Data requirement for effective calibration of process-based crop models

- We assessed data requirement for crop model calibration & reduced uncertainty.
- We used APSIM, canola data from serial sowings and a Bayesian optimisation approach.
- Calibration with data from similar seasons led to large equifinality & uncertainty.
- Calibration with data from contrasting seasons reduced error/uncertainty to minimum.
- Crop modelling need to consider parameter equifinality & simulation uncertainty.

While process-based crop models are increasingly applied in multiple research areas, the role of model calibration and its data requirement have been rarely addressed. We used data for two canola cultivars measured across up to six sowing dates, five plant densities and three years to investigate how different data impacted on the efficacy of calibration of the process-based APSIM-Canola simulation model. A Bayesian optimisation method was used to derive cultivar parameters. Our results showed that model calibration using data from a single or multiple similar seasons would likely result in significant equifinality (i.e. multiple combinations of parameters leading to similar simulation accuracy) and simulation uncertainty. The most effective calibration was to use data from contrasting environments (at least two seasons), particularly with in-season growth measurements. Targeting model calibration to minimise simulation error (NRMSE) any less than the inherent error in measurement (13.5%) appeared to be pointless, because model calibration cannot reduce the simulation error to less than the error in the data used for calibration. Our rigorous calibration and validation procedure helped to identify APSIM's deficiency to simulate the response of canola growth to changes in plant density. Future modelling work is warranted to quantify simulation uncertainty due to equifinality caused by model calibration against limited data, and improve modelling of density effects of canola in APSIM.