of -0.35 (relative bias of -6%). Interestingly, 

149 

an even larger bias (-3.28) was present for reference 

LAI, which was calculated by applying the clumping 

factors derived from DHP to the high resolution LAIe maps (again no correction for WAI was performed). 

This underscores the above-mentioned assumption 

that LAI might be close to “real” LAI (note that the 

e 

other studies also compared MODIS LAI to LAI of e 

in situ measurements). 

For Kakamega Forest, the spatial validation of the 

MODIS LAI product led to a comparatively low 

accuracy of 1.5 (corresponding to a relative accuracy 

of 25%). This is most likely the result of inferior 

field data quality (as mentioned above) for this test 

site and the resulting degradation of accuracy in the 

upscaling process. Consequently, the results must 

be interpreted carefully, as they may be based on 

inaccurate assumptions. The fact that mean in situ 

LAI data for intermediate and late forest stages in 

e 

Kakamega Forest differs by up to 2 LAI units from the 

respective data sampled in Budongo Forest indicates 

the reliability (or non-reliability) of field data in 

Kakamega. Once again this problem emphasises the 

need for accurate field data (and sampling protocols) 

for validation purposes. 

Although a real temporal validation – which would 

require in situ LAI measurements over several 

months – was beyond the scope of this thesis, at 

least the temporal consistency of the MODIS LAI 

product was investigated. An analysis of time series 

for the years 2000-2005 showed that the data was 

reliable and stable, but only if temporal interpolation 

was applied to bad data quality pixels. The observed 

variability in LAI (0.4 for intermediate and late 

forest stages) further corresponds to in situ measured 

seasonal LAI trajectories found by De Wasseige et al. 

(2003) for a semi-deciduous forest in the CAR and 

by Wirth et al. (2001) for a tropical moist rainforest 

in Panama. Maximum LAI values are associated 

with the end of the rainy season, minimum LAI 

values to the dry season. It can thus be assumed

150 

that the MODIS LAI product responds correctly to 

biome-level LAI changes associated with interannual 

climate variability. 

Interestingly, the above-mentioned results contradict 

the findings of Myneni et al (2007), who discovered 

large seasonal swings in leaf area for the Amazonian 

rain forest. In their study, which analysed MODIS 

LAI data from 2000-2005, maximum leaf area was 

always observed during the dry season and decreasing 

approximately 25% during the rainy season. According 

to their reasoning, there is a correspondence between 

maximum leaf area and maximum solar radiation. To 

shed light on these relationships and to facilitate future 

temporal validation campaigns, new sensor systems 

as e.g described by Baret (2007) could be employed. 

They consist of small independent sensor units that 

record above canopy photosynthetically active 

radiation (PAR) and below canopy transmittance. 

These sensors communicate to a master unit that can 

store the collected data. According to Baret (2007) 

these sensors can be produced at low cost and are 

autonomous for up to six months. The sensors offer 

continuous monitoring of PAR balance, with LAI 

being subsequently calculated from transmitted PAR 

(analogous to LAI-2000 PCA measurements). As with 

other indirect optical methods, however, limitations 

in terms of the estimation of WAI or clumping factors 

would persist. 

To summarise, this thesis describes the first validation 

study for the MODIS LAI product in African rain 

forests. Additionally, it has validated the C5 MODIS 

LAI algorithm for the first time ever for the tropical 

forest biome. The outcomes are already expected 

to contribute to CEOS-LPV LAI inter-comparison 

efforts by Garrigues et al. (2006). Field data and 

derived high resolution LAI maps will further be 

made available within the VALERI and CEOS-LPV 

networks for the validation of biophysical products 

derived from other medium resolution satellite 

sensors. 

Although the hypothesis that subtle structural 

changes in rain forests can be monitored through 

repeated measurements of LAI could not be verified 

– as differences in LAI of intermediate and late forest 

changes were not significant – the outcomes of this 

thesis can nevertheless help to improve knowledge 

on tropical rainforests and their response to climatic 

changes and human disturbance. As the differences of 

the temporal LAI signature of intermediate and early 

forest stages in this thesis were tremendous, further 

research should be carried out to better understand 

the link between forest degradation and temporal 

signature of biophysical parameters. 

The outcomes of this thesis may further help to 

reduce predictive uncertainties in biophysical process 

models, as e.g. the CoupModel used by Miao (2008) 

to model the carbon dynamics of Kakamega Forest. 

MODIS LAI data may here serve as supplementary 

data for model calibration. Knowledge about the 

accuracy of this data set is, however, crucial, as it 

is taken as an absolute reference. Also numerical 

weather prediction models need LAI data for the 

calculation of e.g. evapotranspiration processes. 

Therefore LAI was defined as an “essential climate 

variable” by the Global Climate Observing System. 

Validation for different biomes also helps to improve 

later versions of the MODIS LAI product algorithm 

and thus leads to an increase of derivative product 

accuracy. Errors in the MODIS LAI product do not 

only extend into water and energy balance of climate 

models (atmosphere-biosphere exchange component 

of general circulation climate models), but also in 

higher level MODIS products as e.g. the MODIS 

GPP (MOD17A2) product. Further improvements 

of the LAI algorithm can be expected with the new 

version (Collection 6). According to Didan (2007) 

an optimisation of retrievals with SWIR data is 

envisaged, which might possibly lead to a better LAI 

retrieval for broadleaf evergreen forest in Amazonia 

and Central Africa.

References 

Abrams, M. and S. Hook (2002). ASTER User Handbook, Version 2. Accessed 19 October, 2007. 

. 

Abuelgasim, A.A., Fernandes, R.A. and S.G. Leblanc (2006). Evaluation of national and global LAI products 

derived from optical remote sensing instruments over Canada. – In: IEEE Transactions on Geoscience 

and Remote Sensing, 44, pp. 1872−1884. 

Alexandre, D. (1981). L’indice foliaire des forêts tropicales. Analyse bibliographique. – In: Acta Oecologica, 

Oecologia Generalis, 2 (4), pp. 299–312. 

Althof, A. (2005). Human Impact on Flora and Vegetation of Kakamega Forest, Kenya: structure, distribution 

and disturbance of plant communities in an East African rainforest. Accessed 5 September, 2007. 

. 

Anderson, M.C. (1964). Studies of the woodland climate. I. The photographic computation of light conditions. 

– In: Journal of Ecology, 52, pp. 27-41. 

Andrua, H. J. (2003). Tropical Secondary Forest Management in Africa: Reality and perspectives. Uganda 

Country Paper. – In: Proceedings of the “Workshop on Tropical Secondary Forest Management in Africa: 

Reality and Perspectives”, pp. 9-13. December 2002, Nairobi, Kenya. 

Aragão, L.E., Shimabukuro, Y.E., Espírito Santo, F.D.B. and M. Williams (2005). Landscape pattern and spatial 

variability of leaf area index in Eastern Amazonia. – In: Forest Ecology and Management, 211, pp. 240- 

256. 

Aragão, L.E., Shimabukuro, Y.E., Espírito Santo, F.D.B. and M. Williams (2006). Spatial validation of the 

collection 4 MODIS LAI product in eastern Amazonia. – In: IEEE Transactions on Geoscience and 

Remote Sensing, 43 (11), pp. 2526-2534. 

Arvidson, T., Gasch, J. and S.N. Goward (2001). Landsat 7’s long-term acquisition plan – an innovative 

approach to building a global imagery archive. – In: Remote Sensing of Environment, 78, pp. 13-26. 

Aryal, B. (2002). Are trees for the poor? A study from Budongo Forest, Uganda. Unpublished Master Thesis, 

Department of Economics and Social Science, Agricultural University of Norway, Ǻs. 

Asner, G.P., Scurlock, J.M.O. and J.A. Hicke (2003). Global synthesis of leaf area index observations: 

implications for ecological and remote sensing studies. – In: Global Ecology and Biogeography, 12, pp. 

191-205. 

Atzberger, C., Jarmer, T., Kötz, B., Schlerf, M., and W. Werner (2003). Spectroradiometric determination 

of wheat canopy biophysical variables. Comparison of several empirical-statistical methods. – In: R. 

Goossens (Ed.), Remote sensing in transition. Proceedings of the 23rd EARSEeL Symposium on Remote 

Sensing in Transition, Ghent, Belgium, 2-5 June 2003, pp. 463-470. 

Bacour, C., Baret, F., Béal, D., Weiss, M. and K. Pavageau (2006). Neural network estimation of LAI, fAPAR, 

fCover and LAI×Cab, from top of canopy MERIS reflectance data: Principles and validation. – In: 

Remote Sensing of Environment, 105, pp. 313-325. 

Bahrenberg, G., Giese, E. and J. Nipper (2003). Statistische Methoden in der Geographie Band 2. Multivariate 

Statistik. Berlin, Stuttgart. 

Baret, F. (1995). Use of spectral reflectance variation to retrieve canopy biophysical characteristics. – In: Darson, 

M. and S. Plummer (eds.). Advances in environmental remote sensing, Wiley and Sons, pp. 33-51. 

151

152 

Baret, F., Clevers, J.G.P.W. and M.D. Steven (1995). The robustness of canopy gap fraction estimates from red 

and near-infrared reflectances: a comparison of approaches. – In: Remote Sensing of Environment, 54, 

pp. 141-151. 

Baret, F. (2004a). Can-Eye software. Accessed 21 September, 2007. . 

Baret, F. (2004b). A simple method to calibrate hemispherical photographs. Accessed 21 September, 2007. 

. 

Baret, F., Morisette, J.T., Fernandes, R.A., Champeaux, J.L., Myneni, R.B., Chen, J., Plummer, S., Weiss, M., 

Bacour, C., Garrigues, S. and J.E. Nickeson (2006). Evaluation of the Representativeness of Networks 

of Sites for the Global Validation and Intercomparison of Land Biophysical Products: Proposition of 

the CEOS-BELMANIP. – In: IEEE Transactions on Geoscience and Remote Sensing, 44 (7), pp. 1794- 

1802. 

Baret, F. (2007). Sample results for LAI estimates from Hemisphotos and CAN_EYE. Accessed 2 July, 2008. 

 

Baret, F. and P. Rosello (2007). Validation of Land European Remote Sensing Instruments. Accessed 15 August, 

2007. . 

Baret, F., Hagolle, O., Geiger, B., Bicheron, P., Miras, B., Huc, M., Berthelot, B., Niño, F., Weiss, M., Samain, 

O., Roujean, J.L. and M. Leroy (2007). LAI, fAPAR and fCover CYCLOPES global products derived 

from VEGETATION Part 1: Principles of the algorithm. – In: Remote Sensing of Environment, 110 (3), 

pp. 275-286. 

Baret, F., Weiss, M., Garrigues, S., Allard, D., Leroy, M., Jeanjean, H., Fernandes, R., Myneni, R.B., Morisette, 

J.T., Privette, J., Bohbot, H., Bosseno, R., Dedieu, G., Di Bella, C., Espana, M., Gond, V., Gu, X.F., Guyon, 

D., Lelong, C., Maisongrande, P., Mougin, E., Nilson, T., Veroustraete, F. and R. Vintilla (submitted). 

VALERI: A network of sites and a methodology for the validation of medium spatial resolution satellite 

products. – In: Remote Sensing of Environment, submitted for publication. 

Barnes, W.L., Pagano, T.S. and V.V. Salomonson (1998). Prelaunch characteristics of the Moderate Resolution 

Imaging Spectroradiometer (MODIS) on EOS-AM1. – In: IEEE Transactions on Geoscience and Remote 

Sensing, 36 (4), pp. 1088-1100. 

Bartelink, H.H. (1997). Allometric relationships for biomass and leaf area of beech (Fagus sylvatica L). 

– In: Annales des Sciences Forestieres, 54, pp. 39-50. 

Beentje, H.J. (1990). The forests of Kenya. – In: Mitteilungen aus dem Institut für Allgemeine Botanik Hamburg, 

23a, pp. 265-286. 

Beentje, H.J. (1994). Kenya trees, shrubs and lianas, National Museum of Kenya, Nairobi. 

Belward, A.S. (1999). International co-operation in satellite sensor calibration; the role of the CEOS working 

group on calibration and validation. – In: Advances in Space Research, 23 (8), pp. 1443-1448. 

Berk, A., Bernstein L.S., Anderson G.P., Acharya P.K., Robertson D.C., Chetwynd J.H. and S.M. Adler-Golden 

(1998). MODTRAN cloud and multiple scattering upgrades with application to AVIRIS. – In: Remote 

Sensing of Environment, 65, pp. 367-375. 

Berterretche, M. Hudak, A.T., Cohen, W.B., Maiersperger, T.K., Gower, S.T. and J. Dungan (2005). Comparison 

of regression and geostatistical methods for mapping Leaf Area Index (LAI) with Landsat ETM+ data 

over a boreal forest. – In: Remote Sensing of Environment, 96, pp. 49-61.

References 

Bicheron, P., Leroy, M. and O. Hautecoeur (1998). LAI and fAPAR mapping at global scale by model inversion 

against spaceborne POLDER data. – In: Proceedings of IGARSS, 3, 1998, pp. 1228-1230. 

Bicheron, P. and M. Leroy (1999). A Method of Biophysical Parameter Retrieval at Global Scale by Inversion 

of a Vegetation Reflectance Model. – In: Remote Sensing of Environment, 67, pp. 251-266. 

Birth, G. S. and G. McVey (1968). Measuring the color of growing turf with a reflectance spectrophotometer. 

– In: Agronomy Journal, 60, pp. 640-643. 

Bjørlykke, K.O. (1973). Glacial Conglomerates of late Precambrian age from the Bunyoro Series, W. Uganda. 

– In: International Journal of Earth Sciences, 62 (3), pp. 938-947. 

Black, T.A., Chen, J.M., Lee, X. and R.M. Sagar (1991). Characteristics of shortwave and longwave irradiances 

under a Douglas-fir forest stand. – In: Canadian Journal of Forest Research, 21, pp. 1020-1028. 

Blackett, H.L. (1994). Forest Inventory Report No. 3: Kakamega. Kenya Indigenous Forest Conservation 

Programme (KIFCON), Nairobi, Kenya. 

Bleher, B., Uster, D. and T. Bergsdorf (2006). Assessment of threat status and management effectiveness in 

Kakamega Forest, Kenya. – In: Biodiversity and Conservation, 15, pp. 1159-1177. 

Bonhomme, R., Varlet-Grancher, C. and P. Chartier (1974). The use of hemispherical photographs for 

determining the leaf area index of young crops. – In: Photosynthetica, 8 (3), pp. 299-301. 

Bréda, N. (2003). Ground-based measurements of leaf area index: a review of methods, instruments and current 

controversies. – In: Journal of Experimental Botany, 54 (392), pp. 2403-2417. 

Broich, M. (2005). Ableitung des Blattflächenindex aus Landsat ETM+ SLC-off Daten für ostafrikanische 

Regenwaldökosysteme am Beispiel von Kakamega Forest (Westkenia). Unveröffentlichte Diplomarbeit, 

eingereicht am Geographischen Institut der RWTH Aachen. 

Brooks, T.M., Mittermeier, R.A., Mittermeier, C.G., da Fonseca, G.A.B., Rylands, A.B., Konstant, W.R., Flick, 

P., Pilgrim, J., Oldfield, S., Magin, G. and C. Hilton-Taylor (2002). Habitat loss and extinction in the 

hotspots of biodiversity. – In: Conservation Biology, 16, pp. 909-923. 

Brown, L. J., Chen, J. M., Leblanc, S. G. and J. Cihlar (2000). Short wave infrared correction to the simple 

ratio: an image and model analysis. – In: Remote Sensing of Environment, 71, pp. 16-25. 

Burgess, N., D’Amico Hales, J., Underwood, E., Dinerstein, E., Olson, D., Itoua, I., Schipper, J., Ricketts, T. 

and K. Newman. (2004). Terrestrial ecoregions of Africa and Madagascar: a conservation assessment. 

Island Press, Washington, D.C. 

Butson, R. and R.A. Fernandes (2004). A consistency analysis of surface reflectance and leaf area index retrieval 

from overlapping clear-sky Landsat ETM+ imagery. – In: Remote Sensing of Environment, 89, pp. 369- 

380. 

