There exist thousands of water bodies in watersheds, including large-scale water bodies, such as reservoirs, and small-scale water bodies, such as lakes, ponds, etc. In basin flood forecasting and other hydrology-related tasks, water bodies play an important role in the flooding process. The method of efficiently segmenting water bodies from remote sensing images (RSIs) is still a popular research topic in the fields of computer science and remote sensing. We propose a model based on mask R-CNN to automatically detect and segment water bodies in RSIs, thereby avoiding the complex operations of manual feature extraction when processing aerial images or satellite images because these images often have low resolution and complex background. RSIs were obtained from various remote-sensing research datasets and from snapshots from Google Earth. Data augmentation was introduced to enrich the training images dataset. Then, the proposed model was trained on the augmented dataset in two implementations: residual network (ResNet)-50 and ResNet-101. Experimental results show that the proposed method scores 90% on average for regular-shaped water bodies and 76% on average for irregular-shaped water bodies in terms of intersection over union, which indicates that the proposed models offer excellent feasibility and robustness for water-body segmentation.