Urban roads in remote sensing images will be disturbed by surrounding ground features such as building shadows and tree shadows, and the extraction results are prone to problems such as incomplete road structure, poor topological connectivity, and poor accuracy. For mountain roads, there will also be problems such as hill shadow or vegetation occlusion. We propose an improved Deeplabv3+ semantic segmentation network method. This method uses ResNeSt, which introduces channel attention, as the backbone network, and combines the ASPP module to obtain multi-scale information, thereby improving the accuracy of road extraction. Analysis of the experimental results on the Deeplglobe dataset shows that the intersection ratio and accuracy of the method in this paper are 63.15% and 73.16%, respectively, which are better than other methods.