A novel approach for endmember bundle extraction using spectral space splitting

Abstract

The traditional endmember extraction methods search for a fixed set of endmembers, each one assigned to a single material. However, in many real applications, the materials of interest may present spectral variability which is related to subtle absorption features crucial for their discrimination. Thus, extracting multiple spectra or bundles for different materials is considered a more effective approach for data analysis, accounting for intra-class spectral variability. In this work, a novel approach is introduced which aims at obtaining a full representation of materials in a given scene, specifically including those with low spectral contrast. The approach enables a traditional endmember extraction method, the N-FINDR, to extract image endmember bundles exploiting the original spectral bands through a spectral space splitting. Experiments were conducted using an airborne hyperspectral dataset for extracting endmembers of mafic and ultramafic lithological units and preliminary results show the potential usefulness of the new approach.