Cahen, L., Snelling, N.J., Delhal, J. and J.R. Vail (1984). The geochronology and evolution of Africa. Oxford. 

Casa, R. and H.G. Jones (2005). LAI retrieval from multiangular image classification and inversion of a ray 

tracing model. – In: Remote Sensing of Environment, 98, pp. 414-428. 

Chapman, L. (2007). Potential applications of near infra-red hemispherical imagery in forest environments. 

– In: Agricultural and Forest Meteorology, 143, pp. 151-156. 

153

154 

Chason, J.W., Baldocchi, D.D. and M.A. Huston (1991). A comparison of direct and indirect methods for 

estimating forest canopy leaf area. – In: Agricultural and Forest Meteorology, 57, pp. 107-128. 

Chen, J.M., Black, T.A. and R.S. Adams (1991). Evaluation of hemispherical photography for determining plant 

area index and geometry of a forest stand. – In: Agriculture and Forest Meteorology, 56, pp. 129-143. 

Chen, J.M. and T.A. Black (1992). Defining leaf area index for non-flat leaves. – In: Plant Cell Environment, 

15, pp. 421-429. 

Chen, J.M. and J. Cihlar (1995a). Plant canopy gap-size analysis theory for improving optical measurements of 

leaf-area index. – In: Applied Optics, 34 (27), pp. 6211-6222. 

Chen, J.M. and J. Cihlar (1995b). Quantifying the effect of canopy architecture on optical measurements of 

leaf area index using two gap size analysis methods. – In: IEEE Transactions on Geoscience and Remote 

Sensing, 33 (3), pp. 777-787. 

Chen, J. M. (1996). Evaluation of vegetation indices and a modified simple ratio for Boreal applications. 

– In: Canadian Journal of Remote Sensing, 22, 229-241. 

Chen, J. M., and J. Cihlar (1996). Retrieving leaf area index of boreal conifer forests using Landsat TM images. 

– In: Remote Sensing of Environment, 55, pp. 153-162. 

Chen, J.M., Rich P.M., Gower, S.T., Norman J.M. and S. Plummer (1997). Leaf area index of boreal forests: 

Theory, techniques, and measurements. – In: Journal of Geophysical Research, 102 (D24), p. 29429- 

29443. 

Chen, J.M., Pavlic, G., Brown, L., Cihlar, J., Leblanc, S.G., White, H.P., Hall, R.J., Peddle, D.R., King, D.J., 

Trofymow, J.A., Swift, E., Van der Sanden, J. and P.K.E. Pellikka (2002). Derivation and validation of 

Canada-wide coarse-resolution leaf area index maps using high-resolution satellite imagery and ground 

measurements. – In: Remote Sensing of Environment, 80, 165-184. 

Chen, J.M., Menges, C.H. and S.G. Leblanc (2005). Global derivation of the vegetation clumping index from 

multi-angular satellite data. – In: Remote Sensing of Environment, 97, 447-457. 

Clark, D.A. (2007). Detecting Tropical Forests’ Responses to Global Climatic and Atmospheric Change: Current 

Challenges and a Way Forward. – In: Biotropica, 39 (1), pp. 4-19. 

Clark, D.B, Olivas, P.C., Oberbauer, S.F., Clark, D.A. and M.G. Ryan (2008). First direct landscape-scale 

measurement of tropical rain forest Leaf Area Index, a key driver of global primary productivity. 

– In: Ecology Letters, 11 (2), pp. 163-172. 

Cohen, W.B., Maiersperger, T.K., Gower, S.T. and D.P. Turner (2003). An improved strategy for regression of 

biophysical variables and Landsat ETM+ data. – In: Remote Sensing of Environment, 84, pp. 561-571. 

Cohen, W.B., Maiersperger, T.K., Turner, D.P., Ritts, W.D., Pflugmacher, D., Kennedy, R.E., Kirschbaum, 

A., Running, S.W., Costa, M. and S.T. Gower (2006). MODIS land cover and LAI collection 4 product 

quality across nine sites in the Western Hemisphere. – In: IEEE Transactions on Geoscience and Remote 

Sensing, 44, pp. 1843-1857. 

Colditz, R.R., Conrad, C., Wehrmann, T., Schmidt, M. and S.W. Dech (2008). TiSeG – A flexible software tool 

for time series generation of MODIS data utilizing the quality assessment science data set. – In: IEEE 

Transactions on Geoscience and Remote Sensing (accepted). 

Colombo, R., Bellingeri, D., Fasolini, D. and C. Marino (2003). Retrieval of leaf area index in different 

vegetation types using high resolution satellite data. – In: Remote Sensing of Environment, 86, pp. 120-131.

References 

Cuarón, A.D. (2000). A global perspective on habitat disturbance and tropical rainforest mammals. 

– In: Conservation Biology, 14 (6), pp. 1574-1579. 

Curran, P.J., and A. Hay (1986). The importance of measurement error for certain procedures in remote sensing 

at optical wavelengths. – In: Photogrammetric Engineering and Remote Sensing, 52, pp. 229-241. 

Curran, P.J., Dungan, J.L. and H.L. Gholz (1992). Seasonal LAI in Slash Pine Estimated with Landsat TM. 


Cutini, A., Matteucci, G. and G.S. Mugnozza (1998). Estimation of leaf area index with the Li-Cor LAI 2000 in 

deciduous forests. – In: Forest Ecology and Management, 105, pp. 55-65. 

Deng, F., Chen, J.M., Plummer, S. and M. Chen (2006). Algorithm for global leaf area index retrieval using 

satellite imagery. – In: IEEE Transactions on Geoscience and Remote Sensing, 44, pp. 2219-2229. 

De Wasseige, C., Bastin, D. and P. Defourny (2003). Seasonal variation of tropical forest LAI based on field 

measurements in Central African Republic. – In: Agricultural and Forest Meteorology, 119, pp. 181-194. 

Dech, S.W., Tungalagsaikhan, P., Preusser, C. and R.E. Meisner (1998). Operational value-adding to AVHRR 

data over Europe: methods, results, and prospects. – In: Aerospace Science and Technology, 2 (5), 

pp. 335-346. 

Delta-T Devices Ltd. (1999). HemiView Canopy Analysis Software. Version 2.1 SR1. Copyright by Delta-T 

Devices Ltd. 

Didan, K. (2007). MODIS Land Team. Collection 5.0 and beyond. Accessed 3 July 2008 . 

Disney, M. (2001). Improved estimation of surface biophysical parameters through inversion of linear BRDF 

models. PhD thesis. Accessed 15 December, 2007. . 

Dufrêne, E. and N. Bréda (1995). Estimation of deciduous forest leaf area index using direct and indirect 

methods. – In: Oecologia, 104, pp. 156-162. 

Eckert, S. (2006). A Contribution to Sustainable Forest Management in Patagonia. – In: Remote Sensing Series, 

45, 154 p. 

Eggeling, W.J. (1947). Observations on the ecology of Budongo Rain Forest, Uganda. – In: Journal of Ecology, 

34, pp. 20-87. 

Eklundh, L., Hall, K., Eriksson, H., Ardö, J. and P. Pilesjö, 2003. Investigating the use of Landsat thematic 

mapper data for estimation of forest leaf area index in southern Sweden. Canadian Journal of Remote 

Sensing, 29, 349–362. 

Eriksson, H., Eklundh, L., Hall, K. and A. Lindroth (2005). Estimating LAI in deciduous forest stands. 


Eriksson, H., Eklundh, L., Kuusk, A. and T. Nilson (2006). Impact of understory vegetation on forest canopy 

reflectance and remotely sensed LAI estimates. – In: Remote Sensing of Environment, 103, pp. 408-418. 

Ernsting, A. and D. Rughani (2007). ‚Reduced Emissions From Deforestation‘: Can Carbon Trading Save Our 

Ecosystems? Accessed 12 July, 2008. 

155

156 

FAO (2006). Global Forest Resources Assessment 2005. Progress towards sustainable forest management. FAO 

FORESTRY PAPER 147, Rome. 

Fashing, P.J., Forrestel, A., Scully, C. and M. Cords (2004). Long-term tree population dynamics and their 

implications for the conservation of the Kakamega Forest, Kenya. – In: Biodiversity and Conservation, 

13, pp. 753-771. 

Fashing, P.J and J. Mwangi Gathual (2004). Spatial variability in the vegetation structure and composition of an 

East African rain forest. – In: African Journal of Ecology, 42, pp. 189-197. 

Fassnacht, K.S., Gower, S.T., Norman, J.M. and R.E. McMurtrie (1994). A comparison of optical and direct 

methods for estimating foliage surface area index in forests. – In: Agricultural and Forest Meteorology, 

71, pp. 183-207. 

Fassnacht, K.S., Gower, S.T., MacKenzie, M.D., Nordheim, E.V. and T.M. Lillesand (1997). Estimating the leaf 

area index of North Central Wisconsin forests. – In: Remote Sensing of Environment, 61, pp. 229-245. 

Fensholt, R., Sandholt, I. and M. Schultz Rasmussen (2004). Evaluation of MODIS LAI, fAPAR and the relation 

between fAPAR and NDVI in a semi-arid environment using in situ measurements. – In: Remote Sensing 

of Environment, 91, pp. 490-507. 

Ferment, A., Picard, N., Gourlet-Fleury, S. and C. Baraloto (2001). A comparison of five indirect methods for 

characterizing the light environment in a tropical forest. – In: Annals of Forest Science, 58, pp. 877-891. 

Fernandes, R.A., Butson, C., Leblanc, S.G. and R. Latifovic (2003). Landsat-5 and Landsat-7 ETM+ based 

accuracy assessment of leaf area index products for Canada derived from SPOT-4 VEGETATION data. 

– In: Canadian Journal of Remote Sensing, 29 (2), pp. 241-258. 

Fernandes, R., and S.G. Leblanc (2005). Parametric (modified least squares) and non-parametric (Theil– 

Sen) linear regressions for predicting biophysical parameters in the presence of measurement errors. 


Foody, G.M. and P.M. Atkinson (2002). Uncertainty in remote sensing and GIS. Chichester, UK, Wiley and 

Sons, 326 p.. 

Frazer, G.W., Canham, C.D. and K.P. Lertzman (1999). Gap Light Analyzer (GLA), Version 2.0: Imaging 

software to extract canopy structure and gap light transmittance indices from true-colour fisheye 

photographs, users manual and program documentation. Copyright © 1999: Simon Fraser University, 

Burnaby, British Columbia, and the Institute of Ecosystem Studies, Millbrook, New York. 

Frazer, G.W., Fournier, R.A., Trofymow, J.A. and R.J. Hall (2001). A comparison of digital and film fisheye 

photography for analysis of forest canopy structure and gap light transmission. – In: Agricultural and 

Forest Meteorology, 109, pp. 249-263. 

Gao, X., Huete, A.R. and K. Didan (2003). Multisensor Comparisons and Validation of MODIS Vegetation 

Indices at the Semiarid Jornada Experimental Range. – In: IEEE Transactions on Geoscience and Remote 

Sensing, 41 (10), pp. 2368-2381. 

Garrigues, S., Allard, D., Weiss, M. and F. Baret (2002). Comparing VALERI sampling schemes to better 

represent high spatial resolution satellite pixel from ground measurements: How to characterize an ESU. 

Accessed 30 December, 2007. . 

Gascon, F., Gastellu-Etchegorry, J.P., Lefevre-Fonollosa, M.-J. and E. Dufrene (2004). Retrieval of forest 

biophysical variables by inverting a 3-D radiative transfer model and using high and very high resolution 

imagery. – In: International Journal of Remote Sensing, 25 (24), pp. 5601-5616.

References 

Gastellu-Etchegorry, J.P, Zagolski, F. and J. Romier (1996a). A simple anisotropic reflectance model for 

homogeneous multilayer canopies. – In: Remote Sensing of Environment, 57, pp. 22-38. 

Gastellu-Etchegorry, J.P., Demarez, V., Pinel, V. and F. Zagolski (1996b). Modeling radiative transfer in 

heterogeneous 3-D vegetation canopies. – In: Remote Sensing of Environment, 58, pp. 131-156. 

Gobron, N., Pinty, B. and M.M. Verstraete (1997). Theoretical limits to the estimation of the leaf area index on 

the basis of visible and near-infrared remote sensing data. – In: IEEE Transactions on Geoscience and 

Remote Sensing, 35, pp. 1438-1445. 

Goel, N.S. and D.E. Strebel (1983). Inversion of vegetation canopy reflectance models for estimating agronomic 

variables. I. Problem definition and initial results using Suits model. – In: Remote Sensing of the 

Environment, 13, pp. 487-507. 

Goel, N.S. and R.L. Thompson (1984a). Inversion of vegetation canopy reflectance models for estimating 

agronomic variables. III. Estimation using only canopy reflectance data as illustrated by the Suits model. 

– In: Remote Sensing of the Environment, 15, pp. 223-236. 

Goel, N.S. and R.L. Thompson (1984b). Inversion of Vegetation Canopy Reflectance Models for Estimating 

Agronomic Variables. V. Estimation of Leaf Area Index and Average Leaf Angle Using Measured Canopy 

Reflectances. – In: Remote Sensing of the Environment, 16, pp. 69-85. 

Goel, N.S. and T. Grier (1988). Estimation of Canopy Parameters for Inhomogeneous Vegetation Canopies 

from Reflectance Data: III. TRIM: A Model for Radiative Transfer in Heterogeneous Three-Dimensional 

Canopies. – In: Remote Sensing of the Environment, 25, pp. 255-293. 

Gower, S.T., Kucharik, C.J. and J.M. Norman (1999). Direct and indirect estimation of leaf area index, fAPAR, 

and net primary production of terrestrial ecosystems. – In: Remote Sensing of Environment, 70, pp. 29-51. 

Guenther, B., Xiong, X., Salomonson, V.V., Barnes, W.L. and J. Young (2002). On-orbit performance of the 

Earth Observing System Moderate Resolution Imaging Spectroradiometer; first year of data. – In: Remote 

Sensing of Environment, 83 (1-2), pp. 16-30. 

Gullison, R.E., Frumhoff, P.C., Canadell, J.G., Field, C.B., Nepstad, D.C., Hayhoe, K., Avissar, R., Curran, 

L.M., Friedlingstein, P., Jones, C.D. and C. Nobre (2007). Tropical Forests and Climate Policy. 

– In: Science, 316, pp. 985-986 

Hale, S.E. and C. Edwards (2002). Comparison of film and digital hemispherical photography across a wide 

range of canopy densities. – In: Agricultural and Forest Meteorology, 112, pp. 51-56. 

Hall-Beyer, M. (2007). The GLCM Tutorial Home Page. Version 2.10. Accessed 26 October, 2007. 

. 

Hamilton, A.C. (1974). Distribution patterns of forest trees in Uganda and their historical significance. 

– In: Vegetatio, 29, pp. 21-35. 

Hamilton, A.C. (1981). The quaternary history of African Forests: its relevance to conservation. – In: African 

Journal of Ecology, 19, pp. 1-6. 

Hamilton, A.C. (1982). Environmental history of East Africa. A study of the Quaternary. London. 

Hansen, M.C. and R. DeFries (2004). Detecting long-term global forest change using continuous fields of treecover 

maps from 8-km advanced very high resolution radiometer (AVHRR) data for the years 1982-99. 

– In: Ecosystems, 7, pp. 695-716. 

157

158 

Haralick, R.M., Shanmugan, K. and I. Dinstein (1973). Textural features for image classification. – In: IEEE 

Transactions on Systems, Man, and Cybernetics, 3 (6), pp. 610–621. 

Hardisky, M.A., Klemas, V. and R. M. Smart (1983). The influence of soil salinity, growth form, and leaf 

moisture on the spectral radiance of Spartina alterniflora canopies. – In: Photogrammetric Engineering 

and Remote Sensing, 49, pp. 77-83. 

Heiskanen, J. (2006). Estimating aboveground tree biomass and leaf area index in a mountain birch forest using 

ASTER satellite data. – In: International Journal of Remote Sensing, 27 (6), pp. 1135-1158. 

