Feature-Based Bayesian Inference for Seismic Event Monitoring
Date: 4/24/2019
Time: 09:00 AM
Room: Cascade I
Robustly detecting and locating relatively weak seismic events from noisy waveforms requires moving beyond pick-based methods and processing more information from the waveform. Further, because these events push the limits of our detection abilities, new methods must account for modeling and sensor uncertainties when making predictions about an event. This requirement motivates the use of Bayesian methods. An ideal method would assess the likelihood of a candidate event given the data based upon how well that event could predict the observed waveform data. This requires a model that predicts a seismic waveform at a monitoring station given an event hypothesis. Unfortunately, generative waveform models are computationally expensive and have many unquantified uncertainties. Methods like SIGVISA (Moore et al. 2017) have worked to bridge this gap by learning a generative model from historic data that does not require full simulations. This approach has been shown to significantly reduce location errors and improve event detection at the cost of needing to learn and then infer a complex generative signal model that can replicate the seismic waveform.
In contrast to predicting the waveform, we can design a Bayesian method that predicts waveform features. Using historic data or simulations, building models that predict features given a seismic event can be easier and more computationally tractable than direct waveform prediction while doing inference. Inspired by the WCEDS event detection and location method (Arrowsmith et al. 2016), the waveform features we consider are P and S wave arrival time and duration and the integrated power of the signal during the P, S, and pre-signal background intervals of the waveform. Our framework is general, and hence can use different features and/or transformed data, e.g. after STA/LTA processing. We demonstrate the feature-based framework and evaluate its performance on synthetic and real datasets for regional seismic event monitoring.
Presenting Author: Thomas A. Catanach
Authors
Thomas A Catanach tacatan@sandia.gov Sandia National Laboratories, Albuquerque, New Mexico, United States Presenting Author
Corresponding Author
|
Nathan J Downey njdowne@sandia.gov Sandia National Laboratories, Albuquerque, New Mexico, United States |
Christopher J Young cjyoung@sandia.gov Sandia National Laboratories, Albuquerque, New Mexico, United States |
Feature-Based Bayesian Inference for Seismic Event Monitoring
Category
From Drifting to Anchored: Advances in Improving Absolute Hypocenter Location Accuracy for Natural, Induced and Explosion Seismic Events