Endmember extraction analysis considering endmember variability

In recent years, several kinds of endmember extraction algorithms have been proposed from hyperspectral data set which extracts/selects one single standard endmember spectrum for each existing endmember class or scene component. In this article, endmember variability is considered to the mixture spectrum analysis by representing each endmember by a set or a bundle of spectra. Our method contains four steps in order to extract endmember bundles: (1) looking for homogeneous area; (2) threshold segmentation for candidate endmember set; (3) getting initial clustering center with hierarchical clustering; (4) getting endmember bundles and center spectra with mean-shift algorithm. Experiments with real hyperspectral data sets indicate that the proposed strategy has significantly improvement by considering endmember variability to the original hyperspectral data.