基于随机森林的暴雨灾害人口损失预估模型及应用  

Assessing Model of Casualty Loss in Rainstorms Based on Random Forest and Its Application

作　　者：刘扬 王维国 LIU Yang;WANG Weiguo(National Meteorological Centre,Beijing 100081)

机构地区：[1]国家气象中心,北京100081

出　　处：《气象》2020年第3期393-402,共10页Meteorological Monthly

基　　金：国家气象中心青年基金(Q201814)资助。

摘　　要：基于2009—2017年的广西县级暴雨灾情记录,综合考虑致灾因子、孕灾环境和承灾体因素选取7个解释变量,运用随机森林算法,构建暴雨灾害人口损失预估模型;并以精细化网格降水实况分析和预报产品驱动模型,预估是否发生人口损失。研究结果表明:模型训练样本及测试样本的分类准确率均在90%以上,致灾因子(降水情况)是最主要的解释变量,重要性从大到小依次是前10 d降水距平百分率、过程最大日雨量、最大小时雨量和短时强降水频次。应用智能网格降水产品对广西地区近两年的暴雨灾害过程进行回报试验,准确率超过70%。Based on historic casualty loss records of rainstorm that occurred at county level in Guangxi from 2009 to 2017,seven factors were selected as explanatory variables by comprehensively considering the trigger factors,disaster formative environment and exposure units,and the prediction model of casualty loss caused by rainstorms was built up by using random forest algorithms.The refined grid precipitation analysis and forecast products were used to drive the model to predict loss of life.The results showed that the classification accuracies are both above 90%in training and testing samples.Disaster-triggering factors(precipitation)are the most significant explanatory variables.The importances of these precipitation variables in turn are the anomaly percentage of accumulated precipitation over the previous 10 days,the maximum daily precipitation,the maximum hourly precipitation and the frequency of short-time severe rainfall.By applying the intelligent grid precipitation products,several rainstorm processes in Guangxi in recent two years were used to verify the model,showing that prediction accuracies are above 70%.

关 键 词：暴雨 人口损失 随机森林 智能网格 

分 类 号：P49[天文地球—大气科学及气象学]