Evaluation of a probabilistic cloud masking algorithm for ... - CM SAF
Evaluation of a probabilistic cloud masking algorithm for ... - CM SAF Evaluation of a probabilistic cloud masking algorithm for ... - CM SAF
EUMETSAT Satellite Application Facility on Climate Monitoring Visiting Scientist Report Evaluation of a probabilistic cloud masking algorithm for climate data record processing: SPARC: a new scene identification algorithm for MSG SEVIRI CDOP VS Study No 14 Date 15. October 2010 Reference CLM_VS10_03
- Page 2: SPARC: a new scene identification a
- Page 5 and 6: Summary II Overall, SPARCMSG has pr
- Page 7 and 8: Table of Contents IV 4.2.1 ASRB dat
- Page 9 and 10: List of Figures VI 4.10 Histogram o
- Page 11 and 12: List of Tables VIII
- Page 13 and 14: SPARC algorithm 2 haze contaminatio
- Page 15 and 16: SPARC algorithm 4 channel is availa
- Page 17 and 18: SPARC algorithm 6
- Page 19 and 20: SPARC algorithm 8 2.3 Cloud shadow
- Page 21 and 22: SPARC algorithm 10 ct = ct − 5.0s
- Page 23 and 24: SPARC algorithm 12 Figure 3.2: The
- Page 25 and 26: SPARC algorithm 14 where w =weight
- Page 27 and 28: SPARC algorithm 16 had been observe
- Page 29 and 30: SPARC algorithm 18 with Smin=−15.
- Page 31 and 32: SPARC algorithm 20 Figure 3.6: Time
- Page 33 and 34: Chapter 4 Validation Study 4.1 Gene
- Page 35 and 36: SPARC algorithm 24 computational co
- Page 37 and 38: SPARC algorithm 26 Figure 4.1: Scat
- Page 39 and 40: SPARC algorithm 28 Figure 4.3: Obse
- Page 41 and 42: SPARC algorithm 30
- Page 43 and 44: SPARC algorithm 32
- Page 45 and 46: SPARC algorithm 34 Figure 4.7: Simi
- Page 47 and 48: SPARC algorithm 36
- Page 49 and 50: SPARC algorithm 38 Figure 4.10: Sim
- Page 51 and 52: SPARC algorithm 40 Figure 4.11: Com
EUMETSAT Satellite Application Facility on Climate Monitoring<br />
Visiting Scientist Report<br />
<strong>Evaluation</strong> <strong>of</strong> a <strong>probabilistic</strong> <strong>cloud</strong> <strong>masking</strong><br />
<strong>algorithm</strong> <strong>for</strong> climate data record processing:<br />
SPARC: a new scene identification<br />
<strong>algorithm</strong> <strong>for</strong> MSG SEVIRI<br />
CDOP VS Study No 14<br />
Date 15. October 2010<br />
Reference CLM_VS10_03
SPARC: a new scene identification<br />
<strong>algorithm</strong> <strong>for</strong> MSG SEVIRI<br />
Report on Modifications and Validation<br />
Fabio Fontana<br />
University <strong>of</strong> Bern<br />
Reto Stockli<br />
MeteoSwiss<br />
Stefan Wunderle<br />
University <strong>of</strong> Bern<br />
September 2010
Summary<br />
The Separation <strong>of</strong> Pixels Using Aggregated Rating Over Canada (SPARC) algo-<br />
rithm represents an alternative to conventional scene identification <strong>algorithm</strong>s as<br />
it outputs a single <strong>cloud</strong> contamination rating (F) instead <strong>of</strong> a <strong>cloud</strong> mask with<br />
a limited number <strong>of</strong> categories. The rating itself is mainly calculated from three<br />
sub-scores generated based on in<strong>for</strong>mation on the brightness temperature (T-score),<br />
brightness (B-score), and the reflectance in the visible and short-wave infrared por-<br />
tion <strong>of</strong> the electromagnetic spectrum (R-score). Even though the <strong>algorithm</strong> was<br />
originally designed <strong>for</strong> the Advanced Very High Resolution Radiometer (AVHRR),<br />
it may be applied to data from other multispectral sensors.<br />
This report describes a modified version <strong>of</strong> SPARC and its application to data from<br />
the Meteosat Second Generation (MSG) Spinning Enhanced Visible and Infrared<br />
Imager (SEVIRI) sensor. A validation was per<strong>for</strong>med <strong>for</strong> the period from July<br />
2004 to June 2005. SPAR<strong>CM</strong>SG output was compared to <strong>cloud</strong> in<strong>for</strong>mation from<br />
the Alpine Surface Radiation Budget (ASRB) Network in the Swiss Alps: Cloud<br />
Free Index (CFI) and Partial Cloud Amount (PCA; in eighths). Results show a<br />
good agreement between CFI and F mainly <strong>for</strong> daytime observations, with linear<br />
correlation coefficients (r) ranging between 0.67 in fall and 0.78 in the winter months.<br />
At the ASRB site in Payerne, observations <strong>of</strong> complete clear-sky agreed with 85%<br />
(92.3%) during daytime (nighttime) between ASRB PCA and SPAR<strong>CM</strong>SG <strong>cloud</strong><br />
mask. However, results also point to an underestimation <strong>of</strong> nighttime <strong>cloud</strong> cover<br />
by SPAR<strong>CM</strong>SG. If only two SEVIRI spectral channels are fed into SPARC to simulate<br />
data from the Meteosat First Generation (MFG), a good agreement between CFI and<br />
F is obtained during daytime (r=0.65). Finally, application <strong>of</strong> SPARC at full disk<br />
level revealed an underestimation <strong>of</strong> <strong>cloud</strong> cover compared to the Climate Monitoring<br />
(<strong>CM</strong>)-Satellite Application Facility (<strong>SAF</strong>) <strong>cloud</strong> mask.<br />
I
Summary II<br />
Overall, SPAR<strong>CM</strong>SG has proven to provide interesting new opportunities <strong>for</strong> <strong>cloud</strong><br />
detection using the MSG SEVIRI sensor. Results are particularly promising with<br />
regard to the application <strong>of</strong> SPARC <strong>for</strong> data <strong>of</strong> the MFG satellite series.
Contents<br />
Summary I<br />
Table <strong>of</strong> Contents III<br />
List <strong>of</strong> Figures V<br />
List <strong>of</strong> Tables VII<br />
1 Overview 1<br />
2 SPARC <strong>algorithm</strong> 3<br />
2.1 Cloud detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3<br />
2.2 Snow detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7<br />
2.3 Cloud shadow identification . . . . . . . . . . . . . . . . . . . . . . . 8<br />
3 The SPARC <strong>algorithm</strong> <strong>for</strong> MSG SEVIRI 9<br />
3.1 SPAR<strong>CM</strong>SG <strong>cloud</strong> mask . . . . . . . . . . . . . . . . . . . . . . . . . . 9<br />
3.2 Calculation <strong>of</strong> the T-score . . . . . . . . . . . . . . . . . . . . . . . . 13<br />
3.3 The SPAR<strong>CM</strong>SG snow detection scheme . . . . . . . . . . . . . . . . . 16<br />
3.4 Further modifications . . . . . . . . . . . . . . . . . . . . . . . . . . . 20<br />
4 Validation Study 22<br />
4.1 General Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22<br />
4.2 Data and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22<br />
III
Table <strong>of</strong> Contents IV<br />
4.2.1 ASRB data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22<br />
4.2.2 MSG data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23<br />
4.3 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . 25<br />
4.3.1 Comparison with ASRB Cloud Parameters . . . . . . . . . . . 25<br />
4.3.2 MFG Simulation . . . . . . . . . . . . . . . . . . . . . . . . . 34<br />
4.3.3 Validation at Full Disk Level . . . . . . . . . . . . . . . . . . . 39<br />
4.4 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
List <strong>of</strong> Figures<br />
2.1 SPARC output <strong>for</strong> a sample slot. . . . . . . . . . . . . . . . . . . . . 5<br />
3.1 HRV channel <strong>of</strong> a sample slot together with the SPARC output cre-<br />
ated with different thresholds. . . . . . . . . . . . . . . . . . . . . . . 11<br />
3.2 HRV channel <strong>of</strong> a sample slot together with the SPAR<strong>CM</strong>SG output. . 12<br />
3.3 Examples <strong>of</strong> a Mannstein function fitted to clear-sky BT108 values. . 15<br />
3.4 Example <strong>of</strong> a RST map in summer (left) and winter (right) over Swiss<br />
Alps. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15<br />
3.5 Output <strong>of</strong> the novel SPARC snow detection <strong>algorithm</strong>. . . . . . . . . 19<br />
3.6 Time series <strong>of</strong> monthly maximum snow cover in the Swiss Alps. . . . 20<br />
4.1 CFI vs. F <strong>for</strong> all sites and 12 months combined. . . . . . . . . . . . . 26<br />
4.2 CFI vs. F <strong>for</strong> each season separately. . . . . . . . . . . . . . . . . . . 26<br />
4.3 PCA vs. SPAR<strong>CM</strong>SG <strong>cloud</strong> mask <strong>for</strong> the entire 12 month period. . . . 28<br />
4.4 PCA vs. <strong>CM</strong>MSG <strong>for</strong> each season separately. . . . . . . . . . . . . . . 29<br />
4.5 Histogram <strong>of</strong> <strong>CM</strong>MSG output <strong>for</strong> each PCA class at the ASRB site in<br />
Payerne. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31<br />
4.6 Histogram <strong>of</strong> <strong>CM</strong>MSG output <strong>for</strong> each PCA class at the ASRB site at<br />
Weissflujoch. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33<br />
4.7 CFI vs. F <strong>for</strong> all sites and 12 months combined - SPAR<strong>CM</strong>FG mode. . 34<br />
4.8 Output <strong>of</strong> SPAR<strong>CM</strong>SG and SPAR<strong>CM</strong>FG on March 20, 2005 (8:45 am). 35<br />
4.9 Histogram <strong>of</strong> <strong>CM</strong>MFG output <strong>for</strong> each PCA class at the ASRB site in<br />
Payerne - SPAR<strong>CM</strong>FG mode. . . . . . . . . . . . . . . . . . . . . . . . 37<br />
V
List <strong>of</strong> Figures VI<br />
4.10 Histogram <strong>of</strong> <strong>CM</strong>MFG output <strong>for</strong> each PCA class at the ASRB site at<br />
Weissflujoch - SPAR<strong>CM</strong>FG mode. . . . . . . . . . . . . . . . . . . . . 38<br />
4.11 Comparison <strong>of</strong> the <strong>CM</strong>-<strong>SAF</strong>, SPAR<strong>CM</strong>SG, and SPAR<strong>CM</strong>FG <strong>cloud</strong> masks<br />
<strong>for</strong> the full disk. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40<br />
4.12 Comparison <strong>of</strong> the <strong>CM</strong>-<strong>SAF</strong>, SPAR<strong>CM</strong>SG, and SPAR<strong>CM</strong>FG <strong>cloud</strong> masks<br />
<strong>for</strong> five regions <strong>of</strong> interest within the full disk. . . . . . . . . . . . . . 41<br />
4.13 Modification <strong>of</strong> the R-score <strong>for</strong> bright desert surfaces. . . . . . . . . . 42
List <strong>of</strong> Tables<br />
2.1 Boundary coordinates <strong>of</strong> the region <strong>of</strong> interest. . . . . . . . . . . . . . 7<br />
4.1 Sites <strong>of</strong> the ASRB network used <strong>for</strong> the validation. . . . . . . . . . . 23<br />
4.2 Channels considered as SPARC input in MSG and MFG mode. . . . 25<br />
VII
List <strong>of</strong> Tables VIII
Chapter 1<br />
Overview<br />
It is <strong>of</strong> major importance <strong>for</strong> any satellite data processing system to accurately de-<br />
termine the state <strong>of</strong> a pixel. Depending on the application it may be necessary to<br />
assign a pixel to the clear-sky (e.g., <strong>for</strong> land surface applications) or <strong>cloud</strong>y cate-<br />
