Uncertainty Modeling for Domain Generalization in Cross-Scene Hyperspectral Image Classification

Domain generalization is a promising approach for cross-scene hyperspectral image classification. Existing domain generalization-based methods usually neglect uncertainty in feature statistics, which may make the trained model vulnerable to statistical changes. Uncertainty modeling may solve this problem by explicitly characterizing the distribution of feature statistics. However, introducing it directly faces two obstacles: properly describing uncertainty is challenging, and the convergence of the entire model is not guaranteed. To address the above issues, in this paper, we propose an uncertainty modeling method for domain generalization (UMDG) in cross-scene hyperspectral image classification, which consists of a newly designed framework and a convergence theorem. The framework develops two strategies to model uncertainty in input and label spaces, respectively. In the input space, we propose a variational style augmentation approach to model the explicit distribution of feature statistics across domains by sampling from a learned Gaussian distribution. To maintain consistency, in the label space, we introduce a label mixup strategy that leverages the same uncertainty to increase label diversity. Furthermore, to ensure convergence of the entire model, we establish a convergence theorem by demonstrating that the model converges in expectation to a stationary point. The major contributions of UMDG include designing an uncertainty modeling method for domain generalization in cross-scene hyperspectral image classification and providing a convergence theorem. Our experiments on three popular datasets have substantiated the significance of our work.