← Back to Gallery

Machine Learning Applications and Developments at the US Geological Survey's National Earthquake Information Center for Improved Regional-to-global Scale Monitoring

The U.S. Geological Survey’s (USGS) National Earthquake Information Center (NEIC) has a unique monitoring mission that spans local, regional and teleseismic scales. It fulfills its mission by processing waveform data streamed from an evolving inventory of thousands of globally distributed seismic stations. These stations represent a wide range of sensor types, inter-station distances and noise characteristics. To utilize these waveform data, we implement specialized algorithms to derive a rapidly generated and accurate earthquake catalog. These algorithms must generalize across a global range of tectonic settings, recording environments, network configurations, data sampling rates, station quality, etc. Given these circumstances, machine learning (ML) techniques are particularly advantageous in improving the global monitoring capabilities of the NEIC.

In 2021, the NEIC operationalized a suite of machine learning models designed to improve automatic event association and location capabilities. These models improve pick-timing, classify phases and broadly estimate source-station distance which improves associator/location processing. Since the initiation of this project, NEIC has expanded its efforts to produce a reviewed global training dataset that can be leveraged to create new globally generalized teleseismic ML models. We recently created the Machine Learning Asset Aggregation of the Preliminary Determination of Epicenters (MLAAPDE) dataset and software to provide configurable and updatable training datasets tailored for specific ML-driven tasks.

We present the current state of the NEIC operational machine learning applications and the MLAAPDE dataset. Furthermore, we discuss our ongoing development efforts, including improving pick-timing estimates using analysts’ specific models, and our efforts to create models that automatically indicate events that are not of response or cataloging interest to the NEIC. Lastly, we will highlight monitoring specific needs in ML applications, especially when considering their impact on real-time monitoring and the human driven review process.

Session: Machine Learning Techniques for Sparse Regional and Teleseismic Monitoring I

Type: Oral

Room: Regency E-G

Date: 4/21/2022

Presentation Time: 08:45 AM Pacific

Presenting Author: William L. Yeck

Student Presenter: No

Additional Authors