A Local Similarity Driven Model for Blind Hyperspectral Unmixing with Spectral Variability

Abstract

ABSTRACT Blind hyperspectral unmixing is a key technique for mixing spectral analysis, which separate the endmember spectra from hyperspectral image and evaluate their fractional abundances. Recent studies have focused on addressing spectral variability during unmixing procedure, and many new models have been proposed to characterize endmember variation. To further improve the unmixing performance, a local similarity-driven unmixing model is proposed in this paper. We construct a coefficient matrix which can reflect the difference between adjacent spectra of input data, and then take this matrix as pixel-wise weight of spatial constraints of both fractional abundances and scaling factors. By incorporating the proposed local similarity driven constraints with extend linear mixing model, the endmembers extraction and abundances estimation can be better guided by input data. Our model can efficiently solved by alternating non-negative least squares algorithm, the experiments on simulation data and real data show that the proposed method is less time consuming, and our unmixing results are superior to or comparable to the state-of-the-art methods.