Global Sensitivity Analyses of the APSIM-Wheat Model at Different Soil Moisture Levels

The APSIM (Agricultural Production Systems Simulator)-Wheat model has been widely used to simulate wheat growth, but the sensitivity characteristics of the model parameters at different soil moisture levels in arid regions are unknown. Based on 2023~2025 winter wheat field data from the Changji Experimental Site, Xinjiang, China, this study conducted a global sensitivity analysis of the APSIM-Wheat model using Morris and EFAST methods. Twenty-one selected parameters were perturbed at ±50% of their baseline values to quantify the sensitivity of the aboveground total dry matter (WAGT) and yield to parameter variations. Parameters exhibiting significant effects on yield were identified. The calibrated APSIM model performance was evaluated against field observations. The results indicated that the order of influential parameters varied slightly across different soil moisture levels. However, the WAGT output was notably sensitive to accumulated temperature from seedling to jointing stage (T1), accumulated temperature from the jointing to the flowering period (T2), accumulated temperature from grain filling to maturity (T4), and crop water demand (E1). Meanwhile, yield output showed greater sensitivity to number of grains per stem (G1), accumulated temperature from flowering to grain filling (T3), potential daily grain filling rate during the grain filling period (P1), extinction coefficient (K), T1, T2, T4, and E1. The sensitivity indices of different soil moisture levels under Morris and EFAST methods showed highly significant consistency. After optimization, the coefficient of determination (R2) was 0.877~0.974, the index of agreement (d-index) was 0.941~0.995, the root mean square error (RMSE) was 319.45~642.69 kg·ha–1, the mean absolute error (MAE) was 314.69~473.21 kg·ha–1, the residual standard deviation ratio (RSR) was 0.68~0.93, and the Nash–Sutcliffe efficiency (NSE) was 0.26~0.57, thereby enhancing the performance of the model. This study highlights the need for more careful calibration of these influential parameters to reduce the uncertainty associated with the model.