11:00 AM - 1:00 PM
[HDS10-P02] The use of S-net data for tsunami inundation forecasting using machine learning
Keywords:Tsunami inundation forecast, S-net, Machine learning
The Seafloor Observation Network for Earthquakes and Tsunamis along the Japan Trench (S-net) has been established (Aoi et al., 2020: https://doi.org/10.1186/s40623-020-01250-x; Mulia & Satake, 2021: https://doi.org/10.1186/s40623-021-01368-6). We proposed a machine learning-based approach to directly use tsunami data recorded at S-net stations for forecasting tsunami inundation simultaneously at seven coastal cities stretching ~100 km along the southern Sanriku coast: Sanriku, Ofunato, Rikuzentakata, Kesennuma, Motoyoshi, Minamisanriku, and Oppa. We applied a machine learning technique based on artificial neural networks that is expected to exploit the full potential of S-net in real-time tsunami forecasting as illustrated in Figure 1. The direct use of offshore observations can increase the forecast lead time and preclude the uncertainties typically associated with a tsunami source estimate required by the conventional modeling approach.
Since tsunami occurrence is infrequent, our neural networks model was trained on a multitude of precalculated physics-based model results. Using a simple statistical analysis to determine the appropriate number of scenarios, we considered 3060 megathrust events along ~1100 km of the Japan Trench with magnitude ranging from Mw 8.0 to 9.1. Additionally, we included 33 tsunamigenic earthquakes (Mw 7.0–8.7) in the outer rise referring to a study by Baba et al. (2020: https://doi.org/10.1029/2020JB020060). Our machine learning method built upon a fully connected neural networks consists of: (i) an input layer with 150 neurons representing observed tsunami properties at S-net, (ii) two hidden layers each with the same number of neurons as the input layer, (iii) an output layer of more than 200,000 neurons indicating the inundation grid points at the designated locations. A rectified linear unit (ReLU) activation function was used in the hidden layers with a 20% dropout rate. As the flow depths value is always positive, we also found that implementing the ReLU activation function in the output layer helped suppress erroneous inundation patterns. In the training stage we implemented the Adam optimization algorithm with the He normal initialization and a batch size of 20. A mean squared error commonly applied to regression tasks was used as the loss function.
We tested the model against 480 unseen dataset and achieved root mean square errors of flow depths of 0.37, 0.35 and 0.34 m using data obtained within 10, 15, and 20 min after the earthquake origin time, respectively. These results were derived from maximum amplitudes of tsunami pressure waveforms at all S-net stations within the specified windows considered as inputs for the algorithm. Furthermore, we also applied the model to two megathrust earthquakes of the 2011 Tohoku-oki (Mw 9.0) and the 1896 Meiji Sanriku (Mw 8.1) events, and one outer-rise earthquake of the 1933 Showa Sanriku (Mw 8.5) event. From comparisons with observed tsunami inundation heights, we found that the proposed method can produce comparable accuracy to the physics-based model. The main advantage of our method lies in the computing time that took only 0.05 sec, while the conventional physics-based model required approximately 30 min using a standard computer.
Since tsunami occurrence is infrequent, our neural networks model was trained on a multitude of precalculated physics-based model results. Using a simple statistical analysis to determine the appropriate number of scenarios, we considered 3060 megathrust events along ~1100 km of the Japan Trench with magnitude ranging from Mw 8.0 to 9.1. Additionally, we included 33 tsunamigenic earthquakes (Mw 7.0–8.7) in the outer rise referring to a study by Baba et al. (2020: https://doi.org/10.1029/2020JB020060). Our machine learning method built upon a fully connected neural networks consists of: (i) an input layer with 150 neurons representing observed tsunami properties at S-net, (ii) two hidden layers each with the same number of neurons as the input layer, (iii) an output layer of more than 200,000 neurons indicating the inundation grid points at the designated locations. A rectified linear unit (ReLU) activation function was used in the hidden layers with a 20% dropout rate. As the flow depths value is always positive, we also found that implementing the ReLU activation function in the output layer helped suppress erroneous inundation patterns. In the training stage we implemented the Adam optimization algorithm with the He normal initialization and a batch size of 20. A mean squared error commonly applied to regression tasks was used as the loss function.
We tested the model against 480 unseen dataset and achieved root mean square errors of flow depths of 0.37, 0.35 and 0.34 m using data obtained within 10, 15, and 20 min after the earthquake origin time, respectively. These results were derived from maximum amplitudes of tsunami pressure waveforms at all S-net stations within the specified windows considered as inputs for the algorithm. Furthermore, we also applied the model to two megathrust earthquakes of the 2011 Tohoku-oki (Mw 9.0) and the 1896 Meiji Sanriku (Mw 8.1) events, and one outer-rise earthquake of the 1933 Showa Sanriku (Mw 8.5) event. From comparisons with observed tsunami inundation heights, we found that the proposed method can produce comparable accuracy to the physics-based model. The main advantage of our method lies in the computing time that took only 0.05 sec, while the conventional physics-based model required approximately 30 min using a standard computer.