基于DeepLabv3+的高分辨率遥感影像建筑物自动提取  

Automatic extraction of buildings from high-resolution remote sensing images based on DeepLabv3+

作　　者：于明洋 张文焯 陈肖娴 刘耀辉 YU Mingyang;ZHANG Wenzhuo;CHEN Xiaoxian;LIU Yaohui(School of Surveying and Geo-Informatics, Shandong Jianzhu University, Jinan 250101, China;Hebei Key Laboratory of Earthquake Dynamics, Sanhe 065201, China)

机构地区：[1]山东建筑大学测绘地理信息学院,济南250101 [2]河北省地震动力学重点实验室,河北三河065201

出　　处：《测绘工程》2022年第4期1-10,17,共11页Engineering of Surveying and Mapping

基　　金：国家自然科学基金资助项目(41801308,51608309);河北省地震动力学重点实验室开放基金项目(FZ212203);国家对地观测科学数据中心开放基金项目(NODAOP2020008)。

摘　　要：提出一种建筑物自动化提取架构,基于DeepLabv3+网络模型,使用WHU建筑物数据集,完成数据集增强、模型训练、建筑物提取以及精度评估。实验表明,架构中DeepLabv3+模型分类的总体精度为96.3%、准确度为94.2%、召回率为92.5%、F1得分为93.3%、交并比为87.5%,优于基于像素的分类方法(支持向量机、K均值聚类算法(K-Means))和面向对象的分类方法(最邻近节点算法(KNN)、分析与回归树)以及基于深度学习的分类方法(UNet、SegNet、PSPNet)。文中构建的高分辨率遥感影像建筑物自动化提取模式,可以完成建筑物高精度高效率的提取任务。Efficient and accurate extraction of building information from high-resolution images is of great significance to national land planning and so on.This paper proposes an automated building extraction architecture,based on the DeepLabv3+network model,using the WHU building data set to complete data set enhancement,model training,building extraction and accuracy evaluation.Experiments show that the overall accuracy of DeepLabv3+model classification in the architecture of this article is 96.3%,accuracy is 94.2%,recall rate is 92.5%,F1 score is 93.3%,intersection ratio is 87.5%,which is better than pixel-based classification methods(support Vector machine(Support Vector Machine,SVM),K-means clustering algorithm(K-Means))and object-oriented classification methods(K-Nearest Neighbor,KNN),analysis and regression trees(Classification and Regression Trees),CART and classification methods based on deep learning(UNet,SegNet,PSPNet).The high-resolution remote sensing image building automation extraction architecture proposed in this paper can realize accurate and efficient building extraction and provide a reference for urban planning and management.

关 键 词：高分辨率遥感影像 建筑物提取 DeepLabv3+网络模型 深度学习 

分 类 号：P237[天文地球—摄影测量与遥感]