Georgy Manucharyan, Lia Siegelman, Patrice Klein

Satellite altimeters provide global observations of sea surface height (SSH) and present a unique dataset for advancing our theoretical understanding of upper ocean dynamics and monitoring its variability. Considering that mesoscale SSH patterns of 50--300 km in size can evolve on timescales comparable to or shorter than satellite return periods, it is challenging to accurately reconstruct the continuous SSH evolution as currently available altimetry observations are still spatially and temporally sparse. Here we explore the possibility of SSH interpolation via a Deep Learning framework using synthetic observations from a quasigeostrophic model of mesoscale ocean turbulence. We demonstrate that Convolutional Neural Networks with Residual Learning are superior in SSH reconstruction to linear and recently developed dynamical interpolation techniques. In addition, neural networks can provide a skillful state estimate of unobserved deep ocean currents at mesoscales. This conspicuous result suggests that SSH patterns of eddies do contain substantial information about the underlying deep ocean currents that is necessary for SSH prediction. Our framework is highly idealized and several crucial improvements such as transfer learning, diversification of training data, and modification of the loss function would be necessary to implement before its ultimate use with real satellite observations. Nonetheless, by providing a proof of concept based on synthetic data, our results point to Deep Learning as a viable alternative to existing interpolation and, more generally, state estimation methods for satellite observations of eddying currents.