Tensor decomposition-based sparsity divergence index for hyperspectral anomaly detection.

Recently, some methods exploiting both the spatial and spectral features have drawn increasing attention in hyperspectral anomaly detection (AD) and they perform well. In addition, a tensor decomposition-based (TenB) algorithm treating the hyperspectral dataset as a three-order tensor (two modes for space and one mode for spectra) has been proposed to further improve the performance for AD. In this paper, a method using the sparsity divergence index (SDI) based on tensor decomposition (SDI-TD) is proposed. First, three modes of the hyperspectral dataset are obtained by tensor decomposition. Then, low-rank and sparse matrix decomposition is employed separately along the three modes and three sparse matrices are acquired. Finally, SDIs based on the three sparse matrices along the three modes are obtained, and the final result is generated by using the joint SDI. Experiments tested on the real and synthetic hyperspectral dataset reveal that the proposed SDI-TD performs better than the comparison algorithms.