Endmember Extraction with Unknown Number of Sources for Hyperspectral Unmixing

Abstract

Hyperspectral Unmixing requires a prior known number of endmembers present at the scene. Most blind hyperspectral unmixing methods perform the estimation of the number of endmembers as a separate step before extracting those endmembers. Recent approaches to bridge this gap and simultaneously count and extract endmembers increase accuracy of the overall unmixing process and promote the applicability of unsupervised unmixing. In this paper, we propose a new method to extract the endmembers without requiring a prior number of endmembers or a separate step to finding them. The method also attempts to address a few limitations of the existing methods which mainly follow distance-based linear mixing models. The proposed method is based on the fact that the non-endmembers are either linear combinations of the endmembers or a spectral variant of an endmember. The method follows a linear mixing model which imposes the non-negative and sum-to-one constraints on the abundances. Our method first violates the sum-to-one constraint and iteratively adds endmembers to reinforce the constraint. The method removes the redundant spectra by applying a redundancy reduction technique based on correlation. The iterative process and the optimization are applied to a few selected spectra leading to a much faster approach. The proposed method is compared with other competing methods and is found to be faster with increased accuracy.