[Objective] Although deep learning-based semantic segmentation models, such as U-Net and DeepLabv3+, have improved water body extraction performance, they remain limited in delineating continuous boundaries and recognizing narrow water bodies. To address these limitations, this study proposes an improved model named DSP-UNet (U-Net enhanced with Dual Attention and Strip Pooling). The model is designed to improve both local boundary sensitivity and global contextual reasoning, achieving higher precision and robustness while maintaining computational efficiency. The objective is to develop a high-performance deep learning model capable of stable water body extraction under varying geographic and imaging conditions. [Methods] The proposed DSP-UNet model was constructed based on the classical U-Net architecture and integrated three key innovations: strip pooling module (SPM), dual attention mechanism (DAM), and SimAM attention module. The model was trained using the OpenWUSU512 dataset. Comparative experiments with baseline models, including U-Net, PSPNet, HRNet, and DeepLabv3+, were performed under the same training conditions. Ablation studies were conducted to quantify the contribution of each proposed module. Moreover, cross-domain generalization was evaluated on the Satellite Images of Water Bodies dataset to assess model transferability. [Results] Experimental results demonstrated that DSP-UNet achieved higher segmentation accuracy and robustness across all evaluation indicators. On the OpenWUSU512 dataset, DSP-UNet reached a mean intersection over union (MIoU) of 0.962 and an F1 score of 0.966, outperforming PSPNet by 2.4% in MIoU. The precision and recall values were 0.976 and 0.956, respectively, indicating a balanced trade-off between false positives and false negatives. The overall accuracy (OA) was 0.991. Compared with the baseline U-Net, the dual attention mechanism increased MIoU by 1.1%; the integration of SimAM improved edge recognition and reduced boundary noise; and the addition of strip pooling expanded spatial perception, increased MIoU by 2.3%, and reduced total training time by 11.5%. These results confirmed that the progressive introduction of modules effectively enhanced both feature representation capability and computational efficiency. Cross-domain experiments on the Satellite Images of Water Bodies dataset demonstrated strong generalization performance. DSP-UNet achieved an F1 score of 0.909, an IoU of 0.833, and an OA of 0.954, outperforming U-Net and SegNet by more than 5% in IoU. Visual assessment showed that the model accurately distinguished water bodies from shadows and reflections, maintained continuous boundaries, and effectively extracted narrow channels and small ponds, which were often misclassified by conventional models. [Conclusion] The proposed DSP-UNet provides an accurate and robust solution for water body extraction from high-resolution remote sensing imagery. By combining dual attention and strip pooling, the model effectively enhances both global contextual understanding and local boundary refinement. The parameter-free SimAM module further improves spatial discrimination without introducing additional computational burden. DSP-UNet achieves the highest accuracy among the compared models while maintaining stable convergence and reduced training time, demonstrating its efficiency for large-scale applications. The model’s strong generalization ability across datasets indicates its potential application in remote sensing tasks, including wetland mapping, flood detection, and shoreline monitoring. Future research should focus on integrating multi-temporal and multi-sensor data, developing temporal attention mechanisms for dynamic water monitoring, and optimizing the architecture for lightweight and real-time deployment in practical remote sensing applications.