gory (e.g., <strong>for</strong> the retrieval <strong>of</strong> <strong>cloud</strong> properties). Further, assigning <strong>for</strong> instance a<br />
mixed pixel with sub-pixel <strong>cloud</strong> cover to the clear-sky or <strong>cloud</strong>y category largely<br />
depends on the final application. For land surface applications a clear sky con-<br />
servative approach is preferred, where <strong>for</strong> the retrieval <strong>of</strong> <strong>cloud</strong> properties a <strong>cloud</strong><br />
conservative approach has to be used. To achieve discrimination <strong>of</strong> various scene<br />
types (e.g., snow, <strong>cloud</strong>s, water, or vegetation), pixel scene identification <strong>algorithm</strong>s<br />
make use <strong>of</strong> the basic underlying principle that scene types may be distinguished<br />
based on their inherent reflective and emissive properties in certain portions <strong>of</strong> the<br />
electromagnetic spectrum (Lillesand et al., 2004).<br />
Scene identification <strong>algorithm</strong>s traditionally follow a branching/thresholding ap-<br />
proach: a sequence <strong>of</strong> spectral tests is applied to each pixel, assigning the pixel<br />
to a certain class (e.g., <strong>cloud</strong>y, clear-sky, or partly <strong>cloud</strong>y) based on predefined<br />
thresholds. However, such branching/thresholding approaches have two main dis-<br />
advantages (Khlopenkov and Trishchenko, 2007): first, since a yes/no decision is<br />
made at each node <strong>of</strong> the classification tree, a pixel that happens to represent an<br />
intermediate state may be assigned to the wrong category, or its classification may<br />
not be possible. Second, the final result <strong>of</strong> such classification schemes only contains<br />
in<strong>for</strong>mation on a limited number <strong>of</strong> classes, such as ’<strong>cloud</strong>y’, ’clear-sky’, ’partly<br />
<strong>cloud</strong>y’, or ’uncertain’, but does not provide in<strong>for</strong>mation on the degree <strong>of</strong> <strong>cloud</strong> or<br />
1
SPARC <strong>algorithm</strong> 2<br />
haze contamination within the field <strong>of</strong> view. Such in<strong>for</strong>mation may, however, be<br />
valuable <strong>for</strong> the generation <strong>of</strong> clear-sky composites <strong>for</strong> many applications, e.g., land<br />
surface albedo derivation or vegetation dynamics analysis.<br />
In order to address these limitations, the Separation <strong>of</strong> Pixels Using Aggregated<br />
Rating over Canada (SPARC) <strong>algorithm</strong> (Khlopenkov and Trishchenko, 2007) was<br />
developed at the Canada Centre <strong>for</strong> Remote Sensing (CCRS). The SPARC <strong>algorithm</strong><br />
was originally designed <strong>for</strong> the use with Advanced Very High Resolution Radiometer<br />
(AVHRR) data over Canada, but its design should theoretically enable the <strong>algorithm</strong><br />
to be transferred to other multi-spectral sensors such as the Meteosat Second Gen-<br />
eration (MSG) Spinning Enhanced Visible and Infrared Imager (SEVIRI) and other<br />
geographical regions <strong>of</strong> interest, even though some modification might be necessary.<br />
The aim <strong>of</strong> this report is to provide a short overview <strong>of</strong> the SPARC <strong>algorithm</strong> (Sec-<br />
tion 2) and the modifications specific <strong>for</strong> its application with the MSG SEVIRI sensor<br />
(Section 3). The report also presents the results <strong>of</strong> a validation study per<strong>for</strong>med<br />
over the Swiss Alps (Section 4.3.1). The applicability <strong>of</strong> the SPARC <strong>algorithm</strong> to<br />
data from the Meteosat First Generation (MFG) satellite series is evaluated (Section<br />
4.3.2). Finally, preliminary results <strong>of</strong> the applicaiton <strong>of</strong> SPARC at SEVIRI full disk<br />
dimension are presented (Section 4.3.3) and concluding remarks are given in Section<br />
4.4.
Chapter 2<br />
SPARC <strong>algorithm</strong><br />
2.1 Cloud detection<br />
In<strong>for</strong>mation provided in this chapter closely follows the work <strong>of</strong> Khlopenkov and<br />
Trishchenko (2007). The SPARC <strong>algorithm</strong> outputs a score, which can be regarded<br />
as a measure <strong>of</strong> the <strong>cloud</strong> contamination probability <strong>for</strong> each pixel. The SPARC<br />
score (F) is obtained through the summation <strong>of</strong> the output <strong>of</strong> a number <strong>of</strong> individual<br />
tests, which make use <strong>of</strong> all five spectral channels <strong>of</strong> the AVHRR sensor (Eq. 2.1):<br />
where<br />
F = T + B + R + �<br />
Ai, (2.1)<br />
• T is the temperature score, which is based on the comparison <strong>of</strong> the brightness<br />
temperature at 10.8µm and a dynamic threshold derived from an external sur-<br />
face skin temperature field. In the original SPARC version, surface skin tem-<br />
perature is obtained from the North American Regional Reanalysis (NARR)<br />
at a spatial resolution <strong>of</strong> 32 km×32 km (Mesinger and Coauthors, 2006).<br />
• B is the brightness score, calculated from the reflectance in the visible (VIS)<br />
and near infrared (NIR) portion <strong>of</strong> the electromagnetic spectrum over land<br />
and water, respectively,<br />
• R is the reflectance score calculated either from the reflectance values in the<br />
VIS and short wave infrared (SWIR) portions <strong>of</strong> the spectrum (if a SWIR<br />
3<br />
i
SPARC <strong>algorithm</strong> 4<br />
channel is available from AVHRR), or from the solar reflective component at<br />
3.7µm otherwise, and<br />
• Ai represents the added scores <strong>of</strong> four additional tests that are only per<strong>for</strong>med<br />
under certain circumstances due to computational considerations.<br />
These supplementary tests include:<br />
• the simple ratio test (N), exploiting the relatively homogeneous reflective prop-<br />
erties <strong>of</strong> <strong>cloud</strong>y scenes in the VIS and NIR spectrum as opposed to clear-sky<br />
observation <strong>of</strong>, e.g., vegetation cover,<br />
• the texture uni<strong>for</strong>mity test (Utext) based on the NIR spectrum and the ther-<br />
mal uni<strong>for</strong>mity test (Utemp) based on 10.8µm brightness temperature, which<br />
both analyze the variability <strong>of</strong> the pixel values <strong>for</strong> a pixel and its four nearest<br />
neighbors, and finally<br />
• the thin cirrus test based on the brightness temperature difference between<br />
12µm and 10.8µm.<br />
The scores are designed such that strong (weak) <strong>cloud</strong> evidence leads to positive<br />
(negative) values <strong>of</strong> F, whereas values <strong>of</strong> F ≈0 represent intermediate, i.e. uncertain<br />
states. In combination with the linear aggregation principle, this design has the<br />
advantage that a <strong>cloud</strong> mask can be generated even if some <strong>of</strong> the scores are not<br />
calculated. More detailed in<strong>for</strong>mation on the calculation <strong>of</strong> the scores is provided in<br />
Khlopenkov and Trishchenko (2007).<br />
However, some adjustments <strong>of</strong> Equation 2.1 are required in order <strong>for</strong> the single<br />
scores to receive appropriate weights under special observation conditions, namely<br />
snow cover, sun glint, and in the proximity to the terminator line. For this reason,<br />
three correction factors are implemented. These are:<br />
• the correction factor (s) <strong>for</strong> snow conditions, which is necessary since snow-<br />
covered areas observed under clear-sky conditions and <strong>cloud</strong>s can be equally<br />
bright. In the original SPARC scheme, this factor is estimated from the surface<br />
skin temperature fields;
SPARC <strong>algorithm</strong> 5<br />
• the sun glint correction factor (g; calculated from the observation geometry)<br />
to account <strong>for</strong> increased spectral reflectance over sun glint areas;<br />
• the nighttime correction factor (n), which accounts <strong>for</strong> the special circum-<br />
stances at very low Sun zenith angles in the proximity <strong>of</strong> the terminator line.<br />
Equation 2.1 is thus modified as follows:<br />
F = B(1 − g)sn + T (1 + g)(2 − s)(2 − n) + R(1 − 0.6g)(2 − s)n + N(1 − g)sn<br />
+ Utext(1 − g)sn + Utemp(1 + g)(2 − s)(2 − n) + C (2.2)<br />
Note that all correction factors reduce the weight <strong>of</strong> the reflective scores (where<br />
applicable <strong>for</strong> a certain correction factor) and assign more weight to the thermal<br />
tests. By design, reflectance scores are turned <strong>of</strong>f completely at nighttime and the<br />
SPARC output is based only on the test <strong>for</strong> the thermal channels. Figure 2.1 displays<br />
an example <strong>of</strong> all SPARC scores as well as the snow correction factor <strong>for</strong> a MSG<br />
SEVIRI time slot in March 2004. The boundary coordinates <strong>of</strong> the region <strong>of</strong> interest<br />
covering the Swiss Alps are provided in Table 2.1. The HRV channel (1A) shows<br />
significant snow cover over the elevated topography <strong>of</strong> the Swiss Alps under mostly<br />
clear-sky conditions as well as prevalent low stratus <strong>cloud</strong>s over the Swiss Main<br />
Plateau. Note that the absolute values <strong>of</strong> the single contributing scores vary greatly<br />
in magnitude. In particular, scores <strong>of</strong> the supplementary tests display significantly<br />
smaller absolute values compared to the three main scores (T, B, and R). Figure<br />
2.1 also highlights a problematic issue around the calculation <strong>of</strong> the Utext-scores<br />
over the Swiss Alps during wintertime: even under <strong>cloud</strong>-free conditions the high<br />
reflectance contrast between snow-covered mountain tops and low-lying and snow<br />
free valley bottoms results in increased values <strong>of</strong> the Utext-score. The same problem,<br />
Figure 2.1 (following page): High resolution visible (HRV) channel <strong>of</strong> the MSG SE-<br />
VIRI time slot on March 20, 2005 (8:45 am; 1A), together with the corresponding SPARC<br />
scores (see text <strong>for</strong> more details). The correction factor <strong>for</strong> snow conditions (s) is also<br />
provided (1E). Note that s was calculated as described in Khlopenkov and Trishchenko<br />
(2007), however, using a clear-sky BT108 mosaic as input (cp. Section 3.2) instead <strong>of</strong><br />
an external surface skin temperature map. Details on the calculation <strong>of</strong> the scores can be<br />
found in Khlopenkov and Trishchenko (2007).
