Research On Traffic Road Network Extraction Based On Remote Sensing Image Data And Deep Learning

As my country's economy continues to grow,further development has been achieved in urban and rural infrastructure,and the road traffic network has been updated more frequently and rapidly.High-resolution remote sensing image data is an important scientific and technological means serving the information production of modern society.How to use this technology to achieve real-time and automated traffic road network extraction has become an important research content in the field of remote sensing application research.Geographic information system automatic update,urban road planning,vehicle satellite navigation,etc.have huge economic value and farreaching research significance.In recent years,as a research topic that has attracted much attention in the field of artificial intelligence computer vision,deep learning has shown a very superior image segmentation effect and potential compared with traditional machine learning methods.With the continuous improvement of the quality of remote sensing data,the information contained in the image is becoming more and more abundant.The ability of deep learning to mine the in-depth information of the data makes it gradually occupy the mainstream in the research of road extraction in remote sensing images.This paper conducts a systematic and in-depth analysis and research on the automatic extraction of roads in remote sensing images,and summarizes the previous research,combined with the latest research results in the field of deep learning,and finally proposes an improved remote sensing image road extraction algorithm.The main work is as follows:(1)Aiming at the problem of the inconsistent proportions of the positive and negative samples in the remote sensing image road data set,the loss function commonly used in semantic segmentation is improved,and the Binary Cross Entropy(BCE)and the Dice coefficient are combined to form a new loss function that is more suitable for the road extraction task,and the iterative change relationship of the loss value of the two loss functions is visually displayed through the graph.(2)With the help of deep learning theory to improve the codec network,a highresolution remote sensing image road extraction method based on multi-scale feature convolutional neural network is proposed.In the network coding part,the residual network is used to replace the VGG coding network in the original U-Net,and the structural characteristics of the residual network module(Res-Net Block)are used to prevent the gradient explosion phenomenon that may occur when the network layer continues to deepen,And introduce the Dilated Convolutional Module in the center of the network to enhance the multi-feature extraction capabilities of the original network.Based on the Pytorch deep learning framework,this paper uses the Massachusetts Roads Dataset as the training sample to train the original U-Net model and the multi-scale feature-based convolutional neural network model.The experiment proves that the Res Net module is added to the network After adding and expanding the convolution module,the improved network has increased by 2%,10%,and 9% in precision,recall,and F1 indicators,respectively.(3)Aiming at the problem of broken and incomplete extraction results of the proposed multi-scale convolutional neural network in the face of complex scenes,it is done by introducing Spatial Pooling Pyramid and Attention Mechanism(AM).The second improvement has strengthened the generalization ability of the feature extraction network.The three indicators used in Chapter 3 have increased by 4%,13%,and 14% compared with the original U-Net network.