Forecasting of absolute dynamic topography using deep learning algorithm with application to the Baltic Sea

Accurate sea-level forecasting is crucial for navigation, engineering and coastal conservation. One of the major obstacles in obtaining accurate sea-level data, both at coastal and offshore areas, has been in defining a realistic vertical datum. The Baltic Sea countries, however, have collaborated in calculating a high-resolution regional geoid model (NKG2015). This now paves the way for determining accurate sea-levels over the entire sea. Accordingly, this study explores the application of a deep learning two-dimensional Convolutional Neural Network (Conv2D) technique along with using essential inputs (e.g. accurate dynamic topography (DT), wind speed and direction, surface pressure and temperature). The method was tested for a three-year 2017–2019 period in the Baltic Sea. The evaluation was based on two statistical criteria: spatial root mean squared error (RMSE) and R-squared (R2). Results revealed that the proposed Conv2D model allows predicting the DT of the Baltic Sea, with an R2 of 0.91. The spatial RMSE plot also confirms accurate DT predictions for most of the Baltic Sea points, with the discrepancies distributed within ±4 cm. Spatially, some larger RMSE values (∼10 cm) were obtained at particular locations in south-eastern Baltic Sea, which may be due to the input sources utilized. Examination of sea level maxima also showed that the Conv2D model reproduced the maxima events in most scenarios (9 of 10 and 8 of 11 in Gulf of Finland and Riga, respectively) with residuals that varied up to 7 cm–18 cm. For the higher residuals, the Conv2D tended to underestimate the sea level. This suggests the importance of considering the necessary inputs (e.g waves) for forecasting storm surges and that the Conv2D model can still be improved. External validation was also performed using along-track sea level satellite altimetry data, with differences between forecasted model and satellite being within 5 cm. This confirms the validity of the forecasted model and the occurrences of model biases to be minimum. The method utilized shows the potential to contribute toward operational sea level forecasting in the Baltic Sea.