A novel graph convolution and frequency domain filtering approach for hyperspectral anomaly detection

Abstract

Abstract This paper introduces a novel algorithm for hyperspectral anomaly detection (HAD) that combines graph-based representations with frequency domain filtering techniques. In this approach, hyperspectral images (HSIs) are modeled as graphs, where each pixel is treated as a node with spectral features, and the edges capture pixel correlations based on the K-Nearest Neighbor (KNN) algorithm. Graph convolution is employed to extract spatial structural features, enhancing the understanding of spatial relationships within the data. Additionally, the algorithm addresses the ’right-shift’ phenomenon in the spectral domain, often associated with anomalies, by using a beta wavelet filter for efficient spectral filtering and anomaly detection. The key contributions of this work include: 1) the use of a graph-based model for HSI that effectively integrates both spatial and spectral dimensions, 2) employing KNN for edge construction to include distant pixels and mitigate noise, 3) spatial feature extraction via graph convolution to provide detailed insights into spatial interconnections and variations, enhancing the detection process, and 4) leveraging the beta wavelet filter to handle the ’right-shift’ spectral phenomenon and reduce computational complexity. Experimental evaluations on four benchmark datasets show that the proposed method achieves outstanding performance with AUC scores of 0.9986, 0.9975, 0.9859, and 0.9988, significantly outperforming traditional and state-of-the-art anomaly detection techniques.

Keywords

• Hyperspectral anomaly detection (HAD)
• Graph neural networks (GNN)
• Graph convolutional networks(GCN)
• Right-shift phenomenon
• Beta wavelet