Machine Learning Approaches for Early Warning of Tsunami Induced by Volcano Flank Collapse and Implication for Future Risk Management: Case of Anak Krakatau

Abstract

A tsunami triggered by volcanic collapse is a low-probability but high-impact event. Unlike tsunamis triggered by earthquakes which the mechanism is well understood, volcanic tsunami events have complex trigger mechanisms and occur with little to no warning such as the event in December 2018 of the Anak Krakatau volcano tsunami, making it more difficult to detect and issue a warning. We adopted collapse tsunami machine learning (ML) approach model which does not require source information, to predict maximum tsunami amplitude on four coastal stations. Observations from six synthetic observation stations around Anak Krakatau volcano were used as input for collapse tsunami ML. The 320 collapse scenarios triggering tsunamis with various parameters and directions were generated to train model. To evaluate the accuracy and reliability of the tsunami simulations, we conducted a comparison between the simulated waveforms and those recorded at four coastal stations during the December 2018 event. The RMSE values between predicted and actual (via forward tsunami) of Random Forest model consistently provide the most accurate predictions ranging from 0.0586 to 0.1945 across three out of the four stations. We also applied deep learning algorithms, LSTM, and Complex LSTM to predict tsunami full waveform by using short-duration observation as input. Furthermore, we also pointed out the potential of risk management that can be explored and integrated from results of the maximum tsunami amplitude and arrival time predictions for support decision-making. We suggest that the ML approach could be a good alternative for volcanic tsunamis early warning purposes.