Exploring the Use of Deep-Learning to Aid Global Earthquake Monitoring at the NEIC

Date: 4/26/2019

Time: 06:00 PM

Room: Grand Ballroom

Recent research exploring the application of deep learning to seismic problems has demonstrated its vast potential. Here we explore the use of neural networks to predict seismic source characteristics from waveform data to aid in local-to-global real-time earthquake monitoring at the NEIC. In a simple framework, seismic arrival detections made from standard picking methods (e.g., STA/LTA) can be fed to trained neural network models to predict seismic phase type, event distance, and potentially a variety of other source-specific characteristics. We leverage the Advanced National Seismic System comprehensive catalog (ComCAT) to compile a dataset of millions of phase arrivals from the National Earthquake Information Center’s (NEIC) Preliminary Determination of Epicenters Bulletin, with associated phase labels (Pg, Pn, P, Sg, Sn, S, and Noise), event distances, magnitudes, and azimuths. We then extract waveforms from the Incorporated Research Institutions for Seismology Data Management Center, and use this data set to train models to classify various source characteristics. This catalog encompasses a wide range of station-event distances, event magnitudes, event types, and station types. Given a single three-component waveform, our models are successful at classifying P, S, and noise and are able to discern local from teleseismic arrivals, but are inaccurate at estimating intermediate source distances (~5-30 degrees). We discuss the implementation of these classifiers into the NEIC’s real-time systems, including the implementation of these source estimates into our operational associator GLASS3. We find that implementations as simple as removing STA/LTA picks classified with a high probability of being noise can reduce spurious detections from our associator by ~25%.

Presenting Author: William L. Yeck

Authors