Effects of the multiscaled-band partitioning on the abundance estimation

Materials of interest comprised in a hyperspectral image often present intra-class spectral variability inherent to their natural compositional make-up. Obtaining the best spectral representations of such materials with respect to a given application is critical for both identification and spatial mapping. Recently, a multiscaled-band partitioning (MSBP) approach has been developed for detecting and clustering spectrally similar but physically distinct materials. In this work, it is examined 1) whether the endmember clusters of the multiscaled-band partitioning contribute to an improved abundance estimation compared to other endmember extraction methods and, 2) to what extent different unmixing strategies can retain the spectral variability of the extracted endmember clusters in the resulted abundance maps. Experiments were conducted using an airborne hyperspectral dataset highlighting the potential of MSBP for the unmixing process in case of materials with intra-class variability.