Comparative Performance of Regression and Ensemble Learning Algorithms in Precision Irrigation Forecasting of Sweet Potato

Abstract

Precision irrigation is essential for sustainable agriculture under increasing water scarcity. This study compared regression and ensemble learning algorithms for forecasting irrigation requirements in sweet potato, a crop characterized by high variability in water demand. An Internet of Things (IoT)-based prototype was deployed to collect real-time data on soil moisture, temperature, humidity, light intensity, and atmospheric pressure over 42 hours and 50 minutes (August 4-5, 2025), encompassing two complete diurnal cycles at 10-minute intervals and yielding 243 temporal observations. Following preprocessing and feature engineering with lag-based temporal features, the final dataset comprised 240 samples (192 training, 48 testing) using chronological time-based splitting to prevent data leakage. Five algorithms, Support Vector Regression (SVR), AdaBoost, Extreme Gradient Boosting (XGBoost), Random Forest Regressor (RFR), and CatBoost, were evaluated under default and hyperparameter-tuned configurations using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²) as evaluation metrics. Tuned Random Forest achieved superior performance (R² = 0.9802, RMSE = 9.58, MAE = 6.08), followed by default Random Forest (R² = 0.9786) and default CatBoost (R² = 0.9687). XGBoost demonstrated strong performance (R² = 0.9670 tuned) but exhibited overfitting tendencies with near-perfect training scores. SVR improved substantially after tuning (R² = 0.328 to 0.797), although it remained inferior to ensemble methods. Overall, ensemble methods, particularly XGBoost and Random Forest, demonstrated superior efficacy for sweet potato irrigation forecasting. These findings underscore the potential of IoT-integrated machine learning to enhance water-use efficiency and support sustainable smart farming practices.