High-standard farmland construction is an important guarantee for national food security, and the quality assessment of high-standard farmland construction is beneficial to the implementation of farmland planning and government decision-making. As an important project of high-standard farmland construction, the rapid and accurate acquisition of field roads can provide basic data support for the quality assessment and effect evaluation of high-standard farmland construction. Thus, it is necessary to obtain accurate and effective field roads information. However, compared with high-grade roads, the narrow pavement width and easy occlusion by vegetation are the typical characteristics of field roads, which are the main factors leading to the low degree of automation in existing methods. Aiming at the problems of low accuracy and weak generalization ability of traditional recognition methods for narrow field roads, a highstandard farmland road recognition method was proposed based on U-Net network. Firstly, on the basis of analyzing the basic characteristics of the field roads, the GF-2 images were selected as the experimental data, and the object-oriented method was used to segment the image and classify it according to the characteristics of the object, so as to eliminate non-roads such as buildings with similar spectra elements to reduce interference;then, operations such as cropping, labeling, and data enhancement were performed on the image, the U-Net network was used to mine the deep and shallow features of the image, and the network was continuously trained by adjusting parameters to achieve accurate identification of field roads;finally, according to the characteristics of road breakpoints, the local connection method was used to repair the road breakpoints, and the accuracy verification were carried out in Dongting Town, Dingzhou City, Hebei Province as the experimental area. The results showed that by mining the image features of 622 field road samples, the U-Net network could effectively identify high-standard farmland roads in various scenarios. After repairing the road breakpoints, the field road identification precision in the study area reached 96%, and the recall and F1 score were 62% and 75%, respectively. The recognition accuracy could meet the requirements for rapid evaluation of high-standard farmland construction quality. Compared with traditional identification methods, the combination of object-oriented and deep learning methods could quickly identify field roads on the basis of reducing building interference, and could better solve the noise and misidentification issues caused by large differences in field road materials and vegetation occlusion. This method could provide a method reference for the identification of narrow objects in farmland.