Reconstruction of Argo Float Temperature Data in the Mediterranean Sea Using Physics-Informed Deep Learning and Remote Sensing

Accurate reconstruction of subsurface temperature profiles is essential for advancing our understanding of ocean dynamics and climate variability. In the Mediterranean Sea, Argo float temperature observations between 10 dbar and 500 dbar are often sparse or uncertain, limiting their utility for operational oceanography and climate studies. To address this challenge, we propose a physics-informed deep learning framework that leverages spatial-temporal dependencies and integrates auxiliary data from remote sensing and simulation products. The model incorporates wind stress, absolute dynamic topography, sea surface temperature, and simulated temperature and salinity fields from Copernicus datasets to reconstruct and correct uncertainties in Argo float data.

Our framework employs a deep neural network architecture augmented with physics-informed loss functions (PINNs), ensuring consistency with oceanographic principles such as temperature-salinity relationships and geostrophic balance. Evaluation metrics, including root mean-squared error (RMSE), structural similarity index (SSIM), and PINN-based loss terms, are utilized to quantify the model's accuracy and adherence to physical laws. During testing, the model is validated by reconstructing observed Argo temperature profiles and comparing them against independent datasets.

Preliminary results demonstrate that the proposed approach significantly improves the reconstruction of missing temperature profiles, achieving reduced RMSE, high SSIM values, and strong alignment with physical constraints. This study highlights the potential of combining physics-informed deep learning with remote sensing to enhance the reliability and accuracy of observational datasets in complex marine environments like the Mediterranean Sea.