Bayesian estimation of uncertainty in land surface-atmosphere flux predictions

This study addresses the assessment of uncertainty associated with predictions of land surface-atmosphere fluxes using Bayesian Monte Carlo simulation within the generalized likelihood uncertainty estimation (GLUE) methodology. Even a simple soil vegetation-atmosphere transfer (SVAT) scheme is shown to lead to multiple acceptable parameterizations when calibration data are limited to timescales of typical intensive field campaigns. The GLUE methodology assigns a likelihood weight to each acceptable simulation. As more data become available, these likelihood weights may be updated by using Bayes equation. Application of the GLUE methodology can be shown to reveal deficiencies in model structure and the benefit of additional calibration data. The method is demonstrated with data sets taken from FIFE sites in Kansas, and ABRACOS data from the Amazon. Estimates of uncertainty are propagated for each data set revealing significant predictive uncertainty. The value of additional periods of data is then evaluated through comparing updated uncertainty estimates with previous estimates using the Shannon entropy measure.