Evaluating terrestrial ecosystem model performance: An application of uncertainty in eddy covariance CO2 flux measurements

Terrestrial CO2 fluxes from eddy covariance measurements are valuable resources for model identification, calibration, and validation. However, these flux observations are substantially subject to random measurement errors. Accounting for such flux observation errors is a prerequisite for model evaluation and quantitative specification of model errors. We applied a Monte Carlo approach (residual bootstrapping) to estimate the flux variability and the true flux of individual 30- or 60-minute observations. We used multi-model ensemble means of present-day ecosystem model outputs as initial estimates of CO2 fluxes, from which we obtained model residuals against observations. The residuals were then resampled and added back to the ensemble mean values. For each 30- or 60-minute time step we examined whether a 95% confidence interval for a true flux captured an observed flux with a probability of 0.95. We also assessed if estimated true fluxes agree with observations in terms of the magnitude at a monthly scale. We compared synthetic CO2 fluxes from 15 ecosystem models with 'true flux' data from nine Ameriflux towers to identify systematic model errors at different times during the year and at different frequencies using wavelet analysis. The results of this study illustrate the degree to which estimated true fluxes at multiple frequencies can be used as the basis for quantifying uncertainty of present-day ecosystem land models.