Research On Urban Road Extraction Algorithm Based On Deep Learning For High Resolution Remote Sensing Images

The feature information contained in high-resolution remote sensing images is rich,and roads,as an important part of the basic features of geographic information,are widely distributed.Extracting roads from remote sensing images has been a hot research topic in modern society.It is of great significance to the development of cities.Currently,road network extraction from high-resolution remote sensing images has been widely used in urban services such as urban planning,vehicle navigation and geoinformation management.Although it has received considerable attention in the past decade,the extraction of roads from high-resolution remote sensing images is still a challenging task due to the interference of other objects(e.g.,parking lots,building roofs and tree canopies)and the complexity of road networks.Constructing a reasonable network to obtain superior feature representation and focusing on the local distinguishing regions of the image are the key points to improve the accuracy of road extraction results.Therefore,based on the detailed road extraction algorithm,this paper addresses the problem that the segmentation algorithm is not comprehensive enough for the feature representation of high-resolution remote sensing images,and carries out research with the goal of optimizing the feature extraction capability of the network,and the main work accomplished in this paper is as follows:1.The SGE(Spatial Group-wised Enhance)attention mechanism is introduced to improve the U-Net image segmentation algorithm.To address the shortcoming of the U-Net algorithm in capturing insufficient high-dimensional spatial information in the feature extraction stage,the SGE attention mechanism is used to enhance the extraction of spatial features to improve the advantage of feature representation comprehensiveness,and the attention-enhanced part is added to the encoder part and the decoder part before performing feature extraction.The attention-enhancing module can emphasize regions with distinguishability in the display.Finally,the method is compared with several existing commonly used methods,and the results show that the method is feasible.In addition,to improve the segmentation accuracy of the images,data enhancement is used to expand the size of the road dataset,and the experimental results show that the improved algorithm achieves 74.82% and 61.31% of the road extraction accuracy and Io U(intersection-to-merge ratio)on the Massachusetts Road dataset,respectively.2.Based on three different network topologies which are serial,parallel and migrated,we propose a deep learning-based algorithm for automatic extraction of road centerlines.The merging of the road centerline extraction network and the road itself extraction network saves a lot of computational resources for the whole process.By experimenting on the selected Deep Globe road dataset,the correctness,completeness and quality of the algorithm reach 90.2%,95.7% and 91.6%,respectively.