← Back to Sessions

An Explainable Phase-picking Model That Imitates Human Workflow

Description: 

Deep learning is becoming a standard approach for automating the task of picking seismic phase arrival times. Several studies have applied pre-trained models to build event catalogs, and a growing number of seismic network operators are adopting the technology in their monitoring workflows; however, the published deep learning models are often not explainable (and thus not always reliable), which can limit their applicability to mission critical applications where the tolerance to false positive and false negative detections may be low. One of the main factors that limits the explainability of these models is the design itself – the ‘segmentation’ approach that relies on an arbitrary kernel. We demonstrate how the segmentation approach limits model explainability and introduce a multi-step automated phase-picking process that better imitates a human analyst. Instead of relying on a single model for the task, we separate the problem into initial screening (a conventional trigger), arrival detection (analogous to an analyst checking the presence of an arrival from a zoomed-out seismogram), and arrival picking (analogous to an analyst making the pick from a zoomed-in seismogram). The arrival detection and picking rely on two separate neural networks. The detector network is calibrated to output a probability that scales with accuracy, and the picker network is designed to output a probability mass function. We test the effect of feeding seismograms filtered at different frequency bands that analysts commonly choose from. We show how this approach can make the results more explainable, the models more maintainable, and the process more intervenable by analysts.

Session: Network Seismology: Recent Developments, Challenges and Lessons Learned - II

Type: Oral

Date: 4/15/2025

Presentation Time: 11:00 AM (local time)

Presenting Author: Yongsoo

Student Presenter: No

Invited Presentation: 

Poster Number:

Authors