11:15 AM - 11:30 AM
[HDS10-03] Tsunami data assimilation in source region: Separating coseismic deformation and tsunami height from pressure gauge records
Keywords:Data assimilation, Tsunami forecast, Ocean bottom pressure gauge
Tsunami data assimilation uses offshore tsunami data to reconstruct the wavefield and make a forecast (Maeda et al., 2015). It does not require initial source information as long as the true tsunami height are correctly observed at densely distributed stations. Offshore Bottom Pressure Gauge (OBPG) measures water pressure and converts it to tsunami height for data assimilation. Nevertheless, the pressure records of OBPGs in source region contain an offset due to coseismic deformation beneath the sensor, which poses challenge for tsunami data assimilation.
We propose a modified tsunami data assimilation algorithm that separates coseismic deformation and tsunami height from OBPG records. The algorithm contains two steps: 1) invert the coseismic deformation only beneath the sensors; and 2) conduct tsunami data assimilation with corrected OBPG records. In the first step, we estimated the initial sea-surface displacement above the sensor and assumed that the initial coseismic deformation is consistent with the initial sea-surface displacement. We note that the inversion is performed to help the data assimilation in the source region works well; thus, we do not consider the fault geometry or sea-surface displacement not above the sensor. This limited focus enables us to perform the inversion very quickly. In the second step, we assimilated OBPG records corrected by coseismic deformation. The optimal interpolation method was adopted for data assimilation. The tsunami waveforms at coastal stations were forecasted after assimilation time window.
We tested our algorithm with hypothetical earthquakes generated by stochastic rupture simulations (Mori et al., 2017). We assumed a tsunami generated by an M 8.0 earthquake in the Japan Trench. We simulated the records incorporating the coseismic deformation at OBPGs of S-net. Traditional tsunami data assimilation method failed to reconstruct the tsunami wavefield. Our proposed algorithm successfully inverted the coseismic deformation beneath the sensor. Then, we forecasted waveforms at 14 stations in Hokkaido, Tohoku, and Kanto areas and compared with forward simulations. A 15 min time window makes a forecast with 86% accuracy, whereas a 30 min time window makes a forecast with 96% accuracy.
We propose a modified tsunami data assimilation algorithm that separates coseismic deformation and tsunami height from OBPG records. The algorithm contains two steps: 1) invert the coseismic deformation only beneath the sensors; and 2) conduct tsunami data assimilation with corrected OBPG records. In the first step, we estimated the initial sea-surface displacement above the sensor and assumed that the initial coseismic deformation is consistent with the initial sea-surface displacement. We note that the inversion is performed to help the data assimilation in the source region works well; thus, we do not consider the fault geometry or sea-surface displacement not above the sensor. This limited focus enables us to perform the inversion very quickly. In the second step, we assimilated OBPG records corrected by coseismic deformation. The optimal interpolation method was adopted for data assimilation. The tsunami waveforms at coastal stations were forecasted after assimilation time window.
We tested our algorithm with hypothetical earthquakes generated by stochastic rupture simulations (Mori et al., 2017). We assumed a tsunami generated by an M 8.0 earthquake in the Japan Trench. We simulated the records incorporating the coseismic deformation at OBPGs of S-net. Traditional tsunami data assimilation method failed to reconstruct the tsunami wavefield. Our proposed algorithm successfully inverted the coseismic deformation beneath the sensor. Then, we forecasted waveforms at 14 stations in Hokkaido, Tohoku, and Kanto areas and compared with forward simulations. A 15 min time window makes a forecast with 86% accuracy, whereas a 30 min time window makes a forecast with 96% accuracy.