SPARC <strong>algorithm</strong> 6
SPARC <strong>algorithm</strong> 7<br />
Table 2.1: Boundary coordinates <strong>of</strong> the region <strong>of</strong> interest (ROI) covering the Swiss Alps,<br />
where ’lat’ is the latitude [ ◦ ] and ’lon’ the longitude [ ◦ ].<br />
ROI Swiss Alps<br />
latmin<br />
latmax<br />
lonmin<br />
lonmax<br />
yet less pronounced, is observed <strong>for</strong> the Utemp-score. Similarly, the N-score adopts<br />
high values both <strong>for</strong> <strong>cloud</strong> and snow cover. Even though the contributions <strong>of</strong> these<br />
additional scores to the final score F are small, it is suspected that erroneously large<br />
values <strong>of</strong> these scores may lead to some misclassifications in uncertain conditions.<br />
2.2 Snow detection<br />
Snow detection is originally per<strong>for</strong>med in two steps based on the T-, B-, and R-<br />
scores:<br />
45.5<br />
48.0<br />
6.5<br />
10.0<br />
• In step one, a number <strong>of</strong> threshold tests is per<strong>for</strong>med, testing whether the pixel<br />
is sufficiently bright (B-score threshold), limiting the R-score to small values to<br />
exclude <strong>cloud</strong> contamination, and ensuring that the brightness temperature at<br />
10.8µm is close to the surface skin temperature map as outlined above. Once<br />
a pixel is classified as snow-covered in this first step, it undergoes another test<br />
in<br />
• step two, which is designed to detect <strong>cloud</strong>s over snow based on the R- and<br />
T-score. A pixel is classified as clear-sky snow, if both the R- and T-score are<br />
smaller than zero. The pixel is classified as snow-covered and <strong>cloud</strong> contami-<br />
nated, if these criteria are not met.<br />
Additional details as well as a flowchart <strong>of</strong> the SPARC snow <strong>algorithm</strong> can be found<br />
in Khlopenkov and Trishchenko (2007, Figures 4 and 5).
SPARC <strong>algorithm</strong> 8<br />
2.3 Cloud shadow identification<br />
Cloud shadow detection is per<strong>for</strong>med based on geometrical considerations, i.e. by<br />
computing the extent <strong>of</strong> the <strong>cloud</strong> shadow on the Earth’s surface (shadows cast<br />
by high <strong>cloud</strong>s on the top <strong>of</strong> low <strong>cloud</strong>s are not considered). The output mask<br />
includes in<strong>for</strong>mation on three different categories: shadow-free and, according to<br />
two thresholds to F, thin <strong>cloud</strong> shadow and thick <strong>cloud</strong> shadow. For more details<br />
on the <strong>cloud</strong> shadow retrieval it is again referred to Khlopenkov and Trishchenko<br />
(2007).
Chapter 3<br />
The SPARC <strong>algorithm</strong> <strong>for</strong> MSG<br />
SEVIRI<br />
Prior to the application <strong>of</strong> the SPARC <strong>algorithm</strong> to data from the MSG SEVIRI<br />
sensor (hereafter referred to as SPAR<strong>CM</strong>SG), a number <strong>of</strong> modifications relative to<br />
the original version (SPARCORIG) were conducted. These will be explained in the<br />
following sections.<br />
3.1 SPAR<strong>CM</strong>SG <strong>cloud</strong> mask<br />
Other than in SPARCORIG, the primary output <strong>of</strong> the SPAR<strong>CM</strong>SG <strong>cloud</strong> detection<br />
scheme is not the score F itself, but rather a <strong>cloud</strong> mask with three different classes:<br />
clear (0), thin <strong>cloud</strong> (1), and thick <strong>cloud</strong> (2). Similar to the criteria in Khlopenkov<br />
and Trishchenko (2007) <strong>for</strong> the distinction between shadows caused by optically<br />
thin or thick <strong>cloud</strong>s, respectively, this simplified <strong>cloud</strong> mask is obtained by applying<br />
thresholds to F:<br />
0 : F < ct : clear-sky; (3.1)<br />
1 : 4 > F > ct : thin <strong>cloud</strong>;<br />
2 : F > 4 : thick <strong>cloud</strong>,<br />
where ct is the <strong>cloud</strong> threshold, which <strong>for</strong> each pixel is calculated incrementally as<br />
ct = −4.0 : as a basic <strong>cloud</strong> threshold (3.2)<br />
9
SPARC <strong>algorithm</strong> 10<br />
ct = ct − 5.0s : stricter threshold if T-score was not applied<br />
ct = ct − 5.0s : even stricter if C-score was not applied.<br />
Even though SPARC was originally designed to overcome limitations <strong>of</strong> traditional<br />
<strong>cloud</strong> detection schemes as outlined above, the <strong>cloud</strong> mask generation based on the<br />
application <strong>of</strong> thresholds to F represents an interesting feature <strong>of</strong> the new SPAR<strong>CM</strong>SG<br />
<strong>algorithm</strong>: by changing the thresholds in either direction, the resulting <strong>cloud</strong> mask<br />
may be shifted dynamically from a clear-sky conservative to a <strong>cloud</strong> conservative<br />
mode without modifying the retrieval <strong>algorithm</strong> itself. This is illustrated in Figure<br />
3.1 which displays the high resolution visible (HRV) channel <strong>of</strong> a sample slot (A)<br />
covering the Swiss Alps together with F (B) and a number <strong>of</strong> <strong>cloud</strong> masks (D-<br />
I). Cloud masks were generated based on basic <strong>cloud</strong> thresholds varying between<br />
−8≤ct≤4 (thresholds <strong>for</strong> thick <strong>cloud</strong> were changed accordingly and were set to<br />
ct+8). The <strong>cloud</strong> mask displayed in Figure 3.1, panel F, was generated following<br />
Equations 3.1 and 3.2. Note that significant differences are observed depending on<br />
the selected threshold, especially in areas where the <strong>cloud</strong> mask error (see below)is<br />
large.<br />
In a subsequent step, the error (ε) <strong>of</strong> the <strong>cloud</strong> mask is calculated based on F as<br />
follows:<br />
(F −ε)2<br />
−<br />
ε = e (2σ2 ε ) , (3.3)<br />
where ε = 0.0<br />
σε = 10.0<br />
(F − ε) 2 � 2000.0<br />
The <strong>cloud</strong> mask error may adopt values between 0 and 1 is designed to be largest at<br />
values <strong>of</strong> F=0, i.e., in situations where the summation <strong>of</strong> the SPARC scores did not<br />
give clear evidence <strong>for</strong> either <strong>cloud</strong>y or clear-sky conditions. Hence, ε is a potentially<br />
useful parameter, e.g., to set up decision criteria in the clear-sky compositing process.<br />
Figure 3.2 displays the high resolution visible (HRV) channel <strong>of</strong> a MSG SEVIRI time<br />
slot in February, together with the corresponding SPARC score F, <strong>cloud</strong> mask, and<br />
<strong>cloud</strong> mask error ε. Cloud mask error is also displayed in Figure 3.1 (C).
SPARC <strong>algorithm</strong> 11<br />
Figure 3.1: The HRV channel <strong>of</strong> a sample slot on 11 September 2004, 11 am (A), together<br />
with the SPARC score (B), the <strong>cloud</strong> mask error (C; see text <strong>for</strong> explanations) and a<br />
number <strong>of</strong> <strong>cloud</strong> masks based on varying thresholds (C-H; see text <strong>for</strong> further explanations).<br />
Black areas represent clear-sky conditions, optically thick (thin) <strong>cloud</strong>s are delineated in<br />
white (grey).
SPARC <strong>algorithm</strong> 12<br />
Figure 3.2: The HRV channel <strong>of</strong> a sample slot channel on 25 February 2005, 10:15 am<br />
(top, left) together with the SPARC score (top, right), <strong>cloud</strong> mask (bottom, left), and <strong>cloud</strong><br />
mask error (bottom, right). Please note the increased values <strong>of</strong> the error ε along the <strong>cloud</strong><br />
boundaries.
