Enhancing sea surface height estimation using satellite-derived chlorophyll-a and temperature data via machine learning: a case study in the Gulf of Mexico

Satellite observations provide indispensable data that is assimilated into numerical ocean models to correct errors and biases. Traditionally, sea surface height (SSH) from satellite altimeter tracks, sea surface temperature (SST), and more recently, sea surface salinity (SSS), have been assimilated into these models. Temperature and salinity are part of the governing equations of ocean dynamics, and SSH is directly derived from the state of the resolved ocean, making these variables a first choice for data assimilation. However, satellite-derived Chlorophyll-a (Chl-a) data, which offer high-resolution information, is not typically assimilated. This is primarily because this variable is not solved by the physical models, and the biochemical models that simulate broader marine ecosystems, including phytoplankton dynamics and nutrient cycles which do estimate Chl-a, are computationally expensive and not used in operational models.

In this study, we utilize a ten-year free run of a biochemical ocean model of the Gulf of Mexico to simulate satellite observations, including altimeter tracks, SST,  SSS, and Chl-a. We trained and tested various machine learning architectures, including Convolutional Neural Networks (CNNs), Autoregressive Convolutional Neural Networks (AR-CNNs), and Vision Transformers, to learn the relationship between these variables and the SSH. The trained models were then used to estimate sea surface height from the simulated observations to estimate the current and future state of the sea surface height, leveraging the autoregressive properties of one of the tested architectures. Our results demonstrate that this approach outperforms the traditional interpolations in metrics like the RMSE. Finally, we applied the best-performing models to real satellite observations, highlighting the potential of improving SSH estimation quality.