Henry, P. and A. Meygret (2001). Calibration of HRVIR and VEGETATION Cameras on SPOT4. – In: Advances 

in Space Research, 28 (19), pp. 49-58. 

Herwitz, S.R., Peterson, D.L. and J.R. Eastman (1990). Thematic Mapper detection of changes in the leaf area 

of closed canopy pine plantations in central Massachusetts. – In: Remote Sensing of Environment, 29, 

pp. 129-140. 

Hester, B. and W. Boberg (2006). Opportunities for Mining Investment. Department of Geological Survey and 

Mines, Republic of Uganda. 

Hill, M.J, Senarath, U., Lee, A, Zeppel, M., Nightingale, J.M., Williams, R.J. and T.R. McVicar (2006). 

Assessment of the MODIS LAI product for Australian ecosystems. – In: Remote Sensing of Environment, 

101, pp. 495-518. 

Horler, D.N.H. and F. J. Ahern (1986). Forestry information content of thematic mapper data. – In: International 

Journal of Remote Sensing, 7(3), pp. 405-428. 

Howard, P.C. (1991). Nature Conservation in Uganda’s Tropical Forest Reserves. IUCN, Gland, Switzerland 

and Cambridge, UK. 

Howard, P., Davenport, T. and R. Matthews (1996). Budongo Forest Reserve, Biodiversity Report, Report No. 

3, Forest Department, Kampala, Uganda. 

Howard, P., Davenport, T. and F. Kigenyi (1997). Planning conservation areas in Uganda’s natural forests. 

– In: Oryx, 31 (4), pp. 253-264. 

Huang, D., Yang, W., Tan, B., Rautiainen, M., Zhang, P., Hu, J., Shabanov, N.V., Linder, S., Knyazikhin, Y. and 

R.B. Myneni (2006): The Importance of Measurement Errors for Deriving Accurate Reference Leaf Area 

Index Maps for Validation of Moderate-Resolution Satellite LAI Products. – In: IEEE Transactions on 

Geoscience and Remote Sensing, 44, pp. 1866-1871. 

Huemmrich, K.F., Privette, J.L., Mukelabai, M., Myneni, R.B. and Y. Knyazikhin (2005). Time-series validation 

of MODIS land biophysical products in a Kalahari woodland, Africa. – In: International Journal of 

Remote Sensing, 26 (19), pp. 4381-4398. 

Hyer, E.J. and S.J. Goetz (2004). Comparison and sensitivity analysis of instruments and radiometric methods 

for LAI estimation: assessments from a boreal forest site. – In: Agricultural and Forest Meteorology, 122, 

pp. 157-174. 

Inoue, A., Yamamoto, K., Mizoue, N. and Y. Kawahara (2004). Effects of image quality, size and camera type 

on forest light environment estimates using digital hemispherical photography. – In: Agricultural and 


Jacquemoud, S., Baret, F., Andrieu, B., Danson, F.M. and K. Jaggard (1995). Extraction of vegetation 

biophysical parameters by inversion of the PROSPECT + SAIL models on sugar beet canopy reflectance 

data. Application to TM and AVIRIS sensors. – In: Remote Sensing of Environment, 52, pp. 163-172.

References 

Jacquemoud, S., Bacour, C., Poilvé, H. and J.-P. Frangi (2000). Comparison of Four Radiative Transfer Models 

to Simulate Plant Canopies Reflectance: Direct and Inverse Mode. – In: Remote Sensing of Environment, 

74, pp. 471-481. 

Jensen, R.R. and M.W. Binford (2004). Measurement and comparison of Leaf Area Index estimators derived 

from satellite remote sensing techniques. – In: International Journal of Remote Sensing, 25 (20), 

pp. 4251-4265. 

Jensen, J.R. (2007). Remote sensing of the environment: an Earth resource perspective, 2nd ed., Prentice Hall. 

Jin, M. and D.-L. Zhang (2002). Observed variations of leaf area index and its relationship with surface 

temperatures during warm seasons. – In: Meteorology and Atmospheric Physics, 80, pp. 117-129. 

Jonckheere, I., Fleck, S., Nackaerts, K., Muys, B., Coppin, P., Weiss, M. and F. Baret (2004). Review of 

methods for in situ leaf area index determination Part I. Theories, sensors and hemispherical photography. 


Jonckheere I., Nackaerts, K. Muys, B. and P. Coppin (2005). Assessment of automatic gap fraction estimation 

of forests from digital hemispherical photography. – In: Agricultural and Forest Meteorology, 132, pp. 96-114. 

JPL (2001). ASTER Higher-Level Product User Guide. Version 2.0. JPL D-20062, 80 p. Accessed 

6 January, 2008. 

JPL (2004). ASTER Product description AST07. Accessed 10 January, 2008. 

Justice, C.O. and J.R.G. Townshend (1981). A comparison of unsupervised classification procedures using 

Landsat MSS data for an area of complex surface conditions in Basilicata, southern Italy. – In: Remote 


Justice, C., Belward, A., Morisette, J., Lewis, P., Privette, J. and F. Baret (2000). Developments in the ‘validation’ 

of satellite sensor products for the study of the land surface. – In: International Journal of Remote Sensing, 

21 (17), pp. 3383-3390. 

Kalácska, M., Sánchez-Azofeifa, G.A., Rivard, B., Calvo-Alvarado, J.C., Journet, A.R.P., Arroyo-Mora, J.P. 

and D. Ortiz-Ortiz (2004). Leaf area index measurements in a tropical moist forest: A case study from 

Costa Rica. – In: Remote Sensing of Environment, 91, pp. 134-152. 

Kalácska, M., Calvo-Alvarado, J.C. and G.A. Sánchez-Azofeifa (2005). Calibration and assessment of seasonal 

leaf area index of a tropical dry forest in different stages of succession. – In: Tree physiology, 25, pp. 

733-744. 

Kanemasu, E.T., Niblett, C.L., Manges, H., Lenhert, D. and M.A. Newman (1974). Wheat: Its growth and 

disease severity as deduced from ERTS-1. – In: Remote Sensing of Environment, 3, pp. 43-47. 

Kanemasu, E.T., Heilman, J.L., Bagley, J.O. and W.L. Powers (1977). Using LANDSAT data to estimate 

evapotranspiration of winter wheat. – In: Environmental Management, 1, pp. 515-520. 

Kang, S., Running, S.W., Lim, J.-H., Zhao, M., Park, C.R. and R. Loehman (2003). A regional phenology model 

for detecting onset of greenness in temperate mixed forests, Korea: an application of MODIS leaf area 

index. – In: Remote Sensing of Environment, 86, pp. 232-242. 

Kaufmann, M. and C. Troendle (1981) The relationship of leaf area and foliage biomass to sapwood conducting 

area in four subalpine forest tree species. – In: Forest Science, 27 (3), pp. 477-482. 

159

160 

Kendall, R.L. (1969). An Ecological history of the Lake Victoria Basin. – In: Ecological Monographs, 39, 

pp. 121-176. 

Kimes, D.S., Knyazikhin, Y., Privette, J.L., Abuelgasim, A.A. and F. Gao (2000). Inversion methods for 

physically based models. – In: Remote Sensing Reviews, 18, pp. 381-439. 

Kucharik, C.J., Norman, J.M. and S.T. Gower (1998). Measurements of branch area and adjusting leaf area 

index indirect measurements. – In: Agricultural and Forest Meteorology, 91, pp. 69-88 

Knyazikhin, Y., Martonchik, J.V, Myneni, R.B., Diner, D.J. and S.W. Running (1998a). Synergistic algorithm for 

estimating vegetation canopy leaf area index and fraction of absorbed photosynthetically active radiation 

from MODIS and MISR data. – In: Journal of Geophysical Research, 103 (D24), pp. 32,257-32,276. 

Knyazikhin, Y., Martonchik, J.V., Diner, D.J., Myneni, R.B., Verstraete, M.M., Pinty, B. and N. Gobron. 

(1998b). Estimation of vegetation canopy leaf area index and fraction of absorbed photosynthetically 

active radiation from atmosphere-corrected MISR data. – In: Journal of Geophysical Research, 103, 

pp. 32,239-32,256. 

Knyazikhin, Y., Glassy, J., Privette, J.L., Tian, Y., Lotsch, A., Zhang, Y., Wang, Y., Morisette, 

J.T., Votava, P., Myneni, R.B., Nemani, R.R. and S.W. Running (1999). MODIS Leaf Area 

Index (LAI) and Fraction of Photosynthetically Active Radiation Absorbed by Vegetation 

(FPAR) Product (MOD15) Algorithm Theoretical Basis Document. Accessed 8 August, 2007. 

< http://modis.gsfc.nasa.gov/data/atbd/atbd_mod15.pdf >. 

Kowkaro, J.O. (1988). Conservation status of the Kakamega Forest in Kenya: the easternmost relict of the 

equatorial rain forests of Africa. – In: Monographs in Systematic Botany from the Missouri Botanical 

Garden, 25, pp. 471-489. 

Kucharik, C.J., Norman, J.M. and S.T. Gower (1998). Measurements of branch area and adjusting leaf area 

index indirect measurements. – In: Agricultural and Forest Meteorology, 91, 69-88. 

Kucharik, C.J., Norman, J.M. and S.T. Gower (1999). Characterization of radiation regimes in nonrandom 

forest canopies: Theory, measurements and a simplified modeling approach. – In: Tree Physiology, 19, 

pp. 695-706. 

Kuusk, A. (1995). A fast, invertible canopy reflectance model. – In: Remote Sensing of Environment, 51, pp. 

342-350. 

Kuusk, A., Nilson, T. and Paas, M. (2002). Angular distribution of radiation beneath forest canopies using a 

CCD-radiometer. – In: Agricultural and Forest Meteorology, 110, pp. 259-273. 

Lambin, E.F. (1999). Monitoring forest degradation in tropical regions by remote sensing: some methodological issues. 

– In: Global Ecology and Biogeography, 8, pp. 191-198. 

Lang, A.R.G., Xiang, Y. and J.M. Norman (1985). Crop structure and the penetration of direct sunlight. 

– In: Agricultural and Forest Meteorology, Vol. 35, pp. 83-101. 

Lang, A.R.G. (1986). Leaf area and average leaf angle from direct transmission of sunlight. – In: Australian 

Journal of Botany, 34, pp. 349-355. 

Lang, A.R.G. and Y. Xiang (1986). Estimation of leaf area index from transmission of direct sunlight in 

discontinuous canopies. – In: Agricultural and Forest Meteorology, 37, pp. 229-243. 

Lang, A.R.G., McMurtrie, R.E. and M. Benson (1991). Validity of the surface area indices of Pinus radiata 

estimated from transmittance of the sun‘s beam. – In: Agricultural and Forest Meteorology, 57, pp. 157-170.

References 

Langdale-Brown, I., Osmaston, H.A. and J.G. Wilson (1964). The Vegetation of Uganda and its bearing on 

land-use, Entebbe. 

Laumonier, Y., Trichon, V., Gastellu-Etchegorry, J.P. and P. Birnbaum (1994). Reflectance, Leaf Area Index and 

structural studies of a rain forest canopy using the “Opération Canopée” Association’s tree top raft-hot air 

airship combination. – In: Geocarto International, 9 (4), pp. 17-29. 

Leboeuf, A., Beaudoin, A., Fournier, R.A., Guindon, L., Luther, J.E. and M.-C. Lambert (2007). A shadow 

fraction method for mapping biomass of northern boreal black spruce forests using QuickBird imagery. – 

In: Remote Sensing of Environment, 110 (4), pp. 488-500. 

Leblanc, S.G. and J.M. Chen (2001). A practical scheme for correcting multiple scattering effects on optical 

LAI measurements. – In: Agricultural and Forest Meteorology, 110, pp. 125-139. 

Leblanc, S.G., Fernandes, R. and J.M. Chen (2002). Recent advancements in optical field leaf area index, foliage 

heterogeneity, and foliage angular distribution measurements. – In: IEEE International Geoscience and 

Remote Sensing Symposium 2002, IGARSS ’02, 5, pp. 2902-2904. 

Leblanc, S.G., Chen, J.M., Fernandes, R., Deering, D.W. and A. Conley (2005). Methodology comparison 

for canopy structure parameters extraction from digital hemispherical photography in boreal forests. 


Le Dantec, V., Dufrêne, E. and B. Saugier (2000). Interannual and spatial variation in maximum leaf area index 

of temperate deciduous stands. – In: Forest Ecology and Management, 134, pp. 71-81. 

Lee, K.S., Park, Y.I., Kim, S.H., Park, J.H., Woo, C.S. and K.C. Jang (2004). Remote sensing estimation of 

forest LAI in close canopy situation. – In: Proceedings of the XXth ISPRS Congress, 12-23 July 2004, 

Istanbul, Turkey, pp. 400-405. 

Li, X. and A.H. Strahler (1985). Geometric-optical modelling of a conifer forest. – In: IEEE Transactions on 


Li, X. and A.H. Strahler (1986). Geometric-optical bidirectional reflectance modeling of a conifer forest canopy. 

– In: IEEE Transactions on Geoscience and Remote Sensing, 24, pp. 906-919. 

Liang, S., Fang, H., Chen, M., Shuey, C.J., Walthall, C., Daughtry, C., Morisette, J., Schaaf, C. and A. Strahler 

(2002). Validating MODIS land surface reflectance and albedo products: methods and preliminary results. 


Liang, S. (2004). Quantitative Remote Sensing of Land Surfaces, Wiley and Sons, 534 p. 

LI-COR (1992). LAI-2000 plant canopy analyzer operating manual. LI-COR, Lincoln, NE. 

LI-COR (2005). LAI-2000 File Viewer. Version 1.04. Copyright by LI-COR, Lincoln, NE. 

Lieth, H. and M.J.A. Werger (Eds.) (1989). Tropical rain forest ecosystems. Ecosystems of the world, Bd. 14. 

Amsterdam. 

Lind, E.M. and M.E.S. Morrison (1974). East African Vegetation. London. 

Lu, L., Li, X., Ma, M.G., Che, T., Huang, C.L., Veroustraete, F., Dong, Q.H., Bogaert, J. and R. Ceulemans 

(2004). Investigating relationship between Landsat ETM+ data and LAI in a semi-arid grassland of 

Northwest China. – In: IEEE International Geoscience and Remote Sensing Symposium proceedings, 20- 

24 September 2004, Anchorage, Alaska. 

161

162 

Lung, T. (2004). Landbedeckungsänderungen im Gebiet “Kakamega Forest und assoziierte Waldgebiete” 

(Westkenia) - Multispektrale Klassifikation von Landsat-Satellitenbilddaten und Auswertung mittels 

Methoden im Raster-GIS. – In: Karlsruher Geowissenschaftliche Schriften, A (15), 114 p. 

MacDonald, R. (1966). Uganda - Geology: from „The Atlas of Uganda, 2nd Edition, 1969“, Dept. of Lands and 

Surveys, Government of Uganda, Entebbe, Uganda. 

Macfarlane, C., Hoffman, M., Eamus, D., Kerp, N., Higginson, S., McMurtie, R. and M. Adams (2007). 

Estimation of leaf area index in eucalypt forest using digital hemispherical photography. – In: Agricultural 

and Forest Meteorology, 143, pp. 176-188. 

Masuoka, E., Fleig, A.J., Wolfe, R.E. and F.S. Patt (1998). Key characteristics of MODIS data products. 

– In: IEEE Transactions on Geoscience and Remote Sensing, 36 (4), pp. 1313-1323. 

Masson, V., Champeaux, J.-L., Chauvin, F., Meriguet, C. and R. Lacaze (2003). A global database of Land 

Surface Parameters at 1-km Resolution in Meteorological and Climate Models. – In: Journal of Climate, 

16 (9), pp. 1261-1282. 

McCoy, R.M. (2005). Field Methods in Remote Sensing. New York. 159 p. 

