Integrating Remote Sensing, Management Information, and Satellite Embeddings for Machine-Learning Prediction of Field-Level Maize Yield in Kenyan Smallholder Systems

Field-level crop yield prediction is an important application of precision agriculture, but it remains challenging in African smallholder systems because plots are small, management is heterogeneous, and satellite pixels can mix multiple land covers. This study evaluated machine-learning approaches for predicting maize yield from 6,016 field-year observations in Kenya during 2017–2020 using Sentinel-2 remote-sensing variables, climate, soil, terrain and location attributes, survey-based management information, and 64-dimensional satellite embeddings. Eleven candidate algorithms were compared using cross-validation, followed by XGBoost predictor-group and SHAP analyses. XGBoost achieved the strongest pooled performance (R² = 0.383; RMSE = 1,334 kg ha⁻¹; MAE = 1,013 kg ha⁻¹). Conventional remote sensing alone achieved R² = 0.196; adding climate, soil, terrain, and location increased R² to 0.250, while adding management information increased it to 0.386. Management accounted for 32.4% of grouped SHAP importance, with plant density as the most influential individual predictor. Satellite embeddings matched the predictive performance of conventional remote-sensing predictors while reducing satellite predictor dimensionality by approximately 77%. The results show that remote sensing and machine learning can support field-level yield monitoring in heterogeneous smallholder systems, but plot-level management information remains critical for improving prediction.