Comparative Performance Assessments of Machine-Learning Methods for Artificial Seismic Sources Discrimination

Abstract

Mankind is vulnerable to artificial seismic sources and accompanying explosions’ consequences. Recently, seismicity catalog contamination is among the main problems faced by seismologists. Since identifying artificial seismic sources is the first and always challenging stage, it is imperative to develop an automated control system that will discriminate tectonic from non-tectonic events. Detection and removal of the artificial seismic sources have become urgent. Early treatments and cleaning of contaminated seismicity catalogs are crucial to assist in accurate seismic hazard identification and enhance the planning of future urban developments. With the advancement of machine learning (ML) techniques, artificial seismic source detection accuracy has been improved. Today, there are different kinds of methods, ML techniques, and diverse processes like knowledge discovery are developed for discriminating artificial seismic sources and earthquakes. ML techniques offer various probabilistic and statistical methods that allow intelligent systems to learn from reoccurring experiences to detect and identify patterns from a dataset. This study aims to build an automated system that is able to detect the existence of artificial seismic sources in seismicity catalogs. More concretely, we classify seismic activity reports into two classes using classical and ensemble ML algorithms. Classical seismicity parameters or features are supplied to linear and nonlinear ML classifiers. The proposed scheme based on the four features (Latitude, Longitude, depth, and Magnitude) can enhance the performance. To assure the enhanced performance, we have examined the proposed scheme by both the accuracy of each model, ROC curves, Precision-Recall, and Calibration. The obtained results prove that the ensemble learning algorithms exhibit better results compared to other classical ML algorithms by having 98.14% testing accuracy.