SPARC <strong>algorithm</strong> 13<br />
3.2 Calculation <strong>of</strong> the T-score<br />
One <strong>of</strong> the major modifications with regard to SPARCORIG concerns the calculation<br />
<strong>of</strong> the T-score. The high temporal resolution <strong>of</strong> the SEVIRI sensor (15 minutes)<br />
opens up new opportunities <strong>for</strong> surface temperature retrievals, because the chance<br />
<strong>of</strong> a certain location being observed under clear-sky conditions during the course <strong>of</strong> a<br />
single day is significantly increased compared to the AVHRR sensor (1-2 observations<br />
daily). For this reason, an external surface temperature map is not implemented in<br />
SPAR<strong>CM</strong>SG. Instead, a clear sky radiative surface temperature (RST) map directly<br />
derived from SEVIRI 10.8µm channel radiances (BT108) is used as a reference in<br />
the calculation <strong>of</strong> the T-score. As a preliminary solution, brightness temperature is<br />
used instead <strong>of</strong> the radiative surface temperature, because atmospheric correction<br />
is not yet implemented.<br />
The RST map is generated daily as follows: after the processing <strong>of</strong> all 96 daily slots<br />
is completed, the BT108 clear-sky values <strong>for</strong> each pixel and slot are aggregated in<br />
a running array <strong>of</strong> clear sky BT108 values. In this running array, which stores a<br />
maximum <strong>of</strong> two clear-sky values per pixel and slot, the oldest clear-sky BT108<br />
values are replaced by the current observations. However, values observed more<br />
than ten days ago are removed even if there are no current clear-sky observations<br />
available <strong>for</strong> replacement. The resulting array is subsequently used to model the<br />
diurnal evolution <strong>of</strong> surface brightness temperature <strong>for</strong> each pixel on the following<br />
day. This is per<strong>for</strong>med by fitting a modified version <strong>of</strong> the diurnal surface radiative<br />
temperature model (f) by Mannstein et al. (1999). The model is defined as follows<br />
(Dürr, 2009):<br />
2<br />
−2(ωt − a3)<br />
f = a0 + a1(exp( ) + 0.1sin(ωt − a3)) (3.4)<br />
a 2 2<br />
where ω = 2π/96, with 96 the total number <strong>of</strong> slots per day, a0 is the minimium<br />
brightness temperature (BT) <strong>of</strong> the diurnal cycle in Kelvin, a1 is the diurnal ampli-<br />
tude <strong>of</strong> BT in Kelvin, a2 is the half width <strong>of</strong> the day lenght in radians, and a3 is the<br />
true Sun time in radians. Please consult Dürr (2009) <strong>for</strong> more detailed in<strong>for</strong>mation<br />
on the model. Points are weighted as follows, considering the <strong>cloud</strong> mask error <strong>for</strong><br />
the corresponding BT108 clear-sky value as well as the date <strong>of</strong> observation:<br />
w =<br />
1<br />
ε + t , (3.5)<br />
tmax
SPARC <strong>algorithm</strong> 14<br />
where w =weight <strong>for</strong> each point, t =the number <strong>of</strong> days the clear-sky BT108 value<br />
had been stored in the running mosaic, and tmax=10, i.e. the maximum number <strong>of</strong><br />
days considered in the clear-sky compositing (may be adjusted in the configurations<br />
file). Hence, smallest weights are assigned to BT108 values observed early in the<br />
preceeding compositing interval <strong>of</strong> tmax days and under uncertain conditions (F≈0).<br />
Figure 3.3 shows examples <strong>of</strong> a RST fit <strong>for</strong> a location on the Swiss Main Plateau<br />
in summer (left) and winter (right). Please note the difference in the evolution and<br />
the magnitude <strong>of</strong> diurnal brightness temperature.<br />
In addition, a spatial interpolation procedure is implemented in order to fill gaps<br />
in the RST field. The interpolation is based on an elevation-dependent RST filling<br />
using a lapse-rate that is retrieved from neighboring valid pixels with a radius <strong>of</strong> 50<br />
SEVIRI HRV pixels. Figure 3.4 provides an example <strong>of</strong> a RST map <strong>of</strong> the Swiss<br />
Alps in summer 2005 (left) and winter 2004/2005 (right).<br />
The use <strong>of</strong> a RST map derived directly from MSG data has two major advantages:<br />
first, the suggested SPAR<strong>CM</strong>SG <strong>cloud</strong> detection scheme represents a stand-alone ap-<br />
proach that does not rely on any auxillary input data, such as the coarse spatial<br />
resolution NARR surface skin temperature fields as described in Section 2. Sec-<br />
ond, in contrast to these modelled surface temperature fields, the RST map ap-<br />
parently provides substantial spatial detail, which is assumed to be beneficial <strong>for</strong><br />
<strong>cloud</strong> detection over the complex and heterogeneous terrain <strong>of</strong> the Swiss Alps. The<br />
disadvantage <strong>of</strong> modelling the following day’s diurnal BT108 amplitude from the<br />
previous day’s clear-sky BT108 observations is that day-to-day changes in BT108<br />
at a certain time <strong>of</strong> the day cannot be captured. Resulting biases in the T-score<br />
could translate into erroneous <strong>cloud</strong> masks, especially if such day-to-day changes<br />
happen to be significant. However, this problem is likely to occur only in the case <strong>of</strong><br />
warm and low <strong>cloud</strong>s, when the temperature difference between ground and <strong>cloud</strong><br />
top are small. Dectection <strong>of</strong> high and cold <strong>cloud</strong>s should not be affected. A solution<br />
might, however, be to apply a two-stage processing, where first the RST <strong>of</strong> the full<br />
day is calculated, then the <strong>cloud</strong> mask <strong>for</strong> the full day is repeated with the cur-<br />
rent day’s RST. Furthermore, this stand-alone approach implies that, under some<br />
circumstances, the T-score cannot be calculated <strong>for</strong> certain pixels. This is the case<br />
either at the beginning <strong>of</strong> the processing sequence, when the RST map is not defined<br />
due to the lack <strong>of</strong> clear-sky BT values in the mosaic, or if no clear-sky BT values
SPARC <strong>algorithm</strong> 15<br />
Figure 3.3: Examples <strong>of</strong> a Mannstein function fitted to clear-sky BT108 values <strong>for</strong> a<br />
location in the Swiss main land in summer (left) and winter (right). The error bars are<br />
calculated as defined in Equation 3.5.<br />
Figure 3.4: Example <strong>of</strong> a RST map in summer (left) and winter (right) over the Swiss<br />
Alps Region <strong>of</strong> Interest (ROI; see Table 2.1 <strong>for</strong> details on the ROI). Please note the bright-<br />
ness temperature contrast between the low-lying valleys and the more elevated topography,<br />
especially in winter.
SPARC <strong>algorithm</strong> 16<br />
had been observed <strong>for</strong> more than tmax days (certain gaps cannot be filled using the<br />
spatial interpolation scheme if large areas are affected). Limitations at the beginning<br />
<strong>of</strong> a processing sequence can be avoided by adding a spin-up time prior to the actual<br />
period <strong>of</strong> interest. Theoretically, one single clear-sky day would be sufficient to gen-<br />
erate a complete RST field. However, escpecially during the winter months when<br />
low stratus <strong>cloud</strong> cover over the Swiss Main Plateau is observed frequently, gaps in<br />
the RST field may still be observed after several weeks <strong>of</strong> observation and longer<br />
spin-up times are required. However, due to the design <strong>of</strong> the scores and the linear<br />
aggregation principle as mentioned above, SPAR<strong>CM</strong>SG can also be employed if not<br />
all scores are calculated. See Section 4.3.2 <strong>for</strong> a discussion <strong>of</strong> a situation where not<br />
all scores are calculated. Another option could be to consider high resolution surface<br />
temperature data from, e.g., the COSMO-2 model over the Swiss Alps (grid spacing<br />
<strong>of</strong> about 2.2 km; Doms and Schaettler, 2002) as a reference in T-score calculation<br />
where the RST map is not defined.<br />
3.3 The SPAR<strong>CM</strong>SG snow detection scheme<br />
The third major modification <strong>of</strong> the SPARC <strong>algorithm</strong> concerns the calculation <strong>of</strong><br />
the snow mask. A daily snow mask is generated through the aggregation <strong>of</strong> all<br />
snow-covered pixels observed under clear-sky conditions during the course <strong>of</strong> the<br />
day. Snow detection in SPAR<strong>CM</strong>SG includes a number <strong>of</strong> different tests, where<strong>of</strong><br />
one includes the calculation <strong>of</strong> the Normalized Difference Snow Index (NDSI). The<br />
NDSI ([−1.0,1.0]) is calculated by dividing the difference <strong>of</strong> reflectances observed<br />
