In order to improve the recognition accuracy of irrigation water bodies in irrigated farmland, the Jiefangzha Irrigation Area of Hetao Irrigation District was taken as the study area, and the surface water body extraction model (WatNet) was improved based on Sentinel-2 remote sensing images, combined with the actual situation of the irrigation area, to obtain the MWatNet model and extract irrigation water bodies. Overall accuracy (OA), mean intersection over union (MIoU), F1 value and other water body extraction accuracy indicators were used for comprehensive evaluation. The results showed that the improved surface water body extraction model (MWatNet) had good recognition accuracy in the extraction of farmland irrigation water bodies in Jiefangzha Irrigation Area, the overall accuracy of the model reached 96%, the mean interaction over union reached 83%, the F1 value was 80%, and the accuracy of the field research validation was 85.7%;comparing with the original WatNet, the semantic segmentation model of water bodies (Deeplabv3_plus), and the water body extraction model (Deepwatermapv2), MWatNet showed better results and model operation efficiency in terms of connectivity of irrigation water body extraction, and elimination of road and town interference. The quantitative characterization of irrigation water bodies can be achieved by using this model, which provided data support for irrigation water scheduling.