Reflex Project: Using Model-Data Fusion to Characterize Confidence in Analyzes and Forecasts of Terrestrial C Dynamics

The Regional Flux Estimation Experiment, REFLEX, is a model-data fusion inter-comparison project, aimed at comparing the strengths and weaknesses of various model-data fusion techniques for estimating carbon model parameters and predicting carbon fluxes and states. The key question addressed here is: what are the confidence intervals on (a) model parameters calibrated from eddy covariance (EC) and leaf area index (LAI) data and (b) on model analyses and predictions of net ecosystem C exchange (NEE) and carbon stocks? The experiment has an explicit focus on how different algorithms and protocols quantify the confidence intervals on parameter estimates and model forecasts, given the same model and data. Nine participants contributed results using Metropolis algorithms, Kalman filters and a genetic algorithm. Both observed daily NEE data from FluxNet sites and synthetic NEE data, generated by a model, were used to estimate the parameters and states of a simple C dynamics model. The results of the analyses supported the hypothesis that parameters linked to fast-response processes that mostly determine net ecosystem exchange of CO2 (NEE) were well constrained and well characterised. Parameters associated with turnover of wood and allocation to roots, only indirectly related to NEE, were poorly characterised. There was only weak agreement on estimations of uncertainty on NEE and its components, photosynthesis and ecosystem respiration, with some algorithms successfully locating the true values of these fluxes from synthetic experiments within relatively narrow 90% confidence intervals. This exercise has demonstrated that a range of techniques exist that can generate useful estimates of parameter probability density functions for C models from eddy covariance time series data. When these parameter PDFs are propagated to generate estimates of annual C fluxes there was a wide variation in size of the 90% confidence intervals. However, some algorithms were able to make effective estimates of annual fluxes within relatively small CIs. In making predictions of C fluxes most algorithms did not increase the size of confidence intervals relative to analyses. How CIs grow over forecast periods of multiple years needs to be better understood. Data on slow, large C pools should be included in assimilation experiments, even with large confidence intervals, because such data constrain the parameters poorly served by eddy covariance data.