Abstract
Accurate field-level maize-yield estimation in smallholder systems is challenging because plots are small and management is heterogeneous. We analyzed 6016 Kenyan field-years (2017–2020) using conventional Sentinel-2 predictors, climate, soil, terrain, location, survey-based management, and 64-dimensional satellite embeddings. Ten algorithms were compared using random, leave-one-year-out (LOYO), and spatial-block cross-validation, followed by predictor-group comparisons and SHAP analysis. The all-predictor XGBoost model achieved R2 = 0.395, RMSE = 1321 kg ha−1, and MAE = 1010 kg ha−1 under random cross-validation; R2 declined to 0.172 under LOYO and 0.292 under spatial-block validation. Adding management information to conventional remote sensing, climate, soil, terrain, and location raised random-validation R2 from 0.270 to 0.394, with positive gains across all three validation designs. Replacing 252 conventional remote-sensing predictors with 64 embedding dimensions yielded similar accuracy (random-validation R2 = 0.394 and 0.405, respectively) and substantially reduced input dimensionality. These conclusions were broadly consistent with Extra Trees. Because predictors included full-year observations and measurements recorded at harvest, the models provide retrospective, end-of-season yield estimates rather than in-season forecasts. Results highlight management information and compact satellite representations, alongside limitations in transfer to unseen years and locations.

This publication has 38 references indexed in Scilit: