MACHINE LEARNING FOR FAST AND RELIABLE SOURCELOCATION ESTIMATION IN EARTHQUAKE EARLY WARNING

Abstract

This project presents a machine learning-based approach for rapid and accurate earthquake source-location estimation to enhance the efficiency of Earthquake Early Warning (EEW) systems. A Random Forest (RF) model is developed to predict earthquake epicenters using differential P-wave arrival times recorded at the first five seismic stations relative to a reference station. These arrival time differences, combined with the corresponding station coordinates, serve as input features to the RF model. The model is trained and evaluated using an earthquake catalog from Japan. Experimental results show that the RF model achieves a high location prediction accuracy with a Mean Absolute Error (MAE) of 2.88 km. Remarkably, the model maintains reliable performance even when trained on just 10% of the dataset or using data from as few as three seismic stations (MAE < 5 km). The proposed system is efficient, generalizable, and robust, making it a valuable addition to real-time EEW systems for fast and dependable earthquake source localization.