Miao, G. (2008). Modeling of carbon dynamics in an African rainforest. A case study with CoupModel in 

Kakamega Forest, Kenya. Master of Science Thesis at the Department of Land and Water Resources Engineering, 

Royal Institute of Technology (KTH), Stockholm, Sweden, 58 p. 

Miller, J.B. (1967). A formula for average foliage density. – In: Australian Journal of Botany, 15, pp. 141-144. 

Mitchell, N. and G. Schaab (in print). Developing a disturbance index for five East African forests using GIS 

to analyse historical forest use as an important driver of current land use/cover. – In: African Journal of 

Ecology. 

Mitchell, N. (2004). The exploitation and disturbance history of Kakamega Forest, Western Kenya. 

– In: Bielefelder Ökologische Beiträge, 20, 77 p. 

Mitchell, N., Lung, T. and G. Schaab (2006). Tracing significant losses and limited gains in forest cover for 

the Kakamega-Nandi complex in western Kenya across 90 years by use of satellite imagery, aerial 

photography and maps. – In: Proceedings of the ISPRS (TC7) Mid-Term Symposium „Remote Sensing: 

From Pixels to Processes“, 8-11 May 2006, ITC Enschede (The Netherlands); ed. by N. Kerle and A.K. 

Skidmore, pp. 778-785. 

Monsi, M. and T. Saeki (1953). Uber den Lichtfaktor in den Pflanzengesellschaften und seine Bedeutung für die 

Stoffproduktion. – In: Japanese Journal of Botany, 14, pp. 22-52. 

Monteith, J. (1976). Vegetation and the atmosphere. Volume 2 – Case Studies, London, 439 p. 

Moran, P.A.P. (1950). Notes on continuous stochastic phenomena. – In: Biometrika, 37, pp. 17-23. 

Morisette, J.T., Baret, F., Privette, J.L., Myneni, R.B., Nickeson, J.E., Garrigues, S., Shabanov, N.V., Weiss, 

M., Fernandes, R.A., Leblanc, S.G., Kalacska, M., Sánchez-Azofeifa, G.A., Chubey, M., Rivard, B., 

Stenberg, P., Rautiainen, M., Voipio, P., Manninen, T., Pilant, A.N., Lewis, T.E., Iiames, J.S., Colombo, 

R., Meroni, M., Busetto, L., Cohen, W.B., Turner, D.P., Warner, E.D., Petersen, G.W., Seufert, G. and R. 

Cook (2006a). Validation of global moderate-resolution LAI products: A framework proposed within the 

CEOS Land Product Validation Subgroup. – In: IEEE Transactions on Geoscience and Remote Sensing, 

44, pp. 1804-1817. 

Morisette, J.T., Baret, F. and S. Liang (2006b). Special Issue on Global Land Product Validation. 

– In: IEEE Transactions on Geoscience and Remote Sensing, 44 (7), pp. 1695-1698.

References 

Musila, W., Kamoni, P.T., Albrecht, A., Todt, H., Uster, D. and H. Dalitz (2004). Soil characteristics of Kakamega 

Forest, Kenya. – In: H. Dalitz (ed). BIOTA report no. 2, Bielefelder Ökologische Beiträge, 21. 

Mussche, S., Samson, R., Nachtergale, L., De Schrijver, A., Lemeur, R. and N. Lust (2001). A comparison of 

optical and direct methods for monitoring the seasonal dynamics of leaf area index in deciduous forests. 

– In: Silva Fennica, 35 (4), pp. 373-384. 

Mutangah, J.G. (1996). An Investigation of Vegetation Status and Process in Relation to Human Disturbance in 

Kakamega Forest, Western Kenya, Ph.D. thesis, University of Wales Aberystwyth, Great Britain. 

Myers, N., Mittermeier, R.A., Mittermeier, C.G., da Fonseca, G.A.B. and J. Kent (2000). Biodiversity hotspots 

for conservation priorities. – In: Nature, 403, pp. 853-858. 

Myneni, R.B., Asrar, G. and S.A.W. Gerstl (1990). Radiative transfer in three dimensional leaf canopies. 

– In: Transport Theory and Statistical Physics, 19(3-5), pp. 205-250. 

Myneni, R.B. and J. Ross (1991). Photon-Vegetation Interactions: Applications in Optical Remote Sensing and 

Plant Ecology, Springer-Verlag, Heidelberg, Germany, 560 p. 

Myneni, R.B., Maggion, S., Iaquinto, J., Privette, J.L., Gobron, N., Pinty, B., Kimes, D.S., Verstraete, M.M. 

and D.L. Williams (1995). Optical remote-sensing of vegetation-modeling, caveats, and algorithms. 


Myneni, R.B., Nemani, R.R. and S.W. Running (1997). Algorithm for the estimation of global land cover, 

LAI and FPAR based on radiative transfer models. – In: IEEE Transactions on Geoscience and Remote 

Sensing, 35, pp. 1380-1393. 

Myneni, R.B., Hoffman, S., Knyazikhin, Y., Privette, J.L., Glassy, J., Tian, Y., Wang, Y., Song, X., Zhang, Y., 

Smith, Y., Lotsch, A., Friedl, M., Morisette, J.T., Votava, P., Nemani, R.R. and S.W. Running (2002). 

Global products of vegetation leaf area and fraction absorbed PAR from year one of MODIS data. 

– In: Remote Sensing of the Environment, 83, pp. 214-231. 

Myneni, R.B., Shabanov, N., Yang, W., Tan, B., Dong, H., Knyazikhin, Y., Votava, P., Nemani, R. and S.W. 

Running (2004). MODIS LAI and FPAR: Project Status and Validation. Presentation held at LAI 

Intercomparison Group Meeting, Missoula, MT, 16 August 2004. 

Myneni, R.B., Yang, W., Nemani, R., Huete, R., Dickinson, R., Knyazikhin, Y., Didan, K., Fu, R., Negrón 

Juárez, R., Saatchi, S., Hashimoto, H., Ichii, K., Shabanov, N., Tan, B., Ratana, P., Privette, J., Morisette, 

J., Vermote, E., Roy, D., Wolfe, R., Friedl, M., Running, S., Votava, P., El-Saleous, N., Devadiga, S., Su, 

Y., and V. Salomonson (2007): Large seasonal swings in leaf area of Amazon rainforests. – In: PNAS, 

104 (12), pp. 4820–4823. 

NASA (2007a). Specifications of MODIS sensor. Accessed 19 November, 2007. . 

NASA (2007b). Changes in MODIS LAI/FPAR algorithm in collection 5. Accessed 21 November, 2007. 

. 

Nemani, R., Pierce, L., Running, S., and L. Band (1993). Forest ecosystem processes at the watershed scale: 

Sensitivity to remotely-sensed leaf area index estimates. – In: International Journal of Remote Sensing, 

14 (13), 2519-2534. 

Neumann, H.H., Hartog, G.D. and R.H. Shaw (1989). Leaf area measurements based on hemispheric 

photographs and leaf-littercollection in a deciduous forest during autumn leaf-fall. – In: Agricultural and 


163

164 

NIES (2007). Tropical Deforestation. National Institute for Environmental Studies, Japan. Accessed 26 October, 

2007. 

Nilson, T. (1971). A theoretical analysis of the frequency of gaps in plant stands. – In: Agricultural Meteorology, 

8, pp. 25-38. 

Nilson, T. (1999). Inversion of gap frequency data in forest stands. – In: Agricultural and Forest Meteorology, 

98-99, pp. 437-448. 

Nilson, T. and A. Kuusk (2005). Improved algorithm for estimating canopy indices from gap fraction data in 

forest canopies. – In: Agricultural and Forest Meteorology, 124, pp. 157-169. 

O’Hara, K.L. and N.I. Valappil, (1995). Sapwood-leaf are prediction equations for multi-aged ponderosa pine 

stands in western Montana and central Oregon. – In: Canadian Journal of Forest Research, 25, pp. 1553- 

1557. 

Ogawa, H., Yoda, K. and T. Kira (1961). A preliminary survey on the vegetation of Thailand. – In: Kira, T. and 

T. Umesao (eds.). Nature and life in southeast Asia, 1, pp. 21-159. 

Pandya, M.R., Singh, R.P., Chaudhari, K.N., Bairagi, G.D., Sharma, R., Dadhwal, V.K. and J.S. Parihar (2006). 

Leaf Area Index Retrieval Using IRS LISS-III Sensor Data and Validation of the MODIS LAI Product 

Over Central India. – In: IEEE Transactions on Geoscience and Remote Sensing, 44 (7), pp. 1858-1866. 

Panferov, O., Knyazikhin, Y., Myneni, R.B., Szarzynski, J., Engwald, S., Schnitzler, K.G. and G. Gravenhorst 

(2001). The role of canopy structure in the spectral variation of transmission and absorption of solar 

radiation in vegetation canopies. – In: IEEE Transactions on Geoscience and Remote Sensing, 39, 

pp. 241-253. 

Parmesan, C. and G. Yohe (2003). A globally coherent fingerprint of climate change impacts across natural 

systems. – In: Nature, 421, pp. 37-42. 

Peddle, D.R. and R.L. Johnson (2000). Spectral mixture analysis of airborne remote sensing imagery for 

improved prediction of leaf area index in mountainous terrain, Kananaskis, Alberta. – In: Canadian 

Journal of Remote Sensing, 26, pp. 177-188. 

Peterson, D.L., Spanner, M.A., Running, S.W. and K.B. Teuber (1987). Relationship of Thematic Mapper 

Simulator Data to Leaf Area Index of Temperate Coniferous Forests. – In: Remote Sensing of 


Pinty, B., Gobron, N., Widlowski, J.-L., Gerstl, S.A.W., Verstraete, M.M., Antunes, M., Bacour, C., Gascon, F., 

Gastellu, J.-P., Goel, N., Jacquemoud, S., North, P., Qin, W. and R. Thompson (2001). Radiation transfer 

model intercomparison (RAMI) exercise. – In: Journal of Geophysical Research, 106 (D11), pp. 11937- 

11956. 

Pisek, J. and J.M. Chen (2007). Comparison and validation of MODIS and VEGETATION global LAI products 

over four BigFoot sites in North America. – In: Remote Sensing of Environment, 109 (1), pp. 81-94. 

Planchais, I. and J.-Y. Pontailler (1999). Validity of leaf areas and angles estimated in a beech forest from 

analysis of gap frequencies, using hemispherical photographs and a plant canopy analyser. – In: Annals 

of Forest Science, 56, pp. 1–10. 

Plummer, S., Arino, O., Ranera, F., Chen, J., Dedieu, G., Simon, M., Ciais, P., Bondeau, A., Quegan, S., Schultz, 

M. and M. Raupach (2005). The GLOBCARBON initiative: multi-sensor estimation of global biophysical 

products for global terrestrial carbon studies. – In: Proceedings of the MERIS (A)ATSR Workshop 2005 

(ESA SP-597). 26-30 September 2005, ESRIN, Frascati, Italy. Editor: H. Lacoste. Published on CD-ROM.

References 

Plumptre, A.J. (1995). The importance of “seed trees” for the natural regeneration of selectively logged tropical forest. 

– In: Commonwealth Forestry Review, 74, pp. 253-258. 

Plumptre, A.J. (1996). Changes following 60 years of selective timber harvesting in the Budongo Forest Reserve, 

Uganda. – In: Forest Ecology and Management, 89, pp. 101-113. 

Privette, J.L., Myneni, R.B., Knyazikhin, Y., Mukelabai, M., Roberts, G., Tian, Y., Wang, Y. and S.G. Leblanc 

(2002). Early spatial and temporal validation of MODIS LAI product in the Southern Africa Kalahari. 


Pollock, R.B. and E.T. Kanemasu (1979). Estimating leaf-area index of wheat with Landsat data. – In: Remote 


Rautiainen, M. (2005). Retrieval of leaf area index for a coniferous forest by inverting a forest reflectance 

model. – In: Remote Sensing of Environment, 99 (3), pp. 295-303. 

Reynolds, V. (2005). The Chimpanzees of Budongo Forest. Ecology, Behaviour and Conservation. Oxford 

University Press. 312 p. 

Rich, P.M. (1988). Video image analysis of hemispherical canopy photography. – In: Mausel, P.W. (Ed.), 

Proceedings of the First Special Workshop on Videography. American Society for Photogrammetry and 

Remote Sensing, Terre Haute, IN, 19-20 May, pp. 84–95. 

Rich, P.M., (1989). A manual for analysis of hemispherical canopy photography. Los Alamos National 

Laboratory Report LA-11732-M. 

Rich, P.M., (1990). Characterizing plant canopies with hemispherical photographs. – In: Remote Sensing 

Reviews, 5, pp. 13-29. 

Rich, P.M., Clark, D.B., Clark, D.A. and S.F. Oberbauer (1993). Long-term study of solar radiation regimes in 

a tropical wet forest using quantum sensors and hemispherical photography. – In: Agriculture and Forest 

Meteorology, 65, pp. 107-127. 

Rich, P.M., Wood, J., Vieglais, D.A., Burek, K. and N. Webb (1999). Hemiview User Manual Version 2.1. 

Delta-T Devices Ltd., Cambridge, UK. 

Richards, J.A. and X. Jia (2006). Remote Sensing Digital Image Analysis: An Introduction. 4 th edition. Springer. 

469 p. 

Richter, R. (1998). Correction of satellite imagery over mountainous terrain. – In: Applied Optics, 37 (18), 

pp. 4004-4015. 

Richter, R. (2007). Atmospheric / Topographic Correction for Satellite Imagery. ATCOR-2/3 User Guide, 

Version 6.3, DLR-IB 565-01/07, Wessling, Germany, 134 pp. 

Rogers, R. and T.M. Hinckley (1979). Foliar weight and area related to current sapwood area in oak. 

– In: Forest Science, 25 (2), pp. 298-303. 

Root, T.L., Price, J.T., Hall, K.R., Schneider, S.H., Rosenzweigk, C. and J.A. Pounds (2003). Fingerprints of 

global warming on wild animals and plants. – In: Nature, 421, pp. 57-60. 

Rouse, J.W., Haas, R.H., and J.A. Schell (1974). Monitoring the vernal advancement of retrogradiation of 

natural vegetation. NASA/GSFC, type III, final report, 371 p. 

Ross, J. (1981). The Radiation Regime and Architecture of Plant Stands, Den Haag, Netherlands, 391 p. 

165

166 

Rossello, P. (2005). Ground data processing & production of the level 1 high resolution maps for Counami. 

Accessed 3 July, 2008. 

Rozhnyatovsky, I.I. (1954). Determination of leaf area with the help of light sensitive paper. 

– In: Soviet Botanical Journal, 39, 3, pp. 419-420. 

Russell, G., Jarvis, P.G. and J.L. Monteith (1989). Absorption of radiance by canopies and stand growth. 

– In: Russell, G.; Marshall, B.; Jarvis, P.G. (Ed.). Plant canopies: their growth, form and function. 

Cambridge: Cambridge University Press, 1989, pp. 21-40. 

Sachs, L. (2004). Angewandte Statistik: Anwendung statistischer Methoden. Springer, Berlin, Deutschland, 890 p. 

Saeki, T. (1975). Distribution of radiant energy and CO2 in terrestrial communities. – In: Cooper, J.P. (Ed.) 

(1975). Photosynthesis and productivity in different environments, Cambridge, pp. 297–322. 

Schlerf, M., Atzberger, C. and J. Hill (2005). Remote sensing of forest biophysical variables using HyMap 

imaging spectrometer data. – In: Remote Sensing of Environment, 95, pp. 177-194. 

Schlüter, T. (1997). Geology of East Africa, Berlin, 484 p. 

Schmidt, R. (1992). Degradation of a Forest in the Shimba Hills National Reserve, Kenya, as Initiated by Man 

and Maintained by Wildlife. – In: J.G. Goldammer (Ed.) (1992). Tropical Forests in Transition, Basel, 

pp. 85-104. 

