Wagan et al. (2026) Interpreting Irrigation Decisions: Explainable AI Using SHAP and LIME in Agricultural Water Management Models

Identification

• Journal: Water Resources Management
• Year: 2026
• Date: 2026-01-01
• Authors: Shah Mehmood Wagan, Sidra Sidra
• DOI: 10.1007/s11269-025-04468-0

Research Groups

• Sichuan University, Chengdu, PR China
• University of Chinese Academy of Social Sciences, Beijing, P. R. of China

Short Summary

This study systematically compares SHAP and LIME explainable AI (XAI) methods for interpreting machine learning models predicting irrigation decisions in U.S. agriculture, finding LightGBM to be the best predictive model and demonstrating that SHAP provides more robust global explanations while LIME offers intuitive local insights, leading to a proposed hybrid XAI framework for a decision support system.

Objective

• To systematically compare SHAP and LIME explainable AI (XAI) methods for interpreting machine learning (ML)-based irrigation decisions across diverse U.S. agricultural regions using multi-source, high-resolution data.
• To quantify the faithfulness, stability, and human interpretability of SHAP and LIME explanations.
• To develop and integrate a hybrid XAI framework into an operational Decision Support System (DSS) to enhance transparency and trust in agricultural water management.
• To identify policy- and farm-actionable irrigation thresholds derived from XAI insights.

Study Configuration

• Spatial Scale: Three major agricultural areas in the United States: Central Valley of California, High Plains (Ogallala aquifer zone), and Southeast (portions of Georgia and Alabama). Data were harmonized and aggregated to a 1 km x 1 km grid.
• Temporal Scale: Multi-year period from 2018 to 2024. Data sources included 5-day satellite imagery, daily weather data, and monthly irrigation data.

Methodology and Data

• Models used:
  • Machine Learning (ML): Random Forest (RF), XGBoost, LightGBM, Neural Network (NN - 3-layer Multilayer Perceptron).
  • Explainable AI (XAI): SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations).
• Data sources:
  • Satellite: European Space Agency (ESA) Sentinel-2 (10-20 m resolution, resampled to 1 km) for vegetation indices (NDVI, EVI, LSWI, SAVI).
  • Observation/Reanalysis:
    • United States Department of Agriculture (USDA): SSURGO (100 m) for static soil characteristics (texture, organic matter, available water capacity), NASS CDL (30 m) for annual crop type and phenological data, ARMS for monthly irrigation data (volume and method).
    • National Oceanic and Atmospheric Administration (NOAA) NCEI (4 km PRISM, resampled to 1 km) for daily weather data (precipitation, temperature, humidity, wind speed, solar radiation).
    • United States Geological Survey (USGS) NWIS for monthly irrigation data (well meter records).
    • NASA POWER (1 km) for daily reference and crop evapotranspiration (ET0, ETc).
  • Target Variable: Binary irrigation decision (1 if irrigation occurred, 0 otherwise) and continuous irrigation volume (in mm/day).

Main Results

• Model Performance: LightGBM consistently outperformed other ML models, achieving the highest accuracy (0.90), AUC-ROC (0.96), and lowest Brier Score (0.11) for irrigation event prediction. For irrigation volume regression, LightGBM had the lowest RMSE (1.12 mm/day) and MAE (0.83 mm/day), and highest R² (0.86) and NSE (0.84). All ML models significantly surpassed a baseline climatology model (AUC = 0.68).
• Global Feature Importance (SHAP): Vapor Pressure Deficit (VPD) was identified as the most critical factor influencing irrigation decisions (mean absolute SHAP value = 0.30), followed by the 30-day Standardized Precipitation Index (SPI-30), crop stage, and NDVI anomaly. Significant interaction effects were found, such as between NDVI anomaly and VPD (coefficient 0.8), and Soil Moisture Deficit (SMD) and Growing Degree Days (GDD) (coefficient 0.7).
• Local Explanations (LIME): LIME provided intuitive, instance-specific explanations, highlighting key drivers like low NDVI anomaly, high VPD, and soil moisture deficit for irrigation events, or recent heavy precipitation for non-irrigation events.
• XAI Method Comparison:
  • Faithfulness: SHAP explanations were more faithful, leading to a larger mean drop in prediction (0.67) when top features were removed, compared to LIME (0.52).
  • Stability: SHAP demonstrated higher consistency (mean cosine similarity 0.79) across similar instances compared to LIME (mean cosine similarity 0.61).
  • Human Interpretability: Farmer surveys (n=25) indicated LIME explanations were perceived as more intuitive (92% sensible) and actionable (80% actionable) than SHAP (84% sensible, 76% actionable), with LIME being less technical.
• Regional Differences: VPD was most influential in California, while SPI-30 was highly relevant in California and the High Plains, reflecting regional climatic and hydrological variations.
• Actionable Insights: The study derived crop and soil-specific irrigation trigger thresholds (e.g., sandy almond orchards irrigate at VPD > 2.8 kPa). Model-based simulated irrigation showed potential water savings of 22% in almonds and 15% in corn with minimal yield loss (< 2%).
• Hybrid DSS: A prototype Decision Support System (DSS) was developed, integrating both SHAP (for robust global insights) and LIME (for farmer-friendly local explanations) in a two-mode interface, enhancing transparency and trust for diverse stakeholders.

Contributions

• First systematic comparison ("bake-off") of SHAP and LIME on multi-year (2018–2024), high-resolution, multi-source data across three diverse U.S. agro-climates for irrigation decision modeling.
• Comprehensive quantification of explanation quality (faithfulness, stability, human interpretability) using farmer surveys and perturbation analysis, a rare occurrence in agricultural XAI.
• Development and deployment of a novel hybrid XAI framework within a functioning Decision Support System (DSS), combining LIME's intuitive local explanations with SHAP's robust global insights.
• Identification of policy- and farm-actionable irrigation thresholds (e.g., VPD > 2.5 kPa + NDVI anomaly < 0), calibrated by region and crop type.

Funding

No funding was received.

Citation

@article{Wagan2026Interpreting,
  author = {Wagan, Shah Mehmood and Sidra, Sidra},
  title = {Interpreting Irrigation Decisions: Explainable AI Using SHAP and LIME in Agricultural Water Management Models},
  journal = {Water Resources Management},
  year = {2026},
  doi = {10.1007/s11269-025-04468-0},
  url = {https://doi.org/10.1007/s11269-025-04468-0}
}