END-TO-END PHYSICS-INFORMED REPRESENTATION LEARNING FOR SATELLITE OCEAN REMOTE SENSING DATA: APPLICATIONS TO SATELLITE ALTIMETRY AND SEA SURFACE CURRENTS

Ocean sediment traps measure the sequestrated sinking organic carbon. While sinking, the particles are affected by local currents, which presents a challenge with regard to linking the deep flux with the surface. We present a machine learning tool that predicts the source location of the sinking particles based on satellite data. The predictions demonstrate a stronger correlation between surface and deep carbon fluxes, allowing a more comprehensive understanding of the deep carbon sequestration drivers.

The Surface Water and Ocean Topography (SWOT) mission provides unprecedented swath altimetry data. This study examines SWOT's impact on mapping systems, showing a moderate effect with the current nadir altimetry constellation and a stronger impact with a reduced one. Integrating SWOT with dynamic mapping techniques improves the resolution of satellite-derived products, offering promising solutions for studying and monitoring sea-level variability at finer scales.

The biological carbon pump plays a key role in the climate system. Plankton absorb and transform CO₂ into organic carbon, forming particles that sink to the ocean floor. Sediment traps catch these particles and measure the carbon stored in the abyss. However, the particles' surface origin is unknown as ocean currents alter their paths. Here, we train an AI model to predict the origin of these particles. This new tool enables a better link between deep-ocean observations and satellite images.

4DVarNet is a learning-based method based on traditional data assimilation (DA). This new class of algorithms can be used to provide efficient reconstructions of a dynamical system based on single observations. We provide a 4DVarNet application to sea surface height reconstructions based on nadir and future Surface Water and Ocean and Topography data. It outperforms other methods, from optimal interpolation to sophisticated DA algorithms. This work is part of on-going AI Chair Oceanix projects.

Temperature and salinity profiles are essential for studying the ocean’s stratification, but there are not enough of these data. Satellites are able to measure daily maps of the surface ocean. We train a machine to learn the link between the satellite data and the profiles in the Gulf Stream region. We can then use this link to predict profiles at the high resolution of the satellite maps. Our prediction is fast to compute and allows us to get profiles at any locations only from surface data.