9:15 AM - 9:30 AM
[SCG52-02] Automatic seismic arrival-time picking of seismogram sampled at 250 Hz using Deep Learning
Keywords:deep learning, arrival-time picking, Manten network
Accurate seismic arrival-times are required to detect accurate earthquake location and to estimate a velocity structure. In recent years, the number of observation data has been increasing. In particular, We have installed high dense seismic station networks (Manten network) with 250 Hz sampling in Kinki, San-in and Nagano (Miura et al., 2010; Iio., 2010; Iio et al., 2017). The amount of data used for analysis has been increasing as seismic stations have been increasing, but the workload has also been increasing since arrival-time picking is a very time-consuming task.
Automatic arrival-time picking algorithms that can detect arrival-time in a shorter time than humans have been proposed. Conventional automatic arrival-time picking algorithms (Allen., 1978; Sleemen & van Eck., 1999) has poor accuracy, but recently, automatic arrival-time picking algorithms using deep learning (Ross et al., 2018; Zhu and Beroza, 2018; Hara et al., 2020) has been proposed and reported that models made by automatic arrival-time picking algorithms using deep learning are comparable to human accuracy.
In this study, in order to reduce the workload, we made an automatic arrival-time picking model. An algorithm used to make the model is PhaseNet (Zhu and Beroza., 2018). It can output probability values of P-wave, S-wave, and Noise (The Noise probability is calculated by Prob.(Noise) = 1− Prob.(P-wave) – Prob.(S-wave)) for each sampling point of input waveform. Currently, the automatic arrival-time picking model made by PhaseNet is available (https://github.com/wayneweiqiang/PhaseNet), but that model used 100 Hz sampling waveforms for training. Therefore, it is necessary to make the model that is suitable for seismic waveforms obtained with 250 Hz sampling.
For the training of PhaseNet, we used waveforms of M-0.9 to 4.7 recorded by the Manten network in San-in during the period of March to October 2015, whose arrival-times were detected by experts. Although this data set and data set used in Hara et al.(2020) are partially the same, the accuracy of the arrival-time used in this study is better than data used in Hara et al.(2020) due to the redetection of experts. The inputs to PhaseNet are three-component waveforms, and the available model was trained with UD, NS, and EW as input. In this study, we trained PhaseNet for the following three cases, taking into account the fact that when we have detected S-wave arrival-times of seismic waveforms obtained by Manten network, we have usually used SH components.
Case 1 : Model trained by UD, EW and NS
Case 2 : Model trained by UD, SH wave and 0 as missing value
Case 3 : Model trained by UD, SH wave and SH wave (copy)
The accuracies of the three models were evaluated, and the automatic arrival-time picking model of the 250 Hz sampling waveform was made. In this presentation, the accuracy evaluation of the models and the results by the models will be presented.
Automatic arrival-time picking algorithms that can detect arrival-time in a shorter time than humans have been proposed. Conventional automatic arrival-time picking algorithms (Allen., 1978; Sleemen & van Eck., 1999) has poor accuracy, but recently, automatic arrival-time picking algorithms using deep learning (Ross et al., 2018; Zhu and Beroza, 2018; Hara et al., 2020) has been proposed and reported that models made by automatic arrival-time picking algorithms using deep learning are comparable to human accuracy.
In this study, in order to reduce the workload, we made an automatic arrival-time picking model. An algorithm used to make the model is PhaseNet (Zhu and Beroza., 2018). It can output probability values of P-wave, S-wave, and Noise (The Noise probability is calculated by Prob.(Noise) = 1− Prob.(P-wave) – Prob.(S-wave)) for each sampling point of input waveform. Currently, the automatic arrival-time picking model made by PhaseNet is available (https://github.com/wayneweiqiang/PhaseNet), but that model used 100 Hz sampling waveforms for training. Therefore, it is necessary to make the model that is suitable for seismic waveforms obtained with 250 Hz sampling.
For the training of PhaseNet, we used waveforms of M-0.9 to 4.7 recorded by the Manten network in San-in during the period of March to October 2015, whose arrival-times were detected by experts. Although this data set and data set used in Hara et al.(2020) are partially the same, the accuracy of the arrival-time used in this study is better than data used in Hara et al.(2020) due to the redetection of experts. The inputs to PhaseNet are three-component waveforms, and the available model was trained with UD, NS, and EW as input. In this study, we trained PhaseNet for the following three cases, taking into account the fact that when we have detected S-wave arrival-times of seismic waveforms obtained by Manten network, we have usually used SH components.
Case 1 : Model trained by UD, EW and NS
Case 2 : Model trained by UD, SH wave and 0 as missing value
Case 3 : Model trained by UD, SH wave and SH wave (copy)
The accuracies of the three models were evaluated, and the automatic arrival-time picking model of the 250 Hz sampling waveform was made. In this presentation, the accuracy evaluation of the models and the results by the models will be presented.