RANet: Relationship Attention for Hyperspectral Anomaly Detection

Abstract

Hyperspectral anomaly detection (HAD) is of great interest for unknown exploration. Existing methods only focus on local similarity, which may show limitations in detection performance. To cope with this problem, we propose a relationship attention-guided unsupervised learning with convolutional autoencoders (CAEs) for HAD, called RANet. First, instead of only focusing on the local similarity, RANet, for the first time, pays attention to topological similarity by leveraging the graph attention network (GAT) to capture deep topological relationships embedded in a customized incidence matrix from absolutely unlabeled data mixed with anomalies. Notably, the attention intensity of GAT is self-adaptively controlled by adjacency reconstruction ability, which can effectively reduce human intervention. Next, we adopt an unsupervised CAE to jointly learn with the topological relationship attention to achieve satisfactory model performance. Finally, on the basis of background reconstruction, we detect anomalies by the reconstruction error. Extensive experiments on hyperspectral images (HSIs) demonstrate that our proposed RANet outperforms existing fully unsupervised methods.