Seismic Intelligence: Machine Learning Models for Earthquake Magnitude Estimation

Abstract

Earthquakes pose serious risks to human life, infrastructure, and economic stability across the globe. This experiment explores the use of machine learning (ML) to improve the accuracy of earthquake magnitude prediction. A dataset from the U.S. Geological Survey (USGS), containing over 8,000 records with attributes such as time, location, depth, and magnitude type, was used for model development. The data underwent preprocessing to handle missing values, remove anomalies, and eliminate irrelevant features. Feature engineering techniques, including one-hot encoding and normalization, were applied to enhance model performance. Several regression models were evaluated, including Ridge, Lasso, SVR, Random Forest, Gradient Boosting, XGBoost, LightGBM, CatBoost, and ElasticNet. Among these, a stacked ensemble model combining CatBoost, LightGBM, and Random Forest delivered the most accurate results. It achieved a mean squared error (MSE) of 0.0976, a coefficient of determination (R²) of 0.9392, and a normalized root mean squared error (NRMSE) of 4.6%. The findings demonstrate that ensemble learning significantly improves earthquake magnitude estimation. This approach offers valuable insights for seismic risk assessment and supports the development of more effective disaster preparedness strategies.