Partitioning NEE Uncertainty Sources Using a FLUXCOM-X Ensemble

Quantifying uncertainties in data-driven upscaling of biogenic carbon fluxes is essential for improving our understanding of global carbon cycle processes and for providing robust priors to atmospheric inversion systems. However, most existing global carbon flux products derived from data-driven approaches either provide no uncertainty assessments, or are limited to incomplete sources.

In this study, we developed an ensemble of net ecosystem exchange (NEE) estimates within the FLUXCOM-X framework to systematically quantify uncertainty contributions from multiple sources across the entire carbon flux upscaling workflow. These sources include choices regarding eddy covariance (EC) measurement post-processing, meteorological forcing, predictor set, training data splitting, and machine-learning model. Specifically, we post-processed raw EC data using a Monte Carlo approach that randomly selects the friction velocity (u*) threshold for each site to quantify uncertainty related to the EC measurement. Meteorological uncertainty was represented using a 10-member, 3-hourly ERA5 ensemble. Predictor selection uncertainty was assessed by applying a hybrid genetic algorithm to select multiple “equally good” predictor combinations used to train predictor ensembles. In addition, uncertainties related to site representativeness and model structure were captured through alternative training data splits and by training machine-learning models (i.e., xGBoost, RF, and MLP) with different random seeds. As a result, a large ensemble of spatiotemporally explicit NEE estimates at hourly and 0.05 deg resolution was generated.

We further analyzed the relative contributions of these uncertainty sources to the total spatial and temporal uncertainty of NEE. Results for Europe indicate that predictor selection uncertainty dominates the upscaling uncertainty, followed by training data splitting uncertainty and EC post-processing uncertainty. In contrast, the ensemble spread associated with meteorological forcing and XGBoost models is relatively small, whereas MLP models exhibit substantially larger spread. The total uncertainty of the ensemble is not uniformly distributed across the study region; instead, it exhibits spatial hotspots particularly in Ireland, west of the United Kingdom, and the northern coast of Africa. The same ensemble-based methodology will next be applied globally to quantify and attribute regional NEE uncertainties worldwide.