Hyperspectral anomaly detection leveraging spatial attention and right-shifted spectral energy

In this research, we have proposed a novel anomaly detection algorithm for processing hyperspectral images (HSIs), called the Graph Attention Network–Beta Wavelet Graph Neural Network-based Hyperspectral Anomaly Detection (GAN–BWGNN HAD). This algorithm treats each pixel as a node in a graph, where edges represent pixel correlations and node attributes correspond to spectral features. The algorithm integrates spatial and spectral information, utilizing graph neural networks to identify nonlinear relationships within the image, thereby enhancing anomaly detection precision. The K-nearest neighbor (KNN) algorithm facilitates the creation of edges between pixels, enabling the incorporation of distant pixels and improving resilience to noise and local irregularities. The GAN component incorporates an adaptive attention mechanism to dynamically prioritize relevant spatial features. The BWGNN component employs beta wavelets as a localized bandpass filter, effectively identifying spectral anomalies by addressing the right-shifted spectral energy phenomenon. Furthermore, the utilization of beta wavelets obviates the necessity for computationally intensive Laplacian matrix decompositions, thereby enhancing processing efficiency. This approach effectively integrates spatial and spectral information, providing a more accurate and efficient solution for hyperspectral anomaly detection. Experiments on six real-world hyperspectral datasets and one simulated dataset demonstrate the superior performance of our proposed method. It consistently achieved high Area Under the Curve (AUC) values (e.g., 0.9986 on AVIRIS-II, 0.9961 on abu-beach-2, 0.9982 on abu-urban-3, 0.9999 on Salinas-simulate, 0.9872 on Cri), significantly outperforming state-of-the-art methods. The proposed method also exhibited sub-second detection times (0.20–0.28 s) on most datasets, significantly faster than traditional methods (achieving a speedup of 100 to 500 times) and deep learning models (achieving a speedup of 6 to 8 times).