4:30 PM - 4:45 PM
[SSS09-15] Application of seismic interferometry-based Green's Function for real-time forecasts of long-period ground motions
Keywords:Real-time forecast, Seismic interferometry, Long-period ground motions
1. Introduction
We applied seismic interferometry-based Green's functions (GFs) for real-time forecasts of long-period (LP) ground motion (4-10 s in period) by convolving the GFs between two points. Previously, GFs have been obtained from observed waveforms (Nagashima et al., 2008; Kurahashi et al., 2014) or wave propagation simulations (Yoshimoto and Takemura, 2018). Recently, Viens et al. (2017), Denolle et al. (2018), and others have shown the effectiveness of seismic interferometry-based estimation. The merits of this method are the improved S/N ratio of the LP component, the ability to obtain GFs even in areas with few earthquake observation data, and that it is free from uncertainties of the velocity structure models used in simulations. On the other hand, the amplitude of such GFs varies depending on the spatial distribution of ambient noise sources and the calculation methods, resulting in relative amplitudes. Therefore, following Viens and Denolle (2019), we compare the forecasted and observed waveforms of a reference earthquake, and calibrate the amplitude of the GFs to the absolute level.
2. Data and Methods
Data from Hi-net, K-NET, and KiK-net of NIED were used to calculate the GFs of Z components. Considering large earthquakes along the Japan Trench, Hi-net Kita-Ibaraki 2 (KI2H) was selected as the input point and four Hi-net stations located in the Kanto and Nobi Plain (Ichihara: ICHH, Tomioka: YFTH, Minami-Chita: MCTH, Yokkaichi: YOKH) were selected as the forecast points (Fig. 1). Following Viens and Denolle (2019), instrumental response removal, decimating to 4 Hz, and detrending were applied to the waveform data divided into 30 min segments, and the waveforms of the input and forecast points were deconvolved. By stacking the deconvolution data for one year from January to December 2021, the GFs were obtained.
Next, we calibrated the amplitude of the GFs to the absolute level using an M 7.0 earthquake in Hamadori, Fukushima. Waveforms at the four forecasted points are calculated by convolving the GFs with the input waveform at K-NET IBR019 (adjacent to Hi-net KI2H). Following Viens et al. (2014), the sum of Fourier amplitudes of 4-10 s period was compared between forecasted and observed waveforms to obtain the amplitude calibration factor. Different factors (10.1 and 4.73) were used for the forecast points in the Kanto Plain (YFTH and ICHH) and the Nobi Plain (MCTH and YOKH), respectively, considering the effect of the uneven distribution of ambient noise sources.
3. Results and Discussion
In the GFs, we considered sections with velocities faster than 6 km/s as noise and replaced them with 0. The band-pass filtered (4-10 s) GFs from the input point to each forecast point are shown in Fig. 2.
The forecasting method was tested against an M7.4 earthquake which occurred off Fukushima in 2016. Fig. 3 and Fig. 4 show the forecasted waveforms and velocity response spectra (h=5%) at each forecast point, respectively. For comparison, the observed ones are overlaid with red lines.
The correlation coefficients between the envelopes of the forecasted and observed waveforms were 0.49, 0.63, 0.65, and 0.64 for ICHH, YFTH, MCTH, and YOKH, respectively. The agreement was particularly low at ICHH, with a delay of about 1 minute in the arrival time of the large amplitude. Difference is seen in the propagation speeds of the large amplitude phase of the GFs between the Nobi and Kanto Plain, therefore the seismic interferometry-based GFs need to be re-examined. For the response spectra, the forecasted/observed ratios for the 4-10 s period were 0.40, 0.40, 0.47, and 0.64 for ICHH, YFTH, MCTH, and YOKH, respectively. The forecasts underestimated by about a half of the observations, and further study is needed.
We plan to estimate the GFs using the horizontal component, which has a larger amplitude of LP ground motions, and to realize forecasts using multiple stations as the input points.
We applied seismic interferometry-based Green's functions (GFs) for real-time forecasts of long-period (LP) ground motion (4-10 s in period) by convolving the GFs between two points. Previously, GFs have been obtained from observed waveforms (Nagashima et al., 2008; Kurahashi et al., 2014) or wave propagation simulations (Yoshimoto and Takemura, 2018). Recently, Viens et al. (2017), Denolle et al. (2018), and others have shown the effectiveness of seismic interferometry-based estimation. The merits of this method are the improved S/N ratio of the LP component, the ability to obtain GFs even in areas with few earthquake observation data, and that it is free from uncertainties of the velocity structure models used in simulations. On the other hand, the amplitude of such GFs varies depending on the spatial distribution of ambient noise sources and the calculation methods, resulting in relative amplitudes. Therefore, following Viens and Denolle (2019), we compare the forecasted and observed waveforms of a reference earthquake, and calibrate the amplitude of the GFs to the absolute level.
2. Data and Methods
Data from Hi-net, K-NET, and KiK-net of NIED were used to calculate the GFs of Z components. Considering large earthquakes along the Japan Trench, Hi-net Kita-Ibaraki 2 (KI2H) was selected as the input point and four Hi-net stations located in the Kanto and Nobi Plain (Ichihara: ICHH, Tomioka: YFTH, Minami-Chita: MCTH, Yokkaichi: YOKH) were selected as the forecast points (Fig. 1). Following Viens and Denolle (2019), instrumental response removal, decimating to 4 Hz, and detrending were applied to the waveform data divided into 30 min segments, and the waveforms of the input and forecast points were deconvolved. By stacking the deconvolution data for one year from January to December 2021, the GFs were obtained.
Next, we calibrated the amplitude of the GFs to the absolute level using an M 7.0 earthquake in Hamadori, Fukushima. Waveforms at the four forecasted points are calculated by convolving the GFs with the input waveform at K-NET IBR019 (adjacent to Hi-net KI2H). Following Viens et al. (2014), the sum of Fourier amplitudes of 4-10 s period was compared between forecasted and observed waveforms to obtain the amplitude calibration factor. Different factors (10.1 and 4.73) were used for the forecast points in the Kanto Plain (YFTH and ICHH) and the Nobi Plain (MCTH and YOKH), respectively, considering the effect of the uneven distribution of ambient noise sources.
3. Results and Discussion
In the GFs, we considered sections with velocities faster than 6 km/s as noise and replaced them with 0. The band-pass filtered (4-10 s) GFs from the input point to each forecast point are shown in Fig. 2.
The forecasting method was tested against an M7.4 earthquake which occurred off Fukushima in 2016. Fig. 3 and Fig. 4 show the forecasted waveforms and velocity response spectra (h=5%) at each forecast point, respectively. For comparison, the observed ones are overlaid with red lines.
The correlation coefficients between the envelopes of the forecasted and observed waveforms were 0.49, 0.63, 0.65, and 0.64 for ICHH, YFTH, MCTH, and YOKH, respectively. The agreement was particularly low at ICHH, with a delay of about 1 minute in the arrival time of the large amplitude. Difference is seen in the propagation speeds of the large amplitude phase of the GFs between the Nobi and Kanto Plain, therefore the seismic interferometry-based GFs need to be re-examined. For the response spectra, the forecasted/observed ratios for the 4-10 s period were 0.40, 0.40, 0.47, and 0.64 for ICHH, YFTH, MCTH, and YOKH, respectively. The forecasts underestimated by about a half of the observations, and further study is needed.
We plan to estimate the GFs using the horizontal component, which has a larger amplitude of LP ground motions, and to realize forecasts using multiple stations as the input points.