Deep-learning models for predicting high-frequency sea-level oscillations in the Adriatic Sea

Intense high-frequency sea-level oscillations (HFOs) in the Mediterranean Sea, sometimes leading to destructive meteotsunamis, occur due to specific and spatially limited meteorological conditions. Despite the understanding of their physical dynamics, current forecasting systems based on hydrodynamic models are unreliable and computationally demanding. To address this problem, we built deep-learning models of HFOs for the Adriatic tide-gauge stations with long measurement records (Bakar and Ploče) and transferred these models to meteotsunami-prone locations with limited data (Stari Grad, Vela Luka and Sobra). We trained deep convolutional neural networks using simulated data (hourly mean sea-level pressure, geopotential heights, specific humidity, wind speed, air temperature from ERA5, and the calculated Richardson number) alongside measurements (1-min sea levels). We will present the model's architecture, transfer learning results, and predictions of HFO amplitudes based on: (i) forecasting horizons (ranging up to several days with different time windows; 6 h vs. 24 h), (ii) data inputs (total sea level vs. sea level decomposed into components), and (iii) various refinement strategies through inclusions of additional U-net based refinement heads. The results demonstrate that the developed models can predict the highest expected HFO amplitudes for the next three days with reasonable accuracy. Accuracy improves when using the ‘wet’ Richardson number instead of the ‘dry’ version, extending time windows (e.g., targeting the largest amplitude in the overall next 24 h rather than every 6 h), and reducing the input dataset. Performance also varies depending on the station from which the model was transferred. In all cases, the forecast accuracy is higher for smaller HFO amplitudes, with refinements primarily improving predictions of smaller amplitude HFOs.