Teacher-Anchored Prototype-Guided Domain Adaptation for Cross-Scene Hyperspectral Image Classification
TL;DR
Stable teacher-guided contrastive learning and PATS improve the accuracy and robustness of cross-scene HSIC, a coarse-to-fine adaptation framework that combines stable teacher guidance, contrastive structure learning, and reliable target sample mining.
Abstract
Cross-scene hyperspectral image classification (HSIC) remains difficult because large domain shifts make target pseudolabels unreliable, blur class boundaries, and destabilize target supervision. This article proposes teacher-anchored prototype-guided domain adaptation, a coarse-to-fine adaptation framework that combines stable teacher guidance, contrastive structure learning, and reliable target sample mining. A teacher updated by exponential moving average provides stable embeddings, predictions, and anchors for contrastive distillation, which improves intraclass compactness and interclass separation. A prototype-aware target selection (PATS) strategy then evaluates target reliability with teacher confidence and source-prototype consistency, while classwise mining and pair-validity constraints keep the selected target set balanced and usable for contrastive learning. A dual-classifier adversarial alignment module further reduces domain discrepancy and supports feature transfer. Experiments on three representative cross-scene transfer tasks show consistent improvements over strong domain-adaptation baselines, with especially clear gains under severe target imbalance. These results show that stable teacher-guided contrastive learning and PATS improve the accuracy and robustness of cross-scene HSIC. The source code for TAPDA is publicly available online at github.