Islam et al. (2026) Explainable artificial intelligence for hydroclimatic controls and mid-century cotton yield projections across Texas

Identification

• Journal: Agricultural Water Management
• Year: 2026
• Date: 2026-11-04
• Authors: Md Symum Islam, Ali Fares, Ripendra Awal, Anoop Valiya Veettil, Atikur Rahman, Ashok Mishra, Rabi H. Mohtar
• DOI: 10.1016/j.agwat.2026.110920

Research Groups

• College of Agriculture, Food, and Natural Resources, Prairie View A&M University, Prairie View, TX, USA
• Department of Civil & Environmental Engineering, Texas A&M University, College Station, TX, USA
• Department of Biological and Agricultural Engineering, Texas A&M University, College Station, TX, USA
• Department of Biological and Agricultural Engineering, North Carolina State University, Raleigh, NC, USA

Short Summary

This study developed explainable artificial intelligence models to identify hydroclimatic controls on cotton yield and project mid-century yields across Texas. It found that climate impacts on Texas cotton are highly region- and production-system-specific, with some areas experiencing significant yield declines (e.g., High Plains) and others modest increases (e.g., Low Rolling Plains) by 2030–2050.

Objective

• Identify historical cotton yield changes and their correlation with growing-season hydroclimatic conditions across Texas climate divisions.
• Develop Random Forest (RF) and Multiple Linear Regression (MLR) models for overall, irrigated, and non-irrigated cotton yield using detrended county-year data and region-informed growing seasons.
• Quantify the importance, magnitude, and direction of hydroclimatic controls on cotton yield using SHapley Additive exPlanations (SHAP) analysis.
• Quantify regional climate-conditioned cotton-yield responses for 2030–2050 under SSP245 and SSP585 scenarios using GFDL-CM4, GFDL-ESM4, and their two-model ensemble.
• Evaluate intermodel spread and late-century (2080–2100) sensitivity to characterize the strength and limitations of projected regional cotton responses.

Study Configuration

• Spatial Scale: Six major cotton-producing climate divisions of Texas (High Plains, Low Rolling Plains, North Central, South Central, Upper Coast, Lower Valley), analyzed at the county-year level. Climate data at approximately 1/16° (~6 km) resolution.
• Temporal Scale: Historical data from 1968–2014 for cotton yields, with climate data covering 1950–2100. Baseline period for projections was 2000–2014. Mid-century projections focused on 2030–2050, with a late-century sensitivity analysis for 2080–2100. Region-informed growing seasons were used (e.g., February–August, March–September, May–November).

Methodology and Data

• Models used:
  • Random Forest (RF) for nonlinear yield prediction.
  • Multiple Linear Regression (MLR) as a benchmark.
  • SHapley Additive exPlanations (SHAP) for model interpretability.
  • U.S. Geological Survey Monthly Water Balance Model (MWBM) for hydrologic variables.
  • Localized Constructed Analogs Version 2 (LOCA2) for downscaling CMIP6 climate projections.
  • Climate models: GFDL-CM4 and GFDL-ESM4 for future projections.
• Data sources:
  • Historical county-level cotton yield data (1968–2014) from the United States Department of Agriculture National Agricultural Statistics Service (USDA-NASS).
  • LOCA2-downscaled CMIP6 climate data (monthly maximum temperature, minimum temperature, precipitation) under Shared Socioeconomic Pathways (SSPs) SSP245 and SSP585.
  • Hydrologic predictors (potential evapotranspiration, actual evapotranspiration, soil-water storage) from the USGS MWBM framework.

Main Results

• Historical cotton yields (1968–2014) varied significantly by climate division and production system, with mean overall yields ranging from 0.347 t ha⁻¹ (North Central) to 0.625 t ha⁻¹ (Lower Valley).
• Hydroclimate–yield correlations were not spatially uniform and were generally stronger for irrigated cotton.
• Model performance varied regionally; RF validation correlations ranged from -0.009 to 0.867. RF outperformed MLR in some cases (e.g., North Central irrigated cotton, r = 0.867) but not consistently across all regions and production systems. Pooled RF validation showed moderate to strong agreement (r = 0.75 for overall cotton) but tended to underpredict high yields and overpredict low yields.
• SHAP analysis indicated that cotton yield variability was influenced by multiple interacting hydroclimatic factors, with precipitation having the highest global mean absolute SHAP value (0.011 t ha⁻¹), followed by actual evapotranspiration and maximum temperature (0.008 t ha⁻¹ each).
• Mid-century (2030–2050) projections showed heterogeneous climate-conditioned yield responses:
  • Overall cotton in High Plains: declined by 7.2% (SSP245) to 8.3% (SSP585).
  • Overall cotton in Low Rolling Plains: increased by 5.3% (SSP245) to 5.0% (SSP585).
  • North Central showed little change (-0.6% to +0.6%), South Central declined by 2.4–3.4%, and Upper Coast and Lower Valley showed modest increases of approximately 0.6–1.2%.
• Projected growing-season hydroclimatic conditions for 2030–2050 showed consistent warming (TMAX +0.94–2.09 °C, TMIN +0.18–1.58 °C) and declining precipitation (5.83–18.88% decrease), with regional variations in actual evapotranspiration and soil-water storage.
• Late-century (2080–2100) sensitivity analysis revealed more pronounced regional differences and scenario dependence, with some regions experiencing substantial declines (e.g., High Plains overall cotton -20.0% under SSP245) and others significant increases (e.g., North Central overall cotton +17.4% under SSP245).

Contributions

• Developed an integrated regional hydroclimate–cotton modeling framework incorporating explainable artificial intelligence (SHAP) to attribute hydroclimatic controls.
• Utilized region-informed growing seasons to accurately reflect regional differences in cotton phenology, addressing a critical gap in previous studies.
• Applied yield detrending to isolate interannual hydroclimatic variability from long-term technological and management changes, improving model accuracy.
• Employed temporal validation to ensure robust out-of-time predictive performance of the models.
• Quantified future climate-conditioned cotton yield responses using a two-model ensemble (GFDL-CM4 and GFDL-ESM4) under different SSPs, providing insights into intermodel spread and scenario dependence.
• Demonstrated that climate change impacts on Texas cotton yields are highly heterogeneous across climate divisions and production systems, emphasizing the need for region-specific adaptation strategies.

Funding

• Evans-Allen projects (7007152, 1021753)
• Capacity Building Grant (1028616) from the USDA

Citation

@article{Islam2026Explainable,
  author = {Islam, Md Symum and Fares, Ali and Awal, Ripendra and Veettil, Anoop Valiya and Rahman, Atikur and Mishra, Ashok and Mohtar, Rabi H. and Sharara, Mahmoud},
  title = {Explainable artificial intelligence for hydroclimatic controls and mid-century cotton yield projections across Texas},
  journal = {Agricultural Water Management},
  year = {2026},
  doi = {10.1016/j.agwat.2026.110920},
  url = {https://doi.org/10.1016/j.agwat.2026.110920}
}