← Back to Sessions

Towards Building a Machine Learning Based Automatic Detection System for Surface Events in the Pacific Northwest

Description: 

Systematic monitoring of mass movements, such as landslides, avalanches, rockfalls, and debris flows, is vital for mitigating risks in places where such events threaten to impact people or infrastructure. The Pacific Northwest Seismic Network (PNSN) detects a subset of surface events as a byproduct of the routine detection and location workflow for tectonic and volcanic earthquakes.

In this study, we explore a vast range of feature space and model architecture designs for event classification. We found that the random forest algorithm along with a specific set of features is the best model for its generalization over the PNW data sets. We trained our model on a curated set of waveforms from Ni et al. (2023) for earthquakes and other seismic sources recorded by the PNSN over the past 20 years and developed a workflow that can be deployed on continuous data. We then used our model to automatically detect and characterize events at Mount Rainier, identifying surface events in a month of continuous data from three seismic stations. The model successfully identified all surface events manually picked by PNSN analysts for the study period with a high classification probability exceeding 0.9. Additionally, it detected approximately ten times more surface events with a high probability of 0.8, with a subsequent manual review revealing their resemblance to waveforms of verified surface events or volcanic events. In one attempt to validate our new detections, we computed maximum cross-correlation coefficients of detected event waveforms with ground-truthed surface events in the Earthscope Exotic Seismic Event Catalog for the Mount Rainier region and found that many of the detected surface event waveforms exhibited high correlation coefficient values exceeding 0.9. We also report their approximate location using phase picking of the onset of the emergent waveforms and use a grid-search method for determining locations using a uniform velocity model. Ongoing efforts to generate better training catalogs and to verify new detections involve incorporating additional stations, DAS, eyewitness accounts, and infrasound data.

Session: Detecting, Characterizing and Monitoring Mass Movements - II

Type: Oral

Date: 5/2/2024

Presentation Time: 10:30 AM (local time)

Presenting Author: Akash

Student Presenter: Yes

Invited Presentation: 

Authors