Evaluation of Machine Learning Techniques: Artificial Neural Network, Random Forest, and Support Vector Machine in Prediction of Alborz-Azerbaijan Earthquakes

Introduction: One of the most destructive environmental hazards is earthquakes. Therefore, predicting this hazard to reduce its consequences and improve crisis management is one of the most important goals for researchers. By using seismicity indices and applying machine learning techniques, researchers can reveal seismic behavior patterns in a region. These methods have proven especially effective at modeling the nonlinear behavior of seismic data and have thus become important tools for understanding natural phenomena.
 
Materials and methods: In the present study, a catalog of earthquakes from the Alborz-Azerbaijan seismotectonic province covering the period from January 1, 1995, to January 23, 2024, has been compiled. Earthquake magnitudes have been converted to torque magnitudes to homogenize the catalog. Subsequently, based on the temporal variations in the completeness magnitude, the threshold magnitude for the study in the Alborz-Azerbaijan seismotectonic province has been determined to ensure the necessary accuracy for analysis. Three machine learning methods—Artificial Neural Network, Random Forest, and Support Vector Machine—were selected to predict the time and magnitude of earthquakes. Recognizing that some machine learning methods require feature definition, nine representative indices of seismic behavior were estimated for the Alborz-Azerbaijan earthquake catalog to serve as input for the chosen methods. Following the implementation of these techniques, the estimation error rate was calculated and reported using four types of error metrics: F1 Score, Recall, Precision, and Accuracy.

Results and discussion: Machine learning in this study was conducted using 245 vectors formed by 9 indicators. These indices are stored in corresponding two-dimensional arrays, with each column representing a specific set of indices. Each data vector is associated with a binary label of 1 or 0; the label "1" indicates the occurrence of at least one earthquake with a magnitude equal to or greater than the moment magnitude threshold of 5.5, while the label "0" indicates the absence of seismic activity for earthquakes with magnitudes less than 5.5. In this research, 80% of the data vectors were used for model training, and 20% were used for testing. The findings, regarding the estimated true and false alarm error values for each of the machine learning techniques applied to the seismic data of Alborz-Azerbaijan, demonstrate the success of all three techniques in predicting events recorded in the seismic catalog of Alborz-Azerbaijan. Generally, an accuracy exceeding 95% was achieved for all three methods.

Conclusion: The findings indicate the success of these techniques in estimating the cycle of stress accumulation and release associated with seismic activity in the Alborz-Azerbaijan geotechnical province. The accuracy of all three methods shows only a small difference, reflecting the high performance of machine learning techniques. For the seismic data of Alborz-Azerbaijan, the Random Forest method exhibits slightly higher accuracy. The accuracy values obtained from the selected methods in this research suggest that the optimal machine learning method depends on the diversity and quantity of the data. In the context of natural hazard data, particularly geophysical hazards, the differences in success levels among machine learning methods are influenced by the tectonic and geological characteristics of the environment. Furthermore, the results of this study demonstrate that utilizing machine learning techniques for preparedness and mitigation of environmental consequences, as well as for earthquake crisis management, is promising.