Zheng et al. (2026) Stomatal conductance modeling for drip-irrigated kiwifruit in seasonal drought regions of South China: Evaluation of improved empirical models and interpretable machine learning approaches

Identification

• Journal: Agricultural Water Management
• Year: 2026
• Date: 2026-01-16
• Authors: Shunsheng Zheng, NingBo Cui, Quanshan Liu, Shouzheng Jiang, Daozhi Gong, Xiaoxian Zhang
• DOI: 10.1016/j.agwat.2026.110153

Research Groups

• State Key Laboratory of Hydraulics and Mountain River Engineering & College of Water Resource and Hydropower, Sichuan University, Chengdu, China
• Institute of Environment and Sustainable Development in Agriculture, Chinese Academy of Agricultural Sciences, Beijing, China
• Sustainable Soils and Crops, Rothamsted Research, Harpenden, England, UK

Short Summary

This study comprehensively evaluated stomatal conductance (gs) modeling for drip-irrigated kiwifruit in South China by developing and comparing improved Jarvis-type empirical models and interpretable machine learning approaches. It found that the CatBoost model, incorporating soil water content (SWC), achieved superior predictive performance and robust interpretability.

Objective

• To quantify the stage-specific response patterns of kiwifruit gs to varying SWC.
• To develop an improved Jarvis-type empirical model by incorporating a novel, stage-specific nonlinear soil water response function.
• To evaluate the performance of multiple machine learning algorithms in predicting kiwifruit gs.
• To identify the key environmental drivers of gs variation and interpret their effects.

Study Configuration

• Spatial Scale: Kiwifruit experimental station in Chengdu City, Sichuan Province, China (30°19'20''N, 103°25'57''E), characterized by shallow hills and yellow earth soil.
• Temporal Scale: Field experiments conducted from 2017 to 2019 (3 years). Measurements were taken on two days during each of four growth stages (mid-March to late October) at 2-hour intervals from 09:00 to 17:00.

Methodology and Data

• Models used:
  • Empirical: Jarvis (JV), Jarvis with traditional soil water content function (JV1), Jarvis with improved stage-specific nonlinear soil water content function (JV2).
  • Machine Learning: eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Categorical Boosting (CatBoost), Support Vector Regression (SVR), Linear Regression (LR).
• Data sources:
  • Synchronized field measurements of stomatal conductance (gs) and environmental drivers (photosynthetically active radiation (PAR), air temperature (Ta), relative humidity (RH), vapor pressure deficit (VPD)) using a portable photosynthesis system (LCPro-SD, ADC BioScientific, UK).
  • Volumetric soil water content (SWC) monitored using PR2 profile probes (PR2/6, Delta-T Devices Ltd., UK) at five depths (10, 20, 30, 40, 60 cm).
  • Crop water requirements (ETc) estimated using the single crop coefficient method, with reference evapotranspiration (ET0) calculated using the Penman-Monteith function.

Main Results

• Deficit irrigation significantly reduced kiwifruit gs, with the highest sensitivity observed during growth stage II.
• Stage-specific SWC thresholds triggering gs reductions were identified as 0.311, 0.322, 0.331, and 0.326 m³ m⁻³ for stages I–IV, respectively.
• The incorporation of soil water content (SWC) significantly improved the accuracy of both empirical and machine learning models.
• Among empirical models, the improved JV2 model, 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.
• CatBoost, using VPD, Ta, PAR, and SWC as input variables, consistently outperformed all other models (empirical and machine learning) across all growth stages, achieving R² values between 0.815 and 0.839, RMSE between 0.065 and 0.076 mol m⁻² s⁻¹, and MAE between 0.054 and 0.064 mol m⁻² s⁻¹.
• SHapley Additive exPlanations (SHAP) analysis and Partial Dependence Plots (PDPs) identified VPD as the dominant driver of gs variation, followed by SWC. Ta and PAR exhibited more limited and context-dependent influences.

Contributions

• Quantified the stage-specific physiological responses of kiwifruit stomatal conductance to soil water deficit, establishing critical SWC thresholds for different growth stages.
• Developed and validated an improved Jarvis-type empirical model (JV2) by integrating a novel, data-driven, stage-specific nonlinear soil water response function, enhancing model accuracy and physiological realism.
• Provided a comprehensive evaluation of various machine learning algorithms for predicting kiwifruit stomatal conductance, identifying CatBoost as the superior model for its predictive power and generalization capability.
• Utilized interpretable machine learning techniques (SHAP and PDPs) to elucidate the relative importance and marginal effects of environmental drivers on stomatal conductance, offering robust physiological insights into kiwifruit water relations.
• Established a reliable and physiologically informed modeling framework for precision irrigation and water management strategies in drip-irrigated kiwifruit orchards in seasonal drought regions.

Funding

• National Natural Science Foundation of China (51922072, 52109060, 52279041)
• Science and Technology Program of Sichuan (2023YFN0024, 2023NZZJ0015)
• High Impact Scholars Program of Sichuan University (SCU2022CG09)

Citation

@article{Zheng2026Stomatal,
  author = {Zheng, Shunsheng and Cui, NingBo and Liu, Quanshan and Jiang, Shouzheng and Gong, Daozhi and Zhang, Xiaoxian},
  title = {Stomatal conductance modeling for drip-irrigated kiwifruit in seasonal drought regions of South China: Evaluation of improved empirical models and interpretable machine learning approaches},
  journal = {Agricultural Water Management},
  year = {2026},
  doi = {10.1016/j.agwat.2026.110153},
  url = {https://doi.org/10.1016/j.agwat.2026.110153}
}