Tsunami Early Warning From Global Navigation Satellite System Data Using Convolutional Neural Networks

Abstract

Accurate tsunami early warning allows for more effective emergency planning, thereby mitigating the human and economic toll. However, constructing a rapid forecast model is challenging for several reasons. One is that the underlying physical processes are governed by partial differential equations whose solution requires substantial computation that cannot be performed in a short timeframe. Furthermore, determining the proper initial conditions for the differential equations requires solving the earthquake source inversion problem, which itself holds significant uncertainty due to the lack of direct observations. The current US warning system relies on early estimates of earthquake location and magnitude from seismic data, coupled with direct tsunami observations from Deep Ocean Assessment and Reporting of Tsunamis sensors (DART;Titov et al., 2005) in the deep ocean and coastal tide gauges. The sparsity of such sensors limits the amount of data one can collect on the tsunami directly. Moreover, one has to wait for the tsunami to reach these sensors, which can be hours after the earthquake.In previous work (Liu et al., 2021a), hereafter referred to as Liu21, we explored machine learning (ML) techniques to forecast tsunami waveforms at two "forecast gauges" in the Puget Sound (denoted Gauges 901 and 911) shown in Figure 1. The forecasts were based on synthetic tsunami observations from Cascadia Subduction Zone (CSZ) events at an hypothetical "observation gauge" (denoted Gauge 702) in the Strait of Juan de Fuca. This ML approach avoids the need for real-time source inversion and tsunami simulation. We showed that several hours of tsunami waveforms at the forecast gauges could be forecast from shorter time series at the observation gauge, but it still requires 30-60 min of observed data after the tsunami reaches the observation gauge.

Classification of Potential Tsunami Disaster Due to Earthquakes in Indonesia Based on Machine Learning

“…The Confusion Matrix showed that the model correctly classified 232 instances of potential tsunamis and 37 cases of non-potential tsunamis out of 269 actual occurrences. However, 52 instances were falsely International Journal Software Engineering and Computer Science (IJSECS), 4 (1) 2024, [13][14][15][16][17][18][19][20][21][22][23] Classification of Potential Tsunami Disaster Due to Earthquakes in Indonesia Based on Machine Learning predicted as non-potential tsunamis and 77 instances were falsely predicted as potential tsunamis, indicating some misclassifications.…”

“…This aligns with the broader research landscape of improving tsunami predictions through machine learning techniques. Recent studies have explored the use of Convolutional Neural Networks for tsunami early warning systems [21], machine learning-based tsunami inundation prediction from offshore observations [22], and estimation of tsunami bore forces using Extreme Learning Machines [23]. These advances underscore the continuous efforts to leverage sophisticated algorithms and data sources to refine tsunami prediction models and effectively mitigate the impact of earthquake-generated tsunamis.…”

Prediction of Tsunami Alert Levels Using Deep Learning

“…Núñez et al (2022) use convolutional neural networks to predict tsunami time series at specific locations. Rim et al (2022) also use convolutional neural networks to forecast tsunami waveforms from Global Navigation Satellite System (GNSS) data at selected locations. Finally, Mulia et al (2022) use MLP networks to predict flow depths in several coastal locations from offshore tsunami data at 150 offshore stations in the Japan Trench.…”

“…Rim et al. (2022) also use convolutional neural networks to forecast tsunami waveforms from Global Navigation Satellite System (GNSS) data at selected locations. Finally, Mulia et al.…”

A Stability Analysis of Neural Networks and Its Application to Tsunami Early Warning

“…Such an instability would limit their use in applications where stable and accurate prediction is critical. For example, NNbased tsunami early warning models (Makinoshima et al, 2021;Liu et al, 2021;Mulia et al, 2022;Rim et al, 2022) will need to be stable with respect to uncertainty in its input, as the real input measurement is subject to various measurement noise, sensor malfunctions, or other types of anomalies during a tsunamigenic earthquake event (Titov et al, 2005;Kumar and Ahmed, 2011). As NNs are known to produce erroneous predictions when there is a specific type of noise in the input data, analyzing and controlling these instabilities is of utmost importance in order to use them for safety-critical applications.…”