Risk prediction based on oversampling technology and ensemble model optimized by tree-structured parzed estimator

Abstract

High accuracy prediction of urban flood risk is conducive to avoid potential losses, however, it's negatively affected by unbalanced data. Furthermore, ensemble model has been demonstrated to have the ability to improve to prediction accuracy. Nevertheless, the performance of ensemble model is influenced by basic model and ensemble rules, and determining the best ensemble model remains an open issue. To improve the accuracy of flood risk prediction, an approach covering data optimization and ensemble modeling was presented to optimize unbalanced flood data and the selection of various ensemble models based on efficiency and performance. A practical application in Zhengzhou City shows that Borderline-SMOTE2 is the most applicable for optimizing the flood risk data among the state-of-the-art oversampling algorithm utilized, because of the excellent entropy value. The effect of unbalanced data on the performance of the basic models was pervasive according to changes of the common indicators. The optimal ensemble model for flood risk prediction is composed of K-Nearest Neighbor, Decision Tree, Gaussian Naive Bayes and Extreme Gradient Boosting under Stacking rule in the current study. The results of this study supply the valuable reference for the flood prediction and mitigation.