Abstract
Hyperspectral anomaly detection (HAD) aims to identify Earth’s surface anomalies without prior knowledge. However, most existing unsupervised HAD methods rely on training data roughly partitioned into background and anomalies, suffering from severe class imbalance and yielding feature representations heavily tied to scene-specific distributions. In contrast, supervised hyperspectral classification networks can extract well-organized spectral–spatial representations from diverse land-cover types; yet, this valuable knowledge remains underutilized in HAD due to the absence of labeled anomaly samples. To circumvent this limitation, this paper proposes a novel knowledge transfer paradigm from classification tasks to anomaly detection. The proposed framework employs a pretrained convolutional enhanced Transformer (CET) classification network as the teacher model, transferring its rich spectral–spatial context encodings via feature-level distillation. Concurrently, by synergizing latent feature alignment, Gaussian mixture model (GMM) priors, and reconstruction objectives, the student network is equipped with compact background modeling and highly discriminative representation abilities, which provide robust support for feature-space anomaly detection. Extensive experiments on spaceborne and airborne hyperspectral datasets demonstrate that our method substantially improves detection accuracy against complex terrestrial backgrounds with fast inference, while exhibiting superior cross-scene generalization.