A Comparative Analysis of Endmember Extraction Algorithms Using Aviris Hyperspectral Imagery

Spectral unmixing techniques are widely used for hyperspectral data analysis and quantification. Many novel applications have been developed from the unmixing point of view, including surface constituent identification for land use mapping, disaster assessment, geology, biological process analysis and change detection (Keshava and Mustard, 2002). All existing unmixing approaches require a previous step where the spectral signatures of ground constituents (endmembers) are identified (Kruse, 1998; Boardman et al., 1995), and then a mixture model is used to estimate the abundance fractions of these signatures by expressing individual pixels as a linear or non-linear combination of endmembers (Bateson et al., 2000). The accuracy of the quantification depends strongly on how accurate endmembers are identified in the first step.