Spatial sparsity-based blind source separation method including non-negative matrix factorization for multispectral image unmixing

In this paper, we propose an unsupervised spatial method in order to unmix each pixel of a remote sensing multispectral image. This method is related to the blind source separation (BSS) problem, and is based on sparse component analysis (SCA) and non-negative matrix factorization (NMF). Our approach consists in identifying the mixing matrix in the first stages, by using a spatial correlation-based SCA method, combined with clustering. An NMF method is used to extract spatial sources in the last stage. The overall proposed method is applicable to the globally underdetermined BSS model in multispectral remote sensing images. An experiment based on realistic synthetic mixtures is performed to evaluate the feasibility of the proposed approach. We also show that our method significantly outperforms the sequential maximum angle convex cone (SMACC) method.