Picking Regional Seismic Phase Arrival Times with Deep Learning
Albert Leonardo Aguilar Suarez, Gregory C. Beroza
Abstract
Sparse instrumental coverage for much of the Earth requires working with regional seismic phases for effective seismic monitoring. Machine learning phase pickers to date have focused on local earthquake recordings. Here we present deep learning models designed and trained to be effective at picking the arrival times of earthquake phases at distances up to 20 degrees. We trained our models on the CREW dataset, which includes 1.6 million earthquake waveforms with over 3.2 million labeled arrivals on 5 minute long three component seismograms. We present models that accurately pick the first arriving P and S waves and models that pick and classify Pn, Pg, Sn, and Sg phase arrivals. We apply these models in a variety of settings and compare their performance to established machine learning models that were trained on local earthquake recordings. We demonstrate the abilities of our models by finding new earthquakes in the Gorda plate offshore northern California. Finally, we use our multiple phase picker to find new examples with secondary arrivals from our massive training dataset. The goal of this method is to improve automatic earthquake monitoring in regions of sparse instrumental coverage and seismicity in remote regions far from instrumentation.