Sparse Non-negative Matrix Factorization on GPUs for Hyperspectral Unmixing

Hyperspectral unmixing is a typical problem of blind source separation, which can be solved by non-negative matrix factorization (NMF). NMF based on sparsity, which can increase the efficiency of unmixing, is an important topic in hyperspectral unmixing. In this paper, a novel constrained sparse (CS) NMF (CSNMF) method for hyperspectral unmixing is proposed, where a new sparsity term is introduced to improve the stability and accuracy of unmixing model. The corresponding algorithm is designed based on the alternating direction method of multiplies. In order to further enhance the execution speed, parallel optimization of hyperspectral unmixing based on CSNMF on graphics processing units (CSNMF-GPU) is investigated and compared in terms of both accuracy and speed. The realization of the proposed method using compute unified device architecture (CUDA) on GPUs is described and evaluated. Experimental results based on the simulated hyperspectral datasets show that the proposed CSNMF method can improve the unmixing accuracy steadily. The tests comparing the parallel optimization of CSNMF on GPUs with the serial implementation and the multicore implementation, using both simulated and real hyperspectral data, demonstrate the effectiveness of the CSNMF-GPU approach.