Characterizing uncertainty for regional carbon cycle modeling

Abstract

The discrepancy between current estimates of carbon dioxide emissions from natural processes, disturbance of the land surface and fossil fuel combustion, and the magnitude of accumulated CO2 in the atmosphere as measured by the global flask network indicates that there must be a terrestrial and/or oceanic repository for the unaccounted carbon. The imbalance in estimates has become a major focus of research during the last two decades, particularly since our ability to predict and manage CO2 concentrations in the future requires a thorough understanding of this disparity. A major challenge of current carbon cycle research is to reconcile the differences between top-down (eg. global inversions) and bottom-up (eg. process modeling, inventories) approaches. The research presented in this dissertation addresses some of the key aspects of model-data fusion that will be important within the framework of continental scaling efforts such as the North American Carbon Program. The geographical locus of the research is the WLEF-TV tower, a very tall tower located in the Chequamegon National Forest near Park Falls, Wisconsin. Micrometeorological and eddy covariance measurements have been made since 1995 at three levels (30, 122, 396 meters) and many additional studies are underway in the region, forming the Chequamegon Ecosystem-Atmosphere Study (ChEAS). The vegetation of the region is heterogeneous, composed of upland Jack and Red Pine, mixed hardwoods and lowland conifers. In a series of analyses, sources of error, potential for bias and impacts of model and data choices in the context of regional carbon cycle modeling of temperate, forested ecosystems were evaluated. A comprehensive sensitivity analysis was conducted using a Monte-Carlo style framework to evaluate the sensitivity of a commonly used, complex biophysical land surface model (Simple Biosphere Model, v.2; SiB2) to its parameterization through time. The results of the sensitivity analysis were then used within a Bayesian framework to calculate the uncertainty of land surface fluxes predicted by the model attributable to the parameterization. Finally the potential for uncertainty due to the spatial representation of the land surface was evaluated with a version of the SiB2 model coupled to the Regional Atmospheric Modeling System (RAMS). The results of this dissertation suggest that: there is a quantifiable level of model-data mismatch that could be used as a prior uncertainty in regional atmospheric inversions, compensation among parameters in complex biophysical land surface models complicates model optimization, uncertainty due to model parameterization was greater than the uncertainty attributable to mischaracterization of land surface heterogeneity and the potential for reduction in uncertainty appears to be greatest during the periods of rapid land surface change surrounding leaf-on in the spring and leaf-off in the autumn.