Wallach, D. | Search bioRxiv

bioRxiv2019

How well do crop models predict phenology, with emphasis on the effect of calibration?

Predicting phenology is essential for adapting varieties to different environmental conditions and for crop management. Therefore, it is important to evaluate how well different crop modeling groups can predict phenology. Multiple evaluation studies have been previously published, but it is still difficult to generalize the findings from such studies since they often test some specific aspect of extrapolation to new conditions, or do not test on data that is truly independent of the data used for calibration. In this study, we analyzed the prediction of wheat phenology in Northern France under observed weather and current management, which is a problem of practical importance for wheat management. The results of 27 modeling groups are evaluated, where modeling group encompasses model structure, i.e. the model equations, the calibration method and the values of those parameters not affected by calibration. The data for calibration and evaluation are sampled from the same target population, thus extrapolation is limited. The calibration and evaluation data have neither year nor site in common, to guarantee rigorous evaluation of prediction for new weather and sites. The best modeling groups, and also the mean and median of the simulations, have a mean absolute error (MAE) of about 3 days, which is comparable to the measurement error. Almost all models do better than using average number of days or average sum of degree days to predict phenology. On the other hand, there are important differences between modeling groups, due to model structural differences and to differences between groups using the same model structure, which emphasizes that model structure alone does not completely determine prediction accuracy. In addition to providing information for our specific environments and varieties, these results are a useful contribution to a knowledge base of how well modeling groups can predict phenology, when provided with calibration data from the target population.

plant biology↗

bioRxiv2019

A new approach to crop model calibration: phenotyping plus post-processing.

Crop models contain a number of genotype-dependent parameters, which need to be estimated for each genotype. This is a major difficulty in crop modeling. We propose a hybrid method for adapting a crop model to new genotypes. The genotype-dependent parameters of the model could be obtained by phenotyping (or gene-based modeling). Then field data for example from variety trials could be used to provide a simple empirical correction to the model, of the form a+b times an environmental variable. This approach combines the advantages of phenotyping, namely that the genotype-specific parameters have a clear meaning and are comparable between genotypes, and the advantages of fitting the model to field data, namely that the corrected model is adapted to a specific target population. It has the advantage of being very simple to apply, and furthermore gives useful information as to which environmental variables are not fully accounted for in the initial model. In this study, this empirical correction is applied to the SUNFLO crop model for sunflower, using field data from a multi-environment trial network. The empirical correction reduced mean squared error (MSE) on the average by 54% for prediction of yield and by 26% for prediction of oil content, compared to the initial model. Most of the improvement came from eliminating bias, with some further improvement from the environmental term in the regression.

plant biology↗