Research On Road Network Extraction Based On Deep Learning

With the rapid development of remote sensing technology,the remote sensing image data received on the ground is growing at an exponential rate,and the acquired image data contains a large amount of available geographic information.Road is an important basic geographic information,and the extraction of the road is of great significance for GIS data update,map drawing,path analysis and other applications.After decades of research and development,many algorithms related to road extraction have been proposed,and have yielded many results.However,the connectivity of the road network extracted by the current algorithm is generally not high,and the degree of automation is low,all of this cannot meet the needs of quickly and accurately obtaining road information in massive images.Therefore,the road network extraction technology urgently needs to develop in the direction of rapidity,intelligence and precision.With the powerful learning ability,deep learning technology has been widely used in many fields,and is expected to make breakthrough progress in road network extraction tasks.Based on the deep learning methods,this paper conducts an indepth study on the road network extraction algorithm of high-resolution remote sensing image.The main works of the paper are as follows:1.A road intersection detection algorithm based on deep learning is proposed for the problems of high miss rate and low detection of road intersections,which are topological connection nodes in the road network construction.First,the road intersection dataset of remote sensing images was produced to meet the needs of deep network training;then the YOLOv3 network structure was further optimized,and the extracting feature ability of the network was enhanced by fusing multiple resolution feature maps and expanding the width of the convolution module.Experimental results show that the proposed algorithm has higher accuracy and recall rate in the test set compared with YOLOv3 algorithm.The accurate location and type discrimination of road intersections can provide reliable topology nodes for subsequent road network construction tasks.2.Aiming at the problem of low accuracy of road extraction in the process of road network construction,a road extraction algorithm based on deep learning is proposed.Using the classic U-Net as the basic network,the DSU-Net network is constructed by adding feature information transmission channels and adopting a deep supervision mechanism.At the same time,the influence of migration training and loss function on the generalization performance of the model is studied and analyzed,and the obtained influence law is applied to the road extraction task.Experiments show that the proposed algorithm can accurately and quickly extract road data from remote sensing images and provide reliable initial data for road network construction tasks.3.Aiming at the problem of low connectivity of the road network,a road network construction method is designed based on road intersection data and road extraction data.First,the road extraction data and road intersection data are converted into an undirected graph with the form of point and line,and then the roads are connected to the road intersection nodes with topological connection relationships.The broken roads are connected with different algorithms according to the length of the broken area,and finally all of the adjacent road network images are stitched into a large-scale road network image.Experiments show that the proposed method can construct a highly connected road network,which is basically consistent with the real road network image.