Stomatal conductance modeling for drip-irrigated kiwifruit in seasonal drought regions of South China: Evaluation of improved empirical models and interpretable machine learning approaches

Accurate modeling of stomatal conductance (gs) enhances understanding of plant water relations and supports advancements in eco-physiological modeling and adaptive irrigation practices. This study provides a comprehensive evaluation of gs modeling for drip-irrigated kiwifruit through parallel development of three Jarvis-type empirical models (JV, JV1, JV2) and five machine learning algorithms (XGBoost, LightGBM, CatBoost, SVR, LR) based on three years of field measurements comprising synchronized records of gs and key environmental drivers. Models were assessed via year-wise grouped cross-validation, with performance measured by R2, RMSE, and MAE, and interpretability analyzed using SHapley Additive exPlanations (SHAP) and Partial Dependence Plots (PDPs). Results showed that deficit irrigation significantly reduced gs, with sensitivity being most pronounced during stage II. The incorporation of soil water content (SWC) substantially improved the accuracy of both empirical and machine learning models. Among empirical models, JV2, featuring a stage-specific nonlinear SWC response function, demonstrated the highest accuracy (R2 ranging from 0.736 to 0.814) and minimized bias under extreme SWC conditions. Using vapor pressure deficit (VPD), air temperature (Ta), photosynthetically active radiation (PAR), and SWC as input variables, CatBoost outperformed both empirical models and other machine learning algorithms across all growth stages (R2 = 0.815–0.839; RMSE = 0.065–0.076 mol m−2 s−1; MAE = 0.054–0.064 mol m−2 s−1). SHAP analysis and PDPs identified VPD as the dominant driver of gs variation, followed by SWC. Overall, the improved JV2 model offers a structurally transparent framework for gs estimation with acceptable accuracy, while CatBoost combined with SHAP analysis and PDPs provides superior predictive performance and robust interpretability under complex environmental conditions. These findings support the reliable modeling and regulation of kiwifruit gs under varying SWC scenarios in drip-irrigated orchards