Machine learning-based prediction of cereal rye cover crop biomass across diverse agroecosystems

Abstract

Accurate operational predictions of cereal rye (Secale cereale L.) biomass are critical for quantifying the agroecosystem services provided by cover crops and for guiding growers’ management decisions for subsequent cash crops. In this study, we developed machine learning-based biomass prediction models using two advanced gradient-boosted tree algorithms, CatBoost and XGBoost. A comprehensive dataset comprising cereal rye biomass and management information from 24 U.S. states, combined with soil and weather data, were used to train the models. Models relying solely on early spring weather inputs achieved moderate predictive skill (R² ≈ 0.74). Incorporating later-season weather data modestly improved mid-season fits but led to overfitting in late-spring predictions. Extending CatBoost to quantile regression enabled estimation of 10%–90% prediction intervals with moderate pinball loss. Overall, our findings demonstrate that publicly available soil and weather data, supplemented with limited management inputs, can support interpretable, uncertainty-aware biomass predictions suitable for optimal cover crop management.