MSSAAE: multiscale spatial–spectral attention autoencoder framework for hyperspectral anomaly detection

Hyperspectral anomaly detection (HAD) aims to detect objects that are significantly different from their surrounding background. As one important task in the field of remote sensing, HAD has received extensive research attention. In recent years, many researchers have introduced deep-learning-based methods, including autoencoders (AEs), and achieved great detection results. However, most AE-based approaches fail to fully exploit the spatial and spectral information in hyperspectral images (HSIs). To address this issue, we propose a multiscale spatial–spectral attention framework based on AE, which enhances HAD performance through effective feature extraction and fusion. Specifically, the encoder consists of two modules: the first one is the spatial–spectral attention module, which adaptively weights spatial and spectral dimensions to enhance structural and discriminative spectral features and improves the model’s ability to capture critical information across the whole HSI; the second one is the multiscale convolution fusion module, which integrates features at different scales to jointly preserve local details and global contextual information. Furthermore, to improve background reconstruction, we introduce a clustering-based preprocessing step to reduce anomaly interference and enhance background representation. Experimental results on five public datasets demonstrate that the proposed method outperforms several state-of-the-art HAD methods regarding detection performance.