Forms of machine learning are explored as a means of discriminating between earthquake and explosion sources. To start, key waveform features of time series are extracted, allowing a reduction in dimensionality and representation that avoids full-waveform size inputs while still facilitating the automatic recognition of earthquake and explosion sources. This reduction in parameters still allows for an effective source classification methodology, lowering the consequential risk of source misidentification. We then explore the viability of using machine learning for the accurate identification of specific source type exclusively from simulated data: modeling entirely with seismic data waveforms is not possible because of the extreme imbalance between the relative numbers of earthquakes and explosions. Various preprocessing techniques are combined with feature extraction methods to improve the performance of the classifier algorithms. For the particular generated synthetic dataset of earthquakes and explosions, very large numbers of experimental runs show promising results of machine learning discrimination.
Presenting Author: Ghassan Aleqabi
Authors
Ghassan Aleqabi
Presenting Author Corresponding Author
ghassan@seismo.wustl.edu
Washington University in St. Louis, St Louis, Missouri, United States
Presenting Author
Corresponding Author
Michael Wysession
mwysession@gmail.com
Washington University in St. Louis, St. Louis, Missouri, United States