Filtering tide-generated internal waves using Convolutional Neural Networks

Starting from 2021, Surface  Water Ocean Topography (SWOT) satellite altimetry mission will provide an unprecedented amount of Sea Surface Height (SSH) measurements. In addition to allowing for a higher spatial resolution, SWOT will deliver two-dimensional horizontal SSH data thanks to its wide swath capacities, which is a remarkable leap compared to conventional current altimeters.

With the aim of extracting a clean SSH signal from the SWOT measurements, several challenges are expected to be encountered. In this work, we focus on filtering the footprints of Internal Gravity Waves (IGWs), this is of high interest for physical oceanographers who seek to better understand mesoscale and submesoscale ocean physics.

Thanks to recent developments in ocean numerical simulation, we can now have access to a considerable amount of simulation data with exceptional high spatial resolutions up to 1/60° and hourly temporal resolution. Here, we benefit from an advanced North Atlantic simulation of the ocean circulation (eNATL60) that models tidal motions, and design a supervised machine learning experiment that aims to test several techniques for filtering IGWs.

In particular, we show that deep convolutional neural networks are a relevant candidates for this task and presents promising results with regard to conventional linear filtering techniques. We also show how our method can be adapted to the context of the fast-sampling phase of SWOT, and can also take advantage from the presence of additional data such as Sea Surface Temperature.