A Stochastic Method for Crop Models: Including Uncertainty in a Sugarcane Model. (1st March 2017)

• Record Type: Journal Article
• Authors: Marin, Fabio; Jones, James W.; Boote, Kenneth J.
• Is Part Of: Agronomy Journal. Volume 109:Number 2(2017)
• Page Start: 483
• Page End: 495
• Publication Date: 2017-03-01
• Subjects: Agronomy -- Periodicals; 630
• DOI: 10.2134/agronj2016.02.0103
• Languages: English
• ISSNs: 0002-1962

Abstract

Abstract : Core Ideas: We propose an approach for simulating effects of parameter and input variable uncertainties in crop models. The GLUE was useful for estimating the parameter correlation matrix and the best set of parameters. The crop model output uncertainty varied with soil and climate conditions. Crop models are increasingly being used for different purposes, including evaluation of climate change impacts on crop yields and opportunities for adapting management to future conditions. However, past uses of these models have been criticized in part due to a failure of researchers to quantify uncertainties of crop yield prediction. We have developed a method for considering uncertainty in a crop model using a simple sugarcane ( Saccharum spp.) model as a case study. A Bayesian Monte Carlo approach generalized likelihood uncertainty estimation was used to estimate model parameters, their uncertainties, and correlations among them using data from five growing seasons at four locations in Brazil where crops received adequate nutrients and good weed control. Some of the model parameters were assumed to be correlated random variables, based on the literature, and varied across their ranges of uncertainty to estimate posterior distributions of parameters. The mean parameter values, parameter ranges, and the parameter covariance‐correlation matrix are inputs to this Bayesian approach, which includes a Toeplitz–Cholesky factorization to generate correlated random variable samples and then simulate a distribution of state variables on a daily time step. Correlated random simulation, based on posterior distributions of parameters, was an effective method for including uncertainty in the crop growth and yield estimates. We demonstrated that uncertainty can be reduced with respect to model structure and parameter meaning because the optimization process is heavily dependent on prior knowledge of the parameters. Uncertainty varied with environment even though distributions of parameters remained the same across all environments.