Sensitivity, uncertainty and time dependence of parameters in a complex land surface model

Abstract

This paper explores the use of Monte Carlo carbon cycle data assimilation within the generalized likelihood uncertainty estimation (GLUE) framework to evaluate the sensitivities of a well-known complex land surface model (SiB v2.5) to its parameterization and the predictive uncertainty of simulated fluxes on a monthly basis, and for an entire year, at the WLEF tall-tower site in Park Falls, Wisconsin. An analysis is described wherein randomly generated parameter sets were ranked based on their capacity to simulate fluxes of latent (LE) and sensible heat (H) and the net ecosystem exchange of carbon (NEE) for each month of the year 1997 and for the entire year. Two criteria were used to evaluate the success of the simulations; the first evaluated the ability of SiB2.5 to simulate LE and H, the second included NEE as an additional constraint. The best-performing parameter sets for each criterion were used to assess model sensitivity to parameters, to calculate uncertainty bounds for predicted LE, H and NEE and to assess the information content of eddy covariance data on the analyzed time scales. Patterns in the sensitivity of the model to its parameterization and the uncertainty of the predictions were related to the physiological and phenological characteristics of the ecosystem, model structure and the relationship between deterministic models and comparatively stochastic measurements. The results show that model sensitivity varies through time for a larger set of parameters than those typically considered time varying in LSMs, thus optimization of model parameters on tower flux data should allow for variability at sub-annual time scales in order to capture the most information and best simulate fluxes. Further, constraining predictions annually versus monthly showed that some quantities (e.g. nighttime NEE) were on average better constrained annually, whereas other quantities that show more variability with vegetation phenology and structure (e.g. daytime NEE and LE) were better constrained monthly. The addition of the net ecosystem exchange of carbon in the data assimilation scheme improved model results by (a) constraining model parameterization (optimal parameter values), particularly during times of the year when the land surface was rapidly changing (spring and fall), and increasing the number of influential parameters, and (b) decreasing the uncertainty in NEE simulations (but not appreciably reducing uncertainty in LE and H simulations). We also found that there was an irreducible level of mismatch between the simulated and observed fluxes that could not be overcome through optimization due to variability in the observations and/or structural problems with the model. The uncertainty estimates can be used to characterize uncertainty in the simulations at multiple time scales.