Schneider, S.H., Semenov, S., Patwardhan, A., Burton, I., Magadza, C.H.D., Oppenheimer, M., Pittock, A.B., 

Rahman, A., Smith, J.B., Suarez A. and F. Yamin (2007). Assessing key vulnerabilities and the risk from 

climate change. Climate Change 2007: Impacts, Adaptation and Vulnerability. Contribution of Working 

Group II to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change, M.L. Parry, 

O.F. Canziani, J.P. Palutikof, P.J. van der Linden and C.E. Hanson (Eds.), Cambridge University Press, 

Cambridge, UK, pp. 779-810. 

Scholes, R.J., Frost, P.G.H. and Y. Tian (2004). Canopy structure in savannas along a moisture gradient on 

Kalahari sands. – In: Global Change Biology, 10 (3), pp. 292-302. 

Schultka, W. (1974). Möglichkeiten der Gebüsch- und Waldentwicklung in Grasland- und Feuchtsavannen- 

Gebieten von Kenia. Dissertation, Universität Giessen. 

Schultka, W. (1975). Gehölz-Sukzessionen in Regenwald- und Feuchtsavannen-Gebieten bei Kakamega (West- 

Kenia), Oberhessische Naturwissenschaftliche Zeitschrift, 42, pp. 35-48. 

Schulz, A. and T. Wagner (2002). Influence of forest type and tree species on canopy ants (Hymenoptera: 

Formicidae) in Budongo Forest, Uganda. – In: Oecologia, 133, pp. 224-232. 

Scurlock, J.M.O., Asner, G.P. and S.T. Gower (2001). Global Leaf Area Index Data from Field Measurements, 

1932-2000. Data set. Available on-line from the Oak Ridge National Laboratory Distributed Active 

Archive Center, Oak Ridge, Tennessee, U.S.A. Accessed 16 July, 2008. . 

Seed, E.D. and D.J. King (2003). Shadow brightness and shadow fraction relations with effective leaf area 

index: importance of canopy closure and view angle in mixedwood boreal forest. – In: Canadian Journal 

of Remote Sensing, 29 (3), pp. 324-335. 

Sellers, P.J., Meeson, B.W., Hall, F.G., Asrar, G., Murphy, R.E., Schiffer, R.A., Bretherton, F.P., Dickinson, 

R.E., Ellingson, R.G., Field, C.B., Huemmrich, K.F., Justice, C.O., Melack, J.M., Roulet, N.T., Schimel, 

D.S. and P.D. Try (1995). Remote Sensing of the Land Surface for Studies of Global Change: Models 

Algorithms Experiments. – In: Remote Sensing of Environment, 51, pp. 3-26.

References 

Sellers, P.J., Tucker, C.J., Collatz, G.J., Los, S.O., Justice, C.O., Dazlich, D.A. and D.A. Randallet (1994). 

A global l°x1° NDVI data set for climate studies. Part 2: The generation of global fields of terrestrial 

biophysical parameters from the NDVI. – In: International Journal of Remote Sensing, 15 (17), pp. 3519- 

3545. 

Sellers, P.J., Los, S.O., Tucker, C.J., Justice, C.O., Dazlich, D.A., Collatz, G.J. and D.A. Randall (1996). A 

revised land surface parameterization (SiB2) for atmospheric GCM’s. Part II: The generation of global 

fields of terrestrial biophysical parameters from satellite data. – In: Journal of Climate, 9, pp. 706-737. 

Shabanov, N.V., Knyazikhin, Y., Baret, F. and R.B. Myneni (2000). Stochastic Modeling of Radiation Regime 

in Discontinuous Vegetation Canopies. – In: Remote Sensing of Environment, 74, pp. 125-144. 

Shabanov, N.V., Wang, Y., Buermann, W., Dong, J., Hoffman, S., Smith, G.R., Tian, Y., Knyazikhin, Y. and 

R.B. Myneni (2003). Effect of foliage spatial heterogeneity in the MODIS LAI and FPAR algorithm over 

broadleaf forests. – In: Remote Sensing of Environment, 85, pp. 410-423. 

Shabanov, N.V., Huang, D., Yang, W., Tan, B., Knyazikhin, Y., Myneni, R.B., Ahl, D.E., Gower, S.T., Huete, 

A.R., Aragão, L.E. and Y.E. Shimabukuro (2005). Optimization of the MODIS LAI and FPAR algorithm 

performance over broadleaf forests. – In: IEEE Transactions on Geoscience and Remote Sensing, 43 (8), 

pp. 1855-1865. 

Shabanov, N.V., Huang, D., Knjazikhin, Y., Dickinson, R.E. and R.B. Myneni (2007). Stochastic radiative transfer 

model for mixture of discontinuous vegetation canopies. – In: Journal of Quantitative Spectroscopy and 

Radiative Transfer, 107, pp. 236-262. 

Sheil, D. (1995). A critique of permanent plot methods and analysis with examples from Budongo Forest, 

Uganda. – In: Forest Ecology and Management, 77, pp. 11-34. 

Sheil, D. (1998). A half century of permanent plot observation in Budongo Forest, Uganda: histories, highlights 

and hypotheses. – In: Forest Biodiversity Research, Monitoring and Modeling: Conceptual background 

and Old World case studies. Proceedings from the 1995 Smithsonian MAB Washington Symposium, 

M.A.B. UNESCO, Paris, pp. 399-428. 

Sheil, D. (1999). Developing tests of successional hypotheses with size-structured populations, and an 

assessment using long-term data from a Ugandan rain forest. – In: Vegetatio, 140 (1), pp. 117-127. 

Smith F.W., Sampson A.D. and N.J. Long (1991). Comparison of leaf area index estimates from tree allometrics 

and measured light interception. – In: Forest Science, 37, pp. 1682 -1688. 

Smith, N.J. (1993). Estimating plant area index and light extinction coefficients in stands of Douglas-fir 

(Pseudotsuga Menziesii). – In: Canadian Journal of Forest Research, 23, pp. 317-321. 

Smolander, H. and P. Stenberg (1996). Response of LAI-2000 estimates to changes in plant surface area index 

in a Scots pine stand. – In: Tree Physiology, 16, pp. 345-349. 

Song, C., Woodcock, C.E., Pax-Lenny, M. and S.A. Macomber (2001). Classification and change detection using 

Landsat TM date – when and how to correct atmospheric effects. – In: Remote Sensing of Environment, 

75 (2), pp. 230-244. 

Spanner, M.A., Pierce, L.L., Peterson, D.L. and S.W. Running (1990a). Remote sensing of temperate coniferous 

forest leaf area index – the influence of canopy closure, understory vegetation and background reflectance. 

– In: International Journal of Remote Sensing, 11, pp. 95-111. 

Spanner, M.A., Pierce, L.L., Running, S.W. and D.L. Peterson (1990b). The seasonality of AVHRR data of 

temperate coniferous forests: Relationship with leaf area index. – In: Remote Sensing of Environment, 

33, pp. 97-112 

167

168 

Spanner, M.A., Johnson, L. and J. Miller (1994). Remote sensing of seasonal leaf area index across the Oregon 

Transect. - In: Ecological Applications, 4 (2), pp. 258-271. 

Steinbrecher, R. (2004). Impact of changing biodiversity in tropical agro-forest ecosystems in Africa on the 

exchange potential of trace gases between ecosystems and atmosphere. – In: BIOTA East Africa, Final 

report phase I: 2001-2004 

Steininger, M.K. (2000). Satellite estimation of tropical secondary forest above-ground biomass: data from 

Brazil and Bolivia. – In: International Journal of Remote Sensing, 21 (6 & 7), pp. 1139–1157. 

Stenberg, P., Linder, S., Smolander, H. and J. Flower-Ellis (1994). Performance of the LAI-2000 plant canopy 

analyzer in estimating leaf area index of some Scots pine stands. – In: Tree Physiology, 14, pp. 981-995. 

Stenberg, P., Rautiainen, M., Manninen, T., Voipio, P. and H. Smolander (2004). Reduced Simple Ratio Better 

than NDVI for Estimating LAI in Finnish Pine and Spruce Stands. – In: Silva Fennica, 38 (1), pp. 3-15. 

Stork, N.E., Balston, J., Farquhar, G.D., Franks, P.J., Holtum, J.A.M. and M.J. Liddell (2007). Tropical rainforest 

canopies and climate change. – In: Australian Ecology, 32, pp. 105-112. 

Strachan, I.B. and J.H. McCaughey (1996). Spatial and Vertical Leaf Area Index of a Decidouous Forest 

Resolved Using the LAI-2000 Plant Canopy Analyzer. – In: Forest Science, 42 (2), pp. 176-182. 

Strachan, I.B. and J.H. McCaughey (1996). Spatial and vertical leaf area index of a deciduous forest resolved 

using the LAI-2000 Plant Canopy Analyzer. – In: Forest Science, 42, pp. 176-181. 

Suits, G.H. (1972). The Calculation of the Directional Reflectance of a Vegetative Canopy. – In: Remote Sensing 


Tan, B., Hu, J., Zhang, P., Huang, D. and N. Shabanov (2005). Validation of Moderate Resolution Imaging 

Spectroradiometer leaf area index product in croplands of Alpilles, France. – In: Journal of Geophysical 

Research, 110, D01107, doi:10.1029/2004JD004860. 

Thomas, S.C. and W.E. Winner (2000). A rotated ellipsoidal angle density function improves estimation of foliage 

inclination distributions in forest canopies. – In: Agricultural and Forest Meteorology, 100, pp. 19-24. 

Tian, Y., Woodcock, C.E., Wang, Y., Privette, J.L., Shabanov, N.V., Zhou, L., Zhang, Y., Buermann, W., Dong, 

J., Veikkanen, B., Häme, T., Andersson, K., Ozdogan, M., Knyazikhin, Y. and R.B. Myneni (2002a). 

Multiscale analysis and validation of the MODIS LAI product over Maun, Botswana, I. Uncertainty 

Assessment. – In: Remote Sensing of Environment, 83, pp. 414-430. 

Tian, Y., Zhang, Y., Knyazikhin, J., Myneni, R.B., Glassy, J., Dedieu, G. and S.W. Running (2000). Prototyping 

of MODIS LAI and FPAR algorithm with LASUR and LANDSAT data. – In: IEEE Transactions on 

Geoscience and Remote Sensing, 38 (5), pp. 2387-2401. 

Tian, Y., Wang, Y., Zhang, Y., Knyazikhin, Y., Bogaert, J. and R.B. Myneni (2002a). Radiative transfer based 

scaling of LAI/FPAR retrievals from reflectance data of different resolutions. – In: Remote Sensing of the 


Tian, Y., Woodcock, C.E., Wang, Y., Privette, J.L., Shabanov, N.V., Zhou, L., Zhang, Y., Buermann, W., Dong, 

J., Veikkanen, B., Häme, T., Andersson, K., Ozdogan, M., Knyazikhin, Y. and R.B. Myneni (2002b). 

Multiscale analysis and validation of the MODIS LAI product over Maun, Botswana, II. Sampling 

strategy. – In: Remote Sensing of Environment, 83, pp. 431-441. 

Turner, D.P., Cohen, W.B., Kennedy,R.E., Fassnacht, K.S. and J.M. Briggs (1999). Relationships between Leaf 

Area Index and Landsat TM Spectral Vegetation Indices across Three Temperate Zone Sites. – In: Remote 

Sensing of the Environment, 70, pp. 52-68.

References 

USGS (2005). Landsat: A Global Land-Observation Project. Accessed 18 December, 2007. 

. 

USGS (2006a). ASTER Shortwave Infrared (SWIR) Crosstalk-Corrected Products. Accessed 16 November, 

2007. . 

USGS (2006b). Data Set characteristics of MODIS/Terra Leaf Area Index/FPAR 8-day L4 Global 1km SIN 

Grid. Accessed 19 November, 2007. . 

Van Gardingen, P.R., Jackson, G.E., Hernandez-Daumas, S., Russell, G. and L. Sharp (1999). Leaf area index 

estimates obtained for clumped canopies using hemispherical photography. – In: Agricultural and Forest 

Meteorology, 94, pp. 243-257. 

Verhoef, W. and N.J.J. Bunnik (1981). Influence of crop geometry on multispectral reflectance determined 

by the use of canopy reflectance models. – In: NLR MP 81042 U / Proceedings of the International 

Colloquium on spectral signatures of objects in remote sensing, Avignon (France), 8-11 September 1981, 

pp. 273-289. 

Verhoef, W. (1984). Light scattering by leaf layers with application to canopy reflectance. – In: Remote Sensing 


Vermote, E.F., Tanré, D., Deuze, J.L., Herman, M. and J.J. Morcrette (1997). Second simulation of the satellite 

signal in the solar spectrum, 6S: an overview. – In: IEEE Transactions on Geoscience and Remote 

Sensing, 35, pp. 675-686. 

Vertessy, R.A., Benyon, R.G., O’Sullivan, S.K. and P.R. Gribben (1995), Relationships between stem diameter, 

sapwood area, leaf area and transpiration in a young mountain ash forest. – In: Tree Physiology, 15, 

pp. 559-567. 

Walter, J.-M.N., Fournier, R.A., Soudani, K. and E. Meyer (2003). Integrating clumping effects in forest canopy 

structure: an assessment through hemispherical photographs. – In: Canadian Journal of Remote Sensing, 

29 (3), pp. 388-410. 

Wang, Y.S. and D.R. Miller (1987). Calibration of the hemispherical photographic technique to measure leaf 

area index distributions in hardwood forests. – In: Forest Science 33, pp. 210-226. 

Wang, Y., Tian, Y., Zhang, Y., El-Saleous, N., Knyazikhin, Y., Vermote, E. and R.B. Myneni (2001). Investigation 

of product accuracy as a function of input and model uncertainties: Case study with SeaWiFS and MODIS 

LAI/FPAR algorithm. – In: Remote Sensing of Environment, 78, pp. 296-311. 

Wang, Y., Buermann, W., Stenberg, P., Smolander, H., Häme, T., Tian, Y., Hu, J., Knyazikhin, Y. and R.B. 

Myneni (2003). A new parameterization of canopy spectral response to incident solar radiation: Case 

study with hyperspectral data from pine dominant forest. – In: Remote Sensing of the Environment, 85 

(3), pp. 304-315. 

Wang, Y., Woodcock, C.E., Buermann, W., Stenberg, P., Voipio, P., Smolander, H. Häme, T., Tian, Y., Hu, J., Knyazikhin, 

Y. and R.B. Myneni (2004). Evaluation of the MODIS LAI algorithm at a coniferous forest site in Finland. 


Wang, Q., Tenhunen, J., Quoc Dinh, N., Reichstein, M., Otieno, D., Granier, A. and K. Pilegarrd (2005). 

Evaluation of seasonal variation of MODIS derived leaf area index at two European deciduous broadleaf 

forest sites. – In: Remote Sensing of Environment, 96, pp. 475-484. 

Warren Wilson, J. (1959). Analysis of the spatial distribution of foliage by two-dimensional point quadrats. 

– In: New Phytology, 58, pp. 92-101. 

169

170 

Warren Wilson, J. (1960). Inclined point quadrats. – In: New Phytology, 59, pp. 1-8. 

Warren Wilson, J. (1963). Estimation of foliage denseness and foliage angle by inclined point quadrats. 

– In: Australian Journal of Botany, 11, pp. 95-105. 

Watson, D.J. (1947). Comparative physiological studies in the growth of field crops. I. Variation in net 

assimilation rate and leaf area between species and varieties, and within and between years. – In: Annals 

of Botany, 11, pp. 41-76. 

Weiss, M. (2006). CAN-EYE output variables. Definitions and theoretical background. Accessed 21 September, 

2007. . 

Weiss, M. and F. Baret (1999). Evaluation of canopy biophysical variable retrieval performances from the 

accumulation of large swath satellite data. – In: Remote Sensing of Environment, 70, pp. 293-306. 

Weiss, M., Baret, F., Myneni, R.B., Pragnère, A. and Y. Knyazikhin (2000). Investigation of a model inversion 

technique to estimate canopy biophysical variables from spectral and directional reflectance data. 

– In: Agronomie, 20, pp. 3-22. 

Weiss, M., Baret, F., Smith, G. J., Jockheere, I. and P. Coppin (2004). Review of methods for in situ leaf area 

index (LAI) determination: Part II. Estimation of LAI, errors and sampling. – In: Forest Ecology and 

Management, 121, pp. 37-53. 

