| The APSIM model suffers from low optimization efficiency and parameter equifinality in local calibration for dryland spring wheat. To address these issues, this study proposes a hierarchical phenology-yield optimization framework, and adopts Latin Hypercube Sampling (LHS) to verify and visualize the convergence features of parameter space before and after hierarchical optimization. The two-layer framework first identifies core phenological parameters via global sensitivity analysis, and calibrates them with the Improved Sparrow Search Algorithm (ISSA) to lock crop growth stages; it then further optimizes yield-sensitive parameters using ISSA in the second layer. The research is based on 1993–2024 meteorological data of Dingxi, Gansu Province, combined with yield data from the Dingxi Statistical Yearbook (1993–2011, 2024), field-measured yield data of Mazichuan Village (2012–2023) and phenological observation data (2016–2018). The results show that LHS scatter plots present remarkable funnel-shaped convergence after five phenological parameters are fixed. Under this constraint, hierarchical calibration of six yield-sensitive parameters significantly improves the model’s simulation accuracy on both training and test sets of the two data sources: the maximum RMSE reduction reaches 86.12%, the minimum NRMSE drops to 1.38%, and the maximum ME rises to 0.997. This method enhances the prediction accuracy of the APSIM model in Dingxi drylands while ensuring sound biological mechanisms. |