Using Machine Learning to Estimate Source Location Early Earthquake Warning

Abstract

Early warning systems for earthquakes can mitigate their destructive potential by spreading information about the quake's magnitude and location long before destructive waves reach populated areas. Source-location estimations in these systems need to be timely and accurate for them to be useful. This study presents a novel approach for enhancing the precision and speed of seismic early warning using machine learning techniques. Timely warnings may be delayed due to the precision but slowness of traditional seismic techniques for calculating earthquake sites. The purpose of the random forest (RF) model for fast earthquake localization is to aid in the quick decision making required by earthquake early warning (EEW) systems. This approach takes use of the P-wave arrival times recorded by the first five stations to record an earthquake and calculates the variations in these timings with regard to the first station. In order to determine the epicenter, the RF model categorizes these differences in Pwave arrival timings and station locations. The model is used to train and validate the proposed method using a Japanese earthquake dataset. The RF model is quite accurate in predicting earthquake epicenters, with a Mean Absolute Error (MAE) of just 2.88 kilometers. Additionally, the suggested RF model may learn from as little as 10% of the information and as little as three recording stations while still producing usable results (MAE5 km) in most cases. This novel algorithm provides a robust and flexible method for predicting the location of EEW sources in real-time.