Graph Feature Refinement and Adaptive Matching Network for Cross-Scene Hyperspectral Image Classification

Cross-scene hyperspectral image (HSI) classification faces significant challenges due to domain shifts caused by variations in imaging conditions and environmental factors across different scenes. Existing works mainly address this issue through data augmentation and adversarial learning, which often leads to overlooks the structural dependencies among samples and limited adaptability. To overcome these limitations, this study aims to explore feature augmentation as an alternative paradigm for cross-domain generalization in HSI classification, and further integrates graph neural networks to model topological relationships among samples for enhanced structural consistency. Specifically, we design a spectral refinement augmentation module to suppress domain-specific dominant spectral components and enhance domain-invariant features. A structure-guided matching mechanism is further introduced to explicitly align original and enhanced features via structural relationships, thereby mitigating feature bias. Additionally, a multi-graph contrastive learning strategy is employed to enforce class consistency across different graph-based feature representations. Thorough experiments conducted on three well-known HSI benchmark datasets—Houston, Pavia, and Indiana—reveal that the proposed approach markedly surpasses current state-of-the-art DG methods in cross-scene HSI classification. These results highlight its strong robustness and enhanced ability to generalize effectively.