Feature selection and explainable machine learning to identify climatic drivers of alfalfa yield in the Ogallala Region (USA)

Introduction The Ogallala aquifer region supports extensive alfalfa production but faces severe groundwater depletion and intensifying climatic volatility. This study aimed to develop a machine learning (ML) framework to predict regional alfalfa yield variability and identify the climatic factors associated with yield variability across the region. Methods A dual-scenario ML framework was developed using a 38-year dataset (1981–2018). A nested 5-fold stratified spatial block cross-validation with temporal holdout sub-sampling was implemented to eliminate spatial and temporal data leakage. Maximum relevance minimum redundancy (mRMR) feature selection was applied to identify relevant predictors, and model interpretation was performed using SHAP (Shapley Additive exPlanations) analysis. Results Under the full model scenario, extreme gradient boosting achieved the highest predictive performance (R 2 = 0.460 ± 0.127), with latitude emerging as the dominant spatial predictor. Under the climate model scenario, random forest (R 2 = 0.420 ± 0.069) optimized out-of-distribution transferability, reducing prediction uncertainty during severe environmental anomalies. The SHAP analysis revealed that, within the climate-only pipeline, early- and late-season solar radiation were the primary predictors of yield variation, reflecting seasonal variability associated with vegetative growth stages. Discussion These findings demonstrate clear architectural trade-offs between absolute predictive performance and spatial transferability, establishing a robust computational template to interpret sub-regional climate variations. The statistical patterns derived from these models highlight the potential utility of evaluating late-season harvest timing to possibly support root carbohydrate accumulation in hot-summer humid continental (Dfa) zones, while highlighting the sensitivity to moisture retention and thermal constraints across cold semi-arid (BSk) regions. However, the framework remains geographically constrained by peripheral data gaps and the omission of Texas, highlighting the need for future integration of high-resolution spatial data and field-level operational variables to better resolve hyper-local management variations.