Training Neural Mapping Schemes for Satellite Altimetry With Simulation Data.

Satellite altimetry combined with data assimilation and optimal interpolation schemes have deeply renewed our ability to monitor sea surface dynamics. Recently, deep learning schemes have emerged as appealing solutions to address space‐time interpolation problems. However, the training of state‐of‐the‐art neural schemes on real‐world case‐studies is hindered by the sparse space‐time coverage of the sea surface of real altimetry data set. Here, we introduce an innovative approach that leverages state‐of‐the‐art ocean models to train simulation‐based neural schemes for the mapping of sea surface height and demonstrate their performance on real altimetry data sets. We analyze further how the ocean simulation data set used during the training phase impacts this performance. This experimental analysis covers both the resolution from eddy‐present configurations to eddy‐rich ones, forced simulations versus reanalyzes using data assimilation and tide‐free versus tide‐resolving simulations. Our benchmarking framework focuses on a Gulf Stream region for a realistic 5‐altimeter constellation using NEMO ocean simulations and 4DVarNet mapping schemes. All simulation‐based 4DVarNets outperform the operational observation‐driven and reanalysis products, namely DUACS and GLORYS. The more realistic the ocean simulation data set used during the training phase, the better the mapping. The best 4DVarNet mapping was trained from an eddy‐rich and tide‐free simulation data sets. It improves the resolved longitudinal scale from 151 km for DUACS and 241 km for GLORYS to 98 km and reduces the root mean square error by 23% and 61%. These results open research avenues for new synergies between ocean modeling and ocean observation using learning‐based approaches. Plain Language Summary: To train an artificial intelligence (AI) model, one need to describe a task using data and an evaluation procedure. Here we aim at constructing images related to the ocean surface currents. The satellite data we use provide images of the ocean surface with a lot of missing data (around 95% of missing pixels for a given day), and we aim at finding the values of the missing pixels. Because we don't know the full image, it is challenging to train an AI on this task using only the satellite data. However, today's physical knowledge makes it possible to numerically simulate oceans on big computers. For these simulated oceans, we have access to the gap‐free image, so we can train AI models by first hiding some pixels and checking if the model fill the gaps with the correct values. Here, we explore under which conditions AIs trained on simulated oceans are useful for the real ocean. We show that today's simulated oceans work well for training an AI on this task and that training on more realistic simulated oceans improve the performance of the AI! Key Points: We suggest utilizing ocean simulation data sets to train neural schemes for mapping real altimeter dataThe trained neural scheme improves the spatial scales resolved over the operational mapping product on a GulfStream case study by 30%Using more realistic simulation data sets improves the resulting neural mapping scheme by up to 20% in the spatial scales resolved [ABSTRACT FROM AUTHOR]