Constraining ecosystem carbon dynamics in a data-limited world: integrating ecological "common sense" in a model-data-fusion framework.

Many of the key processes represented in global terrestrial carbon models remain largely unconstrained. For instance, plant allocation patterns and residence times of carbon pools are poorly known globally, except perhaps at a few intensively studied sites. As a consequence of data scarcity, carbon models tend to be underdetermined, and so can produce similar net fluxes with very different parameters and internal dynamics. To address these problems, we propose a series of ecological and dynamic constraints (EDCs) on model parameters and initial conditions, as a means to constrain ecosystem variable inter-dependencies in the absence of local data. The EDCs consist of a range of conditions on (a) carbon pool turnover and allocation ratios, (b) steady state proximity, and (c) growth and decay of model carbon pools. We use a simple ecosystem carbon model in a model-data fusion framework to determine the added value of these constraints in a data-poor context. Based only on leaf area index (LAI) time series and soil carbon data, we estimate net ecosystem exchange (NEE) for (a) 40 synthetic experiments and (b) three AMERIFLUX tower sites. For the synthetic experiments, we show that EDCs lead to an an overall 34% relative error reduction in model parameters, and a 65% reduction in the 3 yr NEE 90% confidence range. In the application at AMERIFLUX sites all NEE estimates were made independently of NEE measurements. Compared to these observations, EDCs resulted in a 69–93% reduction in 3 yr cumulative NEE median biases (−0.26 to +0.08 kg C m−2), in comparison to standard 3 yr median NEE biases (−1.17 to −0.84 kg C m−2). In light of these findings, we advocate the use of EDCs in future model-data fusion analyses of the terrestrial carbon cycle.