A proper extraction of internal tidal signals is central to the interpretation of Sea Surface Height (SSH) data. The increased spatial resolution of future wide-swath satellite missions poses a challenge for traditional harmonic analysis, due to prominent and unsteady wave-mean interactions at finer scales. However, the wide swaths will also produce SSH snapshots that are spatially two-dimensional, which allows us to treat tidal extraction as an image translation problem. We design and train a conditional Generative Adversarial Network, which, given a snapshot of raw SSH from an idealized numerical eddying simulation, generates a snapshot of the embedded tidal component. We test it on data whose dynamical regimes are different from the data provided during training. Despite the diversity and complexity of data, it accurately extracts tidal components in most individual snapshots considered and reproduces physically meaningful statistical properties. Predictably, TITE′s performance decreases with the intensity of the turbulent flow.

Plain Language Summary

Wide-swath satellite observations of Sea Surface Height (SSH) data at high spatial resolutions will be available in abundance thanks to advances of instrumental technologies. Embedded in the observed SSH are internal tides, a dynamical component that plays a crucial role in ocean circulation. As they are entangled with background currents and eddies, such tidal signals are challenging to extract. Methods that worked with previous-generation altimeters will break down at the resolutions that the new generation promises. On the other hand, the wide satellite swaths provide new opportunities as they allow us to regard the observations as spatially two-dimensional. Here we treat the tidal extraction solely as an image translation problem. We train a deep neural net so that given a snapshot of a raw SSH signal, it produces a “fake′′ snapshot of the tidal SSH signal that is meant to reproduce the original. The data we use in this article is generated by idealized numerical simulations. Once adapted to realistic data, the network has the potential to become a new tidal extraction tool for satellite observations. More broadly, successes in our experiments can inspire other applications of generative networks to disentangle dynamical components in data where classical analysis may fail.