Spectral Unmixing Applied to Desert Soils for the - Naval ...

also be seen in Figure 23. The leveling off of the pixel purity index curve at
approximately 2000–3000 pixels is the basis that determines the maximum number of
pixels to be used in the visualization step. To be sure that the best subset possible was
used, 3000 pixels were set for the maximum. Once the purest pixels were identified and
previewed, they were then extracted as endmembers.
Two options for endmember extraction are possible, the first is imagery
derived endmembers, which intuitively means that the endmembers are extracted from
the imagery itself and is generally thought of as the best option (Kruse and Boardman,
2011, Kruse and Perry, 2009). The second option is to input user supplied endmembers,
which means one can input field or laboratory collected spectra into the algorithm and try
to detect those as target material. Both options were run on the two data sets with the
highest spatial resolution for the Canyonlands site. The data sets were AVIRIS flight
logs f110512t01p00r07 and f110512t01p00r08 with 2.9 m and 3.3 m spatial resolution,
respectively. In some cases a third data set with 15.2 m resolution was utilized
(f110623t01p00r10). This was done for two reasons; the first is because the BSC
component does not necessarily take up large areas and the bigger the pixel, the less
likely it is that BSCs will be detectable in the imagery. To test how the BSCs and other
endmembers might be distributed in the imagery, they were compared to the results using
imagery derived endmembers. The second reason it was run on two data sets was for a
repeatability test with the same endmembers at the same success rate. To try and create
the most similar conditions, the data sets with the most similar spatial resolution and
topography were selected; the third data set was used when further testing was required.
Figure 24 shows some of the 48 endmembers derived from the imagery sets compared to
the spectra collected using the ASD imaging spectrometer in the Owens/Death Valley
area and input as user supplied endmembers. MF and infeasibility scores were then used
to determine areas with high MF and low infeasibility scores. These correspond to areas
with high abundance and feasible mixtures. These areas were then analyzed to see if the
endmembers could actually be related to any features within the imagery and potentially
mapped as areas of interest.
46