基于随机森林算法的城市内涝风险预估方法及应用—以南昌市为例  

Method and application of urban waterlogging risk prediction based on the random forest algorithm-taking Nanchang as an example

作　　者：盛志军 周雨 孟明华 刘禹杉 傅文兵 SHENG Zhijun;ZHOU Yu;MENG Minghua;LIU Yushan;FU Wenbing(Jiangxi province meteorological observatory,Nanchang 330096;Jiangxi Province meteorological sevice center,Nanchang 330096;Jiangxi Provincial Disaster protection and alleviation engineering technology research center,Nanchang 330096;Nanchang city meteorological bureau,Nanchang 330000)

机构地区：[1]江西省气象台,南昌330096 [2]江西省气象服务中心,南昌330096 [3]江西省防灾减灾工程技术研究中心,南昌330096 [4]南昌市气象局,南昌330000

出　　处：《暴雨灾害》2025年第1期93-101,共9页Torrential Rain and Disasters

基　　金：江西省气象局面上项目(JX2021M03);江西省重点研发计划项目(20202BBGL73100);江西省气象局重点项目(JX2023Z07);江西省气象局指导性计划项目(ZDX2021001);江西省气象局重点项目(JX2020Z09)。

摘　　要：利用南昌市2016—2022年内涝灾情资料及110个国家和区域气象观测站小时降雨数据,首先分析了南昌市内涝灾害的时空分布特征,确定了不同积水深度降雨指标强度临界值,结合承灾体条件,建立内涝风险预估等级;然后进一步基于随机森林算法,构建积水深度预估模型,开展城市内涝风险预评估;最后应用典型个例对风险等级预报的准确性进行检验。结果表明:(1)内涝灾害月变化呈现单峰型分布特征,峰值出现在6月;日变化呈现双峰型分布特征,峰值分别出现在09时和16时,灾害点多集中在城区核心区域。(2)短时强降雨或暴雨是导致南昌市50 cm以上积水深度的主要降雨类型,当1 h降雨量>40 mm、3 h累计降雨量>78 mm、6 h累计降雨量>98 mm、12 h累计降雨量>123 mm或24 h累计降雨量>135 mm时,极易引发50 cm以上的积水深度。(3)积水深度预估模型训练集和测试集平均精确率分别为96%和79%。依据积水深度预估区间[10,25)cm、[25,50)cm、≥50 cm,结合人口和GDP条件,将内涝气象风险划分为低、中、高3级。(4)南昌市两次暴雨过程内涝灾害点风险等级预估准确率分别为67%和56%,提前1~3 h预估的27个内涝点均在预估风险区域内。Waterlogging disaster data from 2016 to 2022 and hourly precipitation data from 110 meteorological observation stations in Nanchang were used to analysis the characteristics of waterlogging disasters,as well as the critical value of the rainfall index of different water depth was determined.Then combined with the disaster-bearing body conditions,the prediction level of waterlogging meteorological risk was furthter established.Moreover,the water depth estimation model was constructed based on the random forest algorithm.Finally,pre-assessment of urban waterlogging risk was carried out and the accuracy of risk level forecast was tested by using the typical cases.The results are as follows.(1)The monthly change of waterlogging disaster demonstrate unimodal distribution with a peak in June.The daily change demonstrate bimodal distribution with two peaks at 09:00 BT and 16:00 BT respectively.The disaster points are concentrated in the core urban area.(2)Water depths above 50 cm in Nanchang are primary caused by short-duration heavy rainfall or rainstorms.It is easy to cause the water depth of more than 50 cm,when 1 h rainfall>40 mm,3 h cumulative rainfall>78 mm,6 h cumulative rainfall>98 mm,12 h cumulative rainfall>123 mm or 24 h cumulative rainfall>135 mm.(3)The average accuracy of the training and tests was 96%and 79%respectively.The meteorological risk of waterlogging is divided into low,medium and high levels according to the estimated range of[10,25)cm,[25,50)cm,≥50 cm combined with the population and GDP conditions.(4)The estimated accuracy rate of waterlogging points in the two rainstorm processes in Nanchang are 67%and 56%respectively.And the 27 waterlogging points estimated 1-3 h in advance are all within the estimated risk area.

关 键 词：内涝 随机森林 风险预估 南昌市 

分 类 号：P429[天文地球—大气科学及气象学]