Abstract: [Purposes]Aiming at the problem that traditional change detection methods struggle to distinguish subtle water body changes from semantic categories,this study proposes a deep learning model capable of accurately identifying change regions and outputting their semantic categories.[Methods] A multi-scale feature-guided semantic change detection network (MFG-SCDNet)is designed,which adopts a dual-stream encoder-decoder architecture and integrates a grouped multi-scale aggregation (GMCA)module,an adaptive feature selection (AFSF)module,and a change-aware attention module (CAAM)to achieve multi-scale feature extraction and semantic classification of change regions.[Findings]Experiments on the SECOND and Landsat-SCD datasets show that MFG-SCDNet outperforms mainstream methods across all evaluation metrics.On the SECOND dataset,it achieves an mIoU of 73.63%,a SeK of 23.04%,and an OA of 87.73%.On the Landsat-SCD dataset,it achieves an mIoU of 88.98%,a SeK of 59.34%,and an OA of 96.14%.Ablation experiments verify the effectiveness of each module.[Conclusions] MFG-SCDNet can effectively improve the accuracy and consistency of water semantic change detection,providing a reliable technical means for water resource remote sensing monitoring.

Key words: semantic change detection, water body change, multi-scale features, attention mechanism, remote sensing images, deep learning