Applying deep learning to teleseismic phase detection and picking: PcP and PKiKP cases

TL;DR

This paper presents a deep learning scheme for fast and accurate detection and picking of teleseismic PcP and PKiKP phases, achieving high recognition rates and low error in a large dataset within a few hours.

Contribution

It introduces a three-step deep learning approach for teleseismic phase detection and picking, improving speed and accuracy over traditional methods.

Findings

Recognition rates of 92.15% for PcP and 94.13% for PKiKP phases.

Mean picking errors of 0.0742 s for PcP and 0.0636 s for PKiKP.

Processing time of approximately 2 hours for 7386 seismograms.

Abstract

The availability of a tremendous amount of seismic data demands seismological researchers to analyze seismic phases efficiently. Recently, deep learning algorithms exhibit a powerful capability of detecting and picking on P- and S-wave phases. However, it is still a challenge to process teleseismic phases fast and accurately. In this study, we detect and pick the PcP and PKiKP phases from a Hinet dataset with 7386 seismograms by applying a deep-learning-based scheme. The scheme consists of three steps: first, we prepare latent phase data, which is truncated from the whole seismogram with the theoretical arrival time; second, we identify and evaluate the latent phase via a convolutional neural network; third, we pick the first break of good or fair phase with a fully convolutional neural network. The detection result shows that the scheme recognizes 92.15% and 94.13% of PcP and PKiKP phases. The picking result has an absolute mean error of 0.0742 s and 0.0636 s for the PcP and PKiKP phases, respectively. The performance of the picking algorithm is compared with the traditional approach of STA/LTA. The scheme processes all 7386 seismograms approximately in 2 hours, especially only cost about five minutes on the last two steps.