MEANet: A Novel Multiscale Edge-Aware Network for Building Change Detection in High-Resolution Remote Sensing Images

Remote sensing building change detection (RSBCD) is critical for land surface monitoring and understanding interactions between human activities and the ecological environment. However, existing deep learning-based RSBCD methods often result in mis-detected pixels concentrated around object boundaries, mainly due to ambiguous object shapes and complex spatial distributions. To address this problem, we propose a new Multiscale Edge-Aware change detection Network (MEANet) that accurately locates edge pixels of changed objects and enhances the separability between changed and unchanged pixels. Specifically, a high-resolution feature fusion network is adopted to preserve spatial details while integrating deep semantic information, and a multi-scale supervised contrastive loss (MSCL) is designed to jointly optimize pixel-level discrimination and embedding space separability. To further improve the handling of difficult samples, hard negative sampling is adopted in the contrastive learning process. We conduct comparative experiments on three benchmark datasets. Both Visual and quantitative results demonstrate that our new MEANet significantly reduces misclassified pixels at object boundaries and achieve superior detection accuracy compared to existing methods. Especially on the GZ-CD dataset, MEANet improves F1-Score and mIoU by more than 2% compared with ChangeFormer, demonstrating strong robustness in complex scenarios. It is worth noting that the performance of MEANet may still be affected by extremely complex edge textures or highly blurred boundaries. Future work will focus on further improving robustness under such challenges and extending the method to broader RSBCD scenarios.