In the context of global warming, extreme droughts are occurring more frequently, and the monitoring of surface water bodies is crucial for the dynamic assessment of water resources. In recent years, the rapid advancement in remote sensing technology has facilitated the acquisition and application of extensive satellite remote sensing data for monitoring surface water bodies. The water index (WI) method, for instance, has several limitations, such as its limitation to a single scenario, difficulty in confirming threshold values and in selecting features using machine learning (ML) methods. While deep learning (DL) methods can yield favorable outcomes, their efficacy is contingent upon the advancement of the network architecture and the availability of well-annotated datasets. Currently, there is a scarcity of publicly accessible water datasets, and the limited data sources pose a challenge for the utilization of large-scale, multi-source satellite data. Accordingly, this study validates the effectiveness of using the ESA WorldCover product as a benchmark for water body masks in the training set, through a series of comparative analyses. Specifically, the ESA WorldCover was used to generate a training set of water body masks. The SWC-1 (Sentinel-1 annotated by ESA WorldCover) and SWC-2 (Sentinel-2 annotated by ESA WorldCover) datasets were constructed using Sentinel-1 and Sentinel-2 data as examples. Several contemporary semantic segmentation networks, such as Deeplabv3+, Swin Transformer, and Segformer, served as benchmark models for water body extraction across various scenarios in Liaoning Province. After conducting a series of experiments, the results show that the water body mask derived from the ESA WorldCover product meets the specified criteria. The F1 scores for all networks on the SWC-1 and SWC-2 datasets exceed 96% and 97%, respectively, and the Intersection over Union (IoU) scores exceed 93% and 95%. The water body extraction results from the model trained with this method are significantly better than those obtained using traditional methods. Moreover, this method has proven capable of delivering optimal results in water body extraction across various scenarios. Utilizing this method, one can rapidly construct a large-scale water body dataset for deep learning training that does not require manual annotation, thereby offering a valuable reference for water body extraction across extensive areas and extended time periods.