Decomposing total uncertainty in rice yield prediction: A quantification of relative contributions of model structure‐driven and parameter value‐driven uncertainties

Abstract Crop models have been widely used to simulate crop growth and yield. However, few studies have investigated factors contributing to uncertainty in crop models. This study evaluated the predictive uncertainty of AquaCrop, WOFOST, and ORYZA version 3 in predicting the early and late season rice ( Oryza sativa L.) yield in Guangxi province, China, and quantified the relative contributions of model structure‐driven (the simplification and/or misrepresentation of crop growth‐defining, ‐limiting, and ‐reducing factors) and parameter value‐driven uncertainties to the total predictive uncertainty. Our results showed that (1) the AquaCrop model achieved the best predictive performance, with root mean square error ranging from 274 to 622 kg ha −1 , mean absolute percentage error ranging from 4% to 15%, and R 2 ranging from 0.50 to 0.63. (2) For each individual crop model, the posterior distribution of rice yield prediction was constructed using 10,000 sets of parameter vectors, which were randomly generated based on the posterior parameter distributions. The 95% confidence interval of the simulated rice yield prediction distribution contained 79.3% of the observed rice yields. (3) The average relative contributions of structure‐driven and parameter value‐driven uncertainties in the early‐season rice yield prediction were 51% and 50%, respectively, and were 39% and 61% in the late‐season rice yield prediction, respectively. This study demonstrated the importance of quantifying structure‐driven and parameter value‐driven uncertainties when evaluating the crop model predictive uncertainty in rice yield prediction.