Error characterization of methane fluxes and budgets derived from a long-term comparison of open- and closed-path eddy covariance systems.

Wetlands represent the dominant natural source of methane (CH) to the atmosphere. Thus, substantial effort has been spent examining the CH budgets of global wetlands continuous ecosystem-scale measurements using the eddy covariance (EC) technique. Robust error characterization for such measurements, however, remains a major challenge. Here, we quantify systematic, random and gap-filling errors and the resulting uncertainty in CH fluxes using a 3.5 year time series of simultaneous open- and closed path CH flux measurements over a sub-boreal wetland. After correcting for high- and low frequency flux attenuation, the magnitude of systematic frequency response errors were negligible relative to other uncertainties. Based on three different random flux error estimations, we found that errors of the CH flux measurement systems were smaller in magnitude than errors associated with the turbulent transport and flux footprint heterogeneity. Errors on individual half-hourly CH fluxes were typically 6%-41%, but not normally distributed (leptokurtic), and thus need to be appropriately characterized when fluxes are compared to chamber-derived or modeled CH fluxes. Integrated annual fluxes were only moderately sensitive to gap-filling, based on an evaluation of 4 different methods. Calculated budgets agreed on average to within 7% (≤ 1.5 g - CH m yr). Marginal distribution sampling using open source code was among the best-performing of all the evaluated gap-filling approaches and it is therefore recommended given its transparency and reproducibility. Overall, estimates of annual CH emissions for both EC systems were in excellent agreement (within 0.6 g - CH m yr) and averaged 18 g - CH m yr. Total uncertainties on the annual fluxes were larger than the uncertainty of the flux measurement systems and estimated between 7-17%. Identifying trends and differences among sites or site years requires that the observed variability exceeds these uncertainties.