Forecasting strong subsequent earthquakes in Japan using an improved version of NESTORE machine learning algorithm

Abstract

In this study, the advanced machine learning algorithm NESTORE (Next STrOng Related Earthquake) was applied to the Japan Meteorological Agency catalog (1973–2024). It calculates the probability that the aftershocks will reach or exceed a magnitude equal to the magnitude of the mainshock minus one and classifies the clusters as type A or type B, depending on whether this condition is met or not. It has been shown useful in the tests in Italy, western Slovenia, Greece, and California. Due to Japan’s high and complex seismic activity, new algorithms were developed to complement NESTORE: a hybrid cluster identification method, which uses both ETAS-based stochastic declustering and deterministic graph-based selection, and REPENESE (RElevant features, class imbalance PErcentage, NEighbour detection, SElection), an algorithm for detecting outliers in skewed class distributions, which takes in account if one class has a larger number of samples with respect to the other (class imbalance).

Trained with data from 1973 to 2004 (7 type A and 43 type B clusters) and tested from 2005 to 2023 (4 type A and 27 type B clusters), the method correctly forecasted 75% of A clusters and 96% of B clusters, achieving a precision of 0.75 and an accuracy of 0.94 six hours after the mainshock. It accurately classified the 2011 Tōhoku event cluster. Near-real-time forecasting was applied to the sequence after the April 17, 2024 M6.6 earthquake in Shikoku, correctly classifying it as a “Type B cluster”. These results highlight the potential for the forecasting of strong aftershocks in regions with high seismicity and class imbalance, as evidenced by the high recall, precision and accuracy values achieved in the test phase.