Background Subspace Projection-based Mahalanobis Distance Method for Hyperspectral Anomaly Detection

Background feature extraction is the core of anomaly detection (AD) in hyperspectral images (HSIs), but it suffers from severe anomaly interference in the complete absence of prior information. As a typical representative, the Mahalanobis distance detector estimates background statistics from the entire scene or local regions, which is subject to anomaly contamination and inverse covariance matrix problems. In this regard, how to separate the background component from HSIs and serve the pure background statistics calculation is a crucial issue that needs to be solved. In this paper, a background subspace projection-based Mahalanobis distance method is proposed for hyperspectral anomaly detection by combining the spectral and spatial information. Under the joint sparsity constraint, joint sparse representation is performed on the local region to select the most representative background bases. Then, the central pixel is projected into the local background subspace to extract the underlying background data and contribute to background statistics calculation. Finally, the Mahalanobis distance operator is applied to detect the probable anomalies. Experiments on five HSI data sets demonstrate the effectiveness of the proposed method over state-of-the-art methods.