TENSOR LOW-RANK SPARSE REPRESENTATION LEARNING FOR HYPERSPECTRAL ANOMALY DETECTION

Some existing anomaly detection methods convert a 3-D hyperspectral data cube into a 2-D matrix, which inevitably destroys the spatial-spectral structure information of the hyperspectral data, resulting in the degradation of detection performance. In this paper, we propose a tensor low-rank sparse representation learning (TLRAD) method for hyperspectral anomaly detection (HAD), which can effectively maintain the spatial-spectral structure of raw hyperspectral data. Specifically, based on tensor low-rank representation (TLRR) learning, both low-rank constraints and sparsity constraints are simultaneously imposed on the coefficient tensor to capture the global and local spatial-spectral structure information of hyperspectral image (HSI), respectively. For anomaly tensor, the tensor l21-norm is applied to encourage the group sparsity of anomalous pixels. Furthermore, the tensor robust principal component analysis (TRPCA) approach is utilized to construct a robust background dictionary tensor. Experimental results gained employing two real hyperspectral datasets prove the superiority of the proposed approach compared to some state-of-the-art algorithms.