Welles, J.M. (1990). Some indirect methods of estimating canopy structure. – In: Remote Sensing Reviews, 5, 

pp. 31-43. 

Welles, J.M. and J.M. Norman (1991). Instrument for indirect measurement of canopy architecture. 

– In: Agronomy Journal, 83 (5), pp. 818-825. 

WGCW (2007). Official Website of the Working Group on Calibration and Validation. Accessed 15 August, 

2007. . 

White, F. (1983). The vegetation of Africa. A descriptive memoir to accompany the Unesco/AETFAT/UNSO 

vegetation map of Africa. Natural resources research (XX). UNESCO, Switzerland, 356 p. 

White, J., Running, S., Nemani, R., Keane, R. and K. Ryan (1997). Measurement and remote sensing of LAI in 

Rocky Mountain montane ecosystems. – In: Canadian Journal of Forest Research, 27, pp. 1714-1727. 

White, M.A., Asner, G.P., Nemani, R.R., Privette, J.L. and S.W. Running (2000). Measuring Fractional Cover 

and Leaf Area Index in Arid Ecosystems: Digital Camera, Radiation Transmittance, and Laser Altimetry 

Methods. – In: Remote Sensing of the Environment, 74, pp. 45-57. 

Whitmore, T.C. (2003). An Introduction to Tropical Rain Forests. Oxford University Press, Oxford, 296 p. 

Wiegand, C.L., Gausman, H.W., Cuellar, J.A., Gerbermann, A.H. and A.J. Richardson (1974). Vegetation 

density as deduced from ERTS-1 MSS response. – In: Third ERTS Symposium, NASA-SP-351, I (A), 

pp. 93-116. 

Wilson, D., Alm, J., Riutta, T., Laine, J., Byrne, K.A., Farrell, E.P. and E.-S. Tuittila (2007). A high resolution 

green area index for modelling the seasonal dynamics of CO2 exchange in peatland vascular plant 

communities. – In: Plant Ecology, 190 (1), pp. 37-51. 

Wirth, R., Weber, B. and R.J. Ryel (2001). Spatial and temporal variability of canopy structure in a tropical 

moist forest. – In: Acta Oecologica, 22, pp. 235-244.

References 

Wright, S.J. (2005). Tropical forests in a changing environment. – In: Trends in Ecology and Evolution, 20 (10), 

pp. 553-560. 

Wright, S.J. and H.C. Muller-Landau (2006). The future of tropical forest species. – In: Biotropica, 38 (3), 

pp. 287-301. 

Wolfe, R.E., Nishihama, M., Fleig, A.J., Kuyper, J.A., Roy, D.P., Storey, J.C. and F.S. Patt (2002). Achieving 

sub-pixel geolocation accuracy in support of MODIS land science. – In: Remote Sensing of Environment, 

83 (1-2), pp. 31-49. 

Wulder, M.A., LeDrew, E.F., Franklin, S.E. and M.B. Lavigne (1998). Aerial image texture information in the 

estimation of northern deciduous and mixed wood forest leaf area index (LAI). – In: Remote Sensing of 

the Environment, 64, pp. 64-76. 

Wulder, M.A. and S.E. Franklin (2003). Remote Sensing of Forest Environments: Concepts and Case Studies. 

Kluwer Academic Publishers, 552 p. 

Xiao, X., Zhang, Q., Saleska, S., Hutyra, L., De Camargo, P., Wofsy, S., Frolking, S., Boles, S., Keller, M. and 

B. Moore (2005). Satellite-based modeling of gross primary production in a seasonally moist tropical 

evergreen forest. – In: Remote Sensing of Environment, 94, pp. 105-122. 

Yamashita, M., Yoshimura, M. and T. Nakashizuka (2004). Cloud cover estimation using multitemporal 

hemisphere images. – In: Proceedings of the XXth ISPRS Congress, Geo-Imagery Bridging Continents, 

12-23 July 2004, Istanbul, Turkey. Accessed 2 January, 2008. < http://www.isprs.org/istanbul2004/ 

comm7/papers/162.pdf>. 

Yang, W., Tan, B., Huang, D., Rautiainen, M., Shabanov, N.V., Wang, Y., Privette, J.L., Huemmrich, K.F., 

Fensholt, I., Weiss, M., Ahl, D.E., Gower, S.T., Nemani, Y., Knyazikhin, Y. and R.B. Myneni (2006). 

MODIS leaf area index products: From validation to algorithm improvement. – In: IEEE Transactions on 


Zhang, Y., Tian, Y., Knyazikhin, J., Martonchik, J.V., Diner, D.J., Leroy, M. and R.B. Myneni (2000). 

Prototyping of MISR LAI and FPAR algorithm with POLDER data over Africa. – In: IEEE Transactions 

on Geoscience and Remote Sensing, 38 (5), pp. 2402-2418. 

Zhang, X., Friedl, M.A., Schaaf, C.B. and A.H. Strahler (2004). Climate controls on vegetation phenological 

patterns in northern mid- and high latitudes inferred from MODIS data. – In: Global Change Biology, 10, 

pp. 1133-1145. 

Zhang, Q., Xiao, X, Braswell, B., Linder, E., Baret, F. and B. Moore III (2005). Estimating light absorption by 

chlorophyll, leaf and canopy in a deciduous broadleaf forest using MODIS data and a radiative transfer 

model. – In: Remote Sensing of Environment, 99, pp. 357-371. 

Zhang, Q., Xiao, X., Braswell, B., Linder, E., Ollinger, S., Smith, M.-L., Jenkins, J.P., Baret, F., Richardson, 

A.D., Moore, B. and R. Minocha (2006). Characterization of seasonal variation of forest canopy in a 

temperate deciduous broadleaf forest, using daily MODIS data. – In: Remote Sensing of Environment, 

105, pp. 189-203. 

171

172

Appendix of Figures 

Figure A-1 

Theoretical field of view of LAI-2000 PCA with 45° view cap in a) vertical projection and b) horizontal projection (solid 

lines represent field of view assuming an average canopy height of 20 m, dashed lines represent average canopy 

height of 10 m). 

173

174 

Figure A-2 

Land cover classification, 2003 

Near natural & old secondary forest 

Secondary forest 

Bushland / shrubs 

Secondary bushland - Psidium guajava 

Grassland with scattered trees 

Grassland 

Plantation forest - Pinus patula 

Plantation forest - Bischoffia javanica 

Tea plantation 

Agricultural land 

Water 

Others 

T. Lung, 2007 

18/05/2003 Landsat 7 (ETM+) 

10/01/2003 Landsat 7 (ETM+) 

0 2.55km Land cover classification for Kakamega Forest, based on Landsat 7 ETM+ data of 2003 (courtesy of T. Lung, BIOTA 

E02, for methodology refer to Lung 2004).


Figure A-3 

Interactive classification process in CAN-EYE. a) unclassified and b) classified hemispherical photographs covering 

one transect. 

175

176 

Left late 

Figure 5. Schema of the experimental design used. 

Diameter of the image 

Left corner 

Left lateral ruler 

Camera 

Background alignment 

Front Nail 

alignement 

 

Optical centre 

Perpandicular ruler 

Right corner 

Right lateral ruler 

Figure 6. Example of an image of the experimental design taken with the hemispherical camera and 

used for the calibration of the projection function. The horizontal doted yellow line corresponds to the 

diameter of the image passing through the optical centre (defined by its coordinates as measured 

previously). The camera is aligned Background thanks to the front nail and background line. 

alignement 

A dedicated matlab code ‘calib_project.m’ was developed to display the results. 

Front nail 

alignement 

Axis of the design 

 

Right la 

Perpandicular ruler 

Figure A-4 Figure 5. Schema of the experimental design used. 

Diameter of the image 

y 

Left lateral ruler 

W 

y 

x 

H 

Background alignment 

x 

Camera 

Experimental design to assess the projection quality of the used camera systems and to retrieve the projection 

function (Baret 2004b). 

Right lateral ruler


Figure A-5 

Figure A-6 

Results of different LAI measures for each ESU in Budongo Forest. 

Results of different LAI measures for each ESU in Kakamega Forest. 

177

178 

Figure A-7 a) ρred 

, b) ρNIR and c) ρSWIR of band 4 and d) ρ of band 5 in [%] derived from ASTER data and plotted per 

SWIR 

forest stage. Upper and lower box boundaries indicate 25th and 75th percentiles, whiskers indicate minimum and 

maximum values except outliers (defined as more than 3 box lengths from 25 th or 75 th percentile, displayed as point). 

Mean is shown as solid line.


Figure A-8 a) ρred 

, b) ρNIR and c) ρ in [%] derived from SPOT data and plotted per forest stage for Kakamega Forest. Upper 

SWIR 

and lower box boundaries indicate 25th and 75th percentiles, whiskers indicate minimum and maximum values except 

outliers (defined as more than 3 box lengths from 25 th or 75 th percentile, displayed as points). Mean is shown as solid 

line. 

Bivariate plot of LAI 

Figure A-9 e (ASTER) in 15 m and 1000 m spatial resolution for the forest compartments included in LAI 

product validation (N=48). 

179

180 

Figure A-10 

not retrieved 

empirical 



Relative frequencies of C4 MODIS LAI algorithm 

usage for the years 2000-2006 over the Budongo 

Forest test site.

Appendix of Tables 

Table A-1 

Logging history of Budongo Forest compartments according to Forest Department Offices, logging return sheets of 

sawmills and old working plans (Plumptre 1996). Note that the Budongo forest block is further split up in the 

Waibira, Nyakafunjo and Biiso blocks (from east to west). The table has not been updated. Various additional logging 

activities could be observed during the time of field work. 

Forest Block Comp. Area 

(ha) 

Date 

logged 

Total 

(m 3 ha -1 ) 

Mahogany 

(m 3 ha -1 ) 

Date 

treated 

181 

Arboricide 

Biiso B1 582 1935 19.9 12.3 1957-58 n/a 

(l ha -1 ) 

n/a 1981-82 21.5 12.9 None n/a 

B2 600 1936-38 40.2 33.9 1958-59 20.4 

B3 867 1939-40 34.9 23.7 1959-60 n/a 

B4 748 1941-42 34.8 19.8 1955-57 20.0 

n/a 1985-92 Some pitsawing n/a n/a 

B5 871 1943-44 34.2 25.2 1960-61 42.6 

B6 394 1943 14.0 10.4 1961-63 42.6 

B7 288 1944 20.2 14.7 n/a n/a 

Nyakafunjo N1 412 1945 58.7 47.1 1962-63 n/a 

N2 630 1945-47 46.2 33.1 1955-56 11-1 

N3 620 1947-52 80.0 39.1 1959-61 22.0 

N4 341 1952-54 94.0 48.0 1960-62 n/a 

N5 568 1954 51.5 41.2 1963-64 38.1 

N6 600 1956-57 59.8 49.4 1956-58 41.6 

N7 474 1958-59 31.9 26.1 1957-58 44.8 

N8 498 1957-58 25.7 20.8 1957 47.0 

N9 498 1958-59 15.2 10.7 1958 42.3 

N10 618 1959-60 30.2 23.5 1958-60 50.4 

N11 564 1960 26.5 21.7 1958-60 42.6 

N12 389 1960 3.0 2.5 1959-60 32.9 

N13 610 1960 13.0 8.5 1960 40.3 

N14 535 1961-62 n/a n/a None n/a 

N15 777 Unlogged n/a n/a None n/a

182 


(ha) 

Date 

logged 

Total 

(m 3 ha -1 ) 

Mahogany 

(m 3 ha -1 ) 

Date 

treated 

Arboricide 

WAIBIRA W16 543 1961-62 21.7 14.3 1961-62 n/a 

(l ha -1 ) 

W17 609 Unlogged n/a n/a None n/a 

W18 681 1960-61 28.8 19.3 1960-62 43.6 

W19 886 1962-63 25.6 16.3 1961-63 41.4 

W20 569 1963/64 30.8 20.3 1962-63 41.4 

W21 1116 1963-64 36.1 23.9 1963-64 29.1 

W22 1036 1965-67 35.9 23.0 1966 n/a 

W23 736 1966-68 26.4 18.5 1966 n/a 

W24 710 1967-69 41.2 9.3 1966-67 n/a 

W25 1181 1968-69 9.6 5.4 1970-71 n/a 

W26 781 1970-71 30.6 19.1 1971-72 36.0 

W27 791 1970-71 n/a n/a 1970-71 46.5 

W28 1010 1972-73 51.4 29.1 1972-73 29.5 

W29 744 1972-74 64.0 40.5 None n/a 

W30 580 1974-75 8.1 7.5 None n/a 





W35 680 Pitsawn n/a n/a None n/a 


W37 566 1978-83 22.2 10.7 None n/a 

W38 422 1983 (Partially pitsawn) None n/a 

W39 667 1974-76 43.5 25.5 None n/a 





KANIYO-PABIDI K1 177 1970-72 11.7 9.4 n/a 

K2 533 1970-72 3.1 2.3 n/a 

K3 330 1977 n/a n/a None 

K4 630 1987-92 20.7 12.6 None n/a 

K5 311 1985-87 25.6 17.2 None n/a 

K7 526 1987-89 11.1 9.0 None n/a 

K8 169 1986-89 31.5 15.1 None n/a 

K11 420 Unlogged n/a n/a None n/a 

K12 1097 Unlogged n/a n/a None n/a 

K13 625 Unlogged n/a n/a None n/a



(ha) 

Date 

logged 

Total 

(m 3 ha -1 ) 

Mahogany 

(m 3 ha -1 ) 

Date 

treated 

183 

Arboricide 

SIBA S1 810 1963-69 40.9 23.5 1972-73 34.4 

Table A-2 

Quality control FPARLAI_QC SDS of MOD15A2 (USGS 2006b). 

Bit range Bit setting Description 

0 - 1 MODLAND QA 

(l ha -1 ) 

S2 602 1969-70 n/a n/a 1971-72 40.0 

S3 528 1966-70 n/a n/a n/a n/a 

S4 799 1972-77 28.9 n/a n/a n/a 

S5 829 1971 n/a n/a n/a n/a 

S6 588 1971 n/a n/a 1972-75 n/a 

S7 686 1990 n/a n/a None n/a 

S8 446 1979/90 48.3 26.3 None n/a 

00 Best possible 

01 OK, but not the best 

10 Not produced due to cloud 

11 Not produced due to other reasons 

2 DEAD DETECTOR 

3 - 4 CLOUD STATE 

5 - 7 SCF_QC 

00 Detectors apparently fine for up to 50% of channels 1,2 

01 Dead detectors caused >50% adjacent detector retrieval 

00 Significant clouds NOT present (clear) 

01 Significant clouds WERE present 

10 Mixed cloud present on pixel 

11 cloud state not defined, assumed clear 

000 Main(RT) method used with the best possible results 

001 Main(RT) method used with saturation 

010 Main(RT) method failed due to geometry problems, empirical method used 

011 Main(RT) method failed due to problems other than geometry, empirical method used 

100 Pixel could not be retrieved

184 

Table A-3 

Additional FparExtra_QC SDS of MOD15A2 (USGS 2006b). 

Bit range Bit setting Description 

0-1 SCF_QC 

Table A-4 

00 Land 

01 Shore 

10 Freshwater 

11 Ocean 

2 SNOW_ICE 

0 No snow, ice detected 

1 Snow, ice detected 

3 AEROSOL 

4 CIRRUS 

0 No or low atmospheric aerosol levels detected 

1 Average or high aerosol levels detected 

0 No cirrus detected 

1 Cirrus was detected 

5 INTERNAL_CLOUD_MASK 

0 No clouds detected 

1 Clouds WERE detected 

6 CLOUD_SHADOW 

0 No cloud shadow detected 

7 SCF_MASK 

1 Cloud shadow was detected 

0 Custom SCF mask, EXCLUDE this pixel 

1 Custom SCF mask, INCLUDE this pixel 

Fill values for non-vegetated surfaces in MOD15A2 (USGS 2006b). 

