Abstract: Monitoring river areas and shorelines is crucial for hydrological stations, as it provides essential data support for water resource management, flood prevention warnings, and ecological conservation.Traditional monitoring methods suffer from inefficiency and insufficient accuracy.Therefore, a lightweight DeepLabV3+ network based on the MobileNetV2 backbone is proposed.By reducing the number of model parameters, this network significantly improves computational speed.Additionally, the introduction of Dice Loss as the loss function optimizes the model′s performance in segmenting river shoreline edge details and small-scale regions.Furthermore, this model incorporates a Spatial Attention Module (SAM) and an Atrous Spatial Pyramid Pooling with Squeeze-and-Excitation (ASPP-SE), further enhancing its ability to capture river features in complex environments.Experimental results show that the improved model achieves increases of 10.43%, 3.41%, and 0.02% in the Kappa coefficient, Dice coefficient, and F1-Score, respectively, outperforming the original DeepLabV3+ network as well as U-Net, PSPNet, and HRNet.This indicates that the improved network offers higher accuracy and efficiency in extracting river areas and shorelines under complex conditions.The research findings provide an efficient and accurate technical solution to support river monitoring tasks at hydrological stations.