10:45 AM - 12:15 PM
[SCG53-P03] Estimation of wave propagation directions from a single observation station with a deep learning algorithm: toward a more sophisticated modeling of the ongoing wavefield
Keywords:Ground motion prediction, Earthquake early warning, Machine learning, Deep learning, Wave propagation
Earthquake early warning (EEW) algorithms that predict ground motions directly from observation of the ongoing wavefield have been proposed to improve the ground motion prediction performance for complex scenarios such as large earthquakes with non-negligible finite faults and multiple earthquakes that occur simultaneously (e.g., Hoshiba and Aoki, 2015; Kodera et al., 2018). This wavefield-based EEW approach is divided into two parts: modeling of the ongoing wavefield and prediction of the future wavefield. In the first part, existing algorithms construct the ongoing wavefield based only on the observation of real-time seismic intensities (Kunugi et al., 2013) at individual stations (i.e., observed wave amplitudes), although there are other physical quantities related to wave propagation such as wave direction and apparent velocity. The accuracy and timeliness of estimating the ongoing wavefield could be improved if observed wave-related quantities are incorporated in addition to wave amplitude. To this end, we developed a method that estimates wave directions from waveforms recorded at a single station with a deep learning technique.
Station Hitachinaka of KiK-net (IBRH18; surface) was selected as a target station for testing our algorithm. We used event waveforms that were observed from January 2010 to July 2022, whose peak real-time seismic intensity was 0.5 or more, and whose initial record time was 2 s before the theoretical P-arrival time or earlier (~1500 waveforms in total). The training data for the deep learning model were 2-s three-component accelerations, obtained by sliding a 2-s time window with a 0.2-s interval from 2 s before the theoretical P-onset to 10 s after the P-onset. We assumed that the ground truth of wave direction was equal to the azimuth from the target station to the epicenter. The wave direction θ was converted into (sinθ,cosθ) for the model training to accommodate the periodicity of angle. The deep learning model employed in this study was a network composed of GRU (Gated Recurrent Unit) and fully connected layers. The training dataset was divided into train, validation, and test with a ratio of 6:2:2.
Application of the deep learning model to the test dataset showed that the model determined wave directions with an accuracy of ±30° for ~50% and ~20% cases using input waveforms within 2 s and 10 s after the theoretical P-arrival, respectively. We also estimated wave directions for synthetic scenarios where two earthquakes occurred simultaneously and their waveforms were overlapped within several seconds. We found that the model could estimate both of the wave directions with a high accuracy if the wave amplitude of the second earthquake was larger than the first one. These results indicate that the proposed model captured waveform features successfully through training and could be useful for complex earthquake scenarios such as multiple occurrences of simultaneous earthquakes.
Acknowledgements: In this study, we used waveform records of KiK-net, maintained by NIED. This study was supported partially by JSPS KAKENHI grant number JP21K03689.
Station Hitachinaka of KiK-net (IBRH18; surface) was selected as a target station for testing our algorithm. We used event waveforms that were observed from January 2010 to July 2022, whose peak real-time seismic intensity was 0.5 or more, and whose initial record time was 2 s before the theoretical P-arrival time or earlier (~1500 waveforms in total). The training data for the deep learning model were 2-s three-component accelerations, obtained by sliding a 2-s time window with a 0.2-s interval from 2 s before the theoretical P-onset to 10 s after the P-onset. We assumed that the ground truth of wave direction was equal to the azimuth from the target station to the epicenter. The wave direction θ was converted into (sinθ,cosθ) for the model training to accommodate the periodicity of angle. The deep learning model employed in this study was a network composed of GRU (Gated Recurrent Unit) and fully connected layers. The training dataset was divided into train, validation, and test with a ratio of 6:2:2.
Application of the deep learning model to the test dataset showed that the model determined wave directions with an accuracy of ±30° for ~50% and ~20% cases using input waveforms within 2 s and 10 s after the theoretical P-arrival, respectively. We also estimated wave directions for synthetic scenarios where two earthquakes occurred simultaneously and their waveforms were overlapped within several seconds. We found that the model could estimate both of the wave directions with a high accuracy if the wave amplitude of the second earthquake was larger than the first one. These results indicate that the proposed model captured waveform features successfully through training and could be useful for complex earthquake scenarios such as multiple occurrences of simultaneous earthquakes.
Acknowledgements: In this study, we used waveform records of KiK-net, maintained by NIED. This study was supported partially by JSPS KAKENHI grant number JP21K03689.