基于数学统计与机器学习模型耦合的滑坡易发性评价方法优化  

Optimization of Landslide Susceptibility Assessment Method Coupling Mathematical Statistics and Machine Learning Models

作　　者：刘山东 李军[2] 江兴元[1,3] 杨义 赵荣乾[3] LIU Shan-dong;LI Jun;JIANG Xing-yuan;YANG Yi;ZHAO Rong-qian(Key Laboratory of Karst Georesources and Environment(Guizhou University),Ministry of Education,Guiyang 550025,China;114 Geological Brigade,Guizhou Geological and Mining Bureau,Zunyi 563000,China;College of Resources and Environmental Engineering,Guizhou University,Guiyang 550025,China)

机构地区：[1]喀斯特地质资源与环境教育部重点实验室(贵州大学),贵阳550025 [2]贵州省地质矿产勘查开发局114地质大队,遵义563000 [3]贵州大学资源与环境工程学院,贵阳550025

出　　处：《科学技术与工程》2025年第5期1827-1839,共13页Science Technology and Engineering

基　　金：国家自然科学基金(42007271);贵州省科技支撑计划项目(黔科合支撑[2023]一般119)。

摘　　要：滑坡地质灾害易发性评价是防灾减灾的一种重要手段,易发性评价模型的选取和优化至关重要。以思南县为研究区,选取高程、坡度、曲率、地层、土地利用、年平均降雨量等16个评价因子,采用频率比(frequency ratio,FR)模型与支持向量机(support vector machine,SVM)模型和随机森林(random forest,RF)模型相耦合,引入网格搜索方法来获取SVM模型、RF模型及其耦合模型最优参数组合并用于模型训练,最终构建SVM、RF、FR-SVM及FR-RF模型对整个研究区进行滑坡易发性预测,并进行了受试者操作特征(receiver operating characteristics,ROC)曲线验证。结果表明:与单一机器学习模型相比,耦合机器学习有更多的滑坡灾害样本落于高易发区和极高易发区,有更高的准确率。单一模型中,RF模型有较多的滑坡灾害样本落于高易发区和极高易发区,耦合模型中,FR-RF模型有较多的滑坡灾害样本落于高易发区和极高易发区,且FR模型和FR-RF模型中没有滑坡灾害样本落在极低易发区,表明无论是单一模型还是耦合模型,RF模型的性能优于SVM模型。4种模型的ROC预测曲线的曲线下面积(area under the curve,AUC)分别为0.8316、0.8439、0.8644、0.9104,说明FR模型与RF模型结合的耦合模型有更高的准确率,该模型更适用于思南县的滑坡易发性评价研究,评价结果可为当地滑坡地质灾害的防灾减灾提供一定的参考。Landslide geological hazard susceptibility assessment is an important means of hazard prevention and reduction.The selection and optimization of susceptibility assessment model is very important.Sinan County was selected as the study area,and 16 assessment factors such as elevation,slope,curvature,lithology,land use,and average annual precipitation were selected.Frequency ratio(FR)model was coupled with support vector machine(SVM)model and random forest(RF)model.Grid search method was introduced to obtain the optimal parameter combination of SVM model,RF model and their coupling model for model training.Finally,SVM,RF,FR-SVM and FR-RF models were constructed to predict landslide susceptibility in the whole study area,and receiver operating characteristics(ROC)curve was performed verification.The results show that compared with the single machine learning model,the coupled machine learning model has more landslide hazard samples fall in the high zone and the very high zone,and has higher accuracy.In the single model,more landslide hazard samples in the RF model fall in the high zone and the extremely high zone.In the coupled model,more landslide hazard samples in the FR-RF model fall in the high zone and the very high zone,and no hazard samples points in the FR model and the FR-RF model fall in the very low zone,indicating that no matter the single model or the coupled model,The performance of RF model is better than that of SVM model.The AUC values of ROC prediction curves of the four models are 0.8316,0.8439,0.8644 and 0.9104,indicating that the coupling model combined with FR model and RF model has a higher accuracy,and this model is more suitable for the assessment of landslide susceptibility in Sinan County.The assessment results can provide some reference for hazard prevention and reduction of local landslide geological hazards.

关 键 词：滑坡易发性评价 频率比模型 机器学习模型 耦合模型 ROC曲线 思南县 

分 类 号：P642.2[天文地球—工程地质学]