Reliable Task-Constrained Band Selection Method for Hyperspectral Anomaly Detection

Hyperspectral band selection utilizes a crucial band subset to represent original data. In hyperspectral anomaly detection tailored for specific tasks, detection performance can be enhanced by pre-selecting a subset of bands that are more representative. However, existing methods remain constrained in modeling spatial–spectral dependencies and simultaneously extracting distinct bands’ contribution from the established model, thus struggling to balance effectiveness and stability. To address these issues, we propose a reliable band selection method for anomaly detection. Concretely, we conduct a convolution–transformer hybrid autoencoder architecture to fully exploit the local and global spatial–spectral interdependencies. Next, we design an anomaly–background separability constraint to seamlessly integrate the task priors of anomaly detection into network optimization. Furthermore, we design a spectral attention module to quantify the contribution of different bands during network optimization. Simultaneously, an adaptive band allocation method is designed to optimize the internal structure of the selected band subset. Extensive experiments demonstrate that the proposed method achieves more robust band selection results compared to existing related methods.