UAFDcd: Uncertainty-Aware Feature Decoupling Network for Remote Sensing Change Detection

Remote sensing change detection (RSCD) aims to identify regions in co-registered bi-temporal images that undergo specific changes over time. However, most existing deep networks for RSCD typically adopt mixed feature extraction and fusion. From a causal perspective, this is detrimental to clear class boundary delineation and thus exerts the negative impact on model performance. To address this issue, this paper proposes an uncertainty-aware cross-domain feature decoupling network for RSCD. Firstly, this paper systematically verifies the rationality of feature disentanglement for RSCD based on causal inference theory. The feature decoupler is constructed through correlation analysis to extract shared features of bi-temporal images, and differential disentanglement is used to separate unique features and obtain change results, which effectively cuts off the causal backdoor path from invariant features to CD outputs. Then, decoupled prototype contrastive learning is introduced to enhance the discriminability of decoupled feature. In the specific decoupled feature space, corresponding pixels are pulled close to positive prototypes and pushed away from negative prototypes, while low attention weights are assigned to heterogeneous pixels. This not only improves intra-class compactness and inter-class separability, but also further blocks the backdoor paths of confounding factors affecting prediction results. Finally, an uncertainty estimation module is introduced to improve the confidence of decoupled features. This module learns change and invariant evidence from decoupled features and constructs their parameterized Dirichlet distribution. It not only produces the probabilities of change and invariance but also quantifies model uncertainty, and reduces the model epistemic uncertainty via evidential loss. Experiments demonstrate that the proposed method achieves significant improvements in both qualitative and quantitative metrics.