Abstract: The image segmentation of water bodies faces a range of challenges, including limited multi-scale adaptability, loss of edge details, and interference from complex backgrounds. To address these issues, this study proposed a dynamically enhanced DeepLabV3+ (DE-DeepLabV3+) model with MobileNetV2 as the backbone network. Specifically, to overcome the limitation caused by the fixed dilation rate of the atrous spatial pyramid pooling (ASPP) module, a dynamic multi-granularity context (DMGC) module was designed. Using an adaptive feature fusion mechanism, the DMGC module enabled multi-scale dynamic fusion of global, regional, and local features, thereby enhancing the adaptability of the DE-DeepLabV3+ model to water body targets at various scales. A dynamic receptive field upsampling (DRFU) module was proposed for the decoding stage, avoiding edge detail loss inherent in conventional models by combining multi-branch atrous convolution and efficient sub-pixel convolution upsampling techniques. Furthermore, a multi-scale edge-aware weighted loss (MEWL) function was designed, aiming to comprehensively improve the model's identification accuracy and boundary segmentation effectiveness for water body targets while also effectively reducing the training errors caused by class imbalance. In this study, all experiments were conducted on a water body segmentation dataset for the Yellow River Basin within Ningxia Hui Autonomous Region, which was constructed using remote sensing images from Sentinel-2. The experimental results indicate that the DE-DeepLabV3+ model significantly outperformed the baseline DeepLabV3+ model. Overall, the optimized model increased the mean intersection over union (mIoU) by 92.78% and exhibited mean recall and accuracy of 96.31% and 97.13%, respectively. Notably, this model delivered significantly high performance in the image segmentation of water bodies, increasing mIoU and recall by 5.13 percentage points and 5.4 percentage points, respectively, compared to its baseline. The DE-DeepLabV3+ model performs well when applied to image segmentation of complex water-land transition zones and small tributaries, exhibiting a remarkable segmentation ability in complex water scenarios.