Yang et al. (2026) Interpreting and forecasting crop-specific irrigation water productivity in an arid irrigated area using explainable machine learning and scenario simulation

Identification

• Journal: Irrigation Science
• Year: 2026
• Date: 2026-01-12
• Authors: Lei Yang, Liuyue He, Shouzheng Jiang, Zailin Huo, Isaya Kisekka, Jingyuan Xue
• DOI: 10.1007/s00271-025-01069-y

Research Groups

• Institute for Disaster Management and Reconstruction, Sichuan University, Chengdu, China
• Ocean College, Zhejiang University, Zhoushan, China
• Donghai Laboratory, Zhoushan, China
• State Key Laboratory of Hydraulics and Mountain River Engineering & College of Water Resource and Hydropower, Sichuan University, Chengdu, China
• Center for Agricultural Water Research in China, China Agricultural University, Beijing, China
• Department of Land Air & Water Resources, Department of Biological and Agricultural Engineering, University of California Davis, Davis, CA, USA

Short Summary

This study developed an explainable machine learning framework to quantify, interpret, and predict crop-specific irrigation water productivity (IWP) for wheat, maize, and sunflower in China's Hetao Irrigation District. It found that a Bayesian-optimized CatBoost model achieved high predictive accuracy and identified irrigation volume, sunshine hours, and groundwater evaporation as key IWP drivers, while also simulating future IWP trajectories under climate change and water conservation measures.

Objective

• Evaluate the performance of six machine learning models (Random Forest, AdaBoost, CatBoost, Deep Neural Networks, K-Nearest Neighbors regression, and Support vector regression) optimized by Bayesian algorithm for modeling the IWP of wheat, maize, and sunflower in the Jiefangzha irrigation district (JFZID) of the Hetao Irrigation District.
• Identify the dominant drivers and spatiotemporal variations of IWP for each crop between 2006 and 2013 using the best-performing explainable machine learning techniques.
• Predict future IWP trajectories under future climate scenarios and water-saving strategies (2023-2030).

Study Configuration

• Spatial Scale: Jiefangzha Irrigation Area (JFZID) in China's Hetao Irrigation District (approximately 2300 km²), divided into 2,485 simulation cells (1 km x 1 km).
• Temporal Scale: Historical analysis from 2006 to 2013 (crop growth period: April-September); Future scenario simulations from 2023 to 2030.

Methodology and Data

• Models used:
  • Machine Learning: CatBoost (Bayesian-optimized, selected as optimal), Random Forest (RF), AdaBoost, Deep Neural Networks (DNN), K-Nearest Neighbors (KNN) regression, Support Vector Regression (SVR).
  • IWP Estimation: Stewart et al. (1977) crop water production function model.
  • Evapotranspiration (ETa) Estimation: Surface Energy Balance Algorithm for Land (SEBAL).
• Data sources:
  • Remote Sensing: MOD11A1 Version 6.1 (land surface temperature, 1 km, daily), MOD13Q1 Version 6.1 (vegetation indices, 250 m, 16-day).
  • Climate Data: National Ecological Science Data Center (net radiation from GLOPEM-CEVSA model), National Earth System Science Data Center (global 0.5° climate data from CRU, high-resolution global climate data from WorldClim, sunshine duration from 824 meteorological stations).
  • Ground-based Measurements: Spatially distributed monitoring network (112 soil points, 53 groundwater points, 22 water salinity points) for soil properties, groundwater, irrigation, and salinity.
  • Agricultural Statistics: Local agricultural statistics for maximum crop yield (Ym).
  • Future Climate Data: CMIP6 (for future climate scenarios).
  • Land Use/Cover: 30 m resolution data (Xue, 2018) for planting structure.

Main Results

• The Bayesian-optimized CatBoost model achieved the highest predictive accuracy for IWP (R² > 0.95 across all three crops: wheat 0.97, maize 0.96, sunflower 0.97).
• SHAP-based interpretation identified irrigation water volume, sunshine hours, and groundwater evaporation as key IWP drivers, with their influence varying among crops. Soil electrical conductivity (SEC) also showed significant influence, particularly for maize.
• Spatially, higher IWP values were consistently found near main canals, while lower IWP was observed near main drainage canals due to salt accumulation. Maize exhibited the highest IWP (0.11 to 3.86 kg/m³), followed by sunflower (0.89 to 3.46 kg/m³), and wheat (0.61 to 2.89 kg/m³).
• Temporally, IWP for all crops declined from 2006 to 2010, then recovered from 2010 to 2013, linked to Yellow River water diversion and soil salinization dynamics.
• By 2030, IWP is projected to decline by an average of 0.105 kg/m³ compared to 2006 under baseline climate change.
• Water conservation measures (canal lining, field irrigation upgrades) can partially offset climate-driven IWP declines by nearly 0.051 kg/m³, with effectiveness being crop- and location-specific. Maize showed a slight overall IWP increase (0.005 kg/m³) when both measures were implemented.
• Optimal integrated management strategies vary by crop: wheat benefits from 10-30% salinity optimization and 30-60% canal lining effect; maize requires precise regulation in low-salt areas (<60% salinity control); sunflower has higher tolerance to salt-water fluctuations.

Contributions

• Developed a comprehensive, robust, and data-driven explainable machine learning framework for quantifying, interpreting, and predicting crop-specific IWP in arid irrigated areas with shallow groundwater.
• Provided transferable insights for sustainable irrigation management by identifying key IWP drivers and their crop-specific variations using SHAP.
• Simulated future IWP trajectories under coupled climate change and water conservation scenarios, offering actionable strategies for adaptive management.
• Demonstrated the superior performance and efficiency of the machine learning approach compared to complex process-based models, requiring fewer parameters and providing higher accuracy.
• Highlighted that soil water-salinity dynamics exert a greater influence on IWP than vegetation indices in arid irrigated areas, challenging traditional assumptions.

Funding

• Youth Program of the National Natural Science Foundation of China (52009132)
• MOE (Ministry of Education in China) Project of Humanities and Social Sciences (24YJCZH367)
• Fundamental Research Funds for the Central Universities (20822041G4066)
• Sichuan University postdoctoral interdisciplinary Innovation Fund

Citation

@article{Yang2026Interpreting,
  author = {Yang, Lei and He, Liuyue and Jiang, Shouzheng and Huo, Zailin and Kisekka, Isaya and Xue, Jingyuan},
  title = {Interpreting and forecasting crop-specific irrigation water productivity in an arid irrigated area using explainable machine learning and scenario simulation},
  journal = {Irrigation Science},
  year = {2026},
  doi = {10.1007/s00271-025-01069-y},
  url = {https://doi.org/10.1007/s00271-025-01069-y}
}