A New Approach for Endmember Extraction and Clustering Addressing Inter- and Intra-Class Variability via Multiscaled-Band Partitioning

Abstract

In this paper, a new method is introduced for detecting and clustering spectrally similar but physically distinct materials. The method exploits the spectral information by dividing the spectral domain into band subsets whose width varies from broad to narrower wavelength ranges. Multiple candidate endmembers containing intraclass spectral variability are extracted using a maximum volume-based endmember extraction method at each band subset. Spectral clustering of the extracted spectra is also accomplished by using a multiscaled-band partitioning approach. This allows for the generation of multiscaled clustering identification vectors that can be used to remove partial mixtures and also be used to derive the final set of endmember bundles which retain interclass endmember variability. The proposed method was evaluated using simulated and real hyperspectral data and in comparison with well-known methods for extracting a fixed set or multiple sets of endmembers. Results revealed the advantages of the multiscaled-band partitioning on both multiple endmember extraction and clustering with the latter being an independent module that can be applicable to endmember candidate libraries derived from other methods.