The SWOT (Surface Water Ocean Topography) mission will provide high-resolution and two-dimensional measurements of sea surface height (SSH). However, despite its unprecedented precision, SWOT’s Ka-band Radar Interferometer (KaRIn) still exhibits a substantial amount of random noise. In turn, the random noise limits the ability of SWOT to capture the smallest scales of the ocean’s topography and its derivatives. In that context, this paper explores the feasibility, strengths and limits of a noise-reduction algorithm based on a convolutional neural network. The model is based on a U-Net architecture and is trained and tested with simulated data from the North Atlantic. Our results are compared to classical smoothing methods: a median filter, a Lanczos kernel smoother and the SWOT de-noising algorithm developed by Gomez-Navarro et al. Our U-Net model yields better results for all the evaluation metrics: 2 mm root mean square error, sub-millimetric bias, variance reduction by factor of 44 (16 dB) and an accurate power spectral density down to 10–20 km wavelengths. We also tested various scenarios to infer the robustness and the stability of the U-Net. The U-Net always exhibits good performance and can be further improved with retraining if necessary. This robustness in simulation is very encouraging: our findings show that the U-Net architecture is likely one of the best candidates to reduce the noise of flight data from KaRIn.

The ODYSEA (Ocean DYnamics and Surface Exchange with the Atmosphere) mission will provide simultaneous two-dimensional measurements of currents and winds for the first time. According to the ODYSEA radar concept, with a high incidence angle, current noise is primarily driven by backscattered power, which is triggered by wind speed. Therefore, random noise will affect the quality of observations. In low wind conditions, the absence of surface roughness increases the noise level considerably, to the point where the measurement becomes unusable, as the error can exceed 3 m/s at 5 km posting compared to mean current amplitudes of tens of cm/s. Winds higher than 7.5 m/s enable current measurements at 5 km posting with an RMS accuracy below 50 cm/s, but derivatives of currents will amplify noise, hampering the understanding of ocean dynamics and the interaction between the ocean and the atmosphere. In this context, this study shows the advantages and limitations of using noise-reduction algorithms. A convolutional neural network, a UNet inspired by the work of the SWOT (Surface Water and Ocean Topography) mission, is trained and tested on simulated radial velocities that are representative of the global ocean. The results are compared with those of classical smoothing: an Adaptive Gaussian Smoother whose filtering transfer function is optimized based on local wind speed (e.g., more smoothing in regions of low wind). The UNet outperforms the kernel smoother everywhere with our simulated dataset, especially in low wind conditions (SNR << 1) where the smoother essentially removes all velocities whereas the UNet mitigates random noise while preserving most of the signal of interest. Error is reduced by a factor of 30 and structures down to 30 km are reconstructed accurately. The UNet also enables the reconstruction of the main eddies and fronts in the relative vorticity field. It shows good robustness and stability in new scenarios.

The launch of the altimetric satellite SWOT (Surface Water and Ocean Topography) was a revolution in oceanography and hydrology. With its120 km swath width, a spatial resolution of 500m² (in Low Resolution acquisition mode) and an instrumental random error significantly lower than the one from nadir altimetry, the Ka-Band Radar Interferometer (KaRIn) onboard SWOT mission also present a huge potential to develop applications in the polar regions. Indeed, the SWOT product enables the observation of leads, icebergs and polynyas (Dibarboure and al 2024) through the measures of surface topography and backscatter coefficient. The surface discrimination between leads and floes is the first step toward polar ocean monitoring, ice thickness and snow depth estimations. However because of the complexity of the surface (different surface roughness properties in the leads, presence of melt pounds) added to residual sensing errors (residual KaRIn random error, residual systematic errors, …) this first required achieved is not straightforward. Therefore several classification approaches were developed : one based on a statistical method (Markov Random Field), one based on an unsupervised machine learning method (Kmeans) and another one based on a supervised machine learning method (XGBoost). The objective of this paper is thus to present the results of these methods (their robustness, strengths and weaknesses) through local comparisons with respect to optical, SAR images and global comparison with existing state of the art products (OSISAF ice concentration products).