Non-negative sparse representation for anomaly detection in hyperspectral imagery

A novel non-negative sparse representation (NSR) model was proposed for hyperspectral anomaly detection. The key idea was that a background pixel can be approximately represented as a sparse linear combination of its surrounding neighbors, while an anomalous pixel cannot. The non-negativity and one-to-one constraints on the sparse vector were imposed for physical meaning and better discrimination power of the algorithm. In order to exclude the potential anomalous pixels presented in the background dictionary, the atoms which were similar to the center pixel was pruned. Then the NSR model was solved by non-negative orthogonal matching pursuit (NOMP) algorithm, and the reconstruction errors were directly used for determining the anomalies. Finally, experimental results on real hyperspectral data set demonstrate the effectiveness of the proposed algorithms by comparing it with state-of-the-art algorithms.