A Bayesian framework for crop model calibration: A case study in the US Corn Belt

Crop models play a key role in simulating crop growth, predicting yield, and assessing interventions for improving production. Nevertheless, their reliability is often hindered by uncertainties in parameterization, soil properties, management practices, and meteorological inputs. These uncertainties can significantly affect model accuracy, especially when models are applied to different crops, cultivars, or fields. This study explores these concepts using the APSIM crop model under varying weather conditions, soil types, and management practices across multiple production years, but with a focus on a single location. Our analysis focuses on three research fields near Lincoln, Nebraska, growing different maize cultivars in either mono-cropping or rotational-crop configurations, and under both rain-fed and irrigated regimes. Initially, we perform a global sensitivity analysis to assess how variations in cultivar parameters affect key model outputs: leaf area index, biomass, and yield. We advance the analysis by conducting an intra-season sensitivity analysis to track the temporal impact of parameters over the growing cycle. Using an MCMC-based Bayesian inference approach, we estimate the most influential parameters. Results indicate that, for this specific location and agronomy, over 50 % (7 out of 13) of cultivar parameters have the greatest impact on model outputs, with the most sensitive parameters varying depending on the model output under investigation. Notably, parameters involved in the early capture of radiation were the most influential across all fields and outputs. The intra-season sensitivity analysis reveals that parameter sensitivity varies across different crop phenological stages, suggesting the potential for a targeted parameter calibration within specific windows of the season. The calibrated model using MCMC in a real-world case scenario delivers a strong agreement between predicted and observed outputs, with R2 values ranging from 0.84 to 0.98, and relative RMSE between 10 % and 34 %. Compared to its uncalibrated counterpart, the calibrated model exhibits improved performance, with at least a 30 % reduction in RMSE values and enhanced correlation with in situ measurements. These findings confirm the robustness of the Bayesian calibration approach and its ability to accurately predict crop development across multiple seasons and maize cultivars. As such, this approach provides a valuable tool for calibrating crop models while simultaneously quantifying the uncertainty associated with input parameters. Extension of this analysis and model to larger regional areas would test its suitability for more generalized application of models at scale.