Variability of the endmembers in spectral unmixing
Abstract
Spectral unmixing is an inverse problem in hyperspectral imaging that aims at recovering the spectra of the pure constituents of an image (called endmembers), as well as at estimating the proportions of said materials in each pixel (called abundances). A linear mixing model is typically used for this purpose, but this approach implicitly assumes that one spectrum can completely characterize each material, while in practice the materials are always subject to intraclass variability. Taking this phenomenon into account within an image amounts to allowing the endmembers to vary on a per-pixel basis. In this chapter, we review and categorize the methods addressing this endmember variability and compare their results on a real data set, thus showing the benefits of incorporating it into the unmixing chain.