Hyperspectral Anomaly Detection Based on a Smooth Prior and Tensor Low-Rank Representation

Hyperspectral anomaly detection is used to separate spatial spectral feature pixels with evident differences from the background. As traditional matrix-based detection methods convert hyperspectral cube data into matrices, they lose spatial spectral information, and noise can interfere with the detection of abnormal information, thereby affecting the detection rate. Therefore, this study proposes a method based on a tensor low-rank representation that retains the geometric features of hyperspectral data and uses a linear iterative clustering algorithm to obtain a dictionary that fully preserves the background feature information of hyperspectral data. Moreover, based on the global and local similarities of the background tensor, this study introduces weighted low-rank and total variation regularization to suppress noise and redundant information. Finally, an efficient solving algorithm was designed using the alternating direction method of multipliers. The experimental results show that the average detection accuracy of this algorithm on four real-scene datasets was 99.44%, verifying its effectiveness.