灾害大数据驱动的县域重大洪涝过程灾害风险评估

Disaster Risk Assessment at County Level of a Heavy Flooding Driven by Disaster Big Data

基于我国南方地区625个重大洪涝过程案例,收集了县域23项指标的近30万条数据,利用XGBoost算法建立了6个重大洪涝过程灾害风险评估模型,用于灾害事件发生前对受灾人口风险、紧急转移安置人口风险、农作物受灾面积风险、倒塌和严重损害房屋风险、直接经济损失风险和灾害综合风险进行评估。通过实际案例对模型进行验证,灾害风险评估结果准确率整体达到80%以上,表明该模型具有较好的泛化能力,能够用于实际灾害评估工作中。实验对比发现,全指标比仅用致灾因子指标可以使评估准确率提升10%~15%;另外,训练样本量提高1~2个数量级也能够使模型评估准确率提升5%~13%,这表明灾害大数据的积累对灾害风险评估工作具有重要意义。

Disaster risk assessment is a key step in the risk management of a heavy flooding.The data-driven machine learning models for risk assessment is not only convenient for modeling,but also can comprehensively consider various indices of the flood disaster system including disaster-causing factors,disaster pregnant environment,disaster-bearing body and disaster losses.The ensemble methods represented by XGBoost can calculate each index importance and thus improve the interpretability of the model.Based on 625 heavy flooding cases in southern China,we collected nearly 300,000 data of 23 county-level indices and established 6 models using the XGBoost algorithm,which can be used before the occurrence of a heavy flooding to assess the risk of the affected population,the population to be transferred and resettled,the affected crop area,the collapse and serious damage houses,the direct economic loss and the comprehensive risk of the flood disaster.The models are verified by a case in 2021 and the overall accuracy of disaster risk assessment exceeded 80%,which proves that the models in this study have a good generalization ability and can be utilized in practical disaster assessment.Compared with the results only using the disaster causing factor index,utilizing the whole index can boost the evaluation accuracy by 10-15%.Moreover,increasing the sample size by 1-2 orders of magnitude can improve the accuracy by 5%-13%.This indicates that the accumulation of disaster big data will be of great significance to disaster risk assessment.