RACL: reliability-aware contrastive learning for weakly supervised change detection

Abstract

Weakly supervised change detection aims to identify land-cover changes from bi-temporal remote sensing images using limited annotations. However, relying solely on image-level supervision often leads to inaccurate change localization and noisy pseudo-labels. To address this issue, we propose an end-to-end framework for weakly supervised change detection. A shared Vision Transformer is first employed to extract deep representations from bi-temporal images and construct difference features to highlight potential changes. Then, an Adaptive Patch Reliability Modeling (APRM) module is designed to estimate patch-level response reliability and filter unreliable features. Furthermore, a Reliability-Aware Contrastive Learning (RACL) strategy is introduced to enhance the separability of change representations. Based on the learned features, class activation maps are generated and refined using DenseCRF to obtain pseudo labels for segmentation training. Experiments demonstrate that the proposed method effectively improves weakly supervised change detection performance.