10:45 AM - 11:15 AM
[SCG51-07] End-to-end tsunami inundation forecasting from observation data using convolutional neural networks
★Invited Papers
Keywords:Tsunami, Real-time forecasting, Machine learning
Fine-scale tsunami inundation forecasting can be an effective risk information supporting evacuation behaviours when the forecasting could be made within a short period of the lead time and promptly delivered. Since the 2011 Tohoku tsunami, intensive investigations on real-time tsunami inundation forecasting have been made in parallel with the enhancements of the tsunami observation networks. Previous studies have proposed several methods such as a real-time tsunami inundation forecasting using supercomputers (e.g., Oishi et al., 2016) with a source inversion (e.g., Tsushima et al., 2014) or a wave field estimation using data assimilation techniques (e.g., Maeda et al., 2015) and a forecasting based on a pre-computed database (e.g., Yamamoto et al., 2016); however, challenges such as difficulties in accurate source estimation within a short period and requirements of relatively large computational infrastructure for the forecasting remain. In this study, we propose an end-to-end tsunami inundation forecasting from observation data using convolutional neural networks (CNNs), which requires small computational resources for inundation forecasting.
In this study, we verified the proposed method with a tsunami inundation waveform forecasting task near the Arahama elementary school (AES) in tsunami scenarios off Tohoku. The fault geometry in Fujii et al. (2013) was used, and various earthquake scenarios were generated by randomly changing the slip amount of sub-faults. Initial water level was calculated with Okada (1992), and the tsunami propagation and inundation were simulated with TUNAMI-N2 (Goto et al., 1997) to generate the training data for CNN. We prepared a total of 12,000 tsunami simulations and split the simulations into 10,000, 1,000, 1,000 cases for the training, validation, and test sets, respectively. The tsunami and geodetic observation points covering the forecasting site were considered as inputs, and we assumed that tsunami signals and ground heights could be obtained at tsunami and geodetic observation points, respectively. CNN in this study consists of 12 convolutional layers and 3 fully-connected layers. Since the consistent CNN architecture was employed, the number of feature values within the CNN increases according to the observation periods. The inputs for the CNN are tsunami signals and the constant vectors of ground heights with the same length, and a MSE between the ground truth waveform and forecasting for 2 hours is considered as the loss function for training.
We first confirmed the capability of the CNN using 1,000 synthetic tsunami data. When only tsunami signals were used as inputs, the forecasting accuracy was improved with longer observation periods; however, equivalent accuracy with only 30 min tsunami observations could be achieved with 5 min tsunami and geodetic observations. An additional analysis found that this improved accuracy corresponded to the improved accuracy in the initial ground height estimations. In this test, the CNN forecasted the maximum tsunami height with the errors of several percent and required only several seconds with 1 node CPU for forecasting including I/O. Additional sensitivity test, which removed certain input signals and investigated changes in forecasting, revealed that the CNN automatically learned the important observation points following tsunami propagations and that the CNN required only a partial of observation points near the forecasting site, demonstrating the feature of the proposed method that does not require source estimations.
We then verified the method using real observation data in the 2011 Tsunami using the CNN trained on the same 10,000 synthetic tsunamis but with the observation setting at that time. In addition to the inundation forecasting near the AES, a complementary forecasting at Sendai New Port (SNP) where had time series data was also conducted. In this test, the forecasting performance of the CNN is observed with changing the observation periods from 10 to 40 minutes. The CNN with 35 min or more observations provided reasonable forecasting both at AES and SNP that are consistent with the tsunami height traces. The accuracy in tsunami arrival time had sufficient quality to provide information for prompting evacuations; however, the forecasted arrival timing was several minutes faster than the actual arrivals. The simplified tsunami data generation process which did not consider effects of rupture propagations of faults can be considered as a cause of this error. CNN trained on much more various tsunami scenarios has the potential to provide more accurate tsunami inundation forecasting with sufficient forecasting speed.
In this study, we verified the proposed method with a tsunami inundation waveform forecasting task near the Arahama elementary school (AES) in tsunami scenarios off Tohoku. The fault geometry in Fujii et al. (2013) was used, and various earthquake scenarios were generated by randomly changing the slip amount of sub-faults. Initial water level was calculated with Okada (1992), and the tsunami propagation and inundation were simulated with TUNAMI-N2 (Goto et al., 1997) to generate the training data for CNN. We prepared a total of 12,000 tsunami simulations and split the simulations into 10,000, 1,000, 1,000 cases for the training, validation, and test sets, respectively. The tsunami and geodetic observation points covering the forecasting site were considered as inputs, and we assumed that tsunami signals and ground heights could be obtained at tsunami and geodetic observation points, respectively. CNN in this study consists of 12 convolutional layers and 3 fully-connected layers. Since the consistent CNN architecture was employed, the number of feature values within the CNN increases according to the observation periods. The inputs for the CNN are tsunami signals and the constant vectors of ground heights with the same length, and a MSE between the ground truth waveform and forecasting for 2 hours is considered as the loss function for training.
We first confirmed the capability of the CNN using 1,000 synthetic tsunami data. When only tsunami signals were used as inputs, the forecasting accuracy was improved with longer observation periods; however, equivalent accuracy with only 30 min tsunami observations could be achieved with 5 min tsunami and geodetic observations. An additional analysis found that this improved accuracy corresponded to the improved accuracy in the initial ground height estimations. In this test, the CNN forecasted the maximum tsunami height with the errors of several percent and required only several seconds with 1 node CPU for forecasting including I/O. Additional sensitivity test, which removed certain input signals and investigated changes in forecasting, revealed that the CNN automatically learned the important observation points following tsunami propagations and that the CNN required only a partial of observation points near the forecasting site, demonstrating the feature of the proposed method that does not require source estimations.
We then verified the method using real observation data in the 2011 Tsunami using the CNN trained on the same 10,000 synthetic tsunamis but with the observation setting at that time. In addition to the inundation forecasting near the AES, a complementary forecasting at Sendai New Port (SNP) where had time series data was also conducted. In this test, the forecasting performance of the CNN is observed with changing the observation periods from 10 to 40 minutes. The CNN with 35 min or more observations provided reasonable forecasting both at AES and SNP that are consistent with the tsunami height traces. The accuracy in tsunami arrival time had sufficient quality to provide information for prompting evacuations; however, the forecasted arrival timing was several minutes faster than the actual arrivals. The simplified tsunami data generation process which did not consider effects of rupture propagations of faults can be considered as a cause of this error. CNN trained on much more various tsunami scenarios has the potential to provide more accurate tsunami inundation forecasting with sufficient forecasting speed.