in MSG SEVIRI channels 1 and 4 by their sum and may be regarded as a measure<br />
<strong>of</strong> the abundance <strong>of</strong> snow and ice within the area covered by a pixel (Salomonson<br />
and Appel, 2004). Even though the spectral channels needed <strong>for</strong> the calculation <strong>of</strong><br />
the NDSI are available on the AVHRR sensors <strong>of</strong> the third generation (AVHRR/3),<br />
historical AVHRR sensors lack the channel at 1.6µm. As a consequence, the NDSI<br />
cannot be implemented systematically <strong>for</strong> AVHRR, which is the reason why an<br />
alternative method was originally developed as described in Section 2.2.<br />
The idea behind the new snow detection scheme is the calculation <strong>of</strong> a S-score that<br />
can be regarded as a measure <strong>of</strong> the snow evidence within a pixel. The S-score is<br />
calculated by summing two newly defined scores, the NDSI-score and the FREEZE-
SPARC <strong>algorithm</strong> 17<br />
score, as well as the T- and R-scores as described above. The NDSI-score is defined<br />
as follows:<br />
NDSI =<br />
REF 006 − REF 016<br />
REF 006 + REF 016<br />
NDSI-score = (NDSI − 0.4)40.0, (3.6)<br />
where REF 006 is the reflectance in SEVIRI channel 1 at 0.6µm and REF 016 the<br />
reflectance in channel 3 at 1.6µm. Positive (negative) values <strong>of</strong> the NDSI-score are<br />
obtained <strong>for</strong> snow-covered (snow-free) pixels. The FREEZE-score is calculate as<br />
follows:<br />
FREEZE-score = −|BT 108 − (Tfreeze) + 5|, (3.7)<br />
where Tfreeze is calculated as a sine function oscillating between a maximmum <strong>of</strong><br />
+2 ◦ C in spring and a minimum <strong>of</strong> −2 ◦ C in fall (Khlopenkov and Trishchenko, 2007).<br />
The FREEZE-score is designed to adopt negative values if the difference between<br />
BT108 and Tfreeze is large (positive or negative), and positive values if BT108 is<br />
close to Tfreeze. The S-score is finally aggregated as<br />
S-score = (R − 3.0)(−1.0) + NDSI-score + (T − 3.0)(−1.0) + FREEZE-score (3.8)<br />
The snow mask is subsequently generated from the S-score by classifying pixels<br />
as snow-covered if the S-score is greater than zero. Again, the sensitivity <strong>of</strong> the<br />
<strong>algorithm</strong> may be adjusted by tuning the selected threshold. Snow masks are only<br />
generated <strong>for</strong> observations made at Sun zenith angles (SZAs) smaller than 75 ◦ , since<br />
the suggested approach in some cases failed to accurately detect snow cover in the<br />
proximity <strong>of</strong> the terminator line, mainly due to erroneously high avalues <strong>of</strong> the<br />
NDSI-score. In addition, two criteria are applied to the snow mask:<br />
• For pixels where BT>tfreeze+10.0, the snow mask is set to zero. This threshold<br />
is needed to account <strong>for</strong> subpixel snow in mountains in summer.<br />
• For pixels with reflectance values in the visible channel (HRV or R0.6) <strong>of</strong>
SPARC <strong>algorithm</strong> 18<br />
with Smin=−15.0 and Smax=10.0. Given the limitations <strong>of</strong> the new snow detection at<br />
high SZAs, the snow correction factor s must be adjusted <strong>for</strong> SZAs greater than 75 ◦ .<br />
The preliminary implementation <strong>of</strong> this correction at high SZAs reads as follows:<br />
tf = e SZA−75.0<br />
10.0<br />
s75 = (s + 0.3)tf, (3.10)<br />
where tf is the correction factor in the proximity <strong>of</strong> the terminator line and s75 is<br />
the adjusted snow correction factor. In order to overcome limitations due to low<br />
SZAs, a possible future solution could be not to generate snow masks at all <strong>for</strong><br />
observations made at SZA>75 ◦ , or, adopting the idea <strong>of</strong> the nighttime correction<br />
factor, to reduce (increase) the weights <strong>of</strong> the reflective (thermal) scores in Equation<br />
3.8 in the proximity to the terminator line.<br />
Figure 3.5 displays the single scores <strong>of</strong> the new snow detection scheme (1A, 2A,1B),<br />
the resulting snow mask (1C) as well as the corresponding snow correction factor<br />
s (2B) <strong>for</strong> a day in March 2005. See Figure 4.8 (left) <strong>for</strong> additional SPAR<strong>CM</strong>SG<br />
output <strong>for</strong> the same time slot. The corresponding snow mask generated based on<br />
the original snow detection scheme is provided as a reference (2C).<br />
Major differences are observed both <strong>for</strong> s and the daily snow mask. Given the<br />
importance <strong>of</strong> the snow correction factor s to reduce the weight <strong>of</strong> the reflective<br />
scores in Equation 2.2, such differences in s inevitably translate into discrepancies<br />
in the magnitude <strong>of</strong> F in SPAR<strong>CM</strong>SG compared to SPARCORIG. Comparing the<br />
values <strong>of</strong> F from SPARCORIG in Figure 2.1 (2E) with SPAR<strong>CM</strong>SG F in Figure 4.8<br />
(1E), significant differences become apparent in the Swiss Alps. Panel 2E in Figure<br />
2.1 shows increased values <strong>of</strong> F over snow-covered areas, except where reflective<br />
scores were turned <strong>of</strong> by s (seen as dark areas; see also Panel 1E in Figure 2.1).<br />
Regarding the comparison <strong>of</strong> the daily snow masks, from a purely visual inspection<br />
the novel snow mask seems to be more accurate, with snow cover being understi-<br />
mated by SPARCORIG. Figure 3.6 (left) displays a preliminary time series <strong>of</strong> monthly<br />
snow cover from SPAR<strong>CM</strong>SG, SPARCORIG, and Moderate Resolution Imaging Spec-<br />
troradiometer (MODIS) data over the Swiss Alps. Monthly maximum snow cover<br />
composites <strong>for</strong> the period between July 2004 and June 2005 were generated <strong>for</strong><br />
both SPARC versions by considering all pixels that were classified as ’snow’ at any<br />
time during the month. Monthly MODIS snow composites provide average monthly
SPARC <strong>algorithm</strong> 19<br />
Figure 3.5: Output <strong>of</strong> the novel SPARC snow detection <strong>algorithm</strong> on March 20, 2005<br />
(8:45am). The corresponding snow mask generated based on the original snow detection<br />
scheme is provided as a reference (2C). See Figure 4.8 <strong>for</strong> the corresponding HRV image<br />
and SPAR<strong>CM</strong>SG.
SPARC <strong>algorithm</strong> 20<br />
Figure 3.6: Time series <strong>of</strong> monthly maximum snow-covered area (SCA; [%]) in the Swiss<br />
Alps region <strong>of</strong> interest <strong>for</strong> SPARCORIG, SPAR<strong>CM</strong>SG, and MODIS (see the text <strong>for</strong> more<br />
in<strong>for</strong>mation on the generation <strong>of</strong> the composites).<br />
snow cover <strong>for</strong> a certain pixel. For comparison with SPARC SCA, MODIS data<br />
were trans<strong>for</strong>med into a binary snow mask by classifying a pixel as snow-covered<br />
if the average monthly snow cover was ≥50%. Differences [%] relative to MODIS<br />
are provided in the right panel. An examination <strong>of</strong> SCA time series reveals that<br />
the underestimation <strong>of</strong> SCA by SPARCORIG relative to SPAR<strong>CM</strong>SG is observed in-<br />
dependently <strong>of</strong> the season. Relative to MODIS, SPAR<strong>CM</strong>SG underestimates (overes-<br />
timates) SCA in summer (fall, winter, and spring), whereas SPARCORIG SCA does<br />
not follow the same pattern. During fall, winter, and spring, an overestimation <strong>of</strong><br />
SCA by SPAR<strong>CM</strong>SG is in accordance with theory since ephemeral snow cover may<br />
not be captured by MODIS due to the lower temporal resolution <strong>of</strong> the MODIS<br />
sensor system. However, during summer we would expect the discrepancies to be<br />
small. Further analysis is required to explain the underestimation <strong>of</strong> SCA by both<br />
SPARC <strong>algorithm</strong>s.<br />
3.4 Further modifications<br />
In the original version <strong>of</strong> SPARC the additional scores are only calculated in uncer-<br />
tain situations to obtain more in<strong>for</strong>mation on the state <strong>of</strong> the pixel. As <strong>of</strong> now, this
SPARC <strong>algorithm</strong> 21<br />
distinction is not made in SPAR<strong>CM</strong>SG and all scores are calculated <strong>for</strong> each pixel<br />
and fed into Equation 2.2.<br />
Furthermore, the coefficients <strong>for</strong> the calculation <strong>of</strong> the R-score needed to be modified,<br />
since the contour lines in Khlopenkov and Trishchenko (2007, Figure 3) could not be<br />
reproduced with MSG SEVIRI reflectances using the original coefficients. The new<br />
coefficients <strong>for</strong> the calculation <strong>of</strong> the R-score are: <strong>of</strong>fset=0.02 and scale factor=320<br />
(original coefficients: <strong>of</strong>fset=0.1, scale factor=160).
