Integrating AquaCrop-OSPy and machine learning to develop transferable surrogates for balancing wheat yield–irrigation trade-offs under deficit irrigation in semi-arid NW India

Process-based crop growth models are widely used for assessment of crop yield response to changes in irrigation management and climate variability. However, users need to undertake elaborate calibration procedures whenever utilizing them for a new location which makes large-scale scenario evaluation and optimization tasks time consuming and computationally expensive. Here, we develop AquaCrop-OSPy based surrogate models to rapidly assess wheat yield–irrigation trade-offs for districts in North-West India using soil moisture threshold (SMT) based irrigation method to control water deficit. Surrogates were trained using two tree-based machine learning algorithms, random forest (RF) and extreme gradient boosting (XGBoost) for two selected source districts, Mahendragarh and Bhilwara for 26 growing seasons. Relative yield and irrigation amount were used as target features. Time based train-test split was used to prevent information leakage across seasons. All yield and irrigation surrogates showed very good generalizability with test set R 2 > 0.90 and small RMSE and MAE for same district. However, XGBoost yield and irrigation surrogates trained on Mahendragarh performed better on cross-district transferability. Further, the underlying non-linear yield-irrigation response behaviour was found to be well preserved when these best performing surrogates were transferred to a reserved unseen target district, Sangrur. Shapley additive explanations (SHAP) analysis showed that SMT and MaxIrr (maximum irrigation per day) features had the maximum contribution to targets. Such computationally-efficient surrogate models can act as effective rapid assessment tools which could be particularly useful for agricultural planners in taking irrigation decisions to achieve adequate yield in data-scarce and water-limited environments.