Developing a machine-learning Surrogate Model to rapidly predict wheat yields under Soil-Moisture-Triggered irrigation strategies in Northwest India

Process-based crop models such as AquaCrop are widely used to assess crop responses to irrigation management and climate variability. However, large-scale scenario exploration and optimization are often constrained by the high computational cost of repeated model simulations. In this study, we develop a machine-learning surrogate model to emulate AquaCrop-OSPy (ACOSP) simulated wheat yields under a wide range of soil-moisture-triggered irrigation (SMT) thresholds and daily maximum irrigation limits (MaxIrr) across 26 growing seasons in Northwest India. A simulation ensemble of 21,840 ACOSP runs was generated by systematically varying SMT (0–100%) and MaxIrr (0–40 mm day⁻¹) for each season. The surrogate model was trained using season-wise climate variability, seasonal precipitation and irrigation strategy parameters (Season, SMT, MaxIrr) as predictors, with wheat yield (t ha⁻¹) as the target variable. We implemented and compared Random Forest and XGBoost regression models using a time-based train–test split to avoid information leakage across seasons. The best-performing XGBoost model explained ~87–90% of the inter-season and management-driven yield variability in the independent test period, while maintaining computational runtimes several orders of magnitude lower than ACOSP. Feature-importance analysis showed that SMT was the dominant explanatory factor, followed by climate-driven seasonal variability, whereas MaxIrr primarily influenced high-stress scenarios. The surrogate model successfully reproduced non-linear yield responses and threshold behaviour, suggesting strong potential for near-real-time decision support and large-scale scenario exploration. This work demonstrates that machine-learning surrogates can complement process-based crop models by enabling rapid evaluation of irrigation strategies, uncertainty assessment, and future climate scenario testing at regional scales. The developed framework is transferable to other regions, crops, and water-limited environments, offering a scalable pathway toward computationally efficient agricultural water-management assessment.