Analysis, development and application of wheat models of differing complexity at the farm and field scale

The development and analysis of crop models of differing complexity is presented, ranging from a simple, static, empirical model, to more complex, dynamic, mechanistic models. Two models, an existing mechanistic model (Sirius) and a new mechanistic model of more intermediate complexity (Intermediate), were developed and analysed using relatively detailed field-level observations. In contrast, two new simpler models, an empirical and a mechanistic model, were developed using relatively coarse farm-level observations. Analysis of the existing mechanistic model (Sirius) revealed that a number of model variables and cultivar-specific parameters were redundant and did not contribute to model performance. The results of this analysis informed the development of mechanistic model of more intermediate complexity (Intermediate), although this was also indicated to contain redundant variables and parameters. Vernalisation simulation was one aspect of the models that was consistently identified as an area of redundancy. These reduced models are a product of the data used to generate the reduced model – in this work, simplified versions of the models were identified that were capable of maintaining the ability to predict differences in cultivar growth and development under different nitrogen treatments. Careful consideration needs to be given to the application of the reduced model, to inform the data used in the reduction. Automated, comprehensive model reduction techniques, such as the one employed here, have the potential to be important tools in reducing unnecessary model complexity, and associated uncertainty, for an application. This unnecessary complexity can act as a barrier to application, analysis and understanding. An empirical (Simple-EMP) and simple mechanistic (Simple-MECH) model were additionally developed to predict wheat yield at the farm-scale, using farm-level observations. The empirical model produced more accurate predictions of wheat yields than the mechanistic design, which was hindered by the monthly time step necessitated by the driving meteorological observations and the lack of traditional mechanistic model input observations. The farm-level survey data from which these models were developed is collected annually, and suggestions for further development of these models are made, including (i) incorporating newly available data on fertiliser application to replace the existing proxy, and (ii) exploring the potential to obtain additional data traditionally required by crop models, primarily the timing of important management decisions, for example, sowing and harvest dates