Global sensitivity analyses prior to parameter calibration significantly improves prediction quality
Wang, C., Harrison, M. T., Brown, H., Qiao, Y., Yin, X., Yang, R., … & Liu, K. (2025). Global sensitivity analyses prior to parameter calibration significantly improves prediction quality. Computers and Electronics in Agriculture, 238, 110750.
Abstract
Effective parameter calibration is vital for accurate crop model simulations but is challenged by extensive parameter sets and inherent uncertainties in agroecosystem models. This study systematically evaluated three global sensitivity analysis (GSA) methods (Morris, Sobol-Martinez, and eFAST) alongside three parameter optimisation algorithms (Nelder–Mead simplex, DREAM-zs, and L-BFGS-B) to enhance parameter estimation and simulation accuracy in APSIM Next Generation, a widely used agroecosystem model. Sensitivity analysis results indicated that the Morris method identified the broadest set of influential parameters due to its inclusive parameter selection strategy, while Sobol-Martinez provided more targeted parameter identification by clearly distinguishing impactful parameters. Conversely, eFAST was highly selective, pinpointing fewer parameters of highest impact, beneficial for computational efficiency. Among the optimisation methods, the Bayesian DREAM-zs algorithm consistently produced superior model predictions across evaluated output variables, including phenology, biomass, leaf area index, and grain yield, outperforming the frequentist Nelder–Mead and gradient-based L-BFGS-B methods. However, DREAM-zs required significantly higher computational resources, particularly at higher iteration settings. The uncertainty analysis revealed that interactions among sensitivity analysis methods, optimisation algorithms, and wheat genotypes dominated the sources of uncertainty in parameter estimation. This underscores the necessity of carefully selecting and integrating complementary sensitivity analysis and optimisation methods tailored to specific modelling objectives. Our findings demonstrated a robust methodological framework to improve calibration accuracy, reduce predictive uncertainty, and ultimately support more reliable agricultural decision-making.