Research On Road Information Extraction Of Remote Sensing Image Based On Deep Learning

In view of the complexity of information in remote sensing(RS)images and the diversity of ground object types,it is an urgent problem to study how to improve the accuracy of complex information extraction in RS images by utilizing the powerful feature expression ability of deep learning,the characteristics of big data and high spatial resolution of RS images.The research of extracting road information on RS image has been carried out for many years.Howerer,different grade roads have the different width and shape characteristics,like the national highway,the provincial highway,the village road and the mountain road and so on.Different road materials have different color and texture characteristics,such as cement,asphalt,clay.Besides,the road areas are often covered by buildings,trees,road central green belts.Because of these factors,road extraction is still the research frontier and technical challenges in the field of RS information extraction.In this paper,the models for road information automatic extraction of RS images are constructed based on deep learning.Also,the post-processing method for experimental results is proposed,and the prototype system for road information automatic extraction of RS images is developed.Specific research contents and results are as follows:(1)Aiming at the problem of precision of road extraction results of D-LinkNet and the defect that the network scale is too large to be applied,this paper proposed new networks,DenseNetPlus,D-DenseNetPlus,DLinkNetPlus,based on new encoding structures which were built on DenseNet structure,modified DenseNet structure,modified ResNet structure.Aiming at the dilated convolution structure in D-LinkNet,this paper introduced bottleneck layer to build the new network B-DLinkNetPlus in order to reduce the parameters.These models were built on D-LinkNet structure.Besides,an improved GAN-Unet model that aggregating multi-scale context is proposed to solve the problem of inconsistent road widths at different road levels.(2)In order to eliminate small independent patches(ESIPs)in the road extraction results,a post-processing method is proposed.Firstly,the connected component and its outer rectangle are detected.Whether the connected component needs to be eliminated by judging three conditions,the number of pixels in the connected component,the ratio of the length and width of the outer rectangle,the ratio of the number of pixels in the connected component and the area of the outer rectangle.This method can effectively eliminate the independent scattered points and patches in the road extraction results,so that the extracted road areas are more integral and have better visual effect.(3)This paper developed a prototype system for road extraction of RS images based on deep learning models and the post-processing method.The system has the image processing basic modules,such as opening images,zoom,browsing image and saving images.It also integrated the road extraction models of RS images proposed in this paper and the road extraction result post-processing method.Thus,the users can get the road extraction results fastly and conveniently by using this prototype system.