Automated Classification of Seismic Signals With Continuous Random Forests

Abstract

Seismic events, such as earthquakes, volcanic activity, and explosions, can have severe con sequences on infrastructure and human life. Therefore, accurate detection and identification of these events are crucial for effective disaster management and emergency response. In recent years, there has been a growing interest in using machine learning algorithms for automated classification of seismic signals. Automated classification can significantly con tribute to seismology by aiding in the identification of earth motions, implementing early warning systems, and assessing seismic hazards.This thesis is centered around utilizing ma chine learning techniques to classify earth movements by analyzing seismic datasets. The dataset used for training is obtained from [1]. Previous works in automated classification have faced challenges, including limited datasets, datasets with limited event classes, and non-vital parameters as attributes. To address these limitations, we have developed a Python code that takes seismic signals as input and generates a dataset consisting of 52 vital pa rameters of seismic signals as output.Among the machine learning models employed in this work, the Continuous Random Forests (CRF) algorithm is particularly suitable for handling time-series data like seismic signals. CRF utilizes a sliding window approach to capture temporal dependencies and can effectively learn complex patterns and features indicative of specific types of seismic events. In our experiments, CRF has demonstrated promising results in accurately distinguishing between different types of seismic events, including big earthquakes, mild earthquakes, and noise.To address the issue of imbalanced datasets, we have applied data balancing techniques, including random sampling, to ensure equal rep resentation of all event classes during the training phase. The performance of the machine learning model has been evaluated before and after the implementation of data balancing techniques, revealing the importance and effectiveness of addressing dataset imbalances in seismic classification tasks.Furthermore, a comparison between CRF and Support Vector Machines (SVM) has been conducted in this work. The evaluation results indicate that the CRF algorithm achieves an accuracy of 99 percent, outperforming the SVM algorithm, which achieves an accuracy of 97 percent. The findings highlight the suitability of auto mated classification methods for effectively categorizing various types of ground motion, such as tele-seismic events, regional events, local earthquakes, explosions, and ambient noise. Overall, this thesis contributes to the advancement of automated seismic event clas sification using machine learning techniques. The results demonstrate the potential of these methods in enhancing seismic analysis, improving early warning systems, and supporting seismic hazard assessment efforts.