Hyperspectral Anomaly Detection Based on Background Purification and Spectral Feature Extraction

Hyperspectral anomaly detection ( HAD) does not require a priori information, and accurate discrimination is made by analyzing the difference between the anomalies and the background pixels. However, the bands of hyperspectral images are highly correlated with each other. There is a lot of redundant information between them, which causes the band selection to be difficult to accurately distinguish between background and anomalies. This paper introduces background purification and feature extraction strategies to increase the distinction between anomalies and background pixels. To be specific, the domain transformation extracts discriminative sample features. The row-constrained low- rank sparse matrix decomposition is utilised to obtain low-rank background matrices to construct purer background to highlight the anomalies. The sliding window strategy is adopted to divide the subspace to reduce the spatial correlation. Highly representative and low redundancy bands are selected for band selection in the local region. Finally, the local region is detected by RX and the map is obtained by domain-valued normalisation of the local results. Experiments on several HSI data sets show that the proposed method can suppress the background well. It can also make full use of the spectral information and achieves acceptable detection accuracy.