Multiscale Feature Enhancement for Water Body Segmentation in High-Resolution Remote Sensing Images

This study addresses the challenges of limited automation in water body extraction, inadequate delineation of edge details, and insufficient feature extraction in multiregional high-resolution remote sensing images. To address these issues, we propose a novel multiscale feature enhancement water body segmentation (MSFE-WaterSeg) framework, tailored for high-resolution remote sensing images. The method uses ConvNeXt as the backbone network, integrates UPerNet for multiscale feature fusion, and introduces the LovaszLoss loss function to enhance segmentation precision in complex water body edge areas. The model is trained on a custom-built dataset of 80 000 high-resolution remote sensing images from multiple countries, and evaluated on 8500 images at a cartographic scale of 1:50 000. The results demonstrate that the MSFE-WaterSeg method achieves an accuracy of 99.43% and an intersection over union (IoU) of 90.20%. Compared to benchmark models UPerNet-Swin-T and UPerNet-ResNet-34, MSFE-WaterSeg yields IoU improvements of 1.29% and 6.8% and F1-Score gains of 0.71% and 4.34%, respectively. The method exhibits superior performance in intricate water body boundaries, narrow fluvial structures, and water–land transition zones. These findings offer efficient and reliable technical support for multiregional water body interpretation and global water resource monitoring, while also providing a feasible optimization strategy for large-scale remote sensing image semantic segmentation tasks.