Research On Direct Economic Loss Assessment Of Earthquake Hazard Based On Machine Learning

Earthquakes are natural disasters caused by tectonic changes and crustal movements inside the earth.When an earthquake occurs,violent vibrations will occur on the surface of the earth,which will cause damage to buildings,bridges,roads and other structures and facilities,and even cause disasters such as landslides.Earthquakes can also trigger secondary disasters such as tsunamis and volcanic eruptions,which pose a great threat to people’s lives and property.After the earthquake,the rapid assessment of the direct economic loss of the earthquake plays a guiding role in earthquake relief.Its significance includes determining the rescue scale according to the degree of loss,and providing the basis for rescue funds for the central and local governments.Assess the impact of earthquake losses on the national economy,and adjust financial input and funding arrangements.In recent years,with the development of computer technology and the application of artificial intelligence in various fields,machine learning algorithms have also become the mainstream method of learning and research,and have been applied to various aspects of earthquake research.In order to solve the need for rapid assessment of direct economic losses of earthquakes,this research proposed an assessment model for the direct economic loss of earthquakes in western.Based on the earthquake damage data from 1993 to 2017,combined with the economic data and seismic design data of each year,two methods of deleting missing feature samples and filling missing feature samples with the median value of sub-seismic magnitude were used for data preprocessing.In the case of data preprocessing,Spearman’s Rank Correlation Coefficient,RFE,and Lasso feature selection algorithm are used to find the optimal feature subset of the model,and nine evaluation models are constructed in combination with decision tree,random forest,and XGBoost machine learning regression tree model for data training and testing,using the combination of grid search and random search to find the optimal combination of model hyperparameters,and realize the rapid assessment of post-earthquake economic losses after model performance optimization.The experimental results show that the evaluation efficiency of the ensemble tree model is better than that of the single decision tree model,and the Spearman correlation analysis and RFE feature selection methods are better than the Lasso algorithm.In the case of reducing the input features of the model,the optimized regression model can obtain better evaluation results.When the data preprocessing method that removes samples with missing features is applied,the XGB-SP model(with an R2 value of 0.8878,MAE of 0.1938,and RMSE of 0.2418)and the XGBRFE model(with an R2 value of 0.8898,MAE of 0.1888,and RMSE of 0.2367)outperform other models in the evaluation of the western region;Under the data preprocessing method of filling missing feature samples with the median value of the magnitude,the XGB-SP model(with an R2 value of 0.8002,MAE of 0.2950,and RMSE of 0.3850)performs the best in the evaluation of the western region,followed by the RF-SP model(with an R2 value of 0.7793,MAE of 0.2967,and RMSE of0.4047).Overall,the data preprocessing method of removing samples with missing features is more suitable for model training with this sample data,and the XGB-SP model performs well in both methods.By using the model constructed in this study to evaluate actual earthquake examples,the evaluation results of the RF-SP,XGB-SP model are consistent with the actual economic losses,which can provide decision support for earthquake relief efforts.