Hyperspectral Anomaly Detection Based on Mean Shift Clustering and Low-rank Decomposition

In the hyperspectral anomaly detection methods based on low-rank representation, the absence of prior information often results in a background dictionary with limited representational capacity, leading to ineffective suppression of background pixels. Moreover, conventional dictionary learning methods fail to address the contamination of anomalous pixels during dictionary construction, further degrading detection accuracy. In order to overcome these shortcomings, this paper proposes a novel hyperspectral anomaly detection algorithm integrating mean-shift clustering and low-rank decomposition. First, mean-shift clustering is applied to segment the hyperspectral data, followed by the RX algorithm to select optimal atoms from each cluster for constructing a purified background dictionary. To account for spectral mixing, matrix constraints are relaxed during decomposition. Finally, low-rank decomposition is performed on the hyperspectral image using the constructed dictionary to derive the detection results. Experimental results demonstrate that the proposed method selects more representative dictionary atoms without prior knowledge while minimizing contamination from anomalous pixels. This approach significantly improves detection performance and robustness compared to existing methods.