Reducing sampling error in Rocky Mountain atmospheric carbon dioxide time series to improve flux retrievals

Establishing accurate CO₂ fluxes by atmospheric inverse approaches in mountainous terrain depends critically on the coverage of observed CO₂ concentrations. The Regional Atmospheric Continuous CO₂ Network in the Rocky Mountains (Rocky RACCOON, www.raccoon.ucar.edu) provides open access to atmospheric carbon dioxide measurements covering the central and southern U.S. Mountain West, a region that is not well represented by models that rely on ground-based measurements of CO₂ to optimize CO₂ fluxes. In terrain, these data are often contaminated by CO₂ observations representative of local air masses that are not well-mixed, thus complicating inversions. Here we test several filtering methods for screening locally biased and transient measurements on Rocky Mountain CO₂ concentration data in order to reduce sampling error (i.e. measurements not indicative of regionally well-mixed air), and to improve flux retrievals.

Reducing sampling error is important to atmospheric budgets and inversions, which are preferred in mountainous regions because measuring CO₂ flux using contemporary eddy covariance techniques in mountainous terrain can lead to regionally unrepresentative results. Coupled transport-biosphere inverse models such as CarbonTracker optimize CO₂ flux estimates by assimilating CO₂ concentration data across a network of in situ sensors. These inverse methods are limited by the representativeness and coverage of assimilated data, which underscores the importance of reducing sampling error in CO₂ concentration data.

To examine our capacity to identify CO₂ observations made during well-mixed conditions we tested three discrete-time filters. The first filter, a combined sample error and detection error filter, is a convenient routine that identifies anomalies by excessive hourly variance and by variance of simultaneous measurements across multiple inlet heights on the tower. The second method, statistical interpolation, filters for near global scale (baseline) CO₂ observations using an iterative approach. The final method is an efficient filter that uses a dynamically determined window size. We use a set of statistical metrics to analyze the differences and agreement among methods, their product time series, their power spectra. Additionally differences for several case studies of synoptic and local scale events are presented.