Parametrization of five classical plant growth models applied to sugar beet and comparison of their predictive capacity on root yield and total biomass

• Evaluation of the predictive capacity of five plant growth models in sugar beet.
• The models had different modelling scales and strategies for biomass repartition.
• A sensitivity analysis was performed for each model to reduce the variability.
• The estimation of the thermal time of initiation is very crucial.
• All models performed very well, but STICS was from far the best for root prediction.

A wide range of models have been proposed and developed for modelling sugar beet growth, each of them with different degrees of complexity and modelling assumptions. Many of them are used to predict crop production or yield, even when they were not originally designed for this purpose, and even though their predictive capacity has never been properly evaluated.In this study, we propose the evaluation and comparison of five plant growth models that rely on a similar energetic concept for the production of biomass, but with different levels of description (individual-based or per square meter) and different ways to describe biomass repartition (empirical or via allocation): Greenlab, LNAS, CERES, PILOTE and STICS. The models were all programmed on the same modelling platform, calibrated on a first set of data, and then their predictive capacities were assessed on an independent data set. First, a sensitivity analysis was carried out on each model to identify a subset of parameters to be estimated, to reduce the variability of the models. We were able to reduce the number of parameters from 10 to 4 for Greenlab, and from 16 to 1 for STICS. Three criteria were then used to compare the predictive capacities of the models: the root mean squared error of prediction and the modelling efficiency for the total dry matter production and the dry matter of root, and the yield prediction error.All the models provided good overall predictions, with high values of the modelling efficiency. The use of sensitivity analysis allowed us to reduce the variability of the models and to enhance their predictive capacities. Models based on an empirical harvest index gave good yield predictions, and similar results compared to allocation models for the total dry matter, but the harvest index might not be very robust. The crucial role of initiation was also pointed out, as well as the need for an accurate estimation and modelling of this early phase of growth.