Abstract: The Danjiangkou Reservoir serves as the water source for the South-to-North Water Diversion Middle Route Project.Accurately and efficiently extracting farmland information around the reservoir is of great significance for ensuring water supply security in the source area.Traditional farmland survey methods primarily rely on manual visual interpretation of remote sensing images, which is highly subjective, time-consuming, and labor-intensive, failing to meet the demand for rapid and large-scale farmland information extraction.To address this, this study proposes a remote sensing farmland semantic segmentation method called MIT-Unet, which extracts deep semantic features of images from both local and global perspectives and achieves the fusion of multi-scale global dependency information through skip connections.A farmland semantic segmentation dataset, DRFL, is constructed to evaluate the method′s effectiveness.Comparing the proposed method with existing classical methods such as Unet, DeepLabV3+, Manet, and Deswin, the results show that MIT-Unet achieves the best extraction performance on DRFL, with IoU and F1-score metrics reaching 82.90% and 90.65%, respectively.The research findings can provide technical support for farmland area statistics and disaster loss assessment within the Danjiangkou Reservoir construction area, and serve as a reference for farmland distribution monitoring, agricultural resource management, and ecological environment protection in other regions.