The chaos in calibrating crop models

Abstract

Calibration, that is the estimation of model parameters based on fitting the model to experimental data, is among the first steps in essentially every application of crop models and process models in other fields and has an important impact on simulated values. The goal of this study is to develop a comprehensive list of the decisions involved in calibration and to identify the range of choices made in practice, as groundwork for developing guidelines for crop model calibration starting with phenology. Three groups of decisions are identified; the criterion for choosing the parameter values, the choice of parameters to estimate and numerical aspects of parameter estimation. It is found that in practice there is a large diversity of choices for every decision, even among modeling groups using the same model structure. These findings are relevant to process models in other fields.

Simultaneous Calibration of Grapevine Phenology and Yield with a Soil–Plant–Atmosphere System Model Using the Frequentist Method

“…However, the estimated parameter values are largely affected by the approach used for the calibration. There is a large diversity of calibration approaches, with different approaches having their own calibration steps and statistical model of errors, thus resulting in different parameter values given the same data [5,7,8]. The fundamental choice in employing a calibration approach depends on whether we consider the simulated values as random variables or not [9][10][11].…”

“…To address the uncertainties, there are commonly two different approaches in statistics based on regression methods: Bayesian and frequentist approaches [12]. A Bayesian approach defines a prior distribution of parameters and estimates the posterior distribution of the parameters and the variance of the model errors [7,13]. The frequentist method is primarily based on the concept of repeated sampling.…”

“…For the parameter estimations, it typically requires defining a mathematical form of an objective function to identify the parameters that can minimize some measures of errors between the observed and simulated values [7,10]. One advantage of the frequentist method is it can avoid specifying the detailed prior information of the parameters as subjective inputs [14], but is still constrained by the imposed parameter values (lower and upper bounds) in many cases [7]. Most modelling groups tend to define an objective function to minimize the Sum of Squared Errors (SSE) of the predictions [8].…”

“…Most modelling groups tend to define an objective function to minimize the Sum of Squared Errors (SSE) of the predictions [8]. The sum is commonly the unweighted sum across different response variables and prediction interests, which essentially corresponds to the ordinary least squares (OLS) criterion [7]. However, it is often not appropriate to combine the prediction errors from variables of a different nature, such as phenology and yield.…”

“…In practice, model calibration has been frequently undertaken using ad hoc or trial and error approaches, which provide little information on parameter uncertainties. As a result, there are currently no standard calibration protocols and a detailed calibration guideline is often separately developed for each model [5,7,8]. In this study, we aim to calibrate the soil-plant-atmosphere system model STICS, developed by INRAE, for simulating grapevine (Vitis vinifera L.) phenology and yield.…”