Hyperspectral Anomaly Detection via Graph Out-of-Distribution Detection

Hyperspectral anomaly detection plays a crucial role in remote sensing applications, yet traditional methods face significant challenges in complex scenarios due to restrictive distributional assumptions and limited feature representation capabilities. This paper presents a novel graph neural network approach that integrates out-of-distribution (OOD) detection theory with hyperspectral anomaly detection through a two-stage framework. We first employ Contextual Residual Decomposition (CRD) for initial anomaly screening and data preprocessing, followed by an energy-based OOD detector that leverages graph structures to capture spectral-spatial relationships between pixels. The proposed method constructs K-nearest neighbor graphs in spectral feature space and utilizes message-passing mechanisms to enhance contextual awareness for anomaly discrimination. An energy function specifically designed for hyperspectral characteristics quantifies anomaly degrees while a belief propagation algorithm leverages neighborhood consensus to improve detection reliability. Comprehensive data augmentation strategies spanning spectral and feature domains enhance model generalization across diverse scenarios. Extensive experiments on six widely-used datasets covering airport, beach, and urban environments demonstrate superior performance with an average AUROC of 0.9978, substantially outperforming state-of-the-art methods. Notably, the method achieves remarkable improvements in challenging scenarios, with AIRPORT-1 showing enhancement from 0.8939 to 0.9976 AUROC compared to the best baseline. The framework maintains consistent high performance across varying scene complexities while requiring reasonable computational resources.