CLOINet: Ocean state reconstructions through remote-sensing, in-situ sparse observations and Deep Learning

Combining remote-sensing data with in-situ observations to obtain a full 3D reconstruction of the ocean state is challenging for classical interpolation techniques. Therefore, we developed a CLuster Optimal Interpolation Neural Network (CLOINet), which leverages the well-consolidated mathematical basis of the Optimal Interpolation (OI) scheme with a"Self-Attention Mechanism": the state-of-the-art neural network technology. Our network segments the remote sensing images into clusters revealing non-local correlations and enhancing ocean fine-scale reconstruction. We trained our neural network with the outputs of an Ocean General Circulation Model (OGCM), which also provided various test scenarios. In this sense, we realized Observing System Simulation Experiments where we aimed to reconstruct the deep salinity fields given the sea surface temperature (SST) or the Sea Surface Height (SSH) together with sparse in-situ salinity observations. As a result, we decreased the reconstruction error of the $25\%$ and resolved $50\%$ smaller scales compared to the baseline OI. Furthermore, even if we trained our neural network purely with simulated data, we improved by a $35\%$ the reconstruction of a real SST field, relying on glider temperature observations and satellite chlorophyll concentration. We thus evidenced how deep-learning networks like CLOINet could be the leading technology to join the efforts of the modeling and observations community toward developing an ocean digital-twin.