基于深度学习和持续同调的LiDAR林下滑坡提取  

Identification of Landslide in Forest Using LiDAR Data Based on Deep Learning and Persistent Homology

作　　者：贺跃光[1] 姜风航 苗则朗 包志轩 易南洲 HE Yueguang;JIANG Fenghang;MIAO Zelang;BAO Zhixuan;YI Nanzhou(School of Traffic and Transportation Engineering,Changsha University of Science and Technology,Changsha 410004,Hunan,China;School of Geosciences and info-Physics,Central South University,Changsha 410083,Hunan,China;Hunan Water Resources and Hydropower Survey,Design,Planning and Research Co Ltd,Changsha 410119,Hunan,China)

机构地区：[1]长沙理工大学交通运输工程学院,湖南长沙410004 [2]中南大学地球科学与信息物理学院,湖南长沙410083 [3]湖南省水利水电勘测设计研究总院,湖南长沙410119

出　　处：《矿冶工程》2023年第5期32-36,共5页Mining and Metallurgical Engineering

基　　金：国家自然科学基金面上项目(42171084)。

摘　　要：为了从森林覆盖区获取林下滑坡信息,采用LiDAR点云技术构建高分辨率数字地形模型,结合Res-Unet网络和持续同调理论提取林下滑坡信息。选取美国华盛顿州风河实验林作为研究区,选择其中3个区域进行定量分析,经计算,区域内提取滑坡信息的准确度均值为79.7%,召回率均值为70.2%,F1均值为65.5%,表明基于Res-Unet和持续同调的提取方法能够准确识别研究区内大部分滑坡;基于深度学习和持续同调的林下滑坡提取方法引入持续同调方法到滑坡提取领域,并与深度学习相结合,弥补了传统遥感方法在植被覆盖区滑坡提取效果方面的不足,可为滑坡分析提供有力的技术支持。In order to obtain the information of landslide in forest area,a high⁃resolution digital terrain model was derived from LiDAR point cloud,and the data of landslide in forest area were extracted by Res⁃Unet network and persistent homology.The Fenghe Experimental Forest in Washington State of USA was selected for study,among which three areas were selected for quantitative analysis.Based on calculation,the extracted data of landslide in the forest area show an average precision of 79.7%,an average recall rate of 70.2%,and average F1 of 65.5%.The extraction method based on Res⁃Unet and persistent homology can accurately identify most landslides in the research area.It is shown that by using deep learning and persistent homology,this extraction method can make up for the weakness of traditional remote sensing methods in extracting landslide information in vegetation covered areas,and also provide a technical support for landslide analysis.

关 键 词：点云数据 滑坡 数字地形模型 深度学习 持续同调 林下滑坡提取 

分 类 号：P237[天文地球—摄影测量与遥感]