基于机器学习算法的热带气旋灾害县级直接经济损失等级评估:以福建省为例OA北大核心CSTPCD

China is frequently affected by tropical cyclones,which can lead to severe economic losses.Rapid disaster loss assessment is crucial for effective emergency response.A variety of factors affect tropical cyclone disaster losses,which can be roughly categorized into hazard,exposure,and vulnerability.In the past,traditional statistical methods were used as the main tools for disaster loss assessment.To explore the potential of machine learning models,we explored five algorithms:the Light Gradient Boosting Machine(LightGBM),Random Forest(RF),eXtreme Gradient Boosting(XGBoost),Support Vector Machine(SVM),and Back-Propagation Neural Network(BP).The maximum gust wind and rainfall of tropical cyclones were selected to represent hazards,fixed capital stock data were used for the valuation of exposure,and the GDP of each county was collected to reflect capacity or vulnerability.In addition,river network density data were used as a simple proxy to demonstrate the contribution of flood-induced tropical cyclone rainfall.The relationship between these input variables and disaster loss at the county level was developed based on the data of 81 tropical cyclone events from 2009 to 2020 in Fujian Province.The performance of these models was compared using accuracy,precision,recall,and F1 scores.The accuracies of the LightGBM,RF,XGBoost,SVM,and BP models were 0.794 6,0.772 6,0.762 8,0.251 8,and 0.268 1,respectively.The main findings are as follows:(1)The performance of the ensemble learning algorithms(RF,XGBoost,and LightGBM)was higher than that of the individual classifiers(BP and SVM).The LightGBM model exhibited the best performance,with accuracy,precision,recall,and F1 scores>79%.(2)Maximum hourly rainfall and maximum wind gust are two of the most important loss-inducing factors,and fixed capital stock is a better proxy for disaster exposure than GDP.(3)The modeled losses are consistent with the actual losses under different but typical tropical cyclone events,indicating that the models can be applied to future tropical cyclone events impacting Fujian Province.However,this study had some limitations.First,some natural hazards,such as floods,storm surges,and waves,were not fully considered,which introduced uncertainty into the model results.Second,the emergency response capacity and actual actions taken among counties may have varied dramatically and were neglected due to data unavailability.In the future,hazard and vulnerability variables should be obtained to extend the model inputs.In addition,whether the model parameters trained with data from Fujian Province can be applied to other provinces remains unaddressed.In the future,to develop an operational model for the whole of coastal China,county-level data of all typhoon-prone areas in China with long-term time series are needed.