Three-dimensional structure reconstruction of ocean mesoscale eddies based on physical process modeling and data-driven machine learning

Abstract

To address the challenge of obtaining ocean mesoscale eddies' refined three-dimensional (3D) structure, we propose a novel 3D structure reconstruction model that combines physical process models with data-driven machine learning. First, based on the universal structure of mesoscale eddies, the 3D density structure of eddies is reconstructed using satellite observations and individual Argo profile observations. These eddy density profiles, along with eddy elements (polarity, eddy center, and radius) and sea surface elements (temperature, salinity, and dynamic height), serve as input data to construct a data-driven machine learning algorithm, which can reconstruct the 3D temperature and salinity structure of the eddies. Using observations of oceanic mesoscale cyclonic and anticyclonic eddies in the northwest Pacific Ocean, we demonstrate that both types of eddies' reconstructed temperature, salinity, and density structures align well with the observations. The root mean square errors (RMSEs) for the anticyclonic eddy are 0.361 °C, 0.0271 PSU, and 0.0570 kg/m³, and for the cyclonic eddy, they are 0.372 °C, 0.0904 PSU, and 0.144 kg/m³, respectively. The correlation coefficients exceed 0.98. Compared to multi-source fusion data (ARMOR 3D) and dynamical statistics data (MODAS), the reconstructed 3D structure from this study shows the closest alignment with observed structures. Furthermore, incorporating physical process model inputs significantly enhances the accuracy of the data-driven machine learning reconstruction of the eddy thermohaline structure, reducing the RMSEs by >40 % and 60 %, respectively.