Remote Sensing Image Road Extraction Based On Deep Learning

Road extraction of remote sensing images is of great significance for GIS data update,map drawing,path analysis and so on.With the development of space technology such as computer technology and satellite,the acquisition of remote sensing images is getting easier and the features in the image are more and more clear,which provides abundant ground information for road extraction,but also brings complex background.Interference,such as trees around roads,urban buildings,vehicles on the road,etc.,makes road extraction difficult.The traditional road extraction method is mainly based on manual operation,and the extraction effect is poor under complex background interference.In response to this problem,this paper mainly studies how to effectively use a large number of high-resolution remote sensing images for road extraction from the following four aspects:(1)The traditional method of traditional road extraction is studied.Through the study of convolutional neural network(CNN)theory,the advantages of applying it to road extraction are analyzed.It has a hierarchical structure and superb learning and expression ability.It can automatically extract road features from massive data and is suitable for road extraction in complex background environments.In this paper,the convolutional neural network is introduced into the road extraction.After comparing the differences of different segmentation networks,this paper chooses an efficient semantic segmentation network—Link Net.(2)Constructed data sets in different backgrounds,mainly covering Thailand,India,Indonesia and other countries,including cities,villages,deserts,beaches,tropical rain forests and other scenarios.The training parameters are trained on the deep learning platform Py Torch.The dataset is used to train the Unet network and the Link Net network.The accuracy of the Unet model is 15% higher than that of the traditional K-means based road extraction.(3)We Analyze the extraction results of the Unet and Link Net models.Due to the incompleteness of the image information during the road feature extraction process,the roads in the extraction result are discontinuous,spots,voids,etc.appear in the road.This paper introduces the deep residual network Res Net50 in the Link Net encoder part to replace the original Res Net18,and learn more image features by increasing the number of layers.At the same time,the cavity convolution is introduced in the central area of Link Net,so that the image perception field can be increased without reducing the resolution of the image.(4)In terms of the loss function,this paper replaces the previous square difference loss function with the binary cross entropy loss function,which can remove the influence of the learning function in the network propagation process.At the same time,in the process of training,for the problem of positive and negative sample imbalance(the road area in the image is much smaller than the area of the background area),this paper introduces the Dice coefficient based on the binary cross entropy loss function,and the two are taken together as new loss function.In this paper,four road extraction methods are compared in the experiment: traditional methods based on K-means,road extraction based on Unet model,road extraction based on Link Net model,deep residual residual network and hole convolution based on Link Net The method of introducing Dice coefficient is carried out on the dataset of this paper.The accuracy rates are 73.6%,88.1%,92.5% and 93.9% respectively.The experimental results show that the proposed road extraction method has the highest accuracy.