11:00 AM - 1:00 PM
[SCG51-P04] Tsunami inundation prediction using seafloor pressure data by power-law regression and multilayer perceptron
Keywords:cluster analysis, neural network, tsunami
In Japan, tsunamis propagating offshore can be observed by seafloor pressure gauges and GPS wavemeters before they reach the coast. Using Green's law obtained from the energy conservation law, the coastal tsunami height can be easily estimated from the offshore tsunami height. A more advanced method is the regression model based on the results of many tsunami simulations (hereafter referred to as regression model). The regression model is simple but practical, and it can make highly accurate predictions despite its processing speed. However, the regression model only predicts the height at one arbitrary point on the coast and does not obtain the areal distribution, such as the maximum inundation depth distribution. For emergency response after a tsunami disaster, it is desirable to be able to predict not only the coastal tsunami height but also the inundation depth distribution. To obtain the inundation depth distribution using the existing regression models is simply a matter of predicting all the points in the space, but this requires a long processing time due to a large number of prediction points. To solve this problem, we propose a method to reduce the number of prediction points by grouping areas with similar inundation depths in advance. This study aims to develop a method for fast prediction of inundation depth distribution. In addition, there is a possibility that the prediction accuracy can be improved by using the machine learning technique, which has been developed remarkably in recent years. Therefore, we applied the multilayer perceptron (MLP) for the tsunami inundation prediction in addition to the regression model.
The target area for analysis was the vicinity of Anan City in Tokushima Prefecture. The tsunami inundation database calculated from Fujiwara et al. (2020) 's 3480-case source fault scenario was used as the inundation depth data for the clustering analysis. The k-means method, a non-hierarchical method, was applied to the inundation depth distributions of 14 randomly selected scenarios from the tsunami inundation database to discriminate the regions with almost the same inundation depth. Since the k-means method requires the analyst to define the number of clusters, the number of clusters was determined by varying the number of clusters and referring to the variation of inundation depth data within the clustered areas.
The characteristics values (mean and standard deviation) of the inundation depth data from the tsunami inundation database were extracted in each clustered area. A power law regression model was constructed using the conjugate gradient method (CG method) with the extracted characteristics as the objective variable and the absolute value of the change in water pressure from the standard at 51 DONET stations as the explanatory variable. The regression model was constructed by the conjugate gradient method (CG method). The tsunami inundation prediction equation by MLP was also constructed. The power-law regression model and the MLP, the tsunami inundation for the 11 cases of the Cabinet Office model for the test data was predicted and compared with the true value by the forward tsunami calculation.
For the Nankai Trough earthquake scenarios 3, 4, and 5, the prediction errors of the CG method were 1.29 m, 1.83 m, and 1.76 m for RMSE, and 0.45 m, 0.93 m, and 1.08 m for MLP. MLP exceeded the prediction accuracy of CG method.
The target area for analysis was the vicinity of Anan City in Tokushima Prefecture. The tsunami inundation database calculated from Fujiwara et al. (2020) 's 3480-case source fault scenario was used as the inundation depth data for the clustering analysis. The k-means method, a non-hierarchical method, was applied to the inundation depth distributions of 14 randomly selected scenarios from the tsunami inundation database to discriminate the regions with almost the same inundation depth. Since the k-means method requires the analyst to define the number of clusters, the number of clusters was determined by varying the number of clusters and referring to the variation of inundation depth data within the clustered areas.
The characteristics values (mean and standard deviation) of the inundation depth data from the tsunami inundation database were extracted in each clustered area. A power law regression model was constructed using the conjugate gradient method (CG method) with the extracted characteristics as the objective variable and the absolute value of the change in water pressure from the standard at 51 DONET stations as the explanatory variable. The regression model was constructed by the conjugate gradient method (CG method). The tsunami inundation prediction equation by MLP was also constructed. The power-law regression model and the MLP, the tsunami inundation for the 11 cases of the Cabinet Office model for the test data was predicted and compared with the true value by the forward tsunami calculation.
For the Nankai Trough earthquake scenarios 3, 4, and 5, the prediction errors of the CG method were 1.29 m, 1.83 m, and 1.76 m for RMSE, and 0.45 m, 0.93 m, and 1.08 m for MLP. MLP exceeded the prediction accuracy of CG method.