Towards a mature framework for integrating bottom-up and top-down constraints in a data-driven ecosystem-level CO2 model

    The net ecosystem exchange of CO₂ (NEE) between the land and the atmosphere is a critical term in the global carbon budget. Because of the complexity of modeling NEE across scales, global estimates of NEE are subject to large uncertainties. The two major data-driven approaches to modeling NEE are commonly described as top-down and bottom-up. Top-down models create a estimate of NEE which is optimally consistent with observations of atmospheric CO₂ from tower, aircraft, and increasingly satellite sensors. Bottom-up NEE models learn a statistical relationship between a set of ecosystem-level biophysical drivers and observations of NEE, often from the global eddy-covariance network. These models are then upscaled to the globe using remotely sensed data. These systems are critical to earth system science. However, both are subject to limitations and disagreement based on the particular view which they represent.

    In previous work we presented two frameworks for integrating top-down and bottom-up approaches. Both studies build a data-driven bottom-up NEE model trained from eddy-covariance data, which is also constrained by atmospheric information. The atmospheric constraint in the first study, derived statistically from an ensemble of atmospheric inversions, created a model which strongly adjusted regional and global model results towards top-down and independent estimates of NEE, albeit with limited improvement in the model’s spatial and temporal representation of NEE. The atmospheric constraint in the second study, derived from direct observations of atmospheric CO2 using a Lagrangian atmospheric transport model, improved the representation of NEE in biomes which are under represented in the eddy-covariance record. This resulted in an improved representation of the dynamics of NEE, producing spatial and temporal variability which better represents independent estimates and our current ecological understanding. However, the second atmospheric constraint produced a model with high internal uncertainty, and which underperformed at the regional and latitudinal scale, producing less plausible annual timeseries and mean seasonal cycles when compared with other bottom-up data-driven models.

    In the current work, we present a model which combines these two constraint techniques: The model uses 1) a core constraint from eddy-covariance, 2) A statistical constraint from atmospheric inversions to limit the possible solutions, reducing uncertainty, and improving regional results, and 3) an atmospheric constraint from direct observations of atmospheric CO₂ to improve the representation of the regional and global dynamics of NEE. Using the three constraints in parallel, the new model produces an estimate of global NEE which preserves the strengths of the two previous studies. When compared with state-of-the-art bottom-up models, it produces improved regional results, consistent spatial and temporal dynamics, and lower internal uncertainty. When transported through the atmosphere, the new model produces realistic estimates of atmospheric CO₂. In this way, we demonstrate the progress towards a mature hybrid framework, which can inherit the strengths of both bottom-up and top-down approaches.