Explainable and statistically validated machine learning for earthquake alert-level classification using seismic intensity features

Abstract

Earthquake alert-level classification can support rapid decision-making by transforming available seismic descriptors into actionable warning categories. However, many machine learning studies in earthquake-related applications report high predictive performance on limited datasets without sufficient statistical validation, which may lead to over-optimistic conclusions. This study presents an explainable and statistically validated machine learning framework for multi-class earthquake alert-level classification using a seismic dataset consisting of 1,300 records and five input variables: magnitude, focal depth, Community Decimal Intensity (CDI), Modified Mercalli Intensity (MMI), and significance (SIG). An Optuna-tuned XGBoost model was developed and compared with conventional machine learning classifiers, including Random Forest, standard XGBoost, K-Nearest Neighbors, Logistic Regression, and Support Vector Machine. To improve the reliability of the evaluation, the experimental protocol was strengthened using stratified repeated cross-validation, macro-averaged performance metrics, bootstrap-based confidence intervals, ablation analysis, and feature-set sensitivity experiments. The optimized XGBoost model achieved strong classification performance; however, the results also indicate that intensity-related variables, particularly MMI, CDI, and SIG, contribute substantially to the predictive outcome. Therefore, the proposed framework should be interpreted as an alert-level classification and decision-support model based on available seismic descriptors rather than as a prospective earthquake prediction system. SHAP-based explainability analysis confirmed that the model decisions were mainly driven by physically meaningful intensity-related features. Overall, the study provides a transparent benchmark framework for earthquake alert classification while explicitly addressing the limitations associated with small datasets, synthetic oversampling, and possible target-related feature dependence.