基于机器学习的降雨—滑坡灾害链直接经济损失脆弱性评估  

Rainfall-induced Landslides Direct Economic Loss Vulnerability Assessment Using Machine Learning Methods

作　　者：李子轩 杜鹃 徐伟[1,2,3] LI Zixuan;DU Juan;XU Wei(Key Laboratory of Environmental Change and Natural Disasters,Ministry of Education,Faculty of Geographical Science,Beijing Normal University,Beijing 100875,China;State Key Laboratory of Earth Surface Processes and Resource Ecology,Beijing Normal University,Beijing 100875,China;Academy of Disaster Reduction and Emergency Management,Ministry of Emergency and Ministry of Education,Beijing 100875,China)

机构地区：[1]北京师范大学地理科学学部,环境演变与自然灾害教育部重点实验室,北京100875 [2]北京师范大学地表过程与资源生态国家重点实验室,北京100875 [3]应急管理部-教育部减灾与应急管理研究院,北京100875

出　　处：《灾害学》2022年第4期220-224,共5页Journal of Catastrophology

基　　金：国家重点研发计划课题“多灾种重大自然灾害承灾体脆弱性与恢复力评估技术”(2018YFC1508802);教育部—国家外国专家局高等学校学科创新引智计划“北京师范大学综合灾害风险管理创新引智基地2.0”(BP0820003)。

摘　　要：承灾体脆弱性定量评估是灾害风险评估中的一项重要任务,也是个难点问题。开展降雨—滑坡直接经济损失脆弱性定量评估能直接为灾害链风险评估提供关键参数,对区域灾害链风险防范具有重要的意义。该文基于区域灾害系统,综合考虑致灾因子危险性、孕灾环境稳定性和承灾体脆弱性要素,采用机器学习算法,构建了降雨—滑坡灾害链直接经济损失脆弱性评估模型,并以贵州省毕节和六盘水两市为例,开展了降雨—滑坡灾害链直接经济损失脆弱性定量评估。结果表明,在三种机器学习算法模型中,随机森林和决策树模型具有相对较好的效果(最优变量组合的R 2分别为0.284和0.342,RMSE分别为7.92和7.59),XGBoost算法的效果相对较差。不同方法在实际损失的极值预测仍然存在偏差。决策树模型中脆弱性贡献变量最为重要的是NDVI、GDP和高程,而随机森林模型中则为累计有效降雨量和距道路距离。The quantitative assessment of vulnerability of cascading disasters is an important task and a difficult problem in disaster risk assessment,which provides key parameters for risk assessment.Therefore,it is vital to regional risk prevention of cascading disaster.In this paper,based on the theory of functional system of disaster system,the vulnerability assessment model of direct economic loss of rainfall-induced landslide is constructed by using machine learning algorithms,and the quantitative assessment of direct economic loss of rainfall-induced landslide is carried out in Bijie and Liupanshui cities of Guizhou Province as an example.The results show that among the three machine learning algorithm models the random forest and decision tree models have relatively good results(R 2 of the optimal variable combination is 0.284 and 0.342,and RMSE is 7.92 and 7.59,respectively),while XGBoost model presents relatively less effective.Biases still exists for the extreme value prediction of the actual losses.The most important vulnerability contributing variables in the decision tree model are NDVI,GDP and elevation,while the cumulative effective rainfall and distance from the road are more important in the random forest model.

关 键 词：降雨—滑坡灾害链 脆弱性 机器学习算法 定量评估 毕节和六盘水市 

分 类 号：X43[环境科学与工程—灾害防治] X915.5[天文地球—工程地质学] P642[天文地球—地质矿产勘探] P694[天文地球—地质学]