Item

Abstract

The accurate determination of dynamic topography (DT) is crucial for various applications, including oceanography (ocean circulation and sea-level rise studies), coastal management, engineering, and navigation. Whilst satellite altimetry (SA) is a vital source for determining both coastal and offshore sea level data, the SA data however may be limited due to its spatial and temporal resolution. This limitation suggests an opportunity to use machine learning approaches to fill these tempo-spatial gaps that may exist. Thus, a deep learning method is tested on the Sentinel 3 data for the entire Baltic Sea for the period 2017‒2021. To fill these gaps in the SA-based dynamic topography, a novel approach is employed through a multivariate deep neural network-based algorithm. Input parameters (such as winds, sea level pressure, significant wave height, etc.) were used and a conditional generative adversarial network (CGAN) was employed by stacking convolutional layers. The concept of the CGAN model comprises two parts: the generator model and the discriminator. The generator is trained to reconstruct a map of DT and mislead the discriminator, while the discriminator is trained simultaneously to differentiate whether the generated map is real or fake. As a result, generated DT is used to complete a grid surface of the SA data. The proposed method is evaluated using tide gauge records and a hydrodynamic model for the Baltic Sea. The method is shown to outperform traditional techniques such as optimum interpolation (IO) in terms of accuracy and computational efficiency.