← Back to Gallery

Automated Real-time Earthquake Energy Discriminator of Deep Earthquakes: A Comparison of Conventional and Ml Methods

Rapid and robust identification of deep earthquakes is useful in the application of more accurate real-time analysis, location and warning, particularly at teleseismic distances where real-time estimates of depth can differ from reviewed calculations by up to tens of kilometers. In Barama and Newman (2018), we developed a method using the first-derivative of the per-station energy time series of earthquakes to identify distinct double-peaks of energy associated with the direct-P phase followed by the energy of the depth phases (pP and sP). This was a promising result from automatic processing of initial energy pulses without any additional processing of the waveforms and allowed Barama and Newman (2021) to apply machine Learning (ML) to deep earthquake detection using a Convolution Neural Network (CNN) trained on both physical features of the energy time series (prominence and peak density) as well as the original waveform. Initial results showed improved results on utilizing the time series and interestingly the peak-density per event time series over the first derivative per-station energy, despite the peak density curves having significantly less training data. In this work we continue testing and complete comparison of the conventional and new ML methods for rapid depth determinations with the inclusion of the smoothed and stacked energy rate determinations per event. Using over 2000 earthquakes (> 70km depth) that occurred between 1989-2019 with moment magnitude greater than 5.5 from the Reviewed International Seismological Centre (ISC) bulletin, we calculated the per-station energy flux (of the P-wave group energy) in the frequency domain. We set a threshold for deep event detection based on time differential between energy peaks in the smoothed energy rate determinations as well as training several CNN models to identify the usefulness of the derivative products over the event time series. We hope to implement these results in the real-time energy determinations operating at Georgia Tech (http://geophysics.eas.gatech.edu/anewman/research/RTerg/).

Session: Machine Learning Techniques for Sparse Regional and Teleseismic Monitoring [Poster]

Type: Poster

Room: Evergreen Ballroom

Date: 4/21/2022

Presentation Time: 08:00 AM Pacific

Presenting Author: Louisa Barama

Student Presenter: Yes

Additional Authors