Parametric Testing of EQTransformer’s Performance against a High-Quality, Manually Picked Catalog for Reliable and Accurate Seismic Phase Picking
Olivia Pita‐Sllim, C. J. Chamberlain, John Townend, Emily Warren‐Smith
Abstract
Abstract This study evaluates EQTransformer, a deep learning model, for earthquake detection and phase picking using seismic data from the Southern Alps, New Zealand. Using a robust, independent dataset containing more than 85,000 manual picks from 13 stations spanning almost nine years, we assess EQTransformer’s performance and limitations in a practical application scenario. We investigate key parameters such as overlap and probability threshold and their influences on detection consistency and false positives, respectively. EQTransformer’s probability outputs show a limited correlation with pick accuracy, emphasizing the need for careful interpretation. Our analysis of illustrative signals from three seismic networks highlights challenges of consistently picking first arrivals when reflected or refracted phases are present. We find that an overlap length of 55 s balances detection consistency and computational efficiency, and that a probability threshold of 0.1 balances detection rate and false positives. Our study thus offers insights into EQTransformer’s capabilities and limitations, highlighting the importance of parameter selection for optimal results.