Transfer Learning-driven Retrieval of Subsurface Temperature and Salinity for Mesoscale Eddies in the Oyashio Current

Understanding mesoscale eddies’ dynamics and their impact on oceanic processes requires accurate retrieval of subsurface temperature and salinity distributions. Traditional methods face challenges in resolving mesoscale features due to limited data availability and complex nonlinear dynamics. In this study, we propose a novel approach leveraging transfer learning and deep learning techniques for the retrieval of subsurface temperature and salinity associated with mesoscale eddies in the Oyashio Current region. Transfer learning enables the utilization of pre-trained deep learning models, adapted to our specific task through fine-tuning on domain-specific data. Results demonstrate remarkable accuracy, with the root mean square error (RMSE) reaching a minimum of 0.09°C (0.01 psu) for temperature (salinity) within anticyclonic eddies and 0.10°C (0.02 psu) within cyclonic eddies in the upper 1000 m. By integrating this approach with oceanographic datasets, our method not only enhances the understanding of eddy dynamics but also holds promise for applications in ocean monitoring, climate studies, and marine resource management.