Cross-domain few-shot hyperspectral classification via orthogonal feature disentanglement

In cross-domain few-shot hyperspectral image classification, the limited availability of labeled target-domain samples renders the model highly sensitive to domain shifts. Existing feature disentanglement approaches struggle to simultaneously suppress domain-specific noise and retain cross-domain discriminative semantics. This leads to the entanglement of domain-shared and domain-specific components, increases the risk of negative transfer, and ultimately becomes the bottleneck limiting further improvements in accuracy and generalization. To address these challenges, this paper presents an Orthogonal Feature Disentanglement Network (OFD-Net). Using orthogonal subspace decomposition, OFD-Net projects features into two mutually exclusive subspaces: domain-shared and domain-specific. The domain-shared subspace focuses on extracting cross-domain invariant features, while the domain-specific subspace retains local domain discriminative information. This dual-stream architecture effectively suppresses interference from irrelevant inter-domain features. Additionally, feature orthogonality constraints enhance the model's adaptability to target domain shifts. OFD-Net also adopts a multi-task learning framework. Cross-domain alignment loss ensures the consistency of shared feature distributions between the source and target domains, while inter-class discriminative loss improves the class separability of specific features, creating a hierarchical feature optimization mechanism. On four public benchmark datasets including Indian Pines, Pavia University, Salinas, and Houston, OFD-Net achieves Overall Accuracy of 80.33%, 85.80%, 93.33%, and 80.05% respectively. Its performance outperforms existing state-of-the-art methods, demonstrating superior cross-domain transfer robustness and feature discriminative capability.