A Deep Learning Approach to Extract Internal Tides Scattered by Geostrophic Turbulence

International audience ; Since the launch of TOPEX/Poseidon, oceanographers have used the geostrophic assumption to infer sea surface velocity from Sea Surface Height (SSH). However, while an estimated 90% of the ocean's kinetic energy exists in the form of currents in quasigeostrophic balance (hereafter qualified as "balanced"; see Ferrari & Wunsch, 2009), one still must account for "unbalanced" flows such as internal tides, hereafter "ITs", for a refined inference of balanced currents (Fu & Ferrari, 2008). Furthermore, ITs play a crucial role in ocean mixing (Lien & Gregg, 2001; Whalen et al., 2020), and are helpful in detecting ocean temperature changes (Zhao, 2016). Therefore, whether ITs are considered "noise" (e.g., for inferring balanced flows) or "signal" (e.g., for inferring tidally induced mixing), their proper extraction from altimetry data is essential. For decades, IT extraction has been conducted via harmonic analysis (Munk & Hasselmann, 1964), a method that relies on a close phase relationship (or coherence) between ITs and astronomical forcings. Departures from this condition are sometimes referred to as "incoherence" (Ponte & Klein, 2015; Zaron & Rocha, 2018). Current altimetry has a typical spatial resolution of O(100) km (Ballarotta et al., 2019), which is sufficient to retrieve mode-1 and some of the mode-2 IT wavelengths of semidiurnal tides, along with the dominant turbulent balanced