Hyperspectral image unmixing via simultaneously sparse and low rank abundance matrix estimation

This paper proposes a semi-supervised unmixing method, which simultaneously exploits the abundance sparsity and spatial correlation that are inherent in hyperspectral images. To capture spatial information, a sliding window containing neighboring image pixels is first defined. Then, a regularized optimization problem is defined that incorporates a mixed penalty consisting of a sparsity inducing ℓ1 norm and a matrix rank-penalizing weighted trace norm. To solve the optimization problem, an efficient alternating direction method of multipliers based algorithm is developed. Experimental results conducted on both synthetic and real data illustrate the effectiveness of the proposed algorithm.