Wheat yield prediction using machine learning, proximal sensing, nitrogen and climate data for Oklahoma

Abstract

Accurate in-season yield prediction is a practical prerequisite for informed nitrogen (N) management decisions in winter wheat production, yet no rigorous framework has been established and tested against the tools producers currently rely on. Our study developed and evaluated machine learning (ML) models for pre-application wheat yield prediction to support in-season top-dress N management in Oklahoma. Four ML algorithms were evaluated using 57 site-years of multi-source data from rainfed Oklahoma environments, including normalized difference vegetation index (NDVI) measured by proximal sensing at multiple growth stages, pre-emergence N rate, soil texture, and monthly climate data (rainfall, temperature, evapotranspiration, and growing degree days and water balance). The In-Season Estimate of Yield (INSEY) and classical statistical approaches i.e., ordinary least squares, linear mixed model, and generalized additive model, served as reference benchmarks. The Categorical Boosting Regression (CatBoost) model achieved the best ML performance (R²= 0.81, RMSE = 0.62 t ha−¹, and MAE = 0.42 t ha⁻¹), substantially outperforming INSEY (R² = 0.12) under the same pre-application data constraint. The final 11feature model identified February–March temperatures, evapotranspiration, water balance, pre-plant N rate, and early-season NDVI as the most influential predictors. Because all climate inputs are freely accessible through the Oklahoma Mesonet, the framework proposed in this study can be integrated into existing N recommendation systems without additional data infrastructure. The use of ML algorithms to combine multi-source data, including remote sensing, climate, and management data, enables accurate winter wheat yield prediction within the in-season N application window in Oklahoma, providing a valuable tool for N management decisions that support better crop productivity and improved N use efficiency. Using 57 site-years of multi-source climate, nitrogen and remote sensing data, CatBoost was chosen as the final model for prediction of in-season winter wheat yield in Oklahoma (R2 = 0.81), which outperformed INSEY, OLS and the linear mixed model, XGBoost and SVR, and showing greater stability across partitions than the additive alternative. CatBoost substantially outperformed the INSEY benchmark (R² = 0.12) and remained superior even when INSEY was augmented with a location-specific intercept (R²= 0.37), rejecting the hypothesis that canopy reflectance and thermal accumulation alone are sufficient for pre-application yield prediction. February–March temperature, evapotranspiration, water balance, pre-plant N rate, and early-season NDVI emerged as the most influential predictors, reflecting the thermal and moisture drivers of tiller survival and kernel set. Non-linear interactions among climate, nutrient, and canopy variables accounted for most of the predictive gain over classical linear approaches. All climate inputs are freely available via the Oklahoma Mesonet, enabling practical deployment alongside existing sensor-based N recommendation systems with no additional data collection.