Hyperspectral Anomaly Detection Via Dual Collaborative Representation

Abstract

Window-based operation is a general technique for hyperspectral anomaly detection. However, the problem remains that background knowledge containing abnormal information often affects the attributes of test pixels. In this article, a dual collaborative representation (DCR)-based hyperspectral anomaly detection method is proposed to solve the above problem effectively, which consists of the following main steps. First, low-rank and sparse matrix decomposition is employed to obtain a low-rank background matrix. Then, the density peak clustering algorithm is applied to the low-rank background matrix to calculate the density information of the pixels in a sliding dual window. Specifically, pixels with the highest density are selected as the pure background pixel set to approximately represent the test pixels in this work. Next, the test pixels are approximated by the linear combination of pixels in the inner window. Finally, a decision function based on the residuals of this dual-stage collaborative representation is utilized to detect abnormal pixels. Experimental results on several hyperspectral datasets demonstrate that the proposed DCR method can both improve the separability between abnormal pixels and their corresponding background and show better detection performance with respect to state-of-the-art anomaly detection methods in terms of detection accuracy.