A framework to quantify uncertainty of crop model parameters and its application in arid Northwest China

Crop modeling is affected by parameter uncertainty. We proposed a framework that integrates sensitivity, uncertainty and parameter calibration of crop models, to provide prediction intervals in place of single values for decision-makers to reduce management risks in agriculture. The framework includes four steps: (1) set prior distributions of parameters and collect measured data, (2) use Morris screening to find out sensitive parameters, (3) adopt Metropolis-Hastings within Gibbs algorithm to calculate posterior distributions of the sensitive parameters and model residual errors, and (4) analyze uncertainties propagation and their applications. The framework was firstly applied on 27 parameters of AquaCrop (version 6.1) on maize in four irrigation scenarios in arid Northwest China, given 5 time series and summary variables including canopy cover (CC), aboveground biomass (Bt), soil water content (SWC), daily evapotranspiration (ET) and final yield (Y) with 1458 measured data points of 27 irrigation treatment-year combinations from 2012 to 2015. The results showed that water stress parameters in AquaCrop were more sensitive in severe drought situations than in full irrigation conditions. The parameter uncertainty brought more variation to simulated final yield than simulated time series variables of maize in arid Northwest China. Model residual error was found to be the major contributor to overall prediction uncertainty, and interannual variation and severe water stress increased its contribution. Adding high-quality measured data of time series variables into MCMC iterations can make the estimated parameters more reliable and more biologically significant. Medians of outputs using the framework were generally closer to the corresponding measurements when compared with the results of using trial and error method. Especially for SWC and Y, Nash–Sutcliffe coefficient (EF) improved from 0.364 to 0.739 and from 0.055 to 0.415, respectively. The framework is straightforward to be applied to other crop models that can be run in batches.