Predicting harvest productivity from field operations data: a random forest approach for the brazilian cerrado

Abstract

Accurate pre-harvest yield prediction enables smarter logistics, targeted agronomic interventions, and evidence-based farm management. Most machine learning (ML) yield models rely on census-level statistics or satellite remote sensing; per-field, per-operation management records remain largely not analyzed as a predictive signal. The study trains Random Forest regressors on 635 harvest records (SOYBEANS n=349, CORN_WET n=239, SORGHUM n=47) gathered  using the John Deere Operations Center (JDOC) in Tocantins, Brazil, integrating NASA POWER weather features and ISRIC SoilGrids soil data into a 59-feature dataset. Cross-validated R² values of 0.499 ± 0.046 for soybeans and 0.551 ± 0.106 for corn (safrinha) demonstrate meaningful predictive capacity at the individual field-operation level. Feature group analysis reveals that weather is the dominant driver for safrinha corn (41.1%), while soybean yield is driven by a mix of weather (34.5%) and agronomic management signals—pesticides, soil preparation, and seeding quality (33.9% combined). SHAP (SHapley Additive exPlanations) interpretability confirms that high vapour pressure deficit (VPD) and heat stress days exert consistent negative effects across crops. Sorghum results are unreliable (n=47) and excluded from inference.