Predicting earthquake magnitude is not an easy but an important activity in seismology, whichenhances disaster risk evaluation and risk reduction measures. In this research, a spatio-temporalregression model is created to predict the magnitude of major earthquakes across the world basedon a comprehensive history of previous earthquakes (18262026) of 102,799 earthquakes.Primary predictors are the spatial codification (latitude, longitude, depth), temporal codification(year, month, day, hour, weekday) of the periodics of a season and of state variance between dayand night, auxiliary seismic parameters (RMS, azimuthal gap), as well as a heuristicallydetermined classification of tectonic zones. Random Forest, lightgboost, xgboost, as well ascatboost were four ensemble machine learning models that were trained on a chronological 80/20Train-test split to ensure temporal integrity. Random Forest model had the best results with aMean Absolute Error (MAE) of 0.2377, root mean square (RMSE) error of 0.3472 andcoefficient of determination (R 2 ) of 0.2635 on the out of sample recent data. The analysis of theimportance of features revealed that the depth, geographic position, and long-term temporalcharacteristics took the leading positions, which is in line with the established seismologicalpatterns. Although the findings indicate the usefulness of ensemble techniques in derivingsignificant signals of heterogeneous historical data, the moderate predictive performanceindicates that there are still systemic constraints associated with observation biases in historicaldata and lack of explicit geophysical constraints. The article ushers a reproducible reference levelof ML-based magnitude estimation, highlighting the importance of domain-driven featureengineering to predictive seismology.
Agha Wafa Abbas (Wed,) studied this question.