ALBERT

Description: A series of synthetic data experiments is performed to investigate the ability of a regional atmospheric inversion to estimate grid-scale CO2 fluxes during the growing season over North America. The inversions are performed within a geostatistical framework without the use of any prior flux estimates or auxiliary variables, in order to focus on the atmospheric constraint provided by the nine towers collecting continuous, calibrated CO2 measurements in 2004. Using synthetic measurements and their associated concentration footprints, flux and model-data mismatch covariance parameters are first optimized, and then fluxes and their uncertainties are estimated at three different temporal resolutions. These temporal resolutions, which include a four-day average, a four-day-average diurnal cycle with 3-hourly increments, and 3-hourly fluxes, are chosen to help assess the impact of temporal aggregation errors on the estimated fluxes and covariance parameters. Estimating fluxes at a temporal resolution that can adjust the diurnal variability is found to be critical both for recovering covariance parameters directly from the atmospheric data, and for inferring accurate ecoregion-scale fluxes. Accounting for both spatial and temporal a priori covariance in the flux distribution is also found to be necessary for recovering accurate a posteriori uncertainty bounds on the estimated fluxes. Overall, the results suggest that even a fairly sparse network of 9 towers collecting continuous CO2 measurements across the continent, used with no auxiliary information or prior estimates of the flux distribution in time or space, can be used to infer relatively accurate monthly ecoregion scale CO2 surface fluxes over North America within estimated uncertainty bounds. Simulated random transport error is shown to decrease the quality of flux estimates in under-constrained areas at the ecoregion scale, although the uncertainty bounds remain realistic. While these synthetic data inversions do not consider all potential issues associated with using actual measurement data, e.g. systematic transport errors or problems with the boundary conditions, they help to highlight the impact of inversion setup choices, and help to provide a baseline set of CO2 fluxes for comparison with estimates from future real-data inversions.

Description: Given the large differences between biospheric model estimates of regional carbon exchange, there is a need to understand and reconcile the predicted spatial variability of fluxes across models. This paper presents a set of quantitative tools that can be applied for comparing flux estimates in light of the inherent differences in model formulation. The presented methods include variogram analysis, variable selection, and geostatistical regression. These methods are evaluated in terms of their ability to assess and identify differences in spatial variability in flux estimates across North America among a small subset of models, as well as differences in the environmental drivers that appear to have the greatest control over the spatial variability of predicted fluxes. The examined models are the Simple Biosphere (SiB 3.0), Carnegie Ames Stanford Approach (CASA), and CASA coupled with the Global Fire Emissions Database (CASA GFEDv2), and the analyses are performed on model-predicted net ecosystem exchange, gross primary production, and ecosystem respiration. Variogram analysis reveals consistent seasonal differences in spatial variability among modeled fluxes at a 1°×1° spatial resolution. However, significant differences are observed in the overall magnitude of the carbon flux spatial variability across models, in both net ecosystem exchange and component fluxes. Results of the variable selection and geostatistical regression analyses suggest fundamental differences between the models in terms of the factors that control the spatial variability of predicted flux. For example, carbon flux is more strongly correlated with percent land cover in CASA GFEDv2 than in SiB or CASA. Some of these factors can be linked back to model formulation, and would have been difficult to identify simply by comparing net fluxes between models. Overall, the quantitative approach presented here provides a set of tools for comparing predicted grid-scale fluxes across models, a task that has historically been difficult unless standardized forcing data were prescribed or a detailed sensitivity analysis was performed.

Description: Robust estimates of regional-scale terrestrial CO2 exchange are needed to support carbon management policies and to improve the predictive ability of models representing carbon-climate feedbacks. Large discrepancies remain, however, both among and between CO2 flux estimates from atmospheric inverse models and terrestrial biosphere models. Improved atmospheric inverse models that provide robust estimates at sufficiently fine spatial scales could prove especially useful for monitoring efforts, while also serving as a validation tool for process-based assumptions in terrestrial biosphere models. A growing network of continental sites collecting continuous CO2 measurements provides the information needed to drive such models. This study presents results from a regional geostatistical inversion over North America for 2004, taking advantage of continuous data from the nine sites operational in that year, as well as available flask and aircraft observations. The approach does not require explicit prior flux estimates, resolves fluxes at finer spatiotemporal scales than previous North American inversion studies, and uses a Lagrangian transport model coupled with high-resolution winds (i.e. WRF-STILT) to resolve near-field influences around measurement locations. The estimated fluxes are used in an inter-comparison with other inversion studies and a suite of terrestrial biosphere model estimates collected through the North American Carbon Program Regional and Continental Interim Synthesis. Differences among inversions are found to be smallest in areas of the continent best-constrained by the atmospheric data, pointing to the value of an expanded measurement network. Aggregation errors in previous coarser-scale inversion studies are likely to explain a portion of the remaining spread. The spatial patterns from a geostatistical inversion that includes auxiliary environmental variables from the North American Regional Reanalysis were similar to those from the median of the biospheric model estimates during the growing season, but diverged more strongly in the dormant season. This could be due to a lack of sensitivity in the inversion during the dormant season, but may also point to a lack of skill in the biospheric models outside of the growing season, particularly in agricultural areas. For the annual continental budget, the boundary conditions used as an input into the inversions were seen to have a substantial impact on the estimated net flux, with a difference of ~0.8 PgC yr−1 associated with results using two different plausible sets of boundary conditions.

