Object-oriented mesoscale eddy prediction

Mesoscale ocean eddies are dynamic structures controlling a significant proportion of water exchanges between the surface and the deep ocean, and therefore of heat, carbon and nutrient transfers. The eddy dynamics, i.e. changes in height, velocity and energy, are classically computed through complex ocean equations such as the quasi-geostrophic balance. However, those computations are time-consuming and slow down decision-making in operational situations. Some recent studies have managed to define eddy dynamics with simple properties - centroid position, amplitude, radius, current velocity, and horizontal displacement - and to predict their future evolution with machine learning models (Wang et al., 2020). We aim to implement a simple machine learning model to predict eddy properties that can reconstruct eddy dynamics and to include it in operational tools.

In this study, we simplified eddy structures, converting their 2D/3D gridded physical space into a parametric space, characterized by the eddy properties obtained with the AMEDA algorithm (Le Vu et al., 2017). Thus we considered eddies as 2D ellipse structures with additional properties - centroid position, amplitude, semi-axis of ellipse, rotation angle, maximal current velocity, and horizontal displacements. Explainable simple ML models were trained to learn the evolution of those parameters between two consecutive time steps. Here we selected two approaches of the least square regression model: the global linear regression on the whole training dataset and the local linear regression based on the nearest neighbors observations. Performances of each model are evaluated with the RMSE metric and compared to identify which model gives the most satisfactory results for eddy prediction. 

Our analysis shows better performances with the local linear regression. However, the choice of more adapted models or a better selection of eddy properties would enhance the prediction of eddies. The next steps to the inclusion of the model in operational tools will be the consideration of eddy interactions - splitting and merging -, the uncertainty quantification and the data assimilation of eddy dynamics with an object-oriented approach.

References

Wang, X., Wang, H., Liu, D., Wang, W., 2020. The Prediction of Oceanic Mesoscale Eddy Properties and Propagation Trajectories Based on Machine Learning. Water 12, 2521. https://doi.org/10.3390/w12092521

Le Vu, B., Stegner, A., Arsouze, T., 2018. Angular Momentum Eddy Detection and Tracking Algorithm (AMEDA) and Its Application to Coastal Eddy Formation. Journal of Atmospheric and Oceanic Technology 35, 739–762. https://doi.org/10.1175/JTECH-D-17-0010.1