The acquisition of field road boundary information is the basis for making high-precision farmland map. In order to solve the problems of inaccurate segmentation of field roads in high-resolution orthophoto maps, such as missed segmentation and false segmentation, a deep learning network model was proposed based on improved U Net. Firstly, the backbone network was replaced with ResNet50 to enhance the ability to extract the features of drivable roads in the field. Secondly, the DSConv module, which can improve the accuracy of tubular structure, improved the accuracy of the field drivable road, and inhibited the feature extraction of the background of field features similar to the field road. Finally, the complete context information was obtained by inserting the ECA Net attention mechanism, and the feature restoration process of the drivable road in the field was optimized, so as to improve the overall segmentation accuracy of the model. Then the traditional image processing method was used to further denoise and eliminate the hole of the segmentation results, and in view of the problem of losing geographic information in the recognition results, so as to obtain high-precision field road boundary information. The experimental results showed that the improved U Net model had the highest evaluation index values in the comparison with the semantic segmentation model in the test set of the constructed dataset with 95.46% MPA and 91.12% MIoU, after post-processing using traditional image processing methods, MIoU and MPA were 92.64% and 96.75% , respectively, the MIoU and MPA were increased by 1.29 percentage points and 1.52 percentage points;and 86.39% and 90.01% in the field drivable road recognition test of high-resolution orthophoto map, respectively, which can clearly identify the field road. After using the traditional image processing method to optimize the obtained high-resolution orthophoto results, the MIoU and MPA were 88.34% and 91.53% , respectively, and the MIoU and MPA were increased by 1.95 percentage points and 1.52 percentage points, respectively. The research result can provide accurate field road boundary information for the subsequent production of high- precision farmland map.