Chapter 4<br />
Validation Study<br />
4.1 General Remarks<br />
The focus <strong>of</strong> the validation study was set on a region <strong>of</strong> interest (ROI) covering the<br />
Swiss Alps. The boundary coordinates <strong>of</strong> the subset are listed in Table 2.1. A time<br />
period <strong>of</strong> 12 months (from July 1, 2004 to June 30, 2005) was considered. Further-<br />
more, seasons were analyzed independently, where winter included the months from<br />
December to February, spring from March to May, summer from June to August,<br />
and fall from September to November. In addition, differences between the Climate<br />
Monitoring (<strong>CM</strong>) Satellite Application Facility (<strong>SAF</strong>) <strong>cloud</strong> mask (EUMETSAT <strong>CM</strong><br />
<strong>SAF</strong> , 2009; Schulz et al., 2009) and SPAR<strong>CM</strong>SG were visually analyzed at full disk<br />
level <strong>for</strong> the 12 UTC slot on June 13, 2006, which is the day selected <strong>for</strong> the <strong>cloud</strong><br />
mask intercomparison study <strong>of</strong> the EUMETSAT workshop (Walther et al., 2009).<br />
4.2 Data and Methods<br />
4.2.1 ASRB data<br />
In<strong>for</strong>mation on <strong>cloud</strong> coverage derived from data <strong>of</strong> the Alpine Surface Radia-<br />
tion Budget (ASRB) network (Philipona et al., 1996; Marty and Philipona, 2000),<br />
was used <strong>for</strong> the validation <strong>of</strong> the SPAR<strong>CM</strong>SG <strong>cloud</strong> mask. Two parameters were<br />
compared to SPAR<strong>CM</strong>SG output at each ASRB site in the Swiss Alps (Dürr and<br />
22
SPARC <strong>algorithm</strong> 23<br />
Philipona, 2004): the Cloud-Free Index (CFI) derived from downward longwave ra-<br />
diation, temperature, and relative humidity in the Swiss Alps, and Partial Cloud<br />
Amount (PCA), which is derived from the CFI. The CFI adopts values >1.0 <strong>for</strong><br />
<strong>cloud</strong> cover and values
SPARC <strong>algorithm</strong> 24<br />
computational considerations, a shorter spin-up time <strong>of</strong> only 5 days was selected <strong>for</strong><br />
the full disk processing.<br />
The SPARC <strong>algorithm</strong> was run both in a MSG as well as a MFG mode. While<br />
the MSG mode includes five spectral channels, the MFG mode only relies on MSG<br />
channels 12 and 9 in order to simulate the spectral properties <strong>of</strong> the MFG sensor<br />
(see Table 4.2 <strong>for</strong> more in<strong>for</strong>mation). The restriction to only two spectral channels<br />
aimed at investigating the <strong>algorithm</strong>’s capability to detect <strong>cloud</strong>s with only a limited<br />
amount <strong>of</strong> spectral in<strong>for</strong>mation, analysis that may be useful with regard to the re-<br />
processing <strong>of</strong> MFG long-term data records. While differences to the MSG mode<br />
should be small at nighttime (when both <strong>algorithm</strong>s rely on thermal scores only),<br />
we would expect some discrepancies during daytime, when in<strong>for</strong>mation from all<br />
scores except the B-, UTEMP, and T-scores is not available in the MFG mode.<br />
As listed in Table 4.2, channels 12 (HRV) and 1 can be used interchangeably in MSG<br />
mode. If the HRV channel is selected, all other bands are resampled from native<br />
0.06 ◦ ×0.06 ◦ to the higher spatial resolution <strong>of</strong> the HRV channel (0.02 ◦ ×0.02 ◦ ). This<br />
implies that scores calculated using the HRV channel (Table 4.2) are influenced by<br />
eventual sub-grid <strong>cloud</strong>s that are not represented in the remaining scores to the same<br />
extent. For the analysis presented here the HRV channel was selected, neglecting<br />
uncertainties due to subgrid-scale features.<br />
For the comparison <strong>of</strong> ASRB measurements with SPARC output, a 9 pixels ×<br />
9 pixels area around each ASRB site was considered. At a spatial resolution <strong>of</strong><br />
0.02 ◦ ×0.02 ◦ this approximately corresponds to an area <strong>of</strong> 15×15 km 2 . As <strong>for</strong> the<br />
comparison with CFI, all values <strong>of</strong> F within the subset were averaged to obtain a<br />
single value <strong>of</strong> F per time slot and ASRB site. In order to compare the <strong>cloud</strong> mask<br />
with ASRB PCA, <strong>for</strong>mer was trans<strong>for</strong>med into partial <strong>cloud</strong> amount in ninths <strong>of</strong><br />
the 15×15 km 2 area. This was achieved by counting the number <strong>of</strong> pixels classified<br />
as ’<strong>cloud</strong>’ (thin and opaque) within the 9 pixels × 9 pixels subset. The resulting<br />
82 classes were then stratified into 10 classes <strong>of</strong> increasing <strong>cloud</strong> contamination (the<br />
0/9 class only contained entirely <strong>cloud</strong> free, i.e. 0/81, scenes).
SPARC <strong>algorithm</strong> 25<br />
Table 4.2: The spectral channels <strong>of</strong> SEVIRI used as input <strong>for</strong> SPARC in MSG and MFG<br />
mode, together with the SPARC scores associated with each spectral channel. Channels<br />
12 (HRV) and 1 can be used interchangeably in MSG mode, however, all channels are<br />
resampled to higher spatial resolution (0.02 ◦ ×0.02 ◦ ) if the HRV channel is selected instead<br />
<strong>of</strong> Channel 1.<br />
Mode<br />
Channel involved scores MSG MFG<br />
HRV (12) B (over land), N X X<br />
1 B (over land), N (X) -<br />
2 B (over water), N, Utext X -<br />
3 R X -<br />
9 T, Utemp, C X X<br />
10 C X -<br />
4.3 Results and Discussion<br />
4.3.1 Comparison with ASRB Cloud Parameters<br />
CFI vs. F<br />
Figure 4.1 displays a scatter-density plot <strong>for</strong> the comparison <strong>of</strong> CFI with F <strong>for</strong> day<br />
(left) and night observations (right; all ASRB sites and months were combined). A<br />
Sun zenith angle (SZA) threshold <strong>of</strong> 88 ◦ was used to separate day and night obser-<br />
vations. The dotted lines indicate the thresholds used <strong>for</strong> CFI (Section 4.2.1) and F<br />
(Section 3.1), respectively, to distinguish between <strong>cloud</strong>y and clear-sky observations<br />
(values <strong>of</strong> CFI
SPARC <strong>algorithm</strong> 26<br />
Figure 4.1: Scatter-density plot <strong>of</strong> CFI vs. F <strong>for</strong> all sites and 12 months combined <strong>for</strong><br />
day (left) and night observations (right). Dotted lines indicate the thresholds used <strong>for</strong><br />
CFI (Section 4.2.1) and F (Section 3.1), respectively, to distinguish between <strong>cloud</strong>y and<br />
clear-sky observations. Colors indicate the frequency <strong>of</strong> occurence <strong>for</strong> a certain CFI/F<br />
combination. A bin size <strong>of</strong> 0.07 (15.0) was selected <strong>for</strong> CFI (F); r: linear correlation<br />
coefficient.<br />
and night. Very large positive values <strong>of</strong> F are also observed, yet at very small<br />
frequencies. The agreement <strong>for</strong> nighttime observations was clearly weaker (r 2 =0.22).<br />
For values <strong>of</strong> CFI>1.0, F mostly adopts slightly negative values, which is indicative<br />
<strong>of</strong> an underestimation <strong>of</strong> nighttime <strong>cloud</strong> cover by SPAR<strong>CM</strong>SG, even though Dürr and<br />
Philipona (2004) report that the application <strong>of</strong> the CFI was found to overestimate<br />
nighttime <strong>cloud</strong> cover. A comparison with synoptic measurements will be needed to<br />
gain additional and more detailed insight into nighttime perormance <strong>of</strong> SPAR<strong>CM</strong>SG.<br />
Similar to Figure 4.1, the CFI/F relationship is shown in Figure 4.2, but <strong>for</strong> each<br />
season separately as defined in Section 4.1. For daytime observations, a strong<br />
Figure 4.2 (following page): Similar to Figure 4.1, but <strong>for</strong> each season separately.
SPARC <strong>algorithm</strong> 27
SPARC <strong>algorithm</strong> 28<br />
Figure 4.3: Observation frequencies <strong>of</strong> all possible PCA (in eights)/<strong>CM</strong>MSG (in ninths)<br />
combinations <strong>for</strong> the 12 month validation period during daytime (left) and nighttime<br />
(right). Colors indicate the frequency <strong>of</strong> occurence <strong>for</strong> a certain PCA/<strong>CM</strong>MSG combi-<br />
nation; corresponding percentage values are also provided.<br />
DFI/F relationship was observed in all seasons, even though better agreements were<br />
found <strong>for</strong> winter and spring (r 2 =0.59 and r 2 =0.61, respectively). Similar to Figure<br />
4.1, predominantly negative values <strong>of</strong> F are observed where CFI>1.0, showing that<br />
the underestimation <strong>of</strong> nighttime <strong>cloud</strong> cover by SPAR<strong>CM</strong>SG relative to ASRB CFI<br />
is observed independently <strong>of</strong> the season.<br />
PCA vs. SPAR<strong>CM</strong>SG Cloud Mask<br />
Two different analyses were per<strong>for</strong>med in order to validate the SPAR<strong>CM</strong>SG <strong>cloud</strong><br />
mask (hereafter referred to as <strong>CM</strong>MSG) with ASRB PCA. Firstly, frequencies <strong>of</strong><br />
occurence <strong>of</strong> certain PCA (in eights)/<strong>CM</strong>MSG (in ninths) <strong>cloud</strong> cover combinations<br />
were analyzed (Figures 4.3 and 4.4). Secondly, the accuracy <strong>of</strong> <strong>CM</strong>MSG was ana-<br />
lyzed separately <strong>for</strong> specific ASRB sites (Figure 4.5 and 4.6). Interestingly, Figure<br />
4.3 shows that in the majority <strong>of</strong> the cases during daytime (left) and nighttime<br />
(right), either complete clear-sky (PCA: 0/8, <strong>CM</strong>MSG: 0/9) or complete <strong>cloud</strong> cov-<br />
erage (PCA: 8/8, <strong>CM</strong>MSG: 9/9) was observed at the ASRB sites. Partial <strong>cloud</strong>
SPARC <strong>algorithm</strong> 29<br />
overcast is only observed in a limited number <strong>of</strong> cases, with
SPARC <strong>algorithm</strong> 30
SPARC <strong>algorithm</strong> 31<br />
plots suggest a clear-sky conservative behavior <strong>of</strong> SPAR<strong>CM</strong>SG, meaning that a pixel<br />
is only classified as ’clear-sky’ where <strong>cloud</strong> evidence is very low. In partial <strong>cloud</strong> cover<br />
situations, there is likely to be some <strong>cloud</strong> evidence at a certain pixel if the neigh-<br />
boring pixel is <strong>cloud</strong> covered, which eventually causes <strong>CM</strong>MSG to assign the pixel<br />
to the ”thin <strong>cloud</strong>” or ”opaque <strong>cloud</strong>” category. This is reflected in the tendency<br />
to classify partial <strong>cloud</strong> cover as completely <strong>cloud</strong> covered. At nighttime, results<br />
<strong>for</strong> partial <strong>cloud</strong> cover exhibit a more bimodal distribution, with partial <strong>cloud</strong> cover<br />
being classified as either clear-sky or complete overcast in most cases. Results <strong>for</strong><br />
the high elevation site at Weissfluhjoch (Figure 4.6) were similar, however, the per-<br />
centage <strong>of</strong> accurately classified daytime clear-sky scenes (70%) was lower compared<br />
to the Payerne site. Furthermore, clear-sky scenes were found to be misclassified as<br />
complete overcast in 8.4% <strong>of</strong> the cases during daytime.<br />
Figure 4.5 (following page): Histogram <strong>of</strong> <strong>CM</strong>MSG output <strong>for</strong> each PCA class <strong>for</strong><br />
daytime (top) and nighttime observations (bottom). Numbers represent relative frequencies<br />
[%]. Results are shown <strong>for</strong> the ASRB site in Payerne (cp. Table 4.1).
