Abstract: The rapid and accurate extraction of rural roads from domestic high-resolution images has important value in fields such as information management and modernization of agriculture and rural areas. However, traditional remote sensing interpretation methods have low efficiency and accuracy in extracting rural road information, due to the rural roads’ background being noisy, the roads being narrow, long and winding, and susceptible to shadow occlusion. In this research is improved the DeepLabv3+ model based on the characteristics of rural roads and is designed a high-resolution image rural road extraction model that is efficient and accurate. Firstly, Mobilenetv2 is used as the backbone of the model to reduce its parameters; secondly, Convolutional Block Attention Module (CBAM) is cascaded in the Atrous Spatial Pyramid Pooling (ASPP) module to enhance the model’s feature perception ability; and finally, Dice Loss function is added to improve the loss function and overcome the sample imbalance. The results show that the improvement of details has significantly improved various indicators, achieving the highest efficiency and accuracy. Compared with the classic model, the improved model has achieved higher scores in Mean Pixel Accuracy (MPA) and Mean Intersection over Union (MIoU). Although deep learning of deep features requires more time, the improved model can achieve higher efficiency and accuracy than other models.