Using Convolutional Neural Network to Determine Time Window for Analyzing Local Shear-Wave Splitting Measurements

Abstract

The time window for analyzing local shear-wave splitting (SWS) phases significantly affects the quality of measurements, revealing a noteworthy domain influence. In this study, an approach using convolutional neural network (CNN) is applied to determine the end of time window (e), which has a similar idea of the phase-picking CNNs. The start of time window is 0.5 s before e. Our data set contains 803 human-labeled measurements, recorded from three stations located in Ridgecrest, California. These measurements are foreshocks and aftershocks of an M 7.1 earthquake on 6 July 2019. After 21 times shifting on each measurement, 90% of the data set is applied as the training data set, with the remaining 10% as the testing data set. The performance of CNN with the testing data set is compared with a nonmachine learning method, multiple filter automatic splitting technique (MFAST). The results reveal that the CNN yields more similar results with human-labeled outcomes than MFAST, as evidenced by lower absolute error and standard deviation for e, SWS time, the orientation of fast-wave polarization, and more consistent results on the map. The CNN also performs well when applied to data recorded by a station in Parkfield, California. This study shows the outstanding performance of CNN in picking the time window and the reliable automatic determination of this time window, and it is also a crucial step for future development of automatic ranking methodologies.