Observation System Simulation Experiments for surface ocean pCO2 reconstructions in the Atlantic Ocean

Global estimates of the ocean carbon sink are released with a yearly frequency as part of the global carbon budget. However, these global estimates hide important spatial and temporal variabilities that can only partly be resolved by direct in situ observations. In this work we explore options for future observational network design combining data streams from various platforms. Our objective is to identify an optimal observational network for surface ocean pCO₂ in the Atlantic Ocean and the Atlantic sector of the Southern Ocean. For this purpose, eleven Observation System Simulation Experiments (OSSEs) were performed. Each OSSE is a Feed-Forward Neural Network (FFNN) that is based on different data distributions and provides ocean surface pCO₂ for the period 2008-2010 with a 5-day time interval. Based on the geographical and time positions from three observational platforms, volunteering observing ships (VOS), Argo floats and OceanSITES moorings, pseudo-observations were constructed using the outputs from an online-coupled physical-biogeochemical global ocean model with a 0.25º nominal spatial resolution. The aim of this work was to find an optimal spatial distribution of observations to supplement the widely used Surface Ocean CO₂ Atlas (SOCAT) and to improve the accuracy of ocean surface pCO₂ reconstructions. OSSEs showed that the additional data from mooring stations and an improved coverage of the southern Hemisphere with biogeochemical ARGO floats corresponding to at least 25% of the density of active floats (2008-2010) would significantly improve the pCO₂ reconstruction and reduce the bias of derived estimates of sea-air CO₂ fluxes by 77%. The use of only SOCAT data results in a correlation coefficient of 0.67 compared to the ocean model output and a 26.08 𝜇atm standard deviation (25.34 𝜇atm for the model reference) over the chosen regions. While the best OSSE has a correlation coefficient of 0.85 and 24.89 𝜇atm for standard deviation. These results are close to the unrealistic benchmark case with total and only Argo float distribution over 2008-2010: 0.87 and 23.79𝜇atm. The reconstructed average pCO₂ over the whole region is also close to the model reference, ~370 𝜇atm and ~371 𝜇atm, respectively. The integrated air-sea fluxes fCO₂ are about -0.83 Pg/yr (best OSSE) and -0.76 Pg/yr (model reference).