Hyperspectral Anomaly Detection by Spatial–Spectral Fusion Based on Extreme Value-Entropy Band Selection and Cauchy Graph Distance Optimization

Hyperspectral technology for detecting camouflaged targets in complex backgrounds represents a current research hotspot. Hyperspectral data often contain numerous spectral bands, which can lead to data redundancy. Traditional anomaly detection methods typically assume that the background spectrum follows a Gaussian distribution, which limits their adaptability to complex and non-Gaussian backgrounds. Furthermore, most existing anomaly detection methods fail to fully leverage the spatial information in hyperspectral images. To address these challenges, we propose a hyperspectral anomaly detection by spatial-spectral fusion based on extreme value-entropy band selection and Cauchy graph distance optimization (EBS-CGD). First, to effectively reduce the redundant bands in hyperspectral images, we introduce a hybrid band selection algorithm. This algorithm combines spectral extremum detection with information entropy filtering to select the most representative bands by considering multi-dimensional information. Second, to enhance the model's adaptability to non-gaussian backgrounds and maximize spectral information utilization, we propose a Cauchy graph distance-optimized RX anomaly detection approach. This method replaces the traditional covariance matrix with a significance-weighted Cauchy distance graph, reducing the influence of the target on the background and improving the model's robustness in complex environments. Subsequently, a low-rank and sparse representation model (LRASR) is employed to construct background dictionaries, low-rank coefficient matrices for the background, and sparse anomaly matrices, enabling effective spatial information utilization for detecting anomalous targets. Finally, to fully exploit spatial-spectral information, we employ a weighted fusion of the detection results to reduce false positives and retain more true anomalies. We tested the proposed method on four real hyperspectral camouflage net datasets, each containing 126 bands. Experimental results demonstrate that the proposed EBS-CGD algorithm outperforms other methods in anomaly detection performance.