Multi-class classification automated machine learning for predicting earthquakes using global geomagnetic field data

The challenging task of earthquake (EQ) prediction has recently gained significant attention, particularly with machine learning techniques. Geomagnetic field analysis has yielded promising results in identifying EQ precursors. However, the complexity of the data has made it difficult to create an accurate model for EQ prediction using this method. This study presents an automated machine learning (AutoML) approach capable of handling the complexity of geomagnetic data and selecting the most suitable model. A dataset containing 50 years of geomagnetic field data was collected, of which the measurements were taken in close proximity to M5.0+ EQs. The study demonstrated that sampling techniques can overcome the problem of an imbalanced dataset from EQ events. Through statistical analysis, important features were extracted and a multi-class classification model using geomagnetic data was created. The extracted features were the input for AutoML, an automatic algorithm selection that was measured by Bayesian Optimization algorithm to select the best performance model. The results indicate that the neural network model outperformed eight other classifiers, achieved an accuracy of 81.19%, F1-score of 80.51%, and a Matthews Correlation Coefficient (MCC) of 77.49%. It is concluded that the neural network multi-class classification model is capable of providing solutions to the challenges faced when using geomagnetic data for EQ prediction.