A Seafloor Pressure Sensor Effectiveness Evaluation in a Tsunami Height Estimation Model Using Gaussian Process Regression with Automatic Relevance Determination

To meet the growing need for accurate and timely tsunami height predictions, a database comprising 3480 high-precision tsunami simulation scenarios for the Nankai Trough was constructed in this study. The database includes the Dense Oceanfloor Network for Earthquakes and Tsunamis (DONET) seafloor pressure sensor data and the maximum tsunami heights for 19 coastal cities. Gaussian process regression with automatic relevance determination (ARD) was used to quantitatively evaluate the effectiveness of each sensor location. In this framework, the ARD assigns a hyperparameter that functions as an indicator of the contribution of each sensor to the prediction. The proportion of this hyperparameter reflects the effectiveness of each sensor during the prediction process. The results demonstrated that removing the 20 least effective sensors (as identified by the ARD) led to an 11% increase in the estimation error, whereas removing the 10 most effective sensors resulted in a 38% increase. These results demonstrate that the proposed method enables the optimization of the sensor placement process and allows for a prior evaluation of the trade-off between the number of sensors and achieved prediction accuracy. This method offers a robust framework for optimizing observation networks, enhancing the accuracy of tsunami predictions, and supporting future disaster risk reduction.