WITHDRAWN Earthquake Detection in Subduction Zones: Transfer Learning With Amphibious Data From the Alaska Amphibious Community Seismic Experiment
Description:
WITHDRAWN We train the machine-learning (ML) earthquake detection and picking algorithms EarthquakeTransformer (EQT) and PhaseNet for use with amphibious or offshore seismic networks in subduction zones, and we apply them to data from the Alaska margin. ML is revolutionizing earthquake detection, with trained models sometimes finding an order of magnitude more events than traditional methods. With better detection, an enhanced picture of small seismicity emerges, yielding better understanding of fault zone processes and seismic hazard. Yet, ML methods are under-utilized in high-hazard subduction zones. We contribute two new subduction-zone oriented training datasets generated from earthquakes recorded by the Alaska Amphibious Community Seismic Experiment (AACSE), 2018-2019, and we use these datasets to train EQT and PhaseNet for subduction zones. One dataset (~52k waveforms) contains ocean-bottom data from 66 sites crossing the trench, and the other (~123k waveforms) contains data from 129 neighboring land sites. We use these datasets to train EQT and PhaseNet, testing various dataset and waveform filtering approaches to generate best models for ocean-bottom and land data separately. We then apply the two models to continuous AACSE data to find more small seismicity.
We provide two important products for earthquake detection in high-hazard subduction zones: large on- and offshore training datasets, and picking models trained on them. These products provide a foundation for future subduction-zone-oriented dataset generation and training, important for SZ4D and similar initiatives. In addition, detection and picking improvement over pre-trained models suggests that transfer learning with known regional earthquakes may be important to customize ML models to regional variability. Last, the improved catalog is a snapshot of Alaska margin seismicity late in the seismic cycle, just prior to the 2020 M7.8 Simeonof, 2020 M7.6 Sand Point, and 2021 M8.2 Chignik sequence of earthquakes.
Session: Opportunities and Challenges for Machine Learning Applications in Seismology [Poster]
Type: Poster
Date: 4/19/2023
Presentation Time: 08:00 AM (local time)
Presenting Author: Grace Barcheck
Student Presenter: No
Invited Presentation:
Authors
Grace Barcheck Presenting Author Corresponding Author grace.barcheck@cornell.edu Cornell University |
Geoffrey Abers abers@cornell.edu Cornell University |
Emily Roland rolande2@wwu.edu Western Washington University |
Susan Schwartz syschwar@ucsc.edu University of California, Santa Cruz |
|
|
|
|
|
WITHDRAWN Earthquake Detection in Subduction Zones: Transfer Learning With Amphibious Data From the Alaska Amphibious Community Seismic Experiment
Category
Opportunities and Challenges for Machine Learning Applications in Seismology