Early detection of earthquake magnitude based on stacked ensemble model

Abstract

A new machine learning model, named, EEWPEnsembleStack has been developed for predicting the magnitude of the earthquake from a few seconds of recorded ground motion after the arrival of the P phase. The testing and training dataset consists of 2360 and 591 strong-motion records from central Japan recorded by the Kyoshin Network. Eight parameters that are well correlated with the magnitude have been used for training and testing of the model. Feature ablation study using several models shows that a minimum mean absolute error of 0.42 has been obtained for the case when the model has been trained by using all parameters rather than by a single parameter. The model ablation study indicates that among all individually trained single models, the minimum error has been obtained for a Decision Tree regression model. However, the error is minimized when all machine learning models have been together utilized in the EEWPEnsembleStack model for the training purposes. The EEWPEnsembleStack model has been used to predict a 6.3 magnitude earthquake by using its 21 records from various stations that lie within 50 to 150 km epicentral distance. The predicted magnitude from the developed model using weighted magnitude prediction is obtained as 6.4, which is close to the actual magnitude. The comparison of the predicted magnitude of this earthquake from the developed model with that predicted by using popular ${\tau }_c$ and $P_d$ methods clearly indicates the suitability of the developed machine learning model over other conventional models.