Data requirements for crop modelling-Applying the learning curve approach to the simulation of winter wheat flowering time under climate change

A prerequisite for application of crop models is a careful parameterization based on observational data. However, there are limited studies investigating the link between quality and quantity of observed data and its suitability for model parameterization. Here, we explore the interactions between number of measurements, noise and model predictive skills to simulate the impact of 2050′s climate change (RCP8.5) on winter wheat flowering time. The learning curve of two winter wheat phenology models is analysed under different assumptions about the size of the calibration dataset, the measurement error and the accuracy of the model structure. Our assessment confirms that prediction skills improve asymptotically with the size of the calibration dataset, as with statistical models. Results suggest that less precise but larger training datasets can improve the predictive abilities of models. However, the non-linear relationship between number of measurements, measurement error, and prediction skills limit the compensation between data quality and quantity. We find that the model performance does not improve significantly with a theoretical minimum size of 7–9 observations when the model structure is approximate. While simulation of crop phenology is critical to crop model simulation, more studies are needed to explore data needs for assessing entire crop models.