← Back to Sessions

Effective u.s. Event Classification Through Model Ensembling

Description: 

Event type discrimination models built with deep learning have proven to be highly effective for identifying known nuisance signals in earthquake catalogs at local scales. We know that deep learning models can also often outperform traditional methods in local-regional discrimination in low signal-to-noise cases. However, models trained in one geographic region on this task typically do not generalize well to new areas. Transfer learning, taking a model trained on data from one area and retraining it on data from a new area is one way to achieve high performance across regions when catalogs from new regions exist. Instead of transfer learning for each new region, in this work we utilize historic phase data available through the NEIC for non-earthquake events, as well as non-earthquake events from the high density Transportable Array, and regional event catalogs across the US to assess the viability of building ensemble models from catalog subsets to achieve high performance on known non-earthquake events and decision abstention when a model is likely to make poor decisions on new signals.

Session: Opportunities and Challenges for Machine Learning Applications in Seismology [Poster]

Type: Poster

Date: 4/19/2023

Presentation Time: 08:00 AM (local time)

Presenting Author: Lisa Linville

Student Presenter: No

Invited Presentation: 

Authors