Characterizing (Sub)mesoscale Ocean Scale Interactions Through Deep Learning-Enhanced Satellite Oceanography

Abstract

The ocean circulation emerges from interacting motions across a vast range of spatial scales, from basin-scale gyres and mesoscale eddies hundreds of kilometers across to kilometer-scale submesoscale fronts and ultimately three-dimensional turbulence. Although each scale is governed by distinct dynamics, under- standing ocean circulation requires considering their mutual interactions. This dissertation focuses on the mesoscale–submesoscale transition, where flows shift from being largely geostrophic and two-dimensional to ageostrophic and three-dimensional, and where cross-scale energy pathways regulate how energy is gen- erated, transferred, and ultimately dissipated in the global ocean. Our understanding of ocean scale interactions must ultimately be grounded in observations. Yet charac-terizing (sub)mesoscale interactions is profoundly challenging, requiring kilometer-scale resolution along- side synoptic coverage over O(1000 km) scales. Satellites provide increasingly detailed observations of sea surface height (SSH), temperature (SST), and salinity (SSS). However, inferring (sub)mesoscale surface currents from these measurements is a major challenge. Put simply, this dissertation aims to estimate sur- face ocean currents from satellites at significantly higher resolution than previously possible and use this to characterize (sub)mesoscale ocean scale interactions. This has been enabled by a convergence of new submesoscale-resolving satellites, high-resolution simulations, and advances in deep learning, creating new opportunities to characterize multiscale ocean dynamics from space. In Chapter 1, we develop a deep learning framework to reconstruct gridded SSH from sparse satellitealtimetry and co-located SST. Tested in the challenging Gulf Stream Extension region, the method reduces SSH mapping error by 17%, resolves spatial scales 30% smaller than objective analysis, and yields surface geostrophic currents that more closely match drifter observations, allowing us to better diagnose crucial diagnostics of mesoscale eddy dynamics. In Chapter 2, we scale this approach to produce Neural Ocean Surface Topography (NeurOST), the firstglobal deep learning-based high-resolution SSH and surface current dataset with state-of-the-art accuracy and resolution, and use this dataset to evaluate the role of the mesoscale kinetic energy cascade in driving seasonality at larger scales. In Chapter 3, we develop a generative data assimilation framework (GenDA) that combines a diffusionprior trained on high-resolution simulations with sparse satellite observations to jointly reconstruct multi- variate, eddy-resolving surface states. By pursuing a generative approach where the model is trained to generate physically plausible multi-variate surface states from a simulation, we are able to target variables and scales not accessible using NeurOST, objective analysis, or physics-based data assimilation. In Chapter 4, we extend this framework to submesoscale-resolving state estimation by retraining the prior on the submesoscale-permitting LLC4320 simulation and incorporating temporal coherence via a video diffusion model. By assimilating high-resolution SST and SSH from the recently-launched SWOT mission, we develop the first submesoscale-resolving surface state estimate in the Agulhas Retroflection, providing an unprecedented view of surface ocean dynamics and allowing us to provide observational constraints on the (sub)mesoscale kinetic energy cascade. This dissertation delivers a step change in our ability to estimate the surface ocean state from satelliteobservations, suggests an increasing role for deep learning methods in ocean state estimation, and provides crucial observational constraints on (sub)mesoscale ocean scale interactions.