Hyperspectral Anomaly Detection Using Tensor Low-Rank Representation

Existing Low-rank (LR) matrix-based approaches have been widely developed for hyperspectral (HS) anomaly detection (AD). However, the 3-D intrinsic LR structure is inevitably destroyed by the conversion of 2-D matrix-form HS images. In this article, we present a novel tensor LR representation method for hyperspectral anomaly detection (TLRR-AD). The background part of HS images can be represented by a designed background dictionary. The dictionary learning of LR background explores the multiple LR tensor subspace property of the complex background. Sparse anomalies are constrained by a $$L_{2,1}$$ sparse norm. Experimental results of AD show the superior performance of the proposed TLRR-AD among the competing algorithms.