Comparison of GLUE and DREAM for the estimation of cultivar parameters in the APSIM-maize model

摘要

Process-based crop models are popular scientific tools to study the impacts of environmental conditions and management decisions on crop growth. Some cultivar parameters in crop models cannot be measured directly and need to be estimated. In this research, two most popular Bayesian methods, namely generalized likelihood uncertainty estimation (GLUE) and Differential Evolution Adaptive Metropolis (DREAM), were used for the first time to estimate parameters of the maize module of the Agricultural Productions Systems sIMulator model (APSIM-maize). Both theoretical and real-world evaluations were conducted to compare the performances of these two methods. The maize yields from 2003 to 2006 were used for model calibration, and the yields from 2007 to 2013 were used for model validation. Both GLUE and DREAM performed well in the theoretical and real-world evaluation. During the validation period (2007-2013), when the heteroscedastic model error assumption was adopted in DREAM, on average approximate 90% of observed yield values were captured in the 95% confidence band of DREAM (P-factor= 90.47%), which was larger than that using GLUE (P-factor= 80.93%). Meanwhile the uncertainty bands of DREAM (R-factor= 4.42) were wider than those of GLUE (R-factor= 2.32). If only one parameter set was allowed to be used in the simulation, the weighted mean parameter values according to the likelihood of each parameter set performed better than the parameter set with maximum likelihoods for GLUE while the opposite is true for DREAM. But considering future analysis in the real-world evaluation, the moderate performance of these two methods suggests that a single parameter set obtained by neither GLUE nor DREAM is satisfactory and ensemble simulation is needed. Overall, GLUE and DREAM had similar performance, but GLUE is more convenient and simpler to use than DREAM. So we think GLUE is a better choice than DREAM for estimating cultivar parameters of APSIM-maize.