Abstract: With the continuous development of high- resolution earth observation system in China, more and more applications for high-resolution image intelligent analysis and processing are required. The semantic segmentation based on deep learning has attracted people's attention too. As a research hotspot of close-range image semantic segmentation, the Deeplab network has achieved good results. In order to solve the problem of multi-scale and high-resolution remote sensing image semantic segmentation, this paper firstly expands the receptive field of Atrous spatial pyramid pooling (ASPP) by using dilated convolution, then improves deeplabv3 model and applies it to the classification processing of high-resolution NO.2 (GF-2) remote sensing images. We take the GF-2 remote sensing image of Chenzhou area as the research objects to verify the method. First, we preprocess the original image, then enhance and expand the data of the preprocessed image. Finally, we compare the classification results under different parameters, and analyze the adaptability and accuracy of the model. In our data set, the experimental classification accuracy of this method is 88.2%, and the MIoU is 72.5%, which is better than that of Deeplab.