Uncertainty in crop phenology simulations is driven primarily by parameter variability

Abstract

AbstractCrop phenology has a major influence on crop yield and is a major aspect of crop response to global warming. Process-based models of phenology are often used to predict the effect of weather on the development rate of crops through their growth phases, but such models are associated with large uncertainties, as demonstrated by the large variability between simulation results of different models. The purpose of this study is to estimate the relative importance of model structure uncertainty (due to uncertainty in the model equations) and model parameter uncertainty. Previous studies have assumed that the choice of model parameters to be fitted to data is fixed, and have evaluated the effect of uncertainty in those parameters. This underestimates parameter variance, as it ignores uncertainty in many aspects of the calibration approach, in particular the choice of parameters to estimate, as well as uncertainty in the parameters not fitted by calibration. Here, we propose and apply two approaches for estimating parameter variance that take into account uncertainty in all aspects of calibration, and, for one approach, uncertainty in parameters not fitted by calibration. Both approaches are based on previously reported large multi-model studies. The first approach uses components of variance analysis, which is possible because these studies included different modeling groups using the same model structure. The second approach is based on a study where multiple modeling groups each applied two different calibration procedures. The variance calculated from the two calibration approaches gives an estimate of parameter variance. Both methods give an overall result of 69% of total variance coming from parameter variance, for those simulated variables for which there are observed data. We conclude that in order to reduce uncertainty in crop phenology predictions, improving and sharing calibration procedures is as or more important than improving model structure.