Effect of calibration data on performance of tsunami early warning model

Abstract

Data-driven tsunami early warning systems can be calibrated using possible wave profiles that are simulated from numerous hypothetical rupture scenarios. However, tsunami wave profiles that are simulated based on a certain synthesis method may not capture future situations comprehensively. To quantify the effects of calibration datasets on tsunami early warning models, a case study focusing on Vancouver Island that faces major tsunami threats from the Cascadia subduction earthquakes is explored. Two tsunami wave databases are generated by considering a logic tree model of potential tsunami sources for probabilistic tsunami hazard analysis and stochastic rupture sources with variable geometry and heterogeneous slip distribution. Tsunami early warning models are developed based on three fitting methods, namely, multiple linear regression, random forest, and neural network. Using consistent and inconsistent training-testing (calibration-evaluation) datasets, performances of the tsunami early warning models are compared. The results of the comparative analyses indicate that the use of random forest and neural network outperform conventional multiple linear regression methods. The effects of calibration data on the model performance are significant and may not be captured well by a conventional cross-validation scheme. This study highlights the importance of epistemic uncertainty of the tsunami early warning model performance.