MLCL: Remote Sensing Change Detection Using Multi-level Contrastive Learning

Yizhou Liang, Chunshi Wang, Bin Zhao

Abstract: Current change detection methods directly use multi-temporal images of the same location as positive sample pairs, which leads neglect of change information in feature extraction, affecting change detection performance. To address this issue, we propose multi-level contrastive learning (MLCL) that utilizes image-level and feature-level differences of bitemporal remote sensing images to guide model learning multi-level difference information to improve accuracy of change detection. Particularly, we use augmented contrastive learning loss and unified contrastive learning loss to reinforce the model’s perception of multi-level difference information. MLCL outperforms the current state-of-the-art methods on three benchmark datasets LEVIR-CD, SYSU-CD and WHU-CD. The code is available at https://github.com/lvvsh/MLCL.

External IDs:dblp:conf/pakdd/LiangWZ25