基于Pytorch框架搭建U-Net网络模型的遥感影像建筑物提取研究  

Building extraction from remote sensing images using U-Net network model based on Pytorch framework

作　　者：焦利伟[1,2] 张敏 麻连伟 秦建辉 JIAO Li-wei;ZHANG Min;MA Lian-wei;QIN Jian-hui(Geophysical Survey Brigade of Henan Coal Geology Bureru,Zhengzhou 450009,China;Research Center for Henan Geological&Geophysical Exploration Engineering Technology,Zhengzhou 450009,China;School of Economics&Management,Henan Polytechnic University,Jiaozuo 454003,China)

机构地区：[1]河南省煤田地质局物探测量队,河南郑州450009 [2]河南省地质物探工程技术研究中心,河南郑州450009 [3]河南理工大学工商管理学院,河南焦作454003

出　　处：《河南城建学院学报》2020年第4期52-57,共6页Journal of Henan University of Urban Construction

基　　金：河南省自然资源科研资助项目(2019-379-15)。

摘　　要：基于深度学习方法,采用Pytorch框架搭建U-Net网络模型,进行了遥感影像建筑物提取研究。首先以建筑物为目标,构建基于光学遥感影像的建筑物样本库,然后进行网络训练建立深度学习模型,并对样本库更新进行模型优化,最后用优化后的模型进行建筑物提取,并与最大似然法、支持向量机法(SVM)进行对比。结果表明:在训练数据集充足的情况下,使用深度学习对台前县建筑物提取总体精度为94.3%、Kappa系数为0.83,罗山县总体精度为97.5%、Kappa系数为0.75,均高于传统方法,说明利用深度学习的方法进行建筑物提取具有一定的有效性和适用性。Based on the deep learning method,the U-Net network model is built by Pytorch framework,and the remote sensing image building extraction research is carried out.Firstly,taking buildings as the target,a building sample database based on optical remote sensing images is constructed,then a depth learning model is established through network training,and the model is optimized for updating the sample database.Finally,buildings are extracted by using the optimized model,and compared with the maximum likelihood method and support vector machine method(SVM).The results show that under the condition of sufficient training data set,the overall accuracy of building extraction by deep learning is 94.3%,Kappa coefficient is 0.83,and that by Luoshan County is 97.5%,Kappa coefficient is 0.75,which is higher than that by traditional methods.This shows that the method of building extraction by deep learning has certain effectiveness and applicability.

关 键 词：遥感影像 建筑物提取 Pytorch U-Net 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]