Earthquake Predictions Using Machine Learning Approaches

Abstract

Abstract: An effective and precise earthquake prediction technique could potentially save numerous lives and economies. Bearing this goal in mind, this thesis conducts a comparative study that explores the application of machine learning (ML) approaches in predicting earthquake characteristics. It utilizes the United States Geological Survey (USGS) dataset, which includes earthquakes of magnitude 3.5 and above from the last year (2023), and the Los Alamos National Laboratory (LANL) datasets for a comprehensive analysis. The USGS dataset is employed to predict the magnitude of earthquakes through regression analysis, while the LANL dataset focuses on forecasting the timing of forthcoming seismic events. A diverse array of machine learning algorithms, including Linear Regression (LR), Support Vector Regression (SVR), Random Forest Regressor (RFR), XGBoost Regressor (XGB), and advanced deep learning models such as Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), are meticulously evaluated for their predictive capabilities. The assessment of model performance leverages several evaluation metrics, namely Mean Absolute Error (MAE), Mean Squared Error (MSE), and R-squared (R²). Among the models evaluated, the RFR demonstrated outstanding performance on the USGS dataset, achieving an MAE of 0.167, an MSE of 0.055, and an impressive R² score of 0.729. This highlights its robust capability in modeling complex relationships within seismic magnitude data. In contrast, the GRU model excelled in the LANL dataset, achieving the best performance with an MAE of 1.777, an MSE of 4.192, and an R² score of 0.610. Future efforts should optimize feature engineering and selection processes by integrating diverse indicators such as seismic activity, historical patterns, and geological data to develop more sophisticated tools for earthquake preparedness and risk mitigation.