Diagnosing uncertainty and improving predictions of terrestrial carbon dioxide fluxes at multiple scales through data assimilation

We use data assimilation techniques to assimilate carbon cycle data over a range of spatial and temporal scales in three models. First, global CO 2 concentration data and ocean flux observations are used to calibrate a simple carbon cycle model, to estimate uncertainty in key carbon cycle parameters, and to predict the behavior of both the terrestrial and ocean carbon sinks. Given equal observation errors, historical temperature forcing, and our assumed model structure, hypothetical annual observations of global terrestrial CO 2 fluxes reduce predictive uncertainties about CO2 sinks more than twice as much as annual observations of global ocean CO2 fluxes. The main reason for this effect lies in the interannual variability of the terrestrial carbon cycle that is resolved by our model and that constrains model parameters. This conclusion hinges on the model structure that assumes much smaller interannual variability of the oceanic carbon cycle.; Direct observations of terrestrial CO2 fluxes do not exist at a global scale, and data assimilation experiments using direct terrestrial flux observations must focus on smaller scales until appropriate upsealing techniques are discovered. Eddy covariance towers measure continuous fluxes of CO2 with a spatial footprint on the order of 1 km2. At the WLEF eddy covariance tower, which is in a northern Wisconsin mixed forest, interannual variations in observed sums of CO2 flux are found to be statistically significant with respect to random error. These flux data were assimilated into a simple climate-driven model. Considerable differences in respiration parameter probability density functions (PDFs) occurred depending on whether daytime, nighttime, or all flux data were used in the assimilation. This simple model was moderately successful in producing statistically significant correlations with interannual variations in annual and growing season net ecosystem exchange (NEE) sums, but was generally unsuccessful in spring and autumn. In all cases, the model underestimated the range of variability in NEE sums.; Upscaling eddy covariance tower fluxes from the scale of an eddy covariance footprint to a regional or global scale is a nontrivial task. Assimilating eddy covariance data into a process-based terrestrial carbon cycle model is one way to attempt this type of upscaling. We use the Top-down Representation of Interactive Foliage and Flora Including Dynamics (TRIFFID) to assimilate eddy covariance data from five sites in eastern North America. Because of the complexity of the model, only a subset of model parameters could be optimized; 22 parameters assumed to contribute the most to observed flux variance are selected. (Abstract shortened by UMI.)...