Abstract
Weakly supervised remote sensing change detection reduces annotation costs but suffers from incomplete class activation maps, unreliable pseudo-label propagation, and insufficient discrimination between adjacent change regions. To address these limitations, we propose PRIWCD, a weakly supervised framework for remote sensing change detection. PRIWCD introduces a dynamic prototype comparison module to enhance high-level differential features by modeling unchanged regions and diverse change patterns with multiple dynamic prototypes. A dual relationship propagation guidance module further refines class activation maps by jointly exploiting changes in bi-temporal attention relationships and feature similarity. In addition, instance separation regularization improves feature compactness within pseudo-instances and enhances the separation between changed regions and the background. Experiments on WHU-CD, CLCD, and DSIFN demonstrate the effectiveness of PRIWCD. The proposed method achieves the best mean F1 scores of 79.10%, 62.33%, and 62.68%, respectively, outperforming the compared weakly supervised methods. The results demonstrate that PRIWCD effectively improves pseudo-label quality and provides robust change detection under image-level supervision.