On the prediction of tropical cyclones wind wave footprint using satellite data and neural networks

Tropical cyclones (TCs) are associated with extreme waves, storm surge and strong winds. As a result, they are considered among nature´s most destructive phenomena. The capability for a rapid estimate of these events is crucial for coastal hazard applications as well as risk preparedness. Numerical simulations of wave generation forced with atmospheric reanalysis can not represent adequately the strength of the TCs due to the insufficient spatial resolution of the wind fields. To overcome this problem, a two-step procedure is usually adopted: (1) Holland-type vortex models to improve the definition of the wind fields and (2) numerical simulations with a wave generation model (e.g., WaveWatch III). This numerical approach can result in a large uncertainty for strong TCs. Therefore, wave measurements (buoy and satellite data) are only used for validation.

We propose a machine learning approach based on satellite data to estimate the wave footprint of tropical cyclones. The methodology for extracting the satellite data associated to each TC assumes a circular shape for the wave influence at each point of the track and defines, in polar coordinates with respect to the forward direction, the maximum significant wave height within a time interval. Following this approach we have been able to build a database with over a million points extracted from TCs developing around the world.

We developed an artificial neural network to predict the significant wave height using the minimum pressure, the TC forward velocity, the curvature of the track, the latitude and the temperature difference between the TC locations with respect to the tropics as input predictors. Using this approach, the wind wave footprint can be predicted for any given track. The model can be applied for fast estimates in operational systems, reconstruction of historical events, risk assessment, or climate change projections.