Applying Machine Learning Salves to Network Build-Out 'Growing Pains' at the Pacific Northwest Seismic Network
Description:
The Pacific Northwest Seismic Network (PNSN) is the authoritative network for near-real-time seismic event monitoring and catalog development in Washington and Oregon. The build-out of the PNSN and surrounding networks over the past decade resulted in a roughly eight-fold increase in the number of seismic channels being continuously analyzed by the PNSN. Presently, PNSN’s Earthworm system only analyzes vertical component data for automated phase detection and only considers P-wave labels when classifying and associating phase detections. All S-wave characterization is conducted by seismic analysts. The increase in data volume per event, combined with the PNSN’s relatively low magnitude threshold (2.95+) for rapid review (less than 10 min post event origin time) produces a substantial increase in workload for PNSN staff.
Machine-learning (ML) seismic analysis tools demonstrate promising accuracy and performance benchmarks for combined phase detection and classification tasks, particularly PhaseNet and EQTransformer. These tools may help soothe growing pains from the PNSN build-out; however, most pretrained models and their applications focus on broadband instrument records, and their performance benchmarks are often based on curated datasets. Data analyzed by the PNSN come from broadband, strong motion, short period, and ocean bottom instruments, have diverse sampling rates and site qualities, and are subject to imperfect data continuity. To bridge this gap, we developed a streaming data workflow that couples SeisBench-hosted ML models to Earthworm’s memory rings, the performance of which we test in two pilot studies. The first study assesses the accuracy of picks from pretrained ML models against analyst picks for 7749 PNSN catalog earthquakes. The second study assesses the computational performance of our workflow on modern PNSN data traffic. Our initial results indicate that these models can reasonably replicate analyst characterization of P- and S-wave arrivals and our workflow can attain faster-than-real-time processing, at scale, on modest computational resources.
Session: Network Seismology: Recent Developments, Challenges and Lessons Learned [Poster Session]
Type: Poster
Date: 5/1/2024
Presentation Time: 08:00 AM (local time)
Presenting Author: Nathan
Student Presenter: No
Invited Presentation:
Authors
Nathan Stevens Presenting Author Corresponding Author ntsteven@uw.edu University of Washington |
Renate Hartog jrhartog@uw.edu University of Washington |
Yiyu Ni niyiyu@uw.edu University of Washington |
Alex Hutko ahutko@uw.edu University of Washington |
Marine Denolle mdenolle@uw.edu University of Washington |
Amy Wright akwright@uw.deu University of Washington |
Jason De Cristofaro jdecristofaro@usgs.gov U.S. Geological Survey |
|
|
Applying Machine Learning Salves to Network Build-Out 'Growing Pains' at the Pacific Northwest Seismic Network
Session
Network Seismology: Recent Developments, Challenges and Lessons Learned