<p>Reliable segmentation of surface water is crucial for monitoring hydrological dynamics, mitigating disaster impacts, and supporting climate-resilient infrastructure development. However, traditional approaches based on spectral indices—such as the Normalized Difference Water Index (NDWI)—often suffer from poor generalization in the presence of spectral ambiguity, seasonal variations, and urban or vegetated interference. Existing deep learning architectures (DeepLabv3+, SegNet) show a trade-off between accuracy and speed, without specific adaptation to the challenges of the domain. This observation justifies the design of a hybrid architecture combining these models to achieve high semantic accuracy, environmental robustness, and computational efficiency. We propose a hybrid deep learning (DL) architecture that integrates two encoder-decoder segmentation frameworks: DeepLabv3+ with residual networks (ResNet18 and ResNet50) and SegNet with Visual Geometry Group (VGG16 and VGG19) backbones. The proposed model seamlessly combines two well-established encoding–decoding architectures, DeepLabv3+ (with ResNet) and SegNet (with VGG), by fusing their segmentation outputs at a decision level to combine the high semantic accuracy of SegNet with the computational efficiency of DeepLabv3+. This hybrid integration thus simultaneously leverages the complementary strengths of both networks to improve the overall quality of water surface segmentation while optimizing training time. This formulation briefly explains that the integration is not limited to a simple choice between models but also involves decision-level fusion or synergistic combination of results, which justifies the term hybrid or dual-stream architecture. These models are optimized for both semantic accuracy and computational efficiency, and trained using multispectral Sentinel-2 satellite imagery sourced from a publicly available Kaggle dataset featuring diverse terrain and spectral complexity. All models were implemented in MATLAB and evaluated using multiple performance metrics: overall accuracy (OA), Intersection over Union (IoU), F1 score, precision, recall, producer accuracy (PA), and user accuracy (UA). Among all configurations, SegNet with VGG16 achieved the best results, delivering An OA of 91%, IoU of 79%, and F1 score of 75%, while effectively minimizing false positives. In contrast, DeepLabv3+ with ResNet18 demonstrated exceptional computational efficiency, completing training in only 95&#xa0;min with a respectable OA of 78%. These results underscore a clear trade-off between segmentation quality and training speed. Compared to baseline DL architectures and conventional NDWI-based methods, the proposed hybrid framework achieved significant gains in segmentation performance: precision improved by 2.64–22.66%, recall by 1.03–29.87%, and F1 score by 1.33–37.33%, with a 25.31–32.91% increase in IoU and up to 80% reduction in training time. Furthermore, the models proved robust under real-world environmental complexities such as cloud occlusion, seasonal transitions, and mixed urban-land–water boundaries. This research presents a scalable and high-accuracy solution for water body extraction, supporting a range of applications including flood risk assessment, environmental monitoring, and water resource planning. Future work will explore the integration of artificial neural networks (ANNs) for enhanced classification, multimodal data fusion with synthetic aperture radar (SAR) for all-weather operation, and real-time deployment using cloud-based platforms for global water surveillance.</p>