Research On Rural Highway Information Extraction Method Based On High Resolution Image

In the development process of national transportation infrastructure construction,highway construction has always played an important role.It has obvious influence on social and economic development.With the policy of revitalising the countryside,rural road construction projects have received strong support from local people’s governments at every level.The rapid advances in remote sensing have considerably contributed to the research of extracting rural road area from High quality remote sensing images with widespread utility and prospects for expansion in the area of rural road management and intelligent transportation.As rural roads are slender targets,it is difficult to automatically and efficiently extract continuous and clear rural roads from images due to the complexity of the environment and distracting factors such as features.In this paper,we use high-resolution remote sensing images as a data source and study the method of extracting rural road information according to deep learning technology.The main work and research content of this article include the following:As rural roads are slender targets,it is difficult to automatically and efficiently extract continuous and clear rural roads from the high resolution remote sensing images because of the complex and distracting factors such as environment and features.In this article,a method of information extraction of countryside roads were conducted using high-resolution remote sensing images as the datasource on the basis of deep learning techniques.The main work and study contents of this paper are as described below.(1)A detailed introduction to the current research status of path acquisition approaches at domestic and international level.The three frequently used models for segmentation of semantics are,U-Net,Deeplabv3+ and CE-Net,are studied and analyzed,and their respective the results were obtained to verify the feasibility of deep learning-based rural road extraction.(2)The public Massachusetts Road set and Deep Globe data set were selected as data sets,and the GF-2 data set was self-built,and the process of label making and data enhancement was introduced.(3)Aiming at the problems of narrow and long rural highway area,incomplete road structure and fuzzy boundary in high-resolution satellite images,A rural road extraction model based on U-Net network integration of Atrous Spatial Pyramid Pooling(ASPP)and Efficient Channel Attention(ECA)was proposed.It improves the access to road details and makes the information extraction of road edge more complete and comprehensive.The improved U-Net network outperformed the original U-Net and Deeplabv3+ models in Recall,Dice and Io U,achieving 84.95%,75.38% and 57.22% accuracy respectively.