Sparse Nonnegative Matrix Factorization based on Double Low-Rank and Sparse Matrix Factorization Model for Spectral Variability

To reduce the interference of complex imaging environment on spectral unmixing, we introduce a double low-rank and sparse matrix factorization model (LRSMD) (DLRSMD) to better characterize the intrinsic and variant endmember components in the scene, and propose a sparse NMF based on DLRSMD (DLRSNMF) by accounting for the sparsity of intrinsic and variant abundances. The sparsity of abundances not only satisfies the practical significance of abundances, but also effectively promotes the representation ability of endmembers. The experimental results also demonstrate the better performance of DLRSNMF, which can provide reliable endmembers and abundances.