SwinMSNet: A Swin Transformer-Based Multiscale Network for Spatio-Environmental Changes
Reliable change detection in remote sensing imagery is critical for supporting applications, such as rapid disaster response and urban development monitoring, where small-scale and incomplete changes are often overlooked. This study proposes SwinMSNet, a Swin transformer-based multiscale network designed to improve recall and robustness in spatio-environmental change detection. The architecture incorporates a cross-scale feature fusion module to capture both large- and small-scale variations, and a contextual feature enhancement block to suppress background noise in high-resolution imagery. To address the problem of incomplete ground truth, a false-positive-guided refinement strategy is introduced to improve annotation quality and enhance the detection of missed changes. Experiments on building and landslide datasets demonstrate that SwinMSNet achieves consistent improvements, with recall enhanced by 17.9% and F1-score by 12.7% over baseline models. These results highlight the potential of SwinMSNet to support real-world Earth observation tasks, including disaster damage assessment, infrastructure monitoring, and land-use planning.