Deep learning-based landslide tsunami run-up prediction from synthetic gage data

Abstract

The present study proposes a deep learning model based on Long-Short Term Memory (LSTM) that uses gage measurements for prediction of landslide-driven maximum tsunami run-up. In an attempt to overcome the limitation of insufficient real-world data in the field, our methodology refers to analytical models to create a comprehensive dataset employing a time series recorded from an offshore gage as input and its corresponding maximum run-up at the shoreline as output, for different landslide scenarios with pre-determined parameters. The LSTM-based model is then trained using this dataset in order to predict the maximum run-up. The results, with mean values of 0.211 m, 0.149 m, 1.745% and 0.9988 for RMSE, MAE, MAPE and $R^2$, respectively, indicate that our model is both accurate and precise. As the data-driven models such as the one proposed here are often utilized to identify relationships that may not be immediately apparent from the physical models alone, our interdisciplinary approach has the potential to foster the development of innovative solutions and methodologies for addressing complex natural hazards by enhancing early warning systems, preparedness and response to tsunamis.