Prototype-Constrained Background Modeling for Hyperspectral Anomaly Detection
Abstract
Hyperspectral anomaly detection aims to identify pixels that significantly deviate from the surrounding background without prior knowledge of anomaly classes. Existing reconstruction-based methods usually learn background representations through autoencoders and detect anomalies using reconstruction residuals. However, in complex scenes, the background often exhibits multimodal distributions, which limits the stability and discriminability of conventional background modeling strategies. To address this issue, this article proposes a prototype-constrained background modeling method for hyperspectral anomaly detection. An orthogonal attention autoencoder is adopted as the reconstruction backbone, and an adaptive prototype aggregation mechanism is introduced in the latent space to organize background features through soft assignment. In addition, compactness and separability constraints are jointly imposed to improve intraprototype consistency and interprototype discrimination. An adaptive reconstruction weighting strategy based on assignment entropy is further employed to emphasize uncertain samples during training. Experiments on five public hyperspectral datasets, including Gulfport, Pavia, Texas Coast, HYDICE, and Salinas, demonstrate that the proposed method achieves competitive performance against several representative methods, with an average AUC of 0.9888. Ablation studies further verify that the prototype mechanism becomes consistently effective when combined with the separability constraint.