河南省滑坡灾害易发性制图研究:多种机器学习模型的对比  

Landslide disaster vulnerability mapping study in Henan Province:Comparison of different machine learning models

作　　者：曹文庚[1] 潘登 徐郅杰 张文培 任宇 南天 CAO Wengeng;PAN Deng;XU Zhijie;ZHANG Wenpei;REN Yu;NAN Tian(Institute of Hydrogeology and Environmental Geology,CAGS,Shijiazhuang 050061,China;Henan Provincial Institute of Natural Resources Monitoring and Land Consolidation with Rehabilitation,Zhengzhou 450016,China;Henan Provincial Key Laboratory of Geological Disaster Prevention and Control,Zhengzhou 450016,China)

机构地区：[1]中国地质科学院水文地质环境地质研究所,石家庄050061 [2]河南省自然资源监测和国土整治院,郑州450016 [3]河南省地质灾害防治重点实验室,郑州450016

出　　处：《地质科技通报》2025年第1期101-111,共11页Bulletin of Geological Science and Technology

基　　金：河南省重点研发与推广专项(科技攻关)(232102321012);国家自然科学基金项目(41972262);河北自然科学基金优秀青年科学基金项目(D2020504032);河南省重点研发专项(221111321500)。

摘　　要：河南省具有复杂的地貌类型,面临着频繁发生滑坡灾害的挑战,因此进行高效准确的滑坡易发性制图对于地方政府和居民具有重要意义。但是,在滑坡易发性制图研究中,如何选取适合河南省滑坡灾害数据集的机器学习模型、提高评价精度的对比研究仍需进一步开展。以河南省为研究区,收集滑坡数据并编录成滑坡灾害数据库。通过递归特征消除法筛选出对滑坡相对影响最高的11个因子(坡度、高程、平面曲率、剖面曲率、土地覆盖、岩性、土壤类型、降水量、道路密度、河流密度、断裂带密度)整合成空间数据集,训练多层感知机(MLP)神经网络、随机森林(RF)、极端梯度提升(XGBoost)和支持向量机(SVM)模型并使用接收者受试特征曲线下面积(AUC)评估各个模型性能,制作高精度滑坡易发性分区图。研究结果表明,多层感知机模型对河南省滑坡灾害数据集适配性最强,AUC达到0.95。相较于支持向量机、极端梯度提升和随机森林模型,MLP模型预测的滑坡灾害高易发区的面积占比最小,能更精确地识别潜在滑坡灾害高风险区域。预测的极高和高易发区主要分布在豫西山地、丘陵地区,地形因素对河南省滑坡灾害发育具有主导作用。研究成果可为大尺度区域开展高精度滑坡灾害易发性评价提供参考。[Objective]Henan Province,with its complex geomorphology,and faces the challenge of frequent.landslide disasters.Therefore,efficient and accurate landslide susceptibility mapping(LSM)is of great significance for local governments and residents.However,further comparative research is needed on how to select machine learning models suitable for the landslide disaster dataset in Henan Province to improve evaluation accuracy in landslide susceptibility mapping research.[Methods]This study takes Henan Province as the research area,collected landslide data and compiled it into a landslide disaster database.By using the recursive feature elimination method,the 11 factors that have the highest relative impact on landslides(slope,elevation,plan curvature,profile curvature,land cover,lithology,soil type,precipitation,road density,river density,fault density)were selected and integrated into a spatial dataset.Multi layer perceptron(MLP)neural network,random forest(RF),extreme gradient Boosting(XGBoost),and support vector machine(SVM)models were trained,and the model performances were evaluated with receiver operating characteristic curves and the area under the curve,finally,a high-precision landslide susceptibility zoning map was created.[Results]The MLP model showed the strongest adaptability to the landslide disaster dataset in Henan Province,with an AUC of 0.95.Compared to SVM,XGBoost,and RF models,the MLP model predicted the smallest landslide proportion in highly susceptible areas,and can more accurately identify potential high-risk areas for landslide disasters.The predicted extremely high and high-risk areas are mainly distributed in the mountainous and hilly areas of western Henan Province,and terrain factors play a dominant role in the development of landslide disasters in Henan Province.[Conclusion]These results provide a high-accuracy reference for landslide susceptibility assessment over large areas.

关 键 词：滑坡 多层感知机 机器学习 易发性制图 河南省 

分 类 号：P642.22[天文地球—工程地质学]