GenerChange: Uncertainty-Aware Domain Generalization Network for Cross-Domain Remote Sensing Change Detection

Remote sensing change detection (RSCD) aims to identify changes in ground objects at the same geographic location across dual-temporal images. Despite deep learning’s success in this field, most existing methods fail to generalize from a source domain (seen) to a target domain (unseen) without retraining. Cross-domain RSCD can learn a transferable change detection model in the source domain, enabling target-domain change detection without retraining. However, existing methods ignore the stylistic variations between dual-temporal images and the lack of prior knowledge about the target domain, which together lead to increased model aleatoric and epistemic uncertainty. To address the above issues, we propose an uncertainty-aware domain generalization network (GenerChange) for cross-domain RSCD. It employs a dual-covariance style-debias mechanism and an evidence-aware uncertainty estimation module to reduce model uncertainty. First, the dual-covariance style-debias mechanism can effectively disentangle content features from style variations, reducing aleatoric uncertainty by aligning the covariance between the enhanced and original images. Furthermore, the evidence-aware uncertainty estimation module models the Dirichlet distribution of pixel change probabilities. This helps the model focus on domain-invariant information during training, which can be generalized to the target domain, enabling the model to reduce the epistemic uncertainty. Extensive experimental results demonstrate that the proposed method outperforms existing domain generalization methods for RSCD.