Research On Road Extraction From High-resolution Remote Sensing Images Based On Deep Learning

With the development of remote sensing satellites,more and more high-resolution satellite data are acquired.Road as the basic feature information is very important,and it is the key content of remote sensing image target extraction.High-resolution remote sensing images not only provide detailed road information,but also provide significant noise.Convolutional neural networks have achieved excellent results in computer vision by automatically extracting deep and shallow features.Therefore,this article is dedicated to the research of road extraction methods on the basis of convolutional neural networks,and uses self-defined algorithm models to achieve accurate road extraction from high-resolution remote sensing images.This paper first studies road extraction under complex terrain conditions.Taking the Sino-Nepalese Highway as the research object,we design a multi-feature road extraction network PPMU-Net that can simultaneously extract road spectral features and terrain features.Next,in order to be able to extract roads in different terrain environments,the reasons for the missed and misjudged phenomena in road extraction in different environments are analyzed,and the road extraction network GMR-Net,which fuses local and global information,is proposed.Finally,a high-resolution remote sensing image road extraction system was developed based on this.The specific research content is as follows:First,a multi-feature road extraction algorithm(PPMU-Net)under complex terrain conditions is proposed.The algorithm first supplements the terrain data of the area on the basis of the spectral data,and becomes a 6-channel multi-feature remote sensing image;Then improve the network on the basis of PSPNet,so that the network can extract spectral features and topographic features at the same time and achieve multi-scale fusion in the process of extraction.The final experiment shows that the precision of PPMU-Net is 84.4%,the recall is80.7%,the accuracy is 94.9%,and the F1-score is 82.5%,which are better than other comparison methods.Second,a convolutional neural network(GMR-Net)combining local and global information is proposed for road extraction in different terrain environments.A residual module with an attention mechanism is first designed to obtain global and other aggregate information for the features of each location.Then,a multi-path dilation convolution approach is used to extract features of roads at different scales,i.e.,to achieve multi-scale road feature extraction.Finally,gated and refinement units are proposed to filter out ambiguous features for gradual refinement of details.Multiple road-extraction methods are compared in this study using the DeepGlobe and Massachusetts datasets.The experimental results show that on the DeepGlobe datasets,the accuracy is 87.97%,the recall is 88.86%,and F1-score is 87.38%.On the Massachusetts datasets,the accuracy is 83.91%,the recall is 87.60%,and the F1-score is85.70%,showing high extraction accuracy and generalization ability.Thirdly,in order to realize the rapid extraction of roads from high-resolution remote sensing images,the system development of road extraction was implemented using Python.Using the GMR-Net road extraction algorithm as the back end of the system,combined with the system interface designed by PyQt5,the functions of four modules are designed for high-resolution remote sensing image input,data preprocessing,road extraction,and superimposition and display of extraction results.Finally,the existing remote sensing data was used to further test the system,which verified the simplicity,convenience and efficiency of the system.