Constraining a data-driven CO ₂ flux model by ecosystem and atmospheric observations using atmospheric transport

Abstract. Global estimates of the net ecosystem exchange of CO2 (NEE) from data-driven models differ widely depending on their underlying data and methodology. Bottom-up models trained on eddy-covariance data are most informative at the ecosystem-level. Top-down models, such as atmospheric inversions, produce regional and global results consistent with the observed atmospheric growth rate, accurately capturing the interannual variability (IAV) of NEE. Both approaches have limitations estimating NEE across scales: Bottom-up models can miss large-scale dynamics of NEE when aggregated globally. Top-down approaches have difficulty relating the large-scale atmospheric signal to biophysical processes at smaller scales. To address these limitations, we create a model that uses a hybrid combination of direct observations and atmospheric dynamics to integrate ecosystem-level eddy-covariance data and atmospheric CO2 mole fraction data into a single coherent ecosystem-level flux model. Aggregated globally, our new model estimates an annual sink with a low bias, and consistent IAV when compared with independent estimates. The IAV of the estimated NEE is closer in magnitude to an ensemble of atmospheric inversions, and our model produces a higher temporal coefficient of correlation with these data than state-of-the-art bottom-up data-driven models. This improvement in IAV is achieved without direct access to the observed variability of the atmosphere: the model is trained using only one year of daytime observations from 3 tall-tower observatories. No atmospheric information is available to the model during the production of global NEE estimates. This shows the efficiency of our method in synthesizing top-down information into bottom-up mapping of flux-environment relationships.