Large and sprawling lakes are of significant importance in the aspects of water supply, ecosystem preservation, flood control, and hydroelectric power. Hence, better lake management is urgently needed. Synthetic Aperture Radar (SAR) remote sensing data are more available nowadays and an imperative source. However, the traditional methods applied to these data for lake information extraction reveal limitations in terms of accuracy and efficiency. In this study, we configure distinctive backbones of ResNet50, ResNet152, EfficientNet-b7, ResNeXt101_32 × 16d, Timm-RegNetX_320, and Timm-RegNetY_320 with the DeepLabV3+ model and fine-tune them for the task of semantic segmentation of large lakes of five continents and downscaling for the national scale of South Korea. We selected the 30 largest lakes of each continent for model training and five lakes for model testing for collecting Sentinel-1 images and generated lake masks to create training and validation datasets. The reults of the 100-epoch model trains for each model show that the mean accuracy was around 0.95, and the use of EfficientNet-b7 was the best model accuracy and efficiency; nonetheless, the lightest networks of ResNet50 and the ResNeXt101_32 × 16d were at the lowest performance. The utilization of European and Australian data- sets resulted in the highest and lowest precision, respectively. The lake segmentations were nicely mapped, preserving the geocoordinates of the input Sentinel-1 images, which can then be used for lateral spatial analyses. The deep learning models presented an efficient tool for large and sprawling segmentation and are highly recommended.