A Surrogate Model for Daily Sea Surface Current Fields Prediction Using CNN-UNET

Precise forecasting of sea surface currents is crucial for diverse applications, including navigation, pollution control, and ecosystem monitoring. Traditional high-resolution hydrodynamic models like NEMO generate detailed short-term forecasts but are computationally expensive and resource-intensive. To overcome these limitations, we present sciCUN: a deep learning framework designed for surface current inference using CNN-U-Net architecture.

In summary, sciCUN utilizes the zonal and meridional wind components, mean sea level pressure, air temperature, and dew point temperature from ECMWF Reanalysis v5 (ERA5) for the current day, along with the high-resolution zonal and meridional sea surface current velocity fields from the Copernicus Marine Service Baltic Sea Physics Reanalysis for the previous day, as input features. It then generates the high-resolution zonal and meridional sea surface current velocity fields for the current day.

As a case study, sciCUN was implemented in the Gulf of Riga domain. The model was trained to capture the influence of atmospheric forcing on preceding sea surface currents over a training period spanning 1993 to 2019. Its predictive performance was subsequently validated through a 4-year testing phase (2020–2023). Results showed that while prediction accuracy was slightly lower in coastal regions near river mouths and the Irbe Strait—areas where hydrodynamic models typically employ boundary conditions—sciCUN exhibited strong overall performance. The model achieved an average Euclidean distance of 2.30 cm/s between its predictions and reference data, with an average component-wise mean absolute error of 1.45 cm/s and correlation coefficient of 92.