融合InSAR与信息量-机器学习耦合模型的黄土滑坡易发性评价  

Evaluation of Loess Landslide Susceptibility by Combining InSAR and Information-Machine Learning Coupling Model

作　　者：胡祥祥 石亚亚 胡良柏 吴涛 庞栋栋 刘帅令 宋宝 HU Xiangxiang;SHI Yaya;HU Liangbai;WU Tao;PANG Dongdong;LIU Shuailing;SONG Bao(School of Resource and Environmental Engineering,Tianshui Normal University,Tianshui 741001,Gansu,China;Gansu Zhenghao Satellite Remote Sensing Science and Technology Center,Tianshui 741001,Gansu,China;Organization Department of CPC Tonggu County Committee,Yichun 336200,Jiangxi,China;School of Automation,Beijing Institute of Technology,Beijing 100081,China)

机构地区：[1]天水师范学院资源与环境工程学院,甘肃天水741001 [2]甘肃正昊地星遥感科技中心,甘肃天水741001 [3]中共铜鼓县委组织部,江西宜春336200 [4]北京理工大学自动化学院,北京100081

出　　处：《西北地质》2025年第2期159-171,共13页Northwestern Geology

基　　金：国家自然科学基金项目“气候暖湿化背景下青藏工程走廊冻土环境工程承载力演化趋势研究(42361020)”;天水师范学院创新基金项目“融合InSAR的天水市滑坡易发性评价研究(CXJ2023-19)”联合资助。

摘　　要：环境因子、气象因子与人类活动之间的相互作用,影响地表形态的变化。尤其对于黄土高原区域,在诸多因子的复杂互馈作用下易导致黄土崩滑灾害,亟需选择适用的影响因子和训练模型开展滑坡易发性评价研究。本研究以天水市为研究区,基于InSAR获取的地表形变信息,综合地形、水文、气候、生态以及人类活动等诸多影响因素,采用信息量模型(IV)分别联接到随机森林模型(RF)、决策树模型(DT)、支持向量机模型(SVM)和BP神经网络模型(BP)构建耦合模型IV-RF、IV-DT、IV-SVM和IV-BP,开展滑坡易发性评价研究。结果表明:耦合模型(IV-RF、IVDT、IV-SVM和IV-BP)的AUC值分别为0.925、0.846、0.883、0.792,IV-RF具有更强的精度。滑坡频率比IV-RF模型从极低易发分区向极高易发区逐渐递增,滑坡易发性分区结果更均匀平稳。IVRF模型具有更强的预测能力和精度,更适合黄土滑坡地质灾害易发性评价。IV-RF模型的极高、高、中、低、极低易发性区域面积占比分别为20.45%、18.28%、22.27%、16.92、22.09%,主要分布在天水市北部地质环境复杂和人类活动强烈的山地、黄土梁峁地区。岩性、坡度、土地利用、降雨、道路密度、InSAR形变在贡献率分析中排在前6位,是影响滑坡发育的主控因子。本研究旨在为黄土高原滑坡灾害的预测和防治工作提供可靠的科学依据,为滑坡易发性评价研究深化建模思路,优化独立模型评价结果不确定性问题。The interaction between environmental factors,meteorological factors,and human activities affects surface morphology change.Especially for the Loess Plateau region,it is easy to cause a loess slide disaster under the complex interaction of many factors,so selecting suitable influencing factors and training models to conduct landslide susceptibility evaluation research is urgent.This study takes Tianshui City as the research area and constructs a multi-factor evaluation system covering terrain scale,basic environmental factors,and human activity scale based on the surface deformation information obtained by InSAR.The coupled models IV-RF,IVDT,IV-SVM,and IV-BP were constructed by connecting the information content model(IV)to the random forest model(RF),decision tree model(DT),support vector machine model(SVM)and BP neural network model(BP),and the landslide susceptibility evaluation was carried out.The results show that the AUC values of the coupled models(IV-RF,IV-DT,IV-SVM,and IV-BP)are 0.925,0.846,0.883,and 0.792,respectively,and IVRF has stronger accuracy.Compared with the IV-RF model,the landslide frequency gradually increases from the very low prone zone to the very high prone zone,and the results of the landslide-prone zone are more uniform and stable.The IV-RF model has stronger prediction ability and accuracy and is more suitable for evaluating the geological hazard susceptibility of loess landslides.The areas of extremely high,high,medium,low,and very low susceptibility in the IV-RF model accounted for 20.45%,18.28%,22.27%,16.92 and 22.09%,respectively,which were mainly distributed in the mountainous and loess ridge areas with complex geological environment and strong human activities in the north of Tianshui City.Lithology,slope,land use,rainfall,road density,and InSAR deformation rank the top 6 in the contribution rate analysis and are the main controlling factors affecting landslide development.This study aims to provide a reliable scientific basis for predicting and preventing landslide disasters in

关 键 词：黄土高原 易发性评价 信息量模型 机器学习方法 INSAR 

分 类 号：P694[天文地球—地质学]