← Back to Sessions

Practical Uncertainty for Network Seismology with Machine Learning

Session: Network Seismology: Keeping the Network Running While Integrating New Technologies I

Type: Oral

Date: 4/22/2021

Presentation Time: 10:30 AM Pacific

Description: 

Real-time seismic monitoring networks strive to create an automated event bulletin by coherently aggregating network-wide single-station predictions, such as arrival time, amplitude, period, backazimuth and slowness. Because these predictions often vary in quality, coherent aggregation necessitates that each prediction be accompanied by some statistical measure of uncertainty, in order to threshold or weight each prediction. Traditional seismic monitoring pipelines utilize carefully-tuned physics-based statistics to estimate prediction uncertainty. However, these physics-based uncertainty measures can be laborious to tune, and are often only loosely covariant with the actual prediction error. Machine Learning models offer an advantage here, in that statistical measures of uncertainty are readily integrated into the model itself, and can be optimized at scale during training. In this work, we present two case studies for machine learning uncertainty based on two single-station predictions: arrival time estimation and backazimuth estimation. In each case, we detail simple code to estimate the prediction uncertainty (quantile regression and softmax probability, respectively) and demonstrate its statistical validity alongside the traditional method.

Presenting Author: Joshua T. Dickey

Student Presenter: No

Authors