Accurate delineation of surface water bodies from satellite imagery is fundamental for hydrological monitoring, disaster management, and environmental assessment. Deep learning–based semantic segmentation has shown great promise for this task, yet the relative performance of widely used encoder–decoder architectures remain insufficiently explored under consistent experimental settings. This study presents a GeoAI-based evaluation of four state-of-the-art segmentation architectures: U-Net, U-Net++, DeepLabV3+, and Feature Pyramid Network (FPN), all employing a ResNet-34 encoder pretrained on ImageNet. Using the Earth Surface Water Dataset, we applied an identical tiling strategy, preprocessing pipeline, data augmentation, optimizer configuration, and training schedule to ensure a fair comparison. Our results show that DeepLabV3+ achieved the highest accuracy with an IoU of 0.955 and Dice of 0.973, followed by U-Net++ (IoU 0.945, Dice 0.969) and U-Net (IoU 0.9352, Dice 0.9598). FPN performed comparatively lower with an IoU of 0.9218 and Dice of 0.9514. These findings demonstrate that architectural choices substantially influence segmentation quality even with identical encoders and training protocols. The study provides benchmark results and practical insights for selecting deep learning architectures for operational water body mapping in remote sensing applications.