Analyzing Susceptibility to Tornado-Induced Injuries Using Hybrid Tree-Based Classifiers and Advanced Resampling Techniques

Tornadoes are among the most catastrophic atmospheric phenomena, causing extensive damage, injuries, and fatalities. This study aims to predict tornado-induced injuries in Texas by employing machine learning (ML) approaches, including Random Forest (RF), Adaptive Boosting (AdaBoost), eXtreme Gradient Boosting (XGBoost), and Support Vector Machines (SVM) in conjunction with advanced resampling methods aiding to address the class imbalance. Accordingly, the present research utilized a total of 2,986 property location characteristics as a result of tornadoes recorded from 2000 to 2023 inclusive, and took meteorological, socioeconomic, and geographical features into account to determine the tornado-induced injuries. Three under-sampling, i.e., random under sampling methods (RUS), condensed nearest neighbor (CNN), and instance hardness threshold (IHT), and three over-sampling, namely synthetic minority oversampling technique (SMOTE), random oversampling methods (ROS), and adaptive synthetic sampling (ADASYN), techniques were tested to enhance models’ predictive accuracy. The validation of these hybrid models demonstrated that the RF-IHT outperformed its counterparts, achieving satisfactory outcomes in identification tornado-induced injuries with an F1-score of 0.965, precision of 0.967, and recall of 0.967. The interpretability of the estimations was enhanced through the Local Interpretable Model-agnostic Explanations (LIME), which further identified the most critical predictors as tornado intensity, land use, and socioeconomic attributes. The susceptibility map developed in this study serves as a crucial decision-support mechanism for disaster risk management strategies, aiming to mitigate tornado-related injuries, and thereby, enhancing community resilience to extreme weather phenomena.