A Shift Reduction Domain Generalization Network for Hyperspectral Image Cross-Domain Classification

In practical applications, varying imaging conditions cause spectral shifts for the same class across hyperspectral image (HSI) domains. Besides, since annotating HSIs is time-consuming, the number of available labeled samples is insufficient, resulting in strong models to overfit during training. To address these issues, a spectral--spatial structure refactoring-based data representation method and a lightweight shift reduction domain generalization network (SRDGnet) are proposed for cross-domain classification. Specifically, a spectral-spatial exchange attribute extraction module (SE-AEM) is designed to capture fine-grained spectral local features, while an integrated feature extraction head (IFEH) is used to extract global features. They both map their respective features into a low-dimensional space. Subsequently, a domain shift reduction module (DSRM) fuses the features for information interaction to reduce domain shift and then feeds them to a lightweight version of the integrated feature extraction module (IFEM)-light for domain-invariant and discriminative features of classes, which are then used for final classification. Moreover, the training leverages multisource domains and a batch-wise traversal constraint strategy to enhance sample diversity and utilization. Experimental results on four public HSI datasets demonstrate that the proposed method can effectively learn a model with higher generalization ability and stability. On the Houston 2013 dataset, it achieves an overall accuracy (OA) of 89.35%, which is 2.67% higher than the best performance of other comparison methods. The code will be released soon on the site of https://github.com/Cherrieqi/SRDGnet