| Urban boundary is the basic measure of planning urban development.Dividing urban boundary is the basic work of modern urban planning and land use.In the process of urban boundary division,identifying urban and non urban areas quickly,accurately and efficiently has become a more and more important requirement in urban research.The continuous growth of high-resolution remote sensing images has brought sufficient data resources to the efficient identification of urban boundaries,which makes the fine management and control of land use in urban development more feasible and practical.Although the traditional urban boundary identification method is easy to operate,it is only suitable for individual urban areas with perfect database because of the inconsistency of basic geographic information data sources,time scales and great difficulty in obtaining.The application of SVM,random forest and other methods in machine learning has made a certain contribution to the automation of urban boundary recognition,but the selection of artificial features has a certain subjectivity,and the recognition accuracy needs to be further improved.Based on high-resolution remote sensing images,using the method of deep learning,mining the deep feature information in remote sensing images through image sample input and model training,so as to avoid the subjectivity of feature extraction.Under the condition of sufficient samples,urban area recognition and extraction can be completed automatically,intelligently and efficiently.Based on this,this paper uses the sentinel-2 data of high-resolution remote sensing image and the method of deep learning semantic segmentation model,and selects Henan Province as the research area to study the urban boundary.The main research contents and results are as follows:(1)According to the characteristics of urban boundary and its spatial distribution law,a set of principles and processes of urban boundary division are constructed,and on this basis,the urban boundary data set of Henan Province(hnub2018)is produced.Data verification shows that hnub2018 has reached a high level of accuracy,with an overall accuracy of 92.82% and a kappa coefficient of 0.8553,which is significantly better than the GUB(Henan)data products in the same period.The establishment of highprecision hnub2018 data set has laid a data foundation for the attempt of urban boundary extraction in the field of deep learning technology.(2)The sample set of urban boundary is constructed according to the hnub2018 data set,and the performance is tested.This article will 256 × The 256 size data sample set is used as the training sample of urban boundary,and is tested in five classical semantic segmentation models: u-net,FPN,pspnet,u-net + +and deeplab V3 +.The test results show that the accuracy of Zhengzhou urban boundary prediction results in the five classical network structures is more than 88%,indicating that hnub2018 data set has good applicability and can be used for urban boundary model training and testing.Among them,the restoration results and accuracy data of deeplab V3 + network are the best in five classical models,which can be used for subsequent urban boundary extraction.(3)Based on the urban boundary sample set constructed in this paper,the deeplab V3 + network is improved,and the urban identification model is constructed combined with conditional random field CRF.Deeplab V3 + model combined with CRF in terms of precision,accuracy Recall、F1-Score、m＿The results of IOU and restoration were improved obviously.The addition of CRF significantly solves the hole phenomenon and "noise" influence in semantic segmentation.Final accuracy index m＿IOU increased significantly,by more than 1.76%,The F1-score reached 96.96%. |
PDF Full Text Request