Machine Learning-Based Reconstruction of T/S Profiles from Satellite-Derived SSH Using Transformer Models

Sea surface height (SSH) data derived from satellite altimetry are widely used in data assimilation to enhance the representations of ocean currents and subsurface temperature and salinity structure. However, accurately projecting SSH onto subsurface temperature and salinity structures presents significant challenges. Consistent adjustment to temperature and salinity profiles are required to conserve the potential vorticity, which depends on the vertical density gradient. Otherwise, SSH assimilation can produce adverse effects (Fu and Zhu, 2011). Several methods have been proposed to address this issue, including the CH96(Cooper and Haines, 1996) method used by Chang et al (2023), which constructs pseudo profile derived from altimetry data by preserving density structures. However, when tidal forcing is applied to an ocean model, the CH96 method becomes challenging to use due to the significant difficulty in removing tidal signals. To overcome these limitations, this study proposes a Transformer-based machine learning approach to reconstruct T/S (Temperature and Salinity) profiles from SSH. Transformers are well-suited for capturing complex correlations through attention mechanisms (Vaswani et al., 2017), making them ideal for learning T/S profiles influenced by diverse and intricate variables. Monthly GLORYS data from 2010 to 2020 was utilized to train a model for reconstructing T/S profiles. The data was structured into 1/2° grids, where learning was conducted grid-by-grid to capture spatiotemporal variability. For improved accuracy and better incorporation of surrounding grid influences, a combination of 4D-Var techniques and CNNs was employed. This approach learns patterns by grouping four neighboring grids into a quadrilateral for joint training, ensuring that the final profiles account for interactions across grids. During prediction, the surface information of a target point is distributed to its four neighboring low-resolution grids to generate profiles, which are then interpolated into a high-resolution 1/12° grid. The final profile is computed using inverse distance weighting (IDW) interpolation, prioritizing the influence of closer profiles for spatial consistency. Model performance was validated by comparing predicted profiles with low-resolution maps for 2021–2022 over the northwest Pacific region (10°S–45°N, 120°–170°E), achieving an RMSE of 0.55 for temperature and 0.12 for salinity. The model will be further validated against in-situ observational data. We plan to conduct experiments to investigate the impact of assimilation of the reconstructed profiles and compared against CH96-derived profiles to evaluate their accuracy and advantages.