Hyperspectral image target detection approach based on local non-negative sparse coding

Sparse representation based hyperspectral image processing methods can excavate potential relationship in high-dimensional hyperspectral data and reveals the essential characteristic of spectral signal. In this paper a novel hyperspectral image target detection algorithm based on non-negative sparse coding is presented. Compared with classical sparse representation methods, the linear coding coefficients are enforced non-negative. On one side the linear coding process has tangible physical interpretation. On the other side the coding coefficients are proved more discriminative and robust. The locally dynamic dictionary is first constructed with atoms which are produced by a sliding dual window strategy. Then non-negative coefficients of each pixel are calculated with the dynamic dictionary. The discrimination between targets and background are based on the sparsity of the coefficients. We carried extensive experiments on both simulated and real data to verify the effectiveness of the proposed method.