Research On High Resolution Remote Sensing Image Road Extraction Based On Deep Learning

Road information extraction is an important work in the field of remote sensing image processing and interpretation,and affected by image resolution.Traditional road extraction technology is difficult to achieve automatic road extraction.With the advent development of aerial satellite in recent years,high-resolution remote sensing images can be obtained more quickly and easily.As the development of abundant,detailed and important feature information in the remote sensing images,road information can be extracted and used efficiently and accurately,having a great impact on traffic,life and other aspects.Recently,as the presence of deep learning,the road information in remote sensing images are well used and greatly improved the effect on automatic road extraction,solving the problems of rough road edge,weak anti-interference and low extraction accuracy in previous methods.Overall,the method of deep learning for extracting road information in high-resolution remote sensing images is proposed and the main work are shown as follows:1.Three high-resolution remote sensing image road extraction models based on deep learning network are compared.More specifically,according to Massachusetts Road data set(Massachusetts Roads Dataset),this paper builds a road extraction model based on full convolution neural network,U-Net network and Deeplab v3 network.After training,verifying,and testing the model,the road extraction results based on three different deep learning networks are derived.These models based on deep learning network can well achieve automatic road extraction,proving the feasibility of deep learning technology in automatic road extraction from high-resolution remote sensing image.Additionally,each of these three methods has its own advantages and disadvantages in road extraction at various background.Generally speaking,Deeplab v3 can not only extract the multi-scale features of the road,but be suitable for the automatic extraction of the road in multi-scene.The accuracy in this paper is about 90 %.2.Compared with the previous road extraction model,the road extraction model from high-resolution remote sensing image is improved based on Deeplab v3 network.Road extraction model based on Deeplab v3 network can extract multi-scale road features,and introduce attention mechanism into Deeplab v3 network to extract high-dimensional road features.Through the training,verifying,and testing of the improved Deeplab v3 road extraction model,it is found that the proposed model combined with attention mechanism refines the edge of the road,which makes the road information more comprehensive and complete.By comparing with the previous Deeplab v3 road extraction model,the accuracy is improved by about 2.4 %.