DN Non-Terrestrial Value Index 

255 Fill Value 

254 Perennial salt or inland fresh water 

253 Barren, sparse vegetation (rock, tundra, desert) 

252 Perennial snow, ice 

251 Permanent wetlands/inundated marshland 

250 Urban/built-up 

249 Unclassified/not able to determine


Table A-5 

Image parameters for camera systems used in Budongo and Kakamega Forest. 

Camera Image size (lines, rows) Optical center (line, row) Diameter 

Nikon Coolpix 4300 1704, 2272 1180, 862 1593 

Nikon Coolpix 4500 1704, 2272 1152, 831 1584 

Table A-6 Spearman’s rank correlation coefficient r for LAI measures and surface reflectance derived from SPOT (N=29) 

s 

and ASTER data (N=26) for all forest stages (** significant at p=0.01, * significant at p=0.05) of Budongo Forest. 

Sensor Band LAI e (LAI2000) LAI e (DHP) LAI (DHP) 

SPOT-HRVIR 1 -0.54(**) -0.46(*) -0.34 

2 -0.48(**) -0.44(*) -0.34 

3 0.10 0.25 0.11 

4 -0.54(**) -0.38(*) -0.33 

ASTER 1 -0.56(**) -0.50(**) -0.33 

2 -0.58(**) -0.51(**) -0.32 

3 -0.14 0.15 -0.01 

4 -0.61(**) -0.46(*) -0.46(*) 

5 -0.58(**) -0.51(**) -0.33 

6 -0.59(**) -0.54(**) -0.35 

7 -0.48(*) -0.49(*) -0.33 

8 -0.59(**) -0.51(**) -0.29 

9 -0.46(*) -0.53(**) -0.35 

Table A-7 Spearman’s rank correlation coefficient r for LAI measures and SVIs derived from SPOT (N=29) and ASTER data 

s 

(N=26) for all forest stages (** significant at p=0.01, * significant at p=0.05) of Budongo Forest. 

Sensor SVI LAI e (LAI2000) LAI e (DHP) LAI (DHP) 

SPOT-HRVIR SR 0.45(*) 0.48(**) 0.37(*) 

NDVI 0.44(*) 0.46(*) 0.36 

NDVI C 0.45(*) 0.42(*) 0.31 

MSR 0.45(*) 0.48(**) 0.37(*) 

RSR 0.43(*) 0.43(*) 0.32 

NDMI 0.42(*) 0.50(**) 0.37 

ASTER SR 0.48(*) 0.62(**) 0.45(*) 

NDVI 0.48(*) 0.62(**) 0.44(*) 

NDVI C 0.53(**) 0.50(**) 0.36 

MSR 0.48(*) 0.60(**) 0.45(*) 

RSR 0.51(**) 0.59(**) 0.40(*) 

NDMI 0.43(*) 0.62(**) 0.46(*) 

185

186 


s 

(N=8) for the late forest stage (** significant at p=0.01, * significant at p=0.05) of Budongo Forest. 


SPOT-HRVIR SR -0.53 -0.44 -0.54 

NDVI -0.55 -0.58 -0.60 

NDVI C -0.60 -0.70(*) -0.72(*) 

MSR -0.53 -0.47 -0.50 

RSR -0.59 -0.60 -0.66(*) 

NDMI -0.62 -0.61 -0.65(*) 

ASTER SR -0.21 -0.10 -0.45 

NDVI -0.14 -0.14 -0.52 

NDVI C -0.52 -0.29 -0.24 

MSR -0.21 -0.10 -0.45 

RSR -0.33 -0.31 -0.62 

NDMI -0.21 -0.10 -0.45 


s 

(N=18) for early and intermediate forest stages (** significant at p< 0.01) of Budongo Forest. 


SPOT SR 0.81 (**) 0.80 (**) 0.64 (**) 

NDVI 0.80 (**) 0.79 (**) 0.64 (**) 

NDVI C 0.75 (**) 0.76 (**) 0.64 (**) 

MSR 0.81 (**) 0.80 (**) 0.65 (**) 

RSR 0.78 (**) 0.78 (**) 0.63 (**) 

NDMI 0.81 (**) 0.84 (**) 0.66 (**) 

ASTER SR 0.84 (**) 0.86 (**) 0.65 (**) 

NDVI 0.84 (**) 0.86 (**) 0.66 (**) 

NDVI C 0.67 (**) 0.72 (**) 0.63 (**) 

MSR 0.84 (**) 0.86 (**) 0.65 (**) 

RSR 0.84 (**) 0.84 (**) 0.62 (**) 

NDMI 0.76 (**) 0.88 (**) 0.71 (**)


Table A-10 Theil-Sen and linear OLS regression models based on LAI and SVIs for early and intermediate forest stages 

e (LAI2000) 

of Budongo Forest together with the respective R2 . 


SPOT-HRVIR SR 3.427+0. 593 X 0.88 3.036+0.669 X 0.86 

NDVI 0. 594+0.026 X 0.86 0.571+0.029 X 0.88 

NDVI c 0. 258+0.050 X 0.83 0.193+0.059 X 0.90 

MSR 0. 897+0.124 X 0.86 0.818+0.139 X 0.87 

RSR 0.912+0.693 X 0.88 0.606+0.759 X 0.89 

NDMI 0. 018+0.043 X 0.86 -0.017+0.049 X 0.88 

ASTER SR 3.009+0.487 X 0.88 3.168+0.468 X 0.89 

NDVI 0.560+0.025 X 0.90 0.557+0.026 X 0.90 

NDVI c 0.207+0.065 X 0.87 0.148+0.073 X 0.91 

MSR 0.787+0.108 X 0.89 0.104+0.819 X 0.90 

RSR 0.849+0.707 X 0.89 0.465+0.743 X 0.92 

NDMI 0.147+0.029 X 0.85 0.139+0.031 X 0.86 

Table A-11 Theil-Sen and linear OLS regression models based on LAI and SVIs for early and intermediate forest stages of 

(DHP) 

Budongo Forest together with the respective R2 . 



NDVI 0.122+0.049 X 0.64 0.463+0.031 X 0.79 

NDVI c 0.424+0.014 X 0.60 -0.020+0.061 X 0.79 

MSR 0.645+0.112 X 0.65 0.325+0.142 X 0.75 

RSR -0.545+0.628 X 0.68 -2.037+0.773 0.76 

NDMI 0.092+0.041 X 0.67 -0.191+0.50 X 0.77 

ASTER SR 2.121+0.426 X 0.76 1.491+0.485 X 0.80 

NDVI 0.503+0.023 X 0.76 0.463+0.027 X 0.83 

NDVI c 0.054+0.061 X 0.65 -0.110+0.075 X 0.81 

MSR 0.605+0.093 X 0.76 0.445+0.108 X 0.81 

RSR -0.523+0.023 X 0.74 -2.133+0.761 X 0.81 

NDMI 0.069+0.028 X 0.76 0.024+0.033 X 0.80 

187

188 

Table A-12 

Spearman’s rank correlation coefficient rs for LAI and texture measures derived from ASTER data (N=8) for the late 

forest stage (** significant at p=0.01, * significant at p=0.05) of the Budongo Forest test site. Due to the large amount 

of data only texture measures with at least one significant correlation to any LAI measure are shown. 

Sensor Kernel Band Texture measure LAI e (LAI2000) LAI e (DHP) LAI (DHP) 

ASTER 3*3 3 Homogeneity -0.88(**) -0.67 -0.48 

3 Entropy 0.64 0.76 (*) 0.29 

4 Variance 0.31 0.81 (*) 0.60 

7 Correlation 0.38 0.81 (*) 0.60 

5*5 3 Homogeneity -0.79 (*) -0.50 -0.29 

4 Variance 0.41 0.83 (*) 0.55 

7*7 3 Homogeneity -0.79 (*) -0.50 -0.29 

4 Variance 0.41 0.83 (*) 0.55 

8 Variance -0.71(*) -0.74 (*) -0.57 

8 Contrast -0.52 -0.71 (*) -0.50 

9*9 4 Variance 0.38 0.86 (**) 0.64 

4 Entropy 0.24 0.74 (*) 0.57 

8 Contrast -0.45 -0.91 (**) -0.69 

11*11 3 Homogeneity -0.81 (*) -0.48 -0.26 

4 Variance 0.41 0.83 (*) 0.55 

8 Variance -0.71(*) -0.74 (*) -0.57 

SPOT-HRVIR 3*3 1 Entropy 0.62 0.71 (*) 0.45 

3 Variance 0.43 0.81 (*) 0.67 

5*5 3 Contrast 0.38 0.71 (*) 0.69 

3 Dissimilarity 0.38 0.71 (*) 0.69 

7*7 3 Contrast 0.38 0.71 (*) 0.69 


9*9 3 Contrast 0.43 0.74 (*) 0.71 (*) 


11*11 3 Contrast 0.38 0.76 (*) 0.62 

Table A-13 Spearman’s rank correlation coefficient r for LAI measures and surface reflectance derived from SPOT data (N=27, 

s 

for LAI N= 25) for all forest stages (** significant at p=0.01, * significant at p=0.05) of Kakamega Forest. 

e (LAI2000) 

Sensor Band LAI e (LAI2000) LAI e (DHP) LAI (DHP) 

SPOT-HRVIR 1 0.51(*) -0.58(**) -0.39(*) 

2 -0.57(**) -0.56(**) -0.32 

3 -0.46(*) -0.24 -0.17 

4 -0.53(**) -0.55(**) -0.36


Table A-14 Spearman’s rank correlation coefficient r for LAI measures and SVIs derived from SPOT data (* significant at p=0.05, 

s 

N =27, for LAI N= 25, all forest stages) for Kakamega Forest. 

e (LAI2000) 


SPOT-HRVIR SR 0.17 0.46(*) 0.43(*) 

NDVI 0.15 0.46(*) 0.41(*) 

NDVI C 0.46(*) 0.55(**) 0.38 

MSR 0.16 0.46(*) 0.42(*) 

RSR 0.36 0.55(**) 0.45(*) 

NDMI 0.12 0.33 0.32 

Table A-15 Spearman’s rank correlation coefficient r for LAI measures and SVIs derived from SPOT data (* significant at p=0.05, 

s 

N =16, early and intermediate forest stages) for Kakamega Forest. 


SPOT-HRVIR SR 0.65(**) 0.75(**) 0.64(*) 

NDVI 0.63(*) 0.75(**) 0.59(*) 

NDVI C 0.66(**) 0.62(*) 0.48 

MSR 0.65(**) 0.75(**) 0.60(*) 

RSR 0.68(**) 0.67(**) 0.55(*) 

NDMI 0.50 0.56(*) 0.42 

Table A-16 Theil-Sen and linear OLS regression models based on LAI and SVIs for early and intermediate forest stages 

e (DHP) 

in Kakamega Forest. 



NDVI 0.739+0.015 X 0.22 0.635+0.039 X 0.63 

NDVI c 0.268+0.059 X 0.51 0.133+0.088 X 0.72 

MSR 3.164+0.125 X 0.32 2.512+0.273 X 0.66 

RSR 1.836+0.949 X 0.58 1.362-0.069 X 0.74 

NDMI 0.643-0.027 X 0.37 0.765-0.057 X 0.57 

189

190 

Table A-17 Spearman’s rank correlation coefficient r for LAI and texture measures derived from SPOT data (N=10) for the late 

s 

forest stage (* significant at p=0.05) of the Kakamega Forest test site. Due to the large amount of data only texture 

measures with at least one significant correlation to any LAI measure are shown. 

Sensor Kernel Band Texture measure LAI e (LAI2000) LAI e (DHP) LAI (DHP) 

SPOT-HRVIR 3*3 1 Variance 0.45 0.73 (*) 0.75 (*) 

1 Entropy 0.50 0.66 (*) 0.72 (*) 

4 Homogeneity -0.43 -0.67 (*) -0.69 (*) 

5*5 1 Entropy 0.21 0.67 (*) 0.75 (*) 

4 Homogeneity -0.57 -0.83 (**) -0.79 (**) 

7*7 4 Homogeneity -0.60 -0.84 (**) -0.75 (*) 

9*9 4 Homogeneity -0.76 (*) -0.67 (*) -0.58

191 

Appendix of Equations 

Atmospheric correction (cf. Chapter 4.2.2) 

Disregarding the adjacency effect and accounting for the directional dependence of direct and diffuse solar 

radiation in rugged terrain, ρ can be approximated iteratively starting with ρ terrain 

0 =0.1 by 



Disregarding the adjacency effect and accounting for the directional dependence of direct and diffuse 

solar radiation in rugged terrain, can be approximated iteratively starting with 

0 

terrain 

=0.1 by 

 

 

 

 

 

y 

x 

v 

z 

E 

z 

y 

x 

E 

y 

x 

z 

E 

y 

x 

b 

z 

z 

L 

y 

x 

DN 

c 

c 

d 

y 

x 

terrain 

i 

terrain 

ground 

d 

s 

s 

v 

v 

v 

, 

, 

, 

, 

cos 

, 

, 

, 

, 

, 

) 

, 

( 

2 

1 

0 

2 

 

 

 

 

 

 

 

 

 

 

 

 

 

, (A.1) 

with 

x, y: horizontal coordinates, corresponding to the georeferenced pixel positions; 

z: elevation in m derived from the DEM; 

 

 

, 

, 

2 v 

z 

L : path radiance for elevation z and viewing geometry 

 

, 

v 

; 

 

v 

v z 

, : ground-to-sensor transmittance for elevation z and viewing angle v 

(direct plus diffuse 

components); 

 

v 

v z 

, : sun-to-ground transmittance (direct beam); 

 

y 

x, 

: angle between the solar ray and the surface normal (illumination angle); 

b(x,y): binary factor: b=1 if pixel receives direct solar beam, otherwise b=0; 

Es: extraterrestrial solar irradiance; 

Ed(x, y, z): diffuse solar flux ; 

Eg(z): global flux (direct plus diffuse solar flux on a horizontal surface); 

terrain 

(i) (x, y): locally varying average terrain reflectance, calculated iteratively (i=1,2,3) and 

terrain 

v (x, y): terrain view factor (range 0-1) calculated from local slope. 

Theoretical field of view of LAI-2000 PCA (cf. Chapter 5.1.1) 

The theoretical field of view of a single LAI-2000 PCA measurement is represented by a sphere segment, 

whose radius can be calculated as 

h 

r 

 

cos 

1 

, (A.2) 

where 

cos is the maximum zenith angle of the outmost ring included in LAI calculation and h is the 

average canopy height. Assuming an average canopy height of 20 m and a maximum zenith angle of 58.1° 

(ring 4) the field of view would thus theoretically include all canopy elements within 37.8 m in horizontal 

direction (cf. Figure A-1a). At 10 m height the theoretical field of view decreases to 18.9 m. The 

horizontal projection in Figure A-1b shows that also the maximum width of a measurement depends on 

the average canopy height. It is represented by a circular sector whose radius r can be calculated by 

equation A.1. Its chord (i.e. is maximum width) is then equivalent to 

) 

2 

(sin 

2 r 

 

, (A.3) 

which in Figure A-1b equals 28.9 m at an average canopy height of 20 m and 14.5 m at an average canopy 

height of 10 m. Theoretically the distance between two measurements should be at least equivalent to r to 

guarantee spatial independence. 

