Causal Disentanglement and Progressive Metric Learning Domain Generalization Network for Cross-Scene Hyperspectral Image Classification

Cross-scene hyperspectral image classification (HSIC) is of great importance for real-world remote sensing applications since it enables models trained on existing scenes to be deployed in new scenes where labeled samples are scarce or unavailable. Existing methods are largely developed based on domain adaptation (DA) techniques, which improve transferability by leveraging target-domain data during training. However, this setting is often impractical in real applications because target-domain samples may be unavailable in advance. In contrast, domain generalization (DG) does not rely on any target-domain data during training and aims to learn models that can be directly transferred to unseen domains, making it a more practical paradigm for HSIC. Nevertheless, existing DG methods still face difficulties in simultaneously enriching source-domain diversity, suppressing environment-related interference, and improving feature discriminability, which limits their generalization ability across scenes. To address these issues, causal disentanglement and progressive metric learning domain generalization network (CDPMNet) is proposed for cross-scene HSIC. The proposed framework improves generalization by jointly enhancing source-domain diversity, disentangling causal semantic features from environment-related interference, and progressively refining feature discriminability. Specifically, a spectral–texture generation module (STGM) is designed to construct semantically consistent pseudo-domain samples through coordinated spatial texture perturbation and frequency-domain spectral modulation, thereby enriching source-domain diversity. A causal disentanglement module (CDM) is further introduced to decouple causal features related to label semantics from noncausal features associated with environmental interference, thereby suppressing shortcut learning caused by scene-dependent variations. In addition, a progressive metric learning (PML) strategy is developed to progressively refine the embedding space by preserving feature structure and enhancing class discrimination, thereby improving both discriminability and generalization capability. Extensive experiments on the Houston13–18 dataset and two underwater coral hyperspectral benchmarks, namely Scene I–Scene II and Scene III–Scene IV, show that the proposed CDPMNet achieves higher classification accuracy than seven state-of-the-art DA and DG methods.