A Physics-informed deep neural network for the joint prediction of chlorophyll-a and hydrographic fields in the Mediterranean Sea

Abstract

Monitoring the ocean's four-dimensional state is essential for marine ecosystem preservation. Artificial Intelligence (AI) algorithms represent promising tools to merge satellite and in situ measurements, improving reconstructions of ocean interior dynamics. Here, we describe 4DMED-bionet, an AI-based model developed under the European Space Agency 4DMED-Sea project, designed to infer subsurface properties from surface observations. Combining Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) layers, the model reconstructs 4D fields of temperature, salinity, density and chlorophyll-a in the Mediterranean Sea. The algorithm includes a physics-informed loss function that imposes constraints on density predictions, improving its accuracy without degrading other outputs. 4DMED-bionet outperforms different deep learning models, providing a high-quality 4D dataset, available at https://doi.org/10.25423/CMCC/4DMEDSEA_BIOPHYS_REP_3D. This dataset includes 4D geostrophic velocities derived from reconstructed physical tracers and surface geostrophic currents. Scientific analysis of 4D data is ongoing, aiming to better understand the processes that couple phytoplankton responses with 3D physical dynamic.