Low-rank and Sparse Representation Hyperspectral Anomaly Detection Based on Spatial-Spectral Dictionary

Low-rank and sparse representations are widely used for hyperspectral anomaly detection. To fully exploit the spatial-spectral information of dictionary atoms, this study proposes a low-rank and sparse representation hyperspectral anomaly detection algorithm based on a spatial-spectral dictionary. To include all background categories in the spatial-spectral background dictionary, K-means clustering was applied. The feature similarity between pixels of each category and their neighboring pixels within a local window was calculated to obtain the residual difference constant matrix for each category, which was then used to compute the anomaly degree of each pixel. Representative atoms from each class were selected to form the spatial-spectral background dictionary, after which the abnormal and background components were separated using low-rank and sparse representations. The original data were reconstructed using this spatial-spectral background dictionary. Experimental results on five hyperspectral datasets demonstrate that the proposed method has good detection performance and can effectively improve the detection accuracy.