Road Extraction From High-Resolution Remote Sensing Images Based On Attention Residual Enhanced Networ

Studying the geographical distribution of roads and extracting road information quickly and accurately is a key part of vehicle navigation,urban emergency response,geospatial planning,and intelligent transportation construction.Due to the complexity of the road itself and the background environment,traditional road extraction methods based on road edge texture,geometric topology or threshold segmentation have a high degree of human participation,time-consuming and laborious,and low efficiency.With the rapid development of remote sensing and computer technology,remote sensing satellite images have gradually developed from medium and low resolution to high resolution remote sensing images,which contain rich detailed information of ground objects and complex road backgrounds.In the images,vegetation cover and building shadows block the spatial information of roads,resulting in road extraction fragmentation,poor continuity,blurred edges and loss of detailed information.In recent years,many deep learning algorithms represented by convolutional neural network have shown extraordinary potential in the era of big data of massive remote sensing images,providing a new idea for road extraction from high-resolution remote sensing images.Road information is extracted from high-resolution remote sensing images based on deep convolutional neural networks.Road features are abstractly extracted from images through convolutional neural network models to realize semantic feature expression of road features on images,and automatic road information extraction is realized.Due to the complexity and long continuity of road information,it is necessary to fully consider the multi-level and multi-scale information characteristics of roads,suppress complex background,strengthen the utilization of significant road features,and improve the accuracy of road extraction.In this paper,based on high resolution remote sensing image,build attention residual road extraction model,enhance network by residual coding improved and embedded channel-space dual attention mechanism,to improve road extraction topological connectivity and integrity.In terms of road extraction,compared with other algorithm models,the F1 and m Io U indexes of the proposed algorithm reach the maximum value year-on-year.DARNet model proposed in this paper,mainly aiming at the shadow and keep out area inaccurate road information extraction,the extraction results of poor continuity,missing details the problems such as roads,using dual attention characteristic expression and the depth of the residual neural network to improve the deep high order road using semantic information,decisive characteristics of road extraction accuracy the path of ascension.Experiments on open source data sets and self-constructed road data sets,ablation experiments and qualitative and quantitative comparative analysis have proved the effectiveness of the modules added in this paper,and effectively verified that the algorithm proposed in this paper can effectively improve the road quality and accuracy.In addition,in view of the lack of high-precision data sets in the current research on road extraction from high-resolution remote sensing images,this paper uses high-resolution remote sensing images for pixel-level annotation and constructs a set of highresolution remote sensing image data sets,which can be used for subsequent research on roads,especially urban roads.The innovative method proposed plays a certain role in promoting the research of highprecision and automatic road extraction from high-resolution remote sensing images,and has certain theoretical research significance and practical application prospects.