SPARC <strong>algorithm</strong> 32
SPARC <strong>algorithm</strong> 33<br />
Figure 4.6: Similar to Figure 4.5, but <strong>for</strong> the ASRB site at Weissfluhjoch (cp. Table<br />
4.1).
SPARC <strong>algorithm</strong> 34<br />
Figure 4.7: Similar to Figure 4.1 (Section 4.3.1), but <strong>for</strong> SPAR<strong>CM</strong>FG mode.<br />
4.3.2 MFG Simulation<br />
Similar to Figure 4.1 in Section 4.3.1, Figure 4.7 displays the comparison <strong>of</strong> CFI<br />
and F <strong>for</strong> the entire 12 month period, however, <strong>for</strong> SEVIRI data processed in the<br />
SPAR<strong>CM</strong>FG mode (see Section 4.2.2 <strong>for</strong> more details). As expected from theory,<br />
differences between both SPARC modes are minor during nighttime. Again, low<br />
values <strong>of</strong> F at values <strong>of</strong> CFI>1.0 are indicative <strong>of</strong> an underestimation <strong>of</strong> nighttime<br />
<strong>cloud</strong> cover. For daytime observations the relationship between CFI and F is weaker<br />
<strong>for</strong> SPAR<strong>CM</strong>FG, with r 2 =0.42 (r 2 =0.52 <strong>for</strong> SPAR<strong>CM</strong>SG; Figure 4.1).<br />
Note that daytime F adopts smaller positive values <strong>for</strong> <strong>cloud</strong>y situations in SPAR<strong>CM</strong>FG<br />
mode due to the smaller number <strong>of</strong> contributing scores. This becomes clear with a<br />
more detailed look at a SEVIRI sample time slot (8:45 am) on March 20, 2005 (Fig-<br />
ure 4.8). The RST mosaics <strong>of</strong> both SPARC modes (B1 and B2) generated based on<br />
clear-sky observations <strong>of</strong> the previous 10 days are quite similar and result in similar<br />
T-scores (C1 and C2; see also Section 2). However, since the R-score is not available<br />
in SPAR<strong>CM</strong>FG, the final SPARC scores (F) clearly differ in magnitide (E1 and E2),<br />
which translates into significantly different <strong>cloud</strong> masks (F1 and F2). Not only does
SPARC <strong>algorithm</strong> 35<br />
SPAR<strong>CM</strong>FG classify larger areas as clear-sky compared to SPAR<strong>CM</strong>SG, but it also<br />
assigns more <strong>cloud</strong>s to the ”thin <strong>cloud</strong>” category.<br />
Following the analysis in Section 4.3.1, the histograms <strong>of</strong> <strong>cloud</strong> mask output in<br />
SPAR<strong>CM</strong>FG mode (<strong>CM</strong>MFG) <strong>for</strong> each PCA class is displayed in Figures 4.9 and 4.10<br />
<strong>for</strong> the ASRB sites in Payerne and Weissfluhjoch, respectively. At the Payerne<br />
site, <strong>cloud</strong> cover appeared to be underestimated by SPAR<strong>CM</strong>FG not only during<br />
nighttime, but also during daytime, with 15% <strong>of</strong> the PCA 8/8 situations being<br />
classified as clear-sky. This is in contrast to the results in Figure 4.5, where a very<br />
good agreement between PCA (8/8) and <strong>CM</strong>MSG (9/9) was demonstrated (Section<br />
4.3.1). The same pattern, yet less pronounced, was observed at the Weissfluhjoch<br />
site, with a misclassification (PCA 8/8 vs. <strong>CM</strong>MFG 9/9) in 3.2% <strong>of</strong> the cases.<br />
Figure 4.8 (following page): Output <strong>of</strong> SPAR<strong>CM</strong>SG (B1-F1) and SPAR<strong>CM</strong>FG (B2-F2)<br />
on March 20, 2005 (8:45 am). The HRV channel (A1) and B-score (A2) are displayed<br />
once as they are identical <strong>for</strong> both SPARC configurations. R-score is not available in<br />
SPAR<strong>CM</strong>FG mode (D2), which results in significantly different <strong>cloud</strong> masks (0: clear-sky,<br />
1: thin <strong>cloud</strong>, 2: thick <strong>cloud</strong>). See the text <strong>for</strong> further explanations.
SPARC <strong>algorithm</strong> 36
SPARC <strong>algorithm</strong> 37<br />
Figure 4.9: Similar to Figure 4.5, but <strong>for</strong> SPAR<strong>CM</strong>FG mode.
SPARC <strong>algorithm</strong> 38<br />
Figure 4.10: Similar to Figure 4.6, but <strong>for</strong> SPAR<strong>CM</strong>FG mode.
SPARC <strong>algorithm</strong> 39<br />
4.3.3 Validation at Full Disk Level<br />
Similar to Figure 1 in Walther et al. (2009) we compared the output <strong>of</strong> SPAR<strong>CM</strong>SG<br />
and SPAR<strong>CM</strong>FG at full disk level to the <strong>cloud</strong> mask produced at the Climate Mon-<br />
itoring Satellite Application Facility (<strong>CM</strong>-<strong>SAF</strong>). The 12:00 UTC time slot on June<br />
13 th 2006 was selected. Prior to comparison, SPAR<strong>CM</strong>SG and SPAR<strong>CM</strong>FG output<br />
was trans<strong>for</strong>med into binary <strong>cloud</strong> masks by combining thin and opaque <strong>cloud</strong>s into<br />
one single class. Similarly, the <strong>CM</strong><strong>SAF</strong> <strong>cloud</strong> in<strong>for</strong>mation was converted to a binary<br />
<strong>cloud</strong> mask as described in Walther et al. (2009). The percentage <strong>of</strong> total <strong>cloud</strong> cover<br />
was then determined. The resulting <strong>cloud</strong> masks are displayed in Figure 4.11. From<br />
a visual comparison <strong>cloud</strong> cover looks quite similar <strong>for</strong> SPAR<strong>CM</strong>SG and SPAR<strong>CM</strong>FG,<br />
which is also confirmed by <strong>cloud</strong> cover percentage <strong>of</strong> 43% and 42%, respectively.<br />
The <strong>CM</strong>-<strong>SAF</strong> <strong>cloud</strong> mask estimates a <strong>cloud</strong> coverage <strong>of</strong> 55% <strong>of</strong> the full disk.<br />
In order to get more detailed insight into SPARC per<strong>for</strong>mance at full disk level,<br />
five subsets were extracted from the full disk, some <strong>of</strong> them similar to the ROIs<br />
in Walther et al. (2009): Central Europe, North Atlantic, Sahara Desert, Western<br />
Africa, and Southern Atlantic. Results are reported in Figure 4.12. At ROI level,<br />
larger differences between SPAR<strong>CM</strong>SG and SPARGMFG are visible. SPAR<strong>CM</strong>FG dis-<br />
plays a higher <strong>cloud</strong> cover percentage than SPAR<strong>CM</strong>SG, however, except <strong>for</strong> the<br />
Sahara ROI, both SPARC <strong>algorithm</strong>s detect significantly less <strong>cloud</strong> cover compared<br />
to the <strong>CM</strong>-<strong>SAF</strong> <strong>cloud</strong> mask. In the Sahara ROI, the extent <strong>of</strong> bright desert surfaces<br />
classified as ’<strong>cloud</strong>’ by both SPARC modes is significant. The reason <strong>for</strong> the over-<br />
estimation <strong>of</strong> <strong>cloud</strong> cover in these areas is tw<strong>of</strong>old: as <strong>for</strong> SPAR<strong>CM</strong>SG, it is mainly<br />
the very large values <strong>of</strong> the R-score over bright desert surfaces that translate into<br />
erroneous values <strong>of</strong> F, since the characteristics <strong>of</strong> the R-score do not account <strong>for</strong> the<br />
spectral properties <strong>of</strong> bright barren land. A possible approach could be to modify<br />
the R-score (Figure 4.13) or to introduce a weighting factor <strong>for</strong> barren land (similar<br />
to s <strong>for</strong> snow conditions). The <strong>for</strong>mer approach was implemented <strong>for</strong> the full disk<br />
processing presented here, however, Figure 4.12 suggests that further improvements<br />
are required. Latter approach may be considered <strong>for</strong> SPAR<strong>CM</strong>FG, as misclassifica-<br />
tion is mainly caused by strongly positive B-scores (R-score cannot be calculated in<br />
MFG mode). The positive B-scores over desert in MFG (and MSG) might also be
SPARC <strong>algorithm</strong> 40<br />
Figure 4.11: Comparison <strong>of</strong> the <strong>CM</strong>-<strong>SAF</strong> (top, right), SPAR<strong>CM</strong>SG (bottom, left), and<br />
SPAR<strong>CM</strong>FG (bottom, right) <strong>cloud</strong> masks <strong>for</strong> the full disk on 13 June, 2008, 12:00 UTC.<br />
The corresponding HRV channel is displayed in the top left panel. The <strong>cloud</strong> mask data<br />
was trans<strong>for</strong>med into a binary <strong>cloud</strong> mask (see the text <strong>for</strong> further explanations). The red<br />
squares (A-E) delineate the selected regions <strong>of</strong> interest (cp. Figure 4.12). Percent values<br />
indicate the <strong>cloud</strong> cover percentage on the full disk. Please note that the top and bottom<br />
212 lines were excluded from analysis due to computational constraints, which is also the<br />
reason why <strong>cloud</strong> cover percentage <strong>for</strong> the <strong>CM</strong>-<strong>SAF</strong> <strong>cloud</strong> mask differs from the value<br />
provided in Walther et al. (2009). Color legend: green: clear-sky land; blue: clear-sky<br />
water surface; yellow: <strong>cloud</strong>.
