← Back to Sessions

Using Deep Learning to Identify Small Magnitude Earthquakes in 32TB of Continuous Seismic Data from the Pacific Northwest

Session: Earthquake Science, Hazards and Policy in Cascadia I

Type: Oral

Date: 4/20/2021

Presentation Time: 02:30 PM Pacific

Description: 

The Cascadia Subduction Zone (CSZ) is known to host both great earthquakes and very small magnitude low-frequency earthquakes. Despite this, there is a remarkable dearth of megathrust earthquakes with magnitudes between these two endmembers making Cascadia one of the least seismic subduction zones in the world. The identification of small magnitude earthquakes in the region could improve knowledge of diagnostic properties of the CSZ and controls on seismogenesis. We develop a machine learning framework that will be used to identify small magnitude earthquakes in 32TB of continuous seismic data from 42 seismic networks that operated in the Pacifc Northwest between 2005 and 2020. We train a U-shaped convolutional neural network (CNN) to pick earthquake body waves using three-component waveforms containing phase picks from the Pacific Northwest Seismic Network (PNSN) and the Northern California Earthquake Data Center (NCEDC). We explore several different network sizes, target functions, and variations in architecture. Our preferred network has a peak accuracy of 95% when identifying signal vs. noise in the testing data. Additionally, 97% of both P- and S-wave picks were made within 10 samples (0.1 seconds) of the true pick. For phase association, we generate synthetic earthquake catalogs with locations, magnitudes, and recurrence intervals representative of seismicity in the Pacific Northwest. We compute travel times using regional velocity models, attenuation from ground motion prediction equations, and false and missed picks by removing true and adding spurious picks to the datasets and use the result to train a long short-term memory network (LSTM). The LSTM performs both classification, to determine whether a given phase pick is associated with an earthquake, and regression to estimate hypocentral location and origin time. The LSTM is over 99% accurate at classifying true and false picks and predicts location and timing with low uncertainty on the training data. As a test, we apply the model to the July 2019 M4.6 Monroe, WA earthquake and aftershock sequence.

Presenting Author: Amanda M. Thomas

Student Presenter: No

Authors