← Back to Sessions

A Deep Learning Approach for Non-binarizing the Impulsive/Emergent Phase Labels

Description: 

The impulsiveness of an arrival is one of the metrics that are often labeled along when picking seismic phase arrival times. This can be considered as the slope of the seismogram at the arrival, which can theoretically (but not realistically) range between 0 (perfectly emergent or parallel to the time axis) and 1 (perfectly impulsive or perpendicular to the time axis). However, due to the broad frequency contents and the noise in seismograms, the slope cannot be measured with confidence in most cases. Hence, this continuous value has been labeled subjectively by choosing from a set of choices – typically either impulsive or emergent (i.e., binary). Here, we propose a simple approach for converting these binary labels to continuous values. We train neural networks for binary classification using the binary impulsive/emergent labels. We show that the two classes cannot be separated even if we optimize the hyperparameters of the models to maximize the classification accuracy, which demonstrates the measure of impulsiveness is indeed a continuous value. We also show that different neural networks output almost consistent classification scores with most samples having a standard deviation of less than 0.1. We further evaluate the approach by altering the impulsiveness and signal-to-noise ratio of an arrival and show that the changes in classification scores align with our intuition. This classification score (or the ‘impulsiveness’) can serve as a measure of difficulty for phase arrivals, which can be used as weights when training phase picking models or for various geophysical inversion problems that rely on arrival time measurements.

Session: Advances in Reliable Earthquake Source Parameter Estimation - III

Type: Oral

Date: 4/16/2025

Presentation Time: 02:00 PM (local time)

Presenting Author: Yongsoo

Student Presenter: No

Invited Presentation: 

Poster Number:

Authors