← Back to Sessions

Feature-based Magnitude Estimates for Small, Nearby Earthquakes in the Yellowstone Volcanic Region

Description: 

Conventional methods of magnitude estimation for small, locally recorded events either rely on measurements of the maximum peak-to-peak amplitude to compute a local magnitude (M_L) or measurements of waveform decay to compute a coda duration magnitude. While at least one of these methods is generally able to produce an event magnitude in an automated framework, small events occurring as part of an earthquake swarm often pose a challenge. These events can occur rapidly and close by, causing overlapped event waveforms that make it challenging to accurately estimate the required magnitude parameters. This is particularly evident in Yellowstone, where ~50% of all seismicity occurs as part of a swarm. Additionally, only a subset of seismometers used to process events in the Yellowstone region have the station corrections needed to compute M_L. This further limits the ability to produce M_L values for very small events (< ~1.5). The need for another approach to estimate magnitude is magnified for large, deep-learning enhanced catalogs.

We introduce a machine learning method that uses features derived from short-duration waveforms capturing individual phase arrivals and event source parameters to predict the catalog network M_L. We train one support vector machine (SVM) for each station-phase pair, resulting in 34 models using P features and 17 using S features. Producing a model for each station allows the SVMs to efficiently learn individual station corrections. We ensemble predictions from multiple models into a network-averaged magnitude. For each phase, we initially examine 48-potential features and narrow the final number used in each model to ~9 by applying feature selection techniques. Using these features, initial results produce an R² value of ~0.9 on the test sets when averaging the model predictions. This method will improve our ability to produce M_L values for Yellowstone swarm events and lower the M_L magnitude of completeness. While not directly transferrable, this approach may be beneficial to other seismic network operators who routinely process earthquake swarms.

Session: Advances in Operational and Research Analysis of Earthquake Swarms [Poster Session]

Type: Poster

Date: 5/3/2024

Presentation Time: 08:00 AM (local time)

Presenting Author: Alysha

Student Presenter: Yes

Invited Presentation: 

Authors