Scott A Martin , Georgy Manucharyan, Patrice Klein

Ocean eddies help shape marine ecosystems, large-scale ocean circulation, and global climate through their non-linear interactions. Observing eddies poses a major challenge due to their chaotic evolution across a wide range of spatio-temporal scales. Satellite-derived estimates of surface ocean currents significantly distort and smooth eddies and, consequently, strongly underestimate the strength of non-linear eddy interactions. Here, we use deep learning to develop a new global estimate of surface currents from satellite observations, capturing ocean eddies with accuracy and resolution that surpasses state-of-the-art reconstructions. We achieved this by synthesising satellite observations of sea surface height with sea surface temperature across the Global Ocean using a single neural network that generalises across regions with diverse eddy dynamics. Our new reconstructions reveal dramatic changes in the inferred eddy dynamics in many regions, highlighting the existence of strongly seasonal non-linear eddy interactions. These eddy interactions transfer large amounts of kinetic energy from small to large scales, emphasising the prominent role of small-scale eddies in setting the seasonal cycle of kinetic energy in the ocean. Our study demonstrates that deep learning can dramatically improve our observations of ocean eddy dynamics from space, providing a new paradigm for satellite oceanography.