Comparative Evaluation of Machine Learning Models for Predicting Crop Yields Across Sixteen Crops in Nigeria

Accurate crop yield prediction supports agricultural planning, food security and the scheduling of harvest-related resources. This study compared machine learning models for predicting the yields of sixteen major crops in Nigeria, using secondary agricultural and climatic data for 2015–2019 from the Agricultural Performance Survey of Nigeria (National Agricultural Extension and Research Liaison Services and Federal Department of Agricultural Extension) and the Nigerian Meteorological Agency. After preprocessing and the removal of 25 extreme yield observations, 1,962 crop–state–year records were analysed. Five approaches were evaluated: Random Forest; XGBoost; CatBoost; Ridge Regression; and a stacked ensemble combining the three tree-based models with a Ridge Regression meta-learner. XGBoost and CatBoost were run with their original libraries. Performance was assessed with crop-specific five-fold cross-validation, repeated three times, using R², RMSE, MSE and MAE. Under the random, record-level With-State design, in which state was an input and almost all test states also appeared in the training data, CatBoost achieved the highest mean R² (0.357; RMSE = 1.031 t/ha; MAE = 0.754 t/ha), followed by XGBoost (0.331) and Random Forest (0.315); its advantage over Random Forest (0.042; 95% CI 0.018 to 0.067) and XGBoost (0.026; 95% CI 0.003 to 0.049) was small relative to differences among crops. Crop-level R² was highest for yam (0.720), soybean (0.549), onion (0.542) and tomato (0.524) and low for groundnut, ginger, benniseed and okra. The stacked ensemble (mean R² = 0.281) did not outperform the strongest individual models and was below the best individual model in 14 of 16 crops