Hybrid Machine Learning Modelling for Rice Yield Forecasting under Climatic Variability in South Sudan

A Hybrid RF–XGBoost Framework for Rice Yield Forecasting under Climatic Variability in South Sudan

Abstract

Improving the reliability of agricultural yield forecasting is essential for strengthening food security planning and agricultural policy in South Sudan, where rice yield remains highly sensitive to climatic variability and structural production constraints. This study evaluates classical statistical and ensemble-based machine learning models for forecasting national rice yield using key agronomic and climatic determinants, including cultivated land area, historical yield trends, rainfall, temperature, and soil pH. The empirical analysis is based on annual time-series data spanning nearly five decades (1961–2010). A supervised learning framework was implemented, with the dataset partitioned chronologically into training (n = 38) and testing (n = 10) subsets to enable disciplined out-of-sample validation. Four modelling approaches were examined: Multiple Linear Regression (MLR), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and a Hybrid RF–XGBoost ensemble constructed through prediction aggregation. Model performance was evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), coefficient of determination (R²), and Mean Absolute Percentage Error (MAPE), computed strictly on the reserved test dataset. Within the evaluated test sample (n = 10), all models exhibited strong correlation (R² > 0.91). While MLR achieved R² = 0.9128, it exhibited higher RMSE (295.60 kg/ha) and MAPE (31.75%) relative to ensemble approaches, suggesting reduced flexibility in modelling nonlinear yield responses within the evaluated test period. Random Forest improved error performance (R² = 0.9647; RMSE = 273.39 kg/ha; MAE = 227.13; MAPE = 28.26%), while XGBoost achieved R² = 0.9360 with RMSE = 289.25 kg/ha, MAE = 182.05, and MAPE = 30.59%. The Hybrid RF–XGBoost model attained the lowest RMSE (259.14 kg/ha), lowest MAE (174.29), lowest MAPE (26.91%), and highest R² (0.9725) within the test sample. Although improvements are moderate in magnitude, the findings suggest that hybrid ensemble aggregation can provide incremental gains in prediction accuracy under data-constrained conditions. Results are interpreted cautiously given the limited sample size. The study contributes an applied comparative assessment of forecasting approaches to support evidence-informed agricultural planning in fragile production environments.