使用深度学习方法实现黄土滑坡自动识别  

Automatic Object Detection of Loess Landslide Based on Deep Learning

作　　者：巨袁臻 许强[1] 金时超 李为乐[1] 董秀军[1] 郭庆华 JU Yuanzhen;XU Qiang;JIN Shichao;LI Weile;DONG Xiujun;GUO Qinghua(State Key Laboratory of Geohazrd Prevention and Geoenvironment Protection,Chengdu University of Technology,Chengdu 610059,China;Plant Phenomics Research Center,Nanjing Agricultural University,Nanjing 210095,China;University of Chinese Academy of Sciences,Beijing 100049,China;State Key Laboratory of Vegetation and Environmental Change,Institute of Botany,Chinese Academy of Sciences,Beijing 100093,China)

机构地区：[1]成都理工大学地质灾害防治与地质环境保护国家重点实验室,四川成都610059 [2]南京农业大学作物表型组学交叉研究中心,江苏南京210095 [3]中国科学院大学,北京100049 [4]中国科学院植物研究所植被与环境变化重点实验室,北京100093

出　　处：《武汉大学学报（信息科学版）》2020年第11期1747-1755,共9页Geomatics and Information Science of Wuhan University

基　　金：国家自然科学基金创新研究群体科学基金(41521002);四川省科技厅重点研发计划(2019YFS0074);四川省科技计划(2018SZ0339)。

摘　　要：区域性滑坡识别是滑坡灾害风险管理的基础,传统的识别工作主要依靠人力完成。在已有的滑坡自动识别研究中,方法上以机器学习为主,数据源上对谷歌地球影像应用较少,识别对象上多以与环境差异较大的新滑坡为主。结合深度学习方法和谷歌地球影像数据对中国典型黄土地区历史滑坡进行自动识别。首先,基于开源谷歌地球影像建立了历史黄土滑坡样本数据库,包含黄土滑坡2498处;然后,利用掩膜区域卷积神经网络(mask region-based convolutional networks,Mask R-CNN)目标检测模块进行黄土滑坡自动识别。识别的准确率为0.56,召回率为0.72,F1值为0.63。结果表明,Mask R-CNN是一种稳健性较好的方法,可以用于以谷歌地球影像为数据源的黄土滑坡自动识别,为快速准确地进行区域滑坡灾害调查提供了可能。Objectives:The knowledge of regional landslides detection plays a fundamental role in the landslide risk management.However,most of that recognition was taken manually in the past,which is rather time-and labor-consuming.As the development of technologies of remote sensing and artificial intelligence,the automatic detection of landslides becomes possible.The previous researches relative to the automatic detection of landslides utilized the machine learning methods to detect these new landslides which were significantly distinguished from their context.Compared to those landslides,the detection of old loess landslides that are not distinct from their context is more challenged.We explore the deep learning to automatically detect the old loess landslides.Methods:Firstly,we build a loess landslide database consists of 2498 which are interpreted from the Google Earth images by experts.Then,we divide the database into three datasets for training,validation and test.Finally,we train Mask R-CNN object detection module with the training dataset,choose the best model by the validation dataset,and apply the best model to the test dataset.Results:The test results of model performance show a precision of 0.56,a recall of 0.72,and a F1-score of 0.63.Conclusions:The results indicate that Mask R-CNN is a robust method even for the detection of loess landslides that are unapparent from the context,and deep learning can provide the possibility for rapid and accurate regional geo-hazard investigation.

关 键 词：黄土滑坡 自动识别 谷歌地球影像 深度学习 Mask R-CNN 目标检测 

分 类 号：P237[天文地球—摄影测量与遥感] P208[天文地球—测绘科学与技术]