Editor's Highlights are summaries of recent papers by AGU journal editors.
sauce: Journal of Advances in Earth System Modeling
A major challenge in working with ocean observations is the lack of observations in all places in space and time. Most users of ocean observations want maps that have no missing areas, so the question of how to fill in these areas has been a long-standing research problem in observational oceanography.
There are various methods to fill in the missing areas, but these methods typically use statistical assumptions (e.g., Gaussian kernels used for best fit interpolation) that are not fully based on or constrained by the underlying physical laws. We recognize that physical processes have important spatiotemporal correlations, but it is not clear how to incorporate all aspects of the correlations into a mapping scheme for ocean observations.
Febvre et al. [2024] We propose a new approach that combines state-of-the-art numerical models with a neural mapping scheme to reconstruct satellite altimetry data. In this method, we use numerical simulations with a dynamic model to train a machine learning method (neural mapping scheme) to map satellite altimetry. The well-measured Gulf Stream region serves as a test case for this analysis. The results of this study show that incorporating simulated ocean data in the training process improves the performance of neural mapping, outperforming traditional methods.
This study is an intriguing example of how the convergence of dynamic modelling, machine learning techniques and oceanographic measurements can enhance our understanding, monitoring and mapping of oceanography.
Quote: Febvre, Q., Le Sommer, J., Ubelmann, C., & Fablet, R. (2024). Training a neural mapping scheme for satellite altimetry using simulated data. Journal of Advances in Earth System Modeling, 16, e2023MS003959. https://doi.org/10.1029/2023MS003959
—Stephen Griffeys, Editor-in-Chief, Oliver Watt Meyer, Deputy Editor James
