Cross-scene Hyperspectral Image Classification Based on Frequency Domain Enhancement and Geometry Preservation

Cross-scene hyperspectral image classification (HSIC) is seriously affected by domain bias, which often leads to the decline of model generalization ability in new scenes. It has become a key challenge in this field. Existing methods mostly rely on cross-entropy loss and contrastive learning strategies. The former only focuses on the classification boundary and is difficult to ensure the domain invariance of features; the latter easily leads to the reduction of intra-class feature space and limits the generalization ability. To this end, this paper proposes a hyperspectral image domain generalization (DG) network that integrates frequency domain enhancement and geometry preservation. The Fourier-based frequency enhancement encoder (FBFE encoder) is designed to extract stable features by perturbing the amplitude and preserving the phase; the geometric divergence minimization constraint is introduced to enhance the local geometric consistency between the extended domain and the source domain; and the domain-adaptive contrastive loss (DAC loss) function is constructed to achieve the coordinated optimization of inter-domain alignment and intra-domain expansion. Experimental results on two real hyperspectral remote sensing datasets show that the proposed method outperforms existing advanced methods in classification accuracy and generalization performance.