Crop model improvement in APSIM: Using wheat as a case study

The process of building and improving a broadly useful set of crop models is a major undertaking and the APSIM (Agricultural Production system SIMulator) development community conduct this work separately from projects that have specific uses of the models. This paper presents a set of standards that a modern crop model should meet, and describes the approaches and software the APSIM development community are using to build and maintain models that meet these standards. The latest version of APSIM combines a range of tools in a single user interface (UI) to assist model developers. It uses a modem version control system to ensure model reliability and a modern distribution system to ensure users can easily access models and receive updates. The wheat model is used as an example to describe the development process using the APSIM platform. Firstly, a test set for the model was developed: Forty-eight experiments, giving 655 treatments, were collated into a database. Each of these treatments was configured as a test simulation using the Experiment component in the UI to hasten the process and reduce human error. The model was implemented in the UI using the plant modelling framework for a visual and flexible approach to model construction. Graphs and statistics for assessing model performance were constructed in the UI. From the UI, changes were made to the model, and its performance was reviewed each time until developers were happy with it. Memo fields were added into the UI to describe experiments and capture the rationale and citations for model structure and parameterisation. This system facilitated a number of science improvements to the wheat model, including new canopy and phenology models and the compilation and more thorough analysis of a large test dataset. The test set was submitted to APSIM governance for review and approval for general release. Once approved, the model was included in the APSIM distribution as a read-only file so users cannot change it. The test set was put into version control to provide a baseline for assessing model performance. Documentation was automatically generated from the test set, providing a full and up-to-date technical description of the model and its evaluation. This demonstrates a robust yet flexible and efficient approach for improving and distributing crop models, facilitated by purpose-designed software.