Constraining biospheric carbon dioxide fluxes by combined top-down and bottom-up approaches

Abstract. While the growth rate of atmospheric CO₂ mole fractions can be measured with high accuracy, there are still large uncertainties in the attribution of this growth to diverse anthropogenic and natural sources and sinks. One major source of uncertainty is the net flux of carbon dioxide from the biosphere to the atmosphere, the Net Ecosystem Exchange (NEE). There are two major approaches to quantifying NEE: top-down approaches that typically use atmospheric inversions, and bottom-up estimates which rely on process-based or data-driven terrestrial biosphere models or inventories. Both approaches have known limitations. Atmospheric inversions produce estimates of NEE that are consistent with the atmospheric CO₂ growth rate at regional and global scales, but are highly uncertain at smaller scales. Bottom-up data-driven flux models match local observations of NEE, but have difficulty in accurately upscaling to a global estimate. We combine the two approaches, constraining a bottom-up data-driven flux model trained on meteorological, remotely-sensed, and eddy-covariance data with regional estimates of NEE derived from an ensemble of atmospheric inversions.

We link the two approaches using a region-specific sparse linear model for 18 regions consistent with the Regional Carbon Cycle Assessment and Processes-2 (RECCAP2) , which allows us to quickly generate regional estimates of NEE based on the data-driven flux model by simulating only a small number of optimally representative pixels. These regional totals then become part of a machine-learning objective function that compares them with top-down regional estimates from an ensemble of atmospheric inverse models. By adding this additional constraint from the top-down objective term, we produce a new “dual-constraint” data-driven flux model that is informative across spatial scales, producing consistent estimates both of the local per-pixel flux and at regional and global scales.

The inferred global terrestrial carbon flux from land, excluding fires and riverine evasion across 2001–2017 is -3.14±1.75 PgC year^-1 (±1 σ). This is a strong improvement over the -20.28±1.75 PgC year^-1 from the exact same data-driven flux model trained without the additional regional top-down constraint (i.e., single constraint) when compared to current best estimates of the global carbon flux from land. The shift in the carbon flux from land estimated by the model with the additional atmospheric constraint occurs largely in tropical regions where the data-driven flux model is poorly constrained, or affected by biased observations of NEE derived from difficult micrometeorological conditions. In extratropical regions, the estimated NEE from dual and single constraint data-driven flux models are very similar, reflecting the denser observational networks of ecosystem fluxes and atmospheric CO₂. Our approach, training a data-driven flux model with multiple constraints at site level and continentally integrated scales, and different temporal resolutions, opens new avenues for data-driven flux models constrained by other observations of atmospheric carbon dioxide, making use of the wealth of available Earth observation data.