Explainable machine learning reveals water-related drivers of sub-field dryland wheat yield variability

Understanding the drivers of within-field yield variability is essential for precision management in dryland agroecosystems where water limitation governs productivity. This study identifies and quantifies controls on sub-field-scale wheat yield variability using an interpretable machine-learning framework. Eighteen dryland wheat fields in northeastern Colorado (2019–2024) were grouped into aspirational (ASP; no-till, diversified rotation with spatial nitrogen zones) and business-as-usual (BAU; reduced tillage, wheat–fallow rotation) systems. Extreme gradient boosting (XGBoost) models were evaluated under eight feature combinations that systematically included precipitation (P), soil moisture (SM), actual evapotranspiration (ET a ), soil properties, and topography. Under random partitioning, models explained over 80% of yield variability in both systems. Performance remained stable even when high-resolution soil properties were excluded, indicating that much of sub-field variability can be captured using spatially explicit water indicators and terrain data. Under field-level holdout, models combining ET a and SM retained strong predictive skill, explaining ∼78% of yield variability in ASP and ∼55% in BAU fields. SHAP analysis showed that early-season water-related variables governed yield variability, with ET a emerging as the most spatially informative indicator of water limitation. Nitrogen-zone classifications were important in ASP fields, likely reflecting structured yield-potential gradients embedded within management zoning rather than isolated fertilizer-rate effects. Elevation acted as a secondary but persistent control. These findings demonstrate that explainable machine learning can provide actionable insight into water–management interactions driving yield variability in semi-arid, dryland wheat systems.