S2DGNet: Spectral-Spatial Decoupled Domain Generalization for Cross-Domain Hyperspectral Image Classification

Traditional hyperspectral image (HSI) classification models often suffer from performance degradation in cross-domain scenarios due to distribution differences between domains. To address this issue, a spectral-spatial decoupled domain generalization (DG) network (S2DGNet) is proposed for cross domain HSI classification. Specifically, a Group-wise Spectral Randomization Module (GSRM) is developed to simulate spectral shifts and introduce controlled group-level perturbations while preserving spectral dependencies, enabling effective characterization of cross-domain spectral shifts. Meanwhile, a Frequency-guided Spatial Perturbation Module (FSPM) is designed with residual-based multi-level perturbation to enhance spatial diversity while maintaining structural consistency. Moreover, a spectral-spatial decoupled discrimina- tor (SSD) is developed to extract spectral-spatial features and adaptively fuse classification results for domain-invariant representation learning. Experimental results on the Houston, Pavia, and HyRank datasets demonstrate that the proposed method achieves superior classification performance and strong cross domain generalization ability in HSI classification tasks.