Regional uncertainty analysis between crop phenology model structures and optimal parameters

Crop phenology models are pivotal for simulating crop development, predicting yields and guiding agricultural practices. However, uncertainties exist in simulations due to different model structures and variability in model parameters. Although quantifying these contributions to total variability is often conducted at a site-specific level, few attempts to address this for regional crop modelling using field-calibrated parameters. Our study employs six crop phenology models (APSIM, CERES, GDD, Richardson, Sigmoid and Wang) for simulating maturity timings of three representative rice cultivars using trial data within the Sichuan Basin, China. The Leave-One-Out Cross-Validation (LOOCV) is applied for model calibration with a global parameter optimization algorithm and evaluation. Calibrated models show robust prediction capabilities during LOOCV with R2 of 0.68–0.95 and RMSE of 2–4 days, though a larger variance is found for evaluation data than for calibration data. Models calibrated with data from sites having frequent high-temperature (Tmax≥32 °C) episodes tend to have better predictability than without high-temperature episodes. Parameter variability, calibrated with different subsets of each cultivar during LOOCV, is low-to-moderate (mostly CV≤20 %) except for the Sigmoid model´s curve steepness parameter. For the early-maturity cultivar, parameter variability is spatially the main uncertainty factor, relating to its greater variability of site-specific calibrated parameter values. For the medium-maturity and late-maturity cultivars, the dominant uncertainty source arises from the interplay between model structures and parameters. Parameter variability notably influences the overall uncertainty more than the model structure variability across the region, except in areas prone to high-temperature extremes where divergent model responses predominate. These findings highlight the cultivar-specific nature of simulation uncertainty, but also the critical need to assess the spatial distribution of uncertainty sources. For parameter uncertainty, a broader conceptualization is essential for more accurate quantifications of uncertainty sources, paving the way for improved ensemble crop modelling, especially at a large spatial scale.