Challenges and Lessons Learned from Applying Machine Learning Models to Seismic Monitoring Across Diverse Tectonic Settings
Description:
Machine learning (ML) is widely used for improving earthquake catalog production and, more recently, real-time seismic monitoring. However, implementing deep learning models in different tectonic settings poses unique challenges due to variations in seismicity rate and density, characteristics of seismic signals and noise, and network coverage. We highlight and address these challenges for two distinct regions: the Northeastern U.S. (NEUS), a stable continental region with sparse station coverage and low seismicity rate, and Axial Seamount, a submarine volcano with high event rates and complex ocean noise. In NEUS, the performance of pre-trained deep learning models such as PhaseNet is degraded due to regional-distance waveforms differing substantially from the local waveforms used in training. We therefore retrained PhaseNet using over two decades of regional analyst picks, and incorporated both time and frequency domain information to create PhaseNet-TF. Using the 2024 Mw4.8 Tewksbury, NJ earthquake sequence, we demonstrate significant improvements over the original PhaseNet model (F1-score increase from ~0.5 to ~0.9). At Axial Seamount, challenges arise from processing continuous OBS data with high event rates while efficiently distinguishing between earthquakes, various seismoacoustic signals (e.g., lava-water interaction, whales), and ocean noise. We designed and operate an ML-based real-time seismic monitoring workflow to detect, classify, and locate events with high precision (MLDD-RT). This framework tracks various source types, including eruption precursory events and impulsive events that track magma outflows. These studies show that, although deep learning models are intended to be generalizable, region-specific retraining is often necessary to adapt to specific tectonic environments with varying event and noise characteristics. We show the benefits of establishing best practices for model adaptation across diverse tectonic settings, as well as developing standardized datasets that better represent the full spectrum of seismic environments.
Session: Network Seismology: Recent Developments, Challenges and Lessons Learned [Poster]
Type: Poster
Date: 4/15/2025
Presentation Time: 08:00 AM (local time)
Presenting Author: Felix
Student Presenter: No
Invited Presentation:
Poster Number: 31
Authors
Kaiwen Wang Corresponding Author kw2988@ldeo.columbia.edu Columbia University |
Felix Waldhauser Presenting Author felixw@ldeo.columbia.edu Columbia University |
Haoyu Wang haoyu.wang@berkeley.edu University of California, Berkeley |
Weiqiang Zhu zhuwq0@gmail.com University of California, Berkeley |
|
|
|
|
|
Challenges and Lessons Learned from Applying Machine Learning Models to Seismic Monitoring Across Diverse Tectonic Settings
Session
Network Seismology: Recent Developments, Challenges and Lessons Learned