A neural network-based monthly climatology of oceanic dissolved inorganic carbon in the upper 2000 m

We present a 2-step neural network method that uses direct measurements to estimate a global monthly climatology of dissolved inorganic carbon (DIC) in the upper 2000 m of the ocean. The method establishes a statistical relationship between better constrained global fields of physical and biogeochemical properties and available observations from the GLODAPv2 database to estimates the global fields of DIC. We test our method with synthetic data from a global hindcast simulation of an ocean biogeochemistry model and with independent time-series and float observations. Our estimate compares well, illustrated by the root mean squared error (RMSE) ranging between 13.1 and 25.5 μmol kg^-1. We find that the surface seasonal cycle of DIC in the northern high latitudes has the largest amplitudes (30 to more than 50 μmol kg^-1), while most of the remaining global ocean, including the southern high latitudes have a mean amplitude of ~5 to 20 μmol kg^-1. In both hemispheres, the months of the highest DIC tend to be in spring, when vertical mixing dominates the seasonal maximum. We further find substantial differences in the nodal depth of DIC, i.e. the depth, where the phase of the seasonal cycle of DIC shifts, varying from less than 50 m in the tropics to ~100 m in the high latitudes. With this study we created the first monthly climatology of DIC based on observations, demonstrating how the seasonal cycle of DIC varies globally and with depth.