4:00 PM - 4:15 PM
[SCG55-12] Prediction of Tsunami Flow Depth Distribution of Nankai Trough Earthquake Tsunami Using Machine Learning and Seafloor Water Pressure Data
Keywords:Tsunami, Machine learning, Multilayer perceptron
Research involving machine learning has been active in various fields in recent years. This study investigates a method for predicting tsunami flow depths using the multilayer perceptron (MLP) in the Nankai trough, Japan. The analysis procedure is as follows. Nonlinear long-wave tsunami calculations were performed for 3,480 earthquake fault scenarios to generate teacher data. As input data to the machine learning model, we assumed to use pressure data observed by an ocean-bottom pressure gauge network in the Nankai trough. We prepared from the 3480 calculated results the maximum or average bottom pressure 3 minutes after the earthquake for 5, 10, and 15 minutes to avoid seismic noise.
The prediction target was tsunami flow depth distribution in a coastal town anticipating severe tsunami damage. The MLP model had three middle layers with 100 nodes, initialized by He normal distribution. ReLU was used as the activation function for the middle layers, and the loss function was a mean square error. We applied L1/L2 regularization to the middle layers and a Gaussian dropout after the middle layers to suppress overlearning. Adam was used as the optimization algorithm. The hyperparameters were determined by trial and error, with L1/L2 regularization of 0.001, Gaussian dropout of 0.2, and Adam of 0.001.
We constructed test data calculated using the same method above from the M9 earthquake 11 scenarios proposed by the Cabinet Office of Japan. We compared tsunami flow depths of 20 cm or greater from the test data and ones predicted by MLP with the fitting index of the geometric mean K and geometric standard deviation κ from Aida (1978). The accuracy (%) was calculated from the geometric mean K.
Using average pressure as input, K-κ were estimated to be 1.57-1.78, 1.21-1.57, and 1.19-1.62 for 5, 10, and 15-minute time windows, respectively. Using the maximum pressure as input, K-κ were estimated to be 1.08-1.67, 1.01-1.65, and 0.89-1.6, respectively. The accuracy obtained from K was 64%, 83%, and 84% for the average pressure input, and 93%, 99%, and 89% for the maximum pressure input. The case of the best accuracy was the maximum pressure used as an input with 10-minute time window. However, the prediction accuracy varied largely among the test scenarios, which were 96%, 96%, 69%, 71%, 78%, 78%, 94%, 81%, 85%, 40%, and 87% from the scenario 1 to 11 in order. We will improve the prediction model to address these issues in the next step.
The prediction target was tsunami flow depth distribution in a coastal town anticipating severe tsunami damage. The MLP model had three middle layers with 100 nodes, initialized by He normal distribution. ReLU was used as the activation function for the middle layers, and the loss function was a mean square error. We applied L1/L2 regularization to the middle layers and a Gaussian dropout after the middle layers to suppress overlearning. Adam was used as the optimization algorithm. The hyperparameters were determined by trial and error, with L1/L2 regularization of 0.001, Gaussian dropout of 0.2, and Adam of 0.001.
We constructed test data calculated using the same method above from the M9 earthquake 11 scenarios proposed by the Cabinet Office of Japan. We compared tsunami flow depths of 20 cm or greater from the test data and ones predicted by MLP with the fitting index of the geometric mean K and geometric standard deviation κ from Aida (1978). The accuracy (%) was calculated from the geometric mean K.
Using average pressure as input, K-κ were estimated to be 1.57-1.78, 1.21-1.57, and 1.19-1.62 for 5, 10, and 15-minute time windows, respectively. Using the maximum pressure as input, K-κ were estimated to be 1.08-1.67, 1.01-1.65, and 0.89-1.6, respectively. The accuracy obtained from K was 64%, 83%, and 84% for the average pressure input, and 93%, 99%, and 89% for the maximum pressure input. The case of the best accuracy was the maximum pressure used as an input with 10-minute time window. However, the prediction accuracy varied largely among the test scenarios, which were 96%, 96%, 69%, 71%, 78%, 78%, 94%, 81%, 85%, 40%, and 87% from the scenario 1 to 11 in order. We will improve the prediction model to address these issues in the next step.