In order to solve the problems of under-utilisation of edge detail features in road extraction, and the difficulty of achieving accurate segmentation of roads in complex background occlusion regions, the study proposed a remote sensing road extraction model based on edge-guidance and multi-scale perception U-Net (EMUNet. Based on U-Net, the Canny edge detection result of remote sensing image is added as input, and the feature guidance of each layer encoder is carried out by designing the edge-guided fusion module combined with the attention, so as to make full use of the edge information to improve the quality of the final road extraction; secondly, in view of the background occlusion problem existing in the image, the multi-scale parallel hollow convolution module is constructed to enhance the multi-scale perception ability of the network, so as to capture more road information and to improve the quality of the road extraction. Secondly, to address the problem of background occlusion in the image, the network is enhanced by constructing a multi-scale parallel dilated convolution module to capture more contextual information and accurately extract the regions that are obscured by the background. The experimental verification is carried out on the Massachusetts road dataset, and compared with U-Net, EMUNet can achieve more accurate segmentation of small roads and occluded regions, and the recall rate, F1 score and intersection ratio are better than other comparative algorithms, so it can achieve more complete and accurate road information extraction.