(A.1) 

with 



1 APPENDIX OF EQUATIONS 





0 

terrain 

=0.1 by 

 

 

 

 

 

y 

x 

v 

z 

E 

z 

y 

x 

E 

y 

x 

z 

E 

y 

x 

b 

z 

z 

L 

y 

x 

DN 

c 

c 

d 

y 

x 

terrain 

i 

terrain 

ground 

d 

s 

s 

v 

v 

v 

, 

, 

, 

, 

cos 

, 

, 

, 

, 

, 

) 

, 

( 

2 

1 

0 

2 

 

 

 

 

 

 

 

 

 

 

 

 

 

, (A.1) 

with 



 

 

, 

, 

2 v 

z 


 

, 

v 

; 

 

v 

v z 



components); 

 

v 

v z 


 

y 

x, 






terrain 


terrain 





h 

r 

 

cos 

1 

, (A.2) 

where 








) 

2 

(sin 

2 r 

 

, (A.3) 




path radiance for elevation z and viewing geometry 






0 

terrain 

=0.1 by 

 

 

 

 

 

y 

x 

v 

z 

E 

z 

y 

x 

E 

y 

x 

z 

E 

y 

x 

b 

z 

z 

L 

y 

x 

DN 

c 

c 

d 

y 

x 

terrain 

i 

terrain 

ground 

d 

s 

s 

v 

v 

v 

, 

, 

, 

, 

cos 

, 

, 

, 

, 

, 

) 

, 

( 

2 

1 

0 

2 

 

 

 

 

 

 

 

 

 

 

 

 

 

, (A.1) 

with 



 

 

, 

, 

2 v 

z 


 

, 

v 

; 

 

v 

v z 



components); 

 

v 

v z 


 

y 

x, 






terrain 


terrain 





h 

r 

 

cos 

1 

, (A.2) 

where 








) 

2 

(sin 

2 r 

 

, (A.3) 




; 






0 

terrain 

=0.1 by 

 

 

 

 

 

y 

x 

v 

z 

E 

z 

y 

x 

E 

y 

x 

z 

E 

y 

x 

b 

z 

z 

L 

y 

x 

DN 

c 

c 

d 

y 

x 

terrain 

i 

terrain 

ground 

d 

s 

s 

v 

v 

v 

, 

, 

, 

, 

cos 

, 

, 

, 

, 

, 

) 

, 

( 

2 

1 

0 

2 

 

 

 

 

 

 

 

 

 

 

 

 

 

, (A.1) 

with 



 

 

, 

, 

2 v 

z 


 

, 

v 

; 

 

v 

v z 



components); 

 

v 

v z 


 

y 

x, 






terrain 


terrain 





h 

r 

 

cos 

1 

, (A.2) 

where 








) 

2 

(sin 

2 r 

 

, (A.3) 




ground-to-sensor transmittance for elevation z and viewing angle 






0 

terrain 

=0.1 by 

 

 

 

 

 

y 

x 

v 

z 

E 

z 

y 

x 

E 

y 

x 

z 

E 

y 

x 

b 

z 

z 

L 

y 

x 

DN 

c 

c 

d 

y 

x 

terrain 

i 

terrain 

ground 

d 

s 

s 

v 

v 

v 

, 

, 

, 

, 

cos 

, 

, 

, 

, 

, 

) 

, 

( 

2 

1 

0 

2 

 

 

 

 

 

 

 

 

 

 

 

 

 

, (A.1) 

with 



 

 

, 

, 

2 v 

z 


 

, 

v 

; 

 

v 

v z 



components); 

 

v 

v z 


 

y 

x, 






terrain 


terrain 





h 

r 

 

cos 

1 

, (A.2) 

where 








) 

2 

(sin 

2 r 

 

, (A.3) 




(direct plus diffuse components); 






0 

terrain 

=0.1 by 

 

 

 

 

 

y 

x 

v 

z 

E 

z 

y 

x 

E 

y 

x 

z 

E 

y 

x 

b 

z 

z 

L 

y 

x 

DN 

c 

c 

d 

y 

x 

terrain 

i 

terrain 

ground 

d 

s 

s 

v 

v 

v 

, 

, 

, 

, 

cos 

, 

, 

, 

, 

, 

) 

, 

( 

2 

1 

0 

2 

 

 

 

 

 

 

 

 

 

 

 

 

 

, (A.1) 

with 



 

 

, 

, 

2 v 

z 


 

, 

v 

; 

 

v 

v z 



components); 

 

v 

v z 


 

y 

x, 






terrain 


terrain 





h 

r 

 

cos 

1 

, (A.2) 

where 








) 

2 

(sin 

2 r 

 

, (A.3) 




a sun-to-ground transmittance (direct beam); 






0 

terrain 

=0.1 by 

 

 

 

 

 

y 

x 

v 

z 

E 

z 

y 

x 

E 

y 

x 

z 

E 

y 

x 

b 

z 

z 

L 

y 

x 

DN 

c 

c 

d 

y 

x 

terrain 

i 

terrain 

ground 

d 

s 

s 

v 

v 

v 

, 

, 

, 

, 

cos 

, 

, 

, 

, 

, 

) 

, 

( 

2 

1 

0 

2 

 

 

 

 

 

 

 

 

 

 

 

 

 

, (A.1) 

with 



 

 

, 

, 

2 v 

z 


 

, 

v 

; 

 

v 

v z 



components); 

 

v 

v z 


 

y 

x, 






terrain 


terrain 





h 

r 

 

cos 

1 

, (A.2) 

where 








) 

2 

(sin 

2 r 

 

, (A.3) 




angle between the solar ray and the surface normal (illumination angle); 





0 

terrain 

=0.1 by 

 

 

 

 

 

y 

x 

v 

z 

E 

z 

y 

x 

E 

y 

x 

z 

E 

y 

x 

b 

z 

z 

L 

y 

x 

DN 

c 

c 

d 

y 

x 

terrain 

i 

terrain 

ground 

d 

s 

s 

v 

v 

v 

, 

, 

, 

, 

cos 

, 

, 

, 

, 

, 

) 

, 

( 

2 

1 

0 

2 

 

 

 

 

 

 

 

 

 

 

 

 

 

, (A.1) 

with 



 

 

, 

, 

2 v 

z 


 

, 

v 

; 

 

v 

v z 



components); 

 

v 

v z 


 

y 

x, 






terrain 


terrain 





h 

r 

 

cos 

1 

, (A.2) 

where 








) 

2 

(sin 

2 r 

 

, (A.3) 




binary factor: b=1 if pixel receives direct solar beam, otherwise b=0; 

E s : extraterrestrial solar irradiance; 

E d (x, y, z): diffuse solar flux ; 

E g (z): global flux (direct plus diffuse solar flux on a horizontal surface); 






0 

terrain 

=0.1 by 

 

 

 

 

 

y 

x 

v 

z 

E 

z 

y 

x 

E 

y 

x 

z 

E 

y 

x 

b 

z 

z 

L 

y 

x 

DN 

c 

c 

d 

y 

x 

terrain 

i 

terrain 

ground 

d 

s 

s 

v 

v 

v 

, 

, 

, 

, 

cos 

, 

, 

, 

, 

, 

) 

, 

( 

2 

1 

0 

2 

 

 

 

 

 

 

 

 

 

 

 

 

 

, (A.1) 

with 



 

 

, 

, 

2 v 

z 


 

, 

v 

; 

 

v 

v z 



components); 

 

v 

v z 


 

y 

x, 






terrain 


terrain 





h 

r 

 

cos 

1 

, (A.2) 

where 








) 

2 

(sin 

2 r 

 

, (A.3) 




locally varying average terrain reflectance, calculated iteratively (i=1,2,3) and 





0 

terrain 

=0.1 by 

 

 

 

 

 

y 

x 

v 

z 

E 

z 

y 

x 

E 

y 

x 

z 

E 

y 

x 

b 

z 

z 

L 

y 

x 

DN 

c 

c 

d 

y 

x 

terrain 

i 

terrain 

ground 

d 

s 

s 

v 

v 

v 

, 

, 

, 

, 

cos 

, 

, 

, 

, 

, 

) 

, 

( 

2 

1 

0 

2 

 

 

 

 

 

 

 

 

 

 

 

 

 

, (A.1) 

with 



 

 

, 

, 

2 v 

z 


 

, 

v 

; 

 

v 

v z 



components); 

 

v 

v z 


 

y 

x, 






terrain 


terrain 





h 

r 

 

cos 

1 

, (A.2) 

where 








) 

2 

(sin 

2 r 

 

, (A.3) 




terrain view factor (range 0-1) calculated from local slope. 


The theoretical field of view of a single LAI-2000 PCA measurement is represented by a sphere segment, whose 

radius can be calculated as 





0 

terrain 

=0.1 by 

 

 

 

 

 

y 

x 

v 

z 

E 

z 

y 

x 

E 

y 

x 

z 

E 

y 

x 

b 

z 

z 

L 

y 

x 

DN 

c 

c 

d 

y 

x 

terrain 

i 

terrain 

ground 

d 

s 

s 

v 

v 

v 

, 

, 

, 

, 

cos 

, 

, 

, 

, 

, 

) 

, 

( 

2 

1 

0 

2 

 

 

 

 

 

 

 

 

 

 

 

 

 

, (A.1) 

with 



 

 

, 

, 

2 v 

z 


 

, 

v 

; 

 

v 

v z 



components); 

 

v 

v z 


 

y 

x, 






terrain 


terrain 





h 

r 

 

cos 

1 

, (A.2) 

where 








) 

2 

(sin 

2 r 

 

, (A.3) 




(A.2) 

where cosθ is the maximum zenith angle of the outmost ring included in LAI calculation and h is the average 

canopy height. Assuming an average canopy height of 20 m and a maximum zenith angle of 58.1° (ring 4) the 

field of view would thus theoretically include all canopy elements within 37.8 m in horizontal direction (cf. 

Figure A-1a). At 10 m height the theoretical field of view decreases to 18.9 m. The horizontal projection in 

Figure A-1b shows that also the maximum width of a measurement depends on the average canopy height. It 

is represented by a circular sector whose radius r can be calculated by equation A.1. Its chord (i.e. is maximum 

width) is then equivalent to 





0 

terrain 

=0.1 by 

 

 

 

 

 

y 

x 

v 

z 

E 

z 

y 

x 

E 

y 

x 

z 

E 

y 

x 

b 

z 

z 

L 

y 

x 

DN 

c 

c 

d 

y 

x 

terrain 

i 

terrain 

ground 

d 

s 

s 

v 

v 

v 

, 

, 

, 

, 

cos 

, 

, 

, 

, 

, 

) 

, 

( 

2 

1 

0 

2 

 

 

 

 

 

 

 

 

 

 

 

 

 

, (A.1) 

with 



 

 

, 

, 

2 v 

z 


 

, 

v 

; 

 

v 

v z 



components); 

 

v 

v z 


 

y 

x, 






terrain 


terrain 





h 

r 

 

cos 

1 

, (A.2) 

where 








) 

2 

(sin 

2 r 

 

, (A.3) 




(A.3) 



guarantee spatial independence.

192 

Curriculum Vitae 

Personal information 

Employment record 

Education 

Publications 

Name 

Date of birth 

Place of birth 

Marital status 

Address 

since 03/2008 

01/2007-02/2008 

10/2002-12/2006 

08/2002 

04/1997-08/2002 

08/1996-02/1997 

1987-1996 

1984-1987 

Tanja Kraus 

May 11, 1977 

Nürnberg, Germany 

married 

Klugstr. 33 

80638 München 

Germany 

tanja@kraus.net 

Scientific Assistant to the Director (Wissenschaftliche Referentin des Direktors) 

at the German Remote Sensing Data Center (DFD) of the German Aerospace 

Center (DLR), Oberpfaffenhofen, Germany 

Research assistant at the Chair for Remote Sensing, University of Würzburg 

Employed at DFD-DLR in the following projects: 

01/2006-12/2006 Research assistant in GSE Forest Monitoring 

04/2005-12/2005 Assistant in the proposal phase of GITEWS (German- 

Indonesian Tsunami Early Warning System) 

06/2004-12/2006 PhD student in BIOTA East Africa, subproject E02 

10/2002-05/2004 Research assistant in BIOTA East Africa, subproject E02 

Degree in Geography (Diplom-Geographin), Thesis on “Quantifizierung von 

Biomasse in Savannenökosystemen mit Landsat-ETM+” 

Studies of Geography at the Friedrich-Alexander-University Erlangen-Nürnberg 

with minors Geology and Biology 

Au Pair in Washington, D.C. 

Studies of English at Georgetown University, Washington, D.C. 

Geschwister-Scholl-Gymnasium, Röthenbach a. d. Pegnitz 

Primary School (Grundschule am Forstersberg, Röthenbach a. d. Pegnitz) 

Kraus, T., Schmidt, M., Dech, S.W. and C. Samimi. The potential of optical high resolution data for 

the assessment of leaf area index in East African rain forest ecosystems. – In: Special Issue of 

the International Journal of Remote Sensing (in review). 

Eisfelder, C., Kraus, T., Bock, M., Werner, M., Buchroithner, M.F. and G. Strunz. Towards automated 

forest-type mapping – a service within GSE Forest Monitoring based on SPOT-5 and IKONOS 

data. – In: Special Issue of the International Journal of Remote Sensing (in review).

Curriculum Vitae 

Publications (continued) 

Kraus, T., Bock, M. and G. Strunz (2007). Forest type mapping based on SPOT-5 and IKONOS data 

as a service within GSE Forest Monitoring. – In: Proceedings of the ForestSAT Conference 

2007, Montpellier, France, November 5-7, 2007, 5 p. 

Kraus, T., Schmidt, M., Dech, S. and C. Samimi (2007). The potential of optical high resolution data 

for the assessment of leaf area index in East African rain forest ecosystems. – In: Proceedings 

of the ForestSAT Conference 2007, Montpellier, France, November 5-7, 2007, 5 p. 

Werner, M., Kraus, T., Bock, M., Strunz, G. and K.-F. Wetzel (2007). Objektbasierte 

Waldflächenkartierung in Schleswig-Holstein mit SPOT-5-Daten. – In: J. Strobl, T. Blaschke 

und G. Griesebner (Hrsg.) 2007: Geoinformatik 2007. Beiträge zum 19. AGIT- Symposium 

Salzburg, pp. 852-857. 

Kraus, T. (2005). LAI measurements in a tropical rainforest: Kakamega Forest, Kenya. 

VALERI Workshop, Avignon, March 10, 2005. Available online 

 

Kraus, T. and R. Ressl (2005). DEMMIN - Testsite for environmental monitoring and validation. 

VALERI Workshop, Avignon, March 10, 2005. Available online 

 

Müller U., Conrad C. and T. Kraus (2004). Analyse der raum-zeitlichen Vegetationsdynamik in 

Ostafrika unter Verwendung von MODIS-Daten. – In: J. Strobl, T. Blaschke und G. Griesebner 

(Hrsg.) 2004: Angewandte Geographische Informationsverarbeitung XVI. Beiträge zum 

AGIT-Symposium Salzburg 2004, pp. 484-489. 

Samimi, C. and T. Kraus (2004). Biomass estimation using Landsat-TM and -ETM+. Towards 

a regional model for Southern Africa? – In: GeoJournal Special Issue: Systems Modelling 

Across Geography´s Interface, 59 (3), pp. 177-187. 

Schaab, G., Kraus, T. and G. Strunz (2004). GIS and remote sensing activities as an integrating 

link within the BIOTA-East Africa project. Sustainable use and conservation of biological 

diversity – A challenge for society. – In: Proceedings of the International Symposium Berlin, 

December 1-4, 2003, Berlin, pp. 161-168. 

Riedlinger, T., Voigt, S., Günther, K.P., Gesell, G., Künzer, C., Zwing, A., Kraus, T., Kiefl, R., Tetzlaff, 

A., Zhang, J., Hasenauer, S., Reinartz, P., Sundermann, D. and H. Mehl (2003). Multisource 

satellite data facilitating disaster management during the 2003 forest fires in Portugal. 

– In: Proceedings of the Mediterranean seminar on new technologies applied to the management 

of disasters risks, Madrid, October 3-6, 2003. Available online at 

 

Kraus T. and G. Schaab (2003). Biodiversitätsforschung in Westkenia – Mit GIS und Fernerkundung 

von lokalen Beobachtungen zu Aussagen in Zeit und Raum. – In: J. Strobl, T. Blaschke und 

G. Griesebner (Hrsg.) 2003: Angewandte Geographische Informationsverarbeitung XV. 

Beiträge zum AGIT-Symposium Salzburg 2003, pp. 244-249. 

Kraus, T. and C. Samimi (2002). Biomass estimation for land use management and fire management 

using Landsat-TM and -ETM+. – In: Erdkunde, 56 (2), pp. 130-143. 

193

Ground-based Validation of the MODIS Leaf Area Index Product for ...

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?