A Weighted Ensemble Learning Approach with Synthetic Oversampling for Maize Yield Modeling in Data-limited Agro-ecological Zones of Uganda

Sparse and irregular agricultural records constrain reliable crop yield modelling in many Sub-Saharan African regions. This study evaluates the use of ensemble machine learning combined with synthetic oversampling for maize yield modelling in data-limited agro-ecological zones of Uganda. Seasonal climatic variables, including rainfall, soil moisture, temperature, and solar radiation, were aggregated from satellite-derived datasets and integrated with maize-yield records from the Uganda Bureau of Statistics for 2018–2020. To address severe data imbalance and scarcity, the Synthetic Minority Oversampling Technique for Regression with Gaussian Noise (SMOGN) was employed to enhance representation of low- and high-yield conditions. An ensemble model combining LightGBM, Random Forest, and Decision Tree algorithms was developed and compared with individual machine-learning and deep-learning models. The ensemble achieved a coefficient of determination of approximately 0.99 and a root-mean-square error of approximately 0.06 t/ha, outperforming individual tree-based models and deep learning baselines (R² ≈ 0.79–0.83). Feature importance analysis indicated that soil moisture, rainfall, and solar radiation were the dominant climatic drivers of yield variability across zones. However, because a substantial proportion of the training samples were synthetically generated and the observations span only three years, the model should be interpreted as a methodological demonstration rather than as an operational yield-forecasting system. The findings indicate that combining ensemble learning with controlled synthetic data augmentation can support agricultural modelling in environments where long-term yield measurements are scarce. The proposed approach provides a foundation for future work incorporating extended field observations and independent validation datasets.