Hybrid CART Model for Adaptive Earthquake Risk Forecasting Under Uncertainty

The challenges in adaptive earthquake and seismic risk assessment underpin the exploration of issues, models and methodologies in Adaptive Earthquake Risk Forecasting, wherein statistical techniques utilise probability quantification to estimate the likelihood of earthquakes within defined spatial, temporal and magnitude parameters, necessitating that each location in a study area be defined by a time-varying conditional intensity. Key building attributes, such as construction materials, number of floors and age, were examined to assess the potential for earthquake-induced damage. Machine learning models were employed, utilising oversampling techniques, specifically Random Oversampling and SMOTE, to address class imbalance and improve predictive accuracy. The findings demonstrate that the CART model, particularly when integrated with SMOTE oversampling, outperformed other methods, attaining a macro F1-score of 95.88. In comparison, Random Oversampling also yielded significant enhancements, achieving a macro F1-score of 92.67 relative to the imbalanced baseline. This study elucidates the comparative strengths and weaknesses of various machine learning methodologies for predicting earthquake damage, underscoring the significance of feature selection and data balancing techniques. It provides critical insights for policymakers and engineers aiming to enhance earthquake preparedness and foster resilient infrastructure design through adaptive data-driven seismic risk forecasting.