Earthquake forecast by imbalance machine learning using geophysical predictors

Aug 4, 2026

Uploaded by Asher

Earthquake forecasting is treated as a binary machine-learning task on an imbalanced dataset across five regions of Georgia. Training data comprise geophysical records collected between 2017 and 2021, including variations of geomagnetic field components, seismic activity, water level in deep boreholes, and tidal signals. This work adds a new predictor: weighted seismic activity for the previous five days, and triples the dataset length. Because earthquakes with M > 3.5 are rare, imbalance mitigation is performed using Matthews’ correlation coefficient and F1 score.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages11

Reading Time28 min

AccessFree to download

Introduction

Machine learning for complex seismic systems

Data and predictor updates

The network of observations, preparation of data bases and methodology of analysis

The network

ANNALS OF GEOPHYSICS, 66, 6, SE636, 2023: doi: 10.4401/ag-8946 OPEN ACCESS Earthquake forecast by imbalance machine learning using geophysical predictors Tengiz Kiria, Tamaz Chelidze, George Melikadze, Tamar Jimsheladze, Gennady Kobzev M. Nodia Institute of Geophysics, Tbilisi State University, Tbilisi, Georgia Article history: received February 1, 2023; accepted October 10, 2023 Abstract In the present paper we consider the earthquake forecast as a binary problem of machine learning on the imbalanced database applied to five regions of Georgia. For the training we used geophysical data base collected in 2017‑2021, namely, variations of statistical characteristics of geomagnetic field components, seismic activity, water level in deep boreholes and tides. In this version a new predictor – the weighted seismic activity for previous 5 days ‑ – is added compared to the predictors’list used in previous papers. Besides, the length of the used database is increased 3 times compared to the earlier results. As in the database the earthquakes of M > 3.5 are rare, the number of negative cases is large (there are many days without EQs of M > 3.5), meaning that there is a strong imbalance between positive and negative cases of the order of 1:20; we apply the specific methodology Matthews’ correlation coefficient (MCC) and F1 score to avoid the strong imbalance effect. Keywords: Earthquake forecast; Water level in wells; Geomagnetic variations; Micro‑seismicity; Machine learning on imbalanced data; Receiver operating characteristics 1. Introduction The seismic process is without doubt a complex process: according to the accepted definition, complexity appears in systems, which are composed of many components interacting nonlinearly. The complexity theory (nonlinear dynamics) approach, requires the detail knowledge of the real process, which allow to describe it by the system of differential equations. Complexity analysis allow revealing the existence of long‑term correlations in the temporal, spatial and energy distributions in dynamical systems such as seismicity using mathematical models of the process [Chelidze et al., 2018] . On the other hand, last years appear modern machine learning (ML) approach, which is concerned with developing algorithms. ML methods improve their performance with increasing the volume of input information [Li, 2020]. This approach gained increasing attention in solving the problems, where it is impossible to formulate exact mathematical models but on the other hand there are a lot of real data measurements. The ML allow to create data‑driven approach to understanding and forecast of behavior of many complex system without constructing the exact analytical model – in contrast to the complexity theory. Tengiz Kiria et al. Last years, ML approach give a lot of promising results in forecasting both laboratory and natural earthquakes [Rouet‑Leduc et al., 2017; Rouet‑Leduc et al., 2018; Ren et al., 2020; Johnson et al., 2021]. It is interesting to note that one of the first publications devoted to application of ML for the EQ forecast belong to Chelidze et al. [1995] . In the paper authors applied the method of Generalized Portrait (now Support Vector Machine, SVM) suggested by Vapnik and Chervonenkis [1974], Vapnik [1984] to forecast the Caucasian EQs of magnitude 5 and more. In the earlier paper [Chelidze et al, 2020] the problem of EQ forecast in the Caucasus region was considered using one‑year only WLand geomagnetic observations. In the present paper a new predictor –the weighted seismic activity for the previous 5 days – is added compared to the predictors used in the previous paper. The parameter reflects the previous seismic activity in the chosen area for a given time interval and isused for improving the model, predicting future relatively strong (M > 3.5) events. Besides, in the previous paper, due to scarcity of data, we apply ML to any of chosen four regions and used fifth regio

FAQs

How is earthquake forecasting formulated in this study?

It is formulated as a binary machine-learning problem using an imbalanced dataset.

What geophysical predictors are used for training?

The study uses geomagnetic variations, seismic activity, water level in deep boreholes, and tides, plus a new predictor: weighted seismic activity over the previous five days.

How does the paper address the imbalance between earthquake and non-earthquake cases?

It applies Matthews’ correlation coefficient (MCC) and F1 score to reduce the impact of strong class imbalance (about 1:20 for M > 3.5).

Aug 4, 2026

Uploaded by Asher