Earthquake Prediction Using Machine Learning: Improved Magnitude Estimation with LSTM over Random Forest

Abstract

Assigning estimated magnitudes for earthquakes and reliably forecasting future events is a critical part of minimizing the overall hazard potential and enhancing planning and preparedness efforts. I this study, a hybrid machine learning system is proposed for predicting earthquakes by combining aspects of the Random Forest (RF) and Long Short - Term Memory (LSTM) models. While the RF is useful for classification of seismic events and discovering spatial patterns, RF models are not useful for making reliable predictions of event magnitudes. The drawback of RF was solved by introducing the LSTM model which is particularly useful for dealing with sequential and time series data. The dataset includes the common seismic parameters of latitude, longitude, depth, and historical magnitudes. The performance of the LSTM model is superior to both the RF for predicting the event magnitude. The LSTM model yields lower and significantly reduced error rates. The experimental results demonstrate that deep learning models, such as LSTM, exceed traditional Machine learning systems for temporal predictions for geophysical data. The findings presented in this work provide an encouraging insight into the enhancement of early warning systems for the potential for improved magnitude predictability.