SPARC <strong>algorithm</strong> 41<br />
Figure 4.12: Comparison <strong>of</strong> the <strong>CM</strong>-<strong>SAF</strong> (second row), SPAR<strong>CM</strong>SG (third row), and<br />
SPAR<strong>CM</strong>FG (fourth row) <strong>cloud</strong> masks <strong>for</strong> five selected Regions <strong>of</strong> Interest (ROIs) within<br />
the full disk. Corresping HRV images are provided in the top row. The ROIs are: A<br />
- Central Europe, B - Northern Atlantic, C - Sahara Desert, D - Western Africa, E<br />
- Southern Atlantic (see also Figure 4.11). Percent values indicate the <strong>cloud</strong> coverage<br />
within the ROI.<br />
overcome by use <strong>of</strong> a similar score <strong>for</strong>mula as <strong>for</strong> the T-score:<br />
B-score = (V IS − ALB − <strong>of</strong>fset)scale (4.1)<br />
instead <strong>of</strong>: B-score = (V IS − <strong>of</strong>fset)scale,<br />
where ALB is the surface albedo and VIS the reflectance in Channel 1 (REF006)<br />
or 12 (HRV). This would mean that a spin-up <strong>for</strong> the surface albedo similar to the<br />
spin-up <strong>of</strong> the surface radiative temperature would need to be implemented. This<br />
would then make the snow-factor obsolete <strong>for</strong> the B-score. The surface albedo spin-<br />
up might, however, be hampered by rapidly changing surface albedo after snowfall<br />
periods.
SPARC <strong>algorithm</strong> 42<br />
Figure 4.13: Similar to Figure 3 in Khlopenkov and Trishchenko (2007), but only <strong>for</strong> the<br />
reflectance in SEVIRI channels 1 (REF006) and 3 (REF016) over bright desert surfaces.<br />
The red line demonstrates how the R-score should be modified in order to account <strong>for</strong> the<br />
spectral properties <strong>of</strong> bright desert surfaces.<br />
4.4 Concluding Remarks<br />
The Separation <strong>of</strong> Pixels Using Aggregated Rating Over Canada (SPARC) algo-<br />
rithm represents an alternative to conventional scene identification <strong>algorithm</strong>s as<br />
it primarily outputs a single <strong>cloud</strong> contamination rating (F) instead <strong>of</strong> a <strong>cloud</strong><br />
mask with a limited number <strong>of</strong> categories. The rating itself isessentially calculated<br />
from three sub-scores generated from in<strong>for</strong>mation on the brightness temperature<br />
(T-score), brightness (B-score), and the reflectance in the visible and short-wave<br />
infrared portion <strong>of</strong> the electromagnetic spectrum (R-score). Even though the algo-<br />
rithm was originally designed <strong>for</strong> AVHRR, it may be applied to data from other<br />
multispectral sensors.
SPARC <strong>algorithm</strong> 43<br />
The application <strong>of</strong> SPARC to data from the MSG SEVIRI sensor was described in<br />
this report. Three major modifications were presented:<br />
1. In addition to F, SPAR<strong>CM</strong>SG outputs a <strong>cloud</strong> mask with three classes (clear,<br />
thin <strong>cloud</strong>, thick <strong>cloud</strong>) based on thresholds applied to F. The application <strong>of</strong><br />
thresholds to F facilitates the trans<strong>for</strong>mation <strong>of</strong> the <strong>cloud</strong> mask from a clear-<br />
sky conservative to a <strong>cloud</strong> conservative state. In addition, the <strong>cloud</strong> mask<br />
error <strong>for</strong> each pixel is calculated based on F, which may be a valuable source<br />
<strong>of</strong> in<strong>for</strong>mation <strong>for</strong> users <strong>of</strong> this simplified <strong>cloud</strong> mask.<br />
2. SPAR<strong>CM</strong>SG is a stand-alone <strong>algorithm</strong> and does not rely on auxiliary data.<br />
In contrast to the original SPARC version, which uses an external surface<br />
skin temperature dataset as a reference in the calculation <strong>of</strong> the T-score,<br />
SPAR<strong>CM</strong>SG takes advantage <strong>of</strong> the high temporal resolution <strong>of</strong> MSG SEVIRI<br />
to derive a brightness temperature mosaic from SEVIRI clear-sky observations<br />
obtained during a preceeding time interval.<br />
3. A new snow detection module is implemented. Based on two newly defined<br />
sub-scores, it outputs both a daily snow mask as well as a snow score (S-score),<br />
which may be regarded as a measure <strong>of</strong> snow contamination within a pixel.<br />
For the period from July 2004 to June 2005, SPAR<strong>CM</strong>SG output was compared to<br />
<strong>cloud</strong> cover in<strong>for</strong>mation from the Alpine Surface Radiation Budget (ASRB) Net-<br />
work in the Swiss Alps: Cloud Free Index (CFI) and Partial Cloud Amount (PCA).<br />
Results show a good agreement between CFI and F mainly <strong>for</strong> daytime observa-<br />
tions. However, results also point to an underestimation <strong>of</strong> nighttime <strong>cloud</strong> cover by<br />
SPAR<strong>CM</strong>SG. Adjustment <strong>of</strong> the <strong>of</strong>fset and scale factors used to calculate the single<br />
scores may be required to account <strong>for</strong> this deficiency.<br />
If only two SEVIRI spectral channels are fed into SPARC to simulate data from<br />
the Meteosat First Generation (MFG), a good agreement between CFI and F is<br />
obtained during daytime. Finally, application <strong>of</strong> SPARC at full disk level revealed an<br />
underestimation <strong>of</strong> <strong>cloud</strong> cover compared to the Climate Monitoring (<strong>CM</strong>)-Satellite<br />
Application Facility (<strong>SAF</strong>) <strong>cloud</strong> mask. Both <strong>cloud</strong> masks should next be compared<br />
over longer time periods to synoptic <strong>cloud</strong> observations to evaluate possible biases<br />
and spatiotemporal inconsistencies.
SPARC <strong>algorithm</strong> 44<br />
Overall, SPAR<strong>CM</strong>SG has proven to provide interesting new opportunities <strong>for</strong> <strong>cloud</strong><br />
detection using the MSG SEVIRI sensor. Results are particularly promising with<br />
regard to the application <strong>of</strong> SPARC to data <strong>of</strong> the MFG satellite series, since the<br />
<strong>algorithm</strong> was found capable <strong>of</strong> generating <strong>cloud</strong> in<strong>for</strong>amtion with only a limited<br />
amount <strong>of</strong> spectral in<strong>for</strong>mation from two SEVIRI channels.
SPARC <strong>algorithm</strong> 45<br />
Bibliography<br />
Doms, G., and U. Schaettler (2002), The nonhydrostatic limited-area model lm<br />
part i: Dynamics and numerics. scientific documentation, deutscher wetterdienst,<br />
<strong>of</strong>fenbach, germany, available online: http://www.cosmo-model.org.<br />
Dürr, B. (2009), Technical report <strong>of</strong> msg-based ir <strong>cloud</strong>-index processing chain, Tech.<br />
rep., Sunergy GmbH.<br />
Dürr, B., and R. Philipona (2004), Automatic <strong>cloud</strong> amount detection by sur-<br />
face longwave downward radiation measurements, J. Geophys. Res., 109 (D5),<br />
D05,201–.<br />
EUMETSAT <strong>CM</strong> <strong>SAF</strong> (2009), Algorithm theoretical basis document.<br />
cm-saf product cm-02, cm-08 and cm-14. <strong>cloud</strong> fraction, <strong>cloud</strong> type<br />
and <strong>cloud</strong> top parameter retrieval from seviri, Available online:<br />
http://www.cmsaf.eu/bvbw/generator/<strong>CM</strong><strong>SAF</strong>/Content/Publication/atbd pdf/<br />
<strong>SAF</strong> <strong>CM</strong> DWD ATBD CFC CTH CTO SEVIRI 1.0,templateId=raw,<br />
property=publicationFile.pdf/<strong>SAF</strong> <strong>CM</strong> DWD ATBD CFC CTH CTO SEVIRI<br />
1.pdf.<br />
Khlopenkov, K. V., and A. P. Trishchenko (2007), SPARC: New <strong>cloud</strong>, snow, and<br />
<strong>cloud</strong> shadow detection scheme <strong>for</strong> historical 1-km AVHRR data over Canada,<br />
Journal Of Atmospheric And Oceanic Technology, 24 (3), 322–343.<br />
Lillesand, T., R. Kiefer, and J. Chipman (2004), Remote Sensing and Image Inter-<br />
pretation, fifth ed., John Wiley & Sons, New York.<br />
Mannstein, H., H. Broesamle, C. Schillings, and F. Trieb (1999), Using a meteosat<br />
<strong>cloud</strong> index to model the per<strong>for</strong>mance <strong>of</strong> solar thermal power stations, in Eumetsat<br />
Conference, p. 239246, Eumetsat, Copenhagen, Denmark.<br />
Marty, C., and R. Philipona (2000), The clear-sky index to separate clear-sky from<br />
<strong>cloud</strong>y-sky situations in climate research, Geophys. Res. Lett., 27 (17), 2649–2652.<br />
Mesinger, F., and Coauthors (2006), North american regional reanalysis, Bulletin <strong>of</strong><br />
the American Meteorological Society, 87, 343360.
SPARC <strong>algorithm</strong> 46<br />
Philipona, R., C. Marty, and C. Fróhlich (1996), Measurement <strong>of</strong> the Longwave<br />
Radiation Budget in the Alps, in IRS96, A. DEEPAK Publishing.<br />
Salomonson, V. V., and I. Appel (2004), Estimating fractional snow cover from<br />
MODIS using the normalized difference snow index, Remote Sensing <strong>of</strong> Environ-<br />
ment, 89 (3), 351–360.<br />
Schulz, J., P. Albert, H.-D. Behr, D. Caprion, H. Deneke, S. Dewitte, B. Dürr,<br />
P. Fuchs, A. Gratzki, P. Hechler, R. Hollmann, S. Johnston, K.-G. Karlsson,<br />
T. Manninen, R. Müller, M. Reuter, A. Riihelä, R. Roebeling, N. Selbach, A. Tet-<br />
zlaff, W. Thomas, M. Werscheck, E. Wolters, and A. Zelenka (2009), Operational<br />
climate monitoring from space: the eumetsat satellite application facility on cli-<br />
mate monitoring (<strong>CM</strong>-<strong>SAF</strong>), Atmospheric Chemistry and Physics, 9 (5), 1687–<br />
1709, doi:10.5194/acp-9-1687-2009.<br />
Walther, A., A. Heidinger, A. Thoss, and R. Roebeling (2009), Report on comparison<br />
study <strong>for</strong> the EUMETSAT workshop, Ascona, Switzerland.