Variation Autoencoder of Spatial-Spectral Joint Mask for Hyperspectral Anomaly Detection

Abstract

In recent years, autoencoders and their variants have emerged as effective tools for hyperspectral anomaly detection. Nevertheless, owing to the complex distribution of anomalous regions and the similarity in spatial-spectral features, these models often reconstruct anomalies and backgrounds simultaneously, hindering their ability to distinguish between them and reducing detection accuracy. To address this issue, we propose a novel hyperspectral anomaly detection method based on a spatial-spectral joint mask variational autoencoder (VAE). By combining the probabilistic modeling capabilities of VAEs with a masking-based attention mechanism, our method enables more precise extraction of essential background information in localized regions. Specifically, the spatial-spectral joint masking technique is proposed to guide the network to concentrate on background features across multiple dimensions, tackling issues of spatial structure approximation and spectral redundancy. To further enhance robustness in noisy and complex environments, we iteratively refine the reconstructed residual image through recursive filtering. Extensive comparative experiments and ablation studies on multiple public datasets demonstrate that our approach consistently outperforms existing methods in detection accuracy.