Proposal and extensive test of a calibration protocol for crop phenology models

Abstract

A major effect of environment on crops is through crop phenology, and therefore, the capacity to predict phenology for new environments is important. Mechanistic crop models are a major tool for such predictions, but calibration of crop phenology models is difficult and there is no consensus on the best approach. We propose an original, detailed approach for calibration of such models, which we refer to as a calibration protocol. The protocol covers all the steps in the calibration workflow, namely choice of default parameter values, choice of objective function, choice of parameters to estimate from the data, calculation of optimal parameter values, and diagnostics. The major innovation is in the choice of which parameters to estimate from the data, which combines expert knowledge and data-based model selection. First, almost additive parameters are identified and estimated. This should make bias (average difference between observed and simulated values) nearly zero. These are “obligatory” parameters, that will definitely be estimated. Then candidate parameters are identified, which are parameters likely to explain the remaining discrepancies between simulated and observed values. A candidate is only added to the list of parameters to estimate if it leads to a reduction in BIC (Bayesian Information Criterion), which is a model selection criterion. A second original aspect of the protocol is the specification of documentation for each stage of the protocol. The protocol was applied by 19 modeling teams to three data sets for wheat phenology. All teams first calibrated their model using their “usual” calibration approach, so it was possible to compare usual and protocol calibration. Evaluation of prediction error was based on data from sites and years not represented in the training data. Compared to usual calibration, calibration following the new protocol reduced the variability between modeling teams by 22% and reduced prediction error by 11%.

Why is there so much variability in crop multi-model studies?

“…This study revisits three simulation experiments on wheat phenology that have been previously reported (Wallach et al, 2023, 2021a, 2021b). In the previous studies, the emphasis was on prediction accuracy, whereas in the present study the simulation results are reanalyzed with the objective of estimating the contribution of parameter uncertainty to the overall uncertainty of the phenology prediction.…”

“…Our second estimate of parameter uncertainty is based on a comparison, for multiple modeling groups, between two different calibration procedures, their usual procedure, and a new protocol procedure. In general, there were major differences between the two calibration procedures for each group (Wallach et al, 2023). Both are plausible in the sense that they were developed and used by modeling groups well-versed in the use of their model structure.…”

“…The proposed calibration protocol is described in detail by Wallach et al (2023). The following is a brief summary of that protocol.…”

“…How to sample from plausible parameter vectors on the other hand is not clear. Wallach et al (2023) identified several steps in crop model calibration, including the choice of default parameter values, the choice of observed data to use in the objective function for frequentist calibration, the choice of parameters to be estimated from the data and finally the choice of algorithm and software for doing the calibration. They showed that different modeling groups made different choices, and that even different modeling groups using the same model structure made different choices.…”

A calibration protocol for soil-crop models

“…None of these approaches specifically addresses the problem of over-parameterization, or the risk of highly correlated and therefore highly uncertain parameter estimators. The AgMIP study on calibration using only phenology data proposed an original approach to choice of parameters to estimate, based on a standard model selection criterion (Wallach et al, 2023b). An analogous approach is used here, but extended to multiple measured variables.…”