The Impact of Error Accounting in a Bayesian Approach to Calibrating Modeled Turbulent Fluxes in an Open-Canopy Forest

Abstract

AbstractLand surface models are notorious for containing many parameters that control the exchange of heat and moisture between land and atmosphere. Properly modeling the partitioning of total evapotranspiration (ET) between transpiration and evaporation is critical for accurate hydrological modeling, but depends heavily on the treatment of turbulence within and above canopies. Previous work has constrained estimates of evapotranspiration and its partitioning using statistical approaches that calibrate land surface model parameters by assimilating in situ measurements. These studies, however, are silent on the impacts of the accounting of uncertainty within the statistical calibration framework. The present study calibrates the aerodynamic, leaf boundary layer, and stomatal resistance parameters, which partially control canopy turbulent exchange and thus the evapotranspiration flux partitioning. Using an adaptive Metropolis?Hastings algorithm to construct a Markov chain of draws from the joint posterior distribution of these resistance parameters, an ensemble of model realizations is generated, in which latent and sensible heat fluxes and top soil layer temperature are optimized. A set of five calibration experiments demonstrate that model performance is sensitive to the accounting of various sources of uncertainty in the field observations and model output and that it is critical to account for model structural uncertainty. After calibration, the modeled fluxes and top soil layer temperature are largely free from bias, and this calibration approach successfully informs and characterizes uncertainty in these parameters, which is essential for model improvement and development. The key points of this paper are 1) a Markov chain Monte Carlo calibration approach successfully improves modeled turbulent fluxes; 2) ET partitioning estimates hinge on the representation of uncertainties in the model and data; and 3) despite these inherent uncertainties, constrained posterior estimates of ET partitioning emerge.