AquaCrop-derived surrogate model for crop yield response to monthly irrigation deficit

Abstract

Context: The Upper Yellow River Basin faces acute water competition between cascade hydropower operations and large irrigation districts. Traditional multi-objective reservoir operation leads to distorted energy-food benefit trade-offs from oversimplified linear representation of agricultural objectives, while process-based crop models are infeasible for direct embedding into reservoir operation frameworks requiring long-sequence iterative optimization due to excessive computational cost. Objective: This study aims to develop an AquaCrop-derived and computationally efficient agricultural surrogate model that accurately maps monthly water deficits to nonlinear crop yield responses, thereby providing a refined agricultural loss function for reservoir operation frameworks. Methods: Focusing on spring wheat, spring maize, and sunflower in the Hetao Irrigation District, a locally calibrated AquaCrop model was applied to simulate yield responses under full-gradient deficit irrigation scenarios. On this basis, we developed high-order polynomial ridge regression surrogate models driven by growing-season Standardized Precipitation Evapotranspiration Index and monthly irrigation deficit ratios. Finally, the model performance was validated using an offline dual validation framework covering generalization performance and decision consistency, and critical irrigation months and yield-loss thresholds were further quantified. Results and conclusions: The surrogate model achieved a global coefficient of determination above 0.93 across dual held-out test sets, with median yield-loss deviations from the sample-best AquaCrop scheme of less than 2%, and outperformed benchmark irrigation allocation schemes. However, prediction accuracy declined substantially under severe water stress, especially for spring maize. Furthermore, computational efficiency was improved by over five orders of magnitude compared with the process-based model. Under the locally parameterized AquaCrop framework, this study identified April as the most irrigation-sensitive month for spring wheat, May for spring maize, and June for sunflower. Compared with interval median scheme, the surrogate model-recommended scheme delayed physically feasible yield-loss thresholds by 6.7 to 14.7 percentage points, with bootstrap 95% confidence intervals entirely above zero. Significance: This study highlights that the AquaCrop-derived surrogate model bridges the gap between crop modeling and reservoir operation by mapping monthly water-allocation decisions to nonlinear crop yield responses. Aligned with the monthly decision cycle of reservoir operation and with ultra-high computational efficiency, this model is expected to serve as a crucial methodological component for future multi-objective reservoir operation frameworks.