Convolutional Neural Network for Seismic Phase Picking, Performance Demonstration in the Absence of Extensive Training Data

Date: 4/24/2019

Time: 06:00 PM

Room: Grand Ballroom

We present a Convolutional Neural Network (CNN) for classifying seismic phase onsets over local seismic networks, where the CNN is trained over a relatively small dataset for deep-learning purposes. The training catalogue consists of 411 events located throughout northern Chile. CNN based classifiers have recently achieved unprecedented success in seismic phase picking, outperforming traditional autopicking methods and achieving results comparable to the picks of an expert seismologist; these previous studies utilised extensive catalogues (~millions of examples) during the training process. We now investigate the limiting case of supervised learning-based methods such as the CNN approach, where extensive training data are not available. Our results show that in the absence of extensive training data, with appropriate regularisation constraints, the CNN approach to seismic phase picking still demonstrates exceptional performance, outperforming traditional methods in seismic phase classification. The CNN method is compared against an optimised STA/LTA autopicker applied to the same region. We perform a further test again comparing the trained CNN against the STA/LTA approach in picking phases for a separate catalogue of events throughout northern Chile. Based on station travel-time residuals, the CNN outperforms the STA/LTA approach, achieving a location residual distribution close to the manual picks of an expert seismologist. These results further corroborate the potential of supervised-learning based methods in solving the problem of seismic phase classification.

Presenting Author: Andreas Rietbrock

Authors