9:00 AM - 9:15 AM
[SCG50-01] Retraining of Neuro picker based on JMA unified seismic catalog
Keywords:JMA catalog, deep learning, Neuro picker
Several travel-time reading models based on deep learning techniques (neuro pickers) have recently been developed and are publicly available (e.g., Zhu & Beroza, 2019; Mousavi et al., 2020). Those models are trained by large datasets, mainly from outside Japan. These are often used to develop seismic catalogs in regions other than the training data without retraining and have revealed seismicity in high resolution. It is, however, known that higher performance can be achieved by retraining or transfer learning using the dataset of the target observation network (e.g., Kim et al., 2023). For datasets with many users, such as the seismic network for the JMA unified catalog, the models trained on these data sets should be publicly available to promote related research. In this study, we retrained PhaseNet (Zhu & Beroza, 2019), the widely used neuro picker model, based on a travel-time-reading dataset for the JMA catalog and compared the resultant performance with the original model trained on California data by Zhu & Beroza (2019).
We prepared the training dataset from the arrival-time dataset for the JMA routine catalog between 2018 and 2020. Since September 1, 2020, arrival time data for S-net records have begun to be used for the cataloging process, and these data were included in the training data. The waveforms of S-net stations were used after rotation into NS, EW, and UD components based on the attitude information investigated by Takagi et al. (2019). Of the 10,546,390 observations obtained during this period, we selected 1,295,195 three-component waveforms that satisfy the following criteria: 1) both P and S arrival-time readings are available, 2) all of the three-component waveform records are available, and 3) manual checks were performed on both P and S waves. Of the 1,295,195 records, we used 80% for the training and the remaining 20% for the validation.
We applied the PhaseNet models trained on the JMA dataset and California dataset (i.e., the original model by Zhu and Beroza, 2019) to the JMA dataset obtained in 2021, which was not used for our training process. The obtained precision-recall curves clearly show that the retraining contributes to the performance improvement for JMA datasets.
The degree of improvement was limited for the Hi-net data but was significant for the S-net and DONET data. In addition, the residual distribution of travel times showed that the original model of Zhu and Beroza (2019) tended to read slightly delayed points compared to JMA measurements, while the retrained model reduced this bias
We prepared the training dataset from the arrival-time dataset for the JMA routine catalog between 2018 and 2020. Since September 1, 2020, arrival time data for S-net records have begun to be used for the cataloging process, and these data were included in the training data. The waveforms of S-net stations were used after rotation into NS, EW, and UD components based on the attitude information investigated by Takagi et al. (2019). Of the 10,546,390 observations obtained during this period, we selected 1,295,195 three-component waveforms that satisfy the following criteria: 1) both P and S arrival-time readings are available, 2) all of the three-component waveform records are available, and 3) manual checks were performed on both P and S waves. Of the 1,295,195 records, we used 80% for the training and the remaining 20% for the validation.
We applied the PhaseNet models trained on the JMA dataset and California dataset (i.e., the original model by Zhu and Beroza, 2019) to the JMA dataset obtained in 2021, which was not used for our training process. The obtained precision-recall curves clearly show that the retraining contributes to the performance improvement for JMA datasets.
The degree of improvement was limited for the Hi-net data but was significant for the S-net and DONET data. In addition, the residual distribution of travel times showed that the original model of Zhu and Beroza (2019) tended to read slightly delayed points compared to JMA measurements, while the retrained model reduced this bias