[A-TT35] Machine Learning Techniques in Weather, Climate, Ocean, Hydrology and Disease Predictions

[ATT35-P02] Machine Learning Approach to Reconstruct Vertical T/S Profiles from Satellite-Derived SSH.

*Geonmin Lee¹, YoungHo Kim (1.Pukyong University)

Keywords:Temperature and Salinity profile, Transformer model

Satellite observations of sea surface height (SSH) have beenwidely used for data assimilation to improve the representation of ocean currents and vertical hydrographic structures. However, accurately converting SSH into vertical temperature and salinity (T/S) profiles remains a significant challenge, as it requires precise adjustments to density structures in order to preserve potential vorticity. The CH96 method (Cooper and Haines, 1996), employed by Chang et al. (2023), offers a solution to this challenge but faces limitations, particularly in removing tidal signals. Instead, we propose a machine learning-based approach to directly reconstruct T/S profiles. This method leverages machine learning techniques, including Transformer models and Convolutional Neural Networks (CNN). The training dataset comprises the GLORYS monthly data from 2010 to 2020, along with surface variables such as absolute dynamic topography (ADT), sea surface temperature (SST), sea surface salinity (SSS), and geostrophic u and v components. For reconstructing T/S profiles at a single point, the Transformer model is trained on a 0.5° resolution grid. The trained model generates low-resolution prediction maps and then interpolated to high-resolution (1/12°) maps using a 4D-Variational (4D-Var) assimilation basedconvolution neuron network (CNN). This interpolation captures interactions between neighboring grids and introduces additional spatiotemporal effects. Finally, the profile is computed using Inverse Distance Weighting (IDW) interpolation with the four nearest grids. Model performance was validated by comparing the low-resolution maps, which achieved a temperature RMSE of 0.45°C and a salinity RMSE of 0.09 in the northwestern Pacific (5°S–65°N, 99°–170°E), with the IDW-interpolated maps, which resulted in a temperature RMSE of 0.60°C and a salinity RMSE of 0.09 when compared to the 1,085 observations from the World Ocean Database. Future research will focus on further validating the model against in-situ observational data and conducting assimilation experiments using profiles, which will be applied to the ocean circulation model to evaluate the improvements in themodel’s initial state.