Description: Given the large differences between biospheric model estimates of regional carbon exchange, there is a need to understand and reconcile the predicted spatial variability of fluxes across models. This paper presents a set of quantitative tools that can be applied to systematically compare flux estimates despite the inherent differences in model formulation. The presented methods include variogram analysis, variable selection, and geostatistical regression. These methods are evaluated in terms of their ability to assess and identify differences in spatial variability in flux estimates across North America among a small subset of models, as well as differences in the environmental drivers that best explain the spatial variability of predicted fluxes. The examined models are the Simple Biosphere (SiB 3.0), Carnegie Ames Stanford Approach (CASA), and CASA coupled with the Global Fire Emissions Database (CASA GFEDv2), and the analyses are performed on model-predicted net ecosystem exchange, gross primary production, and ecosystem respiration. Variogram analysis reveals consistent seasonal differences in spatial variability among modeled fluxes at a 1° × 1° spatial resolution. However, significant differences are observed in the overall magnitude of the carbon flux spatial variability across models, in both net ecosystem exchange and component fluxes. Results of the variable selection and geostatistical regression analyses suggest fundamental differences between the models in terms of the factors that explain the spatial variability of predicted flux. For example, carbon flux is more strongly correlated with percent land cover in CASA GFEDv2 than in SiB or CASA. Some of the differences in spatial patterns of estimated flux can be linked back to differences in model formulation, and would have been difficult to identify simply by comparing net fluxes between models. Overall, the systematic approach presented here provides a set of tools for comparing predicted grid-scale fluxes across models, a task that has historically been difficult unless standardized forcing data were prescribed, or a detailed sensitivity analysis performed.

Description: Atmospheric inversion models have the potential to quantify CO2 fluxes at regional, sub-continental scales by taking advantage of near-surface CO2 mixing ratio observations collected in areas with high flux variability. This study presents results from a series of regional geostatistical inverse models (GIM) over North America for 2004, and uses them as the basis for an inter-comparison to other inversion studies and estimates from biospheric models collected through the North American Carbon Program Regional and Continental Interim Synthesis. Because the GIM approach does not require explicit prior flux estimates and resolves fluxes at fine spatiotemporal scales (i.e. 1° × 1°, 3-hourly in this study), it avoids temporal and spatial aggregation errors and allows for the recovery of realistic spatial patterns from the atmospheric data relative to previous inversion studies. Results from a GIM inversion using only available atmospheric observations and a fine-scale fossil fuel inventory were used to confirm the quality of the inventory and inversion setup. An inversion additionally including auxiliary variables from the North American Regional Reanalysis found inferred relationships with flux consistent with physiological understanding of the biospheric carbon cycle. Comparison of GIM results with bottom-up biospheric models showed stronger agreement during the growing relative to the dormant season, in part because most of the biospheric models do not fully represent agricultural land-management practices and the fate of both residual biomass and harvested products. Comparison to earlier inversion studies pointed to aggregation errors as a likely source of bias in previous sub-continental scale flux estimates, particularly for inversions that adjust fluxes at the coarsest scales and use atmospheric observations averaged over long periods. Finally, whereas the continental CO2 boundary conditions used in the GIM inversions have a minor impact on spatial patterns, they have a substantial impact on the continental carbon budget, with a difference of 0.8 PgC yr−1 in the total continental flux resulting from the use of two plausible sets of boundary CO2 mixing ratios. Overall, this inter-comparison study helps to assess the state of the science in estimating regional-scale CO2 fluxes, while pointing towards the path forward for improvements in future top-down and bottom-up modeling efforts.

Description: Using synthetic continuous CO2 measurements from the nine sampling locations operational across North America in 2004, this paper investigates the optimal setup for, and constraint on fluxes achieved by, a regional geostatistical atmospheric CO2 inversion over the continent. The geostatistical framework does not require explicit prior flux estimates, nor any other process-based information, and is therefore particularly well suited for investigating the information content of the atmospheric CO2 measurements from a limited network. The atmospheric data are first used with the Restricted Maximum Likelihood (RML) algorithm to infer the model-data mismatch and a priori spatial covariance parameters applied in the inversion. The implemented RML algorithm is found to infer robust spatial covariance parameters from the atmospheric data, as compared to the "true" solution, for cases where the flux and measurement timescales match, while model-data mismatch variances are inferred correctly across all examined cases. A series of analyses is also performed investigating the impact of the temporal scale of concentration measurements and fluxes on inversion results. Inversions using measurement data at sub-daily resolution are found to yield fluxes with a lower Root Mean Square Error (RMSE) relative to inversions using coarser-scale observations, whereas the flux resolution appears to have a lesser impact on the inversion quality. In addition, night-time data for the tall and marine boundary layer towers are found to help constrain fluxes across the continent, although they can potentially bias near-field fluxes. These general conclusions are likely to also be applicable to inversions using a synthesis Bayesian inversion approach. Overall, despite the relatively sparse and unevenly distributed network of nine towers across the North American continent, a geostatistical inversion using an optimal setup and relying solely on the atmospheric data constraint is found to estimate the North American sink for June 2004 to within approximately 10%.