Rahmah et al. (2025) Comparative Performance of Regression and Ensemble Learning Algorithms in Precision Irrigation Forecasting of Sweet Potato

Identification

• Journal: Jurnal Elektronika dan Telekomunikasi
• Year: 2025
• Date: 2025-12-31
• Authors: Muthia Rahmah, Indra Maulana
• DOI: 10.55981/jet.799

Research Groups

• Institut Prima Bangsa, Informatics and Computer Engineering Education, Cirebon, Indonesia

Short Summary

This study systematically compared five machine learning algorithms for precision irrigation forecasting in sweet potato using real-time Internet of Things (IoT) sensor data, finding that a hyperparameter-tuned Random Forest Regressor achieved the highest predictive accuracy (R² = 0.9802).

Objective

• To systematically compare the performance of five regression and ensemble learning algorithms (Support Vector Regression, AdaBoost, Extreme Gradient Boosting, Random Forest, CatBoost) for forecasting sweet potato irrigation requirements using real-time IoT sensor data.

Study Configuration

• Spatial Scale: A sweet potato field in Halimpu Village, Beber District, Cirebon Regency, Indonesia (6°44′S, 108°33′E).
• Temporal Scale: 42 hours and 50 minutes (August 4-5, 2025), with data collected at 10-minute intervals, totaling 243 observations across two complete diurnal cycles.

Methodology and Data

• Models used: Support Vector Regression (SVR), AdaBoost Regressor, Extreme Gradient Boosting (XGBoost) Regressor, Random Forest Regressor (RFR), CatBoost Regressor.
• Data sources: Real-time sensor data collected via an IoT prototype deployed in a sweet potato field. Sensors included a capacitive soil moisture sensor (0-1023 analog-to-digital converter (ADC) units), a DHT22 temperature and humidity sensor (±0.5 K and ±2% accuracy), a BH1750 light intensity sensor (1-65,535 lux range), and a BMP180 atmospheric pressure sensor (±100 Pa accuracy). Data was transmitted via MQTT to the ThingsBoard IoT Platform.

Main Results

• The hyperparameter-tuned Random Forest Regressor achieved the highest predictive accuracy with R² = 0.9802, Root Mean Squared Error (RMSE) = 9.58 ADC units, and Mean Absolute Error (MAE) = 6.08 ADC units.
• The default Random Forest Regressor also performed exceptionally well (R² = 0.9786), demonstrating robustness with minimal hyperparameter sensitivity.
• XGBoost (tuned R² = 0.9670) and CatBoost (default R² = 0.9687) showed strong performance but exhibited overfitting tendencies with near-perfect training scores.
• Support Vector Regression (SVR) significantly improved after tuning (R² from 0.328 to 0.797) but remained inferior to ensemble methods.
• Random Forest Default offered the best balance of high accuracy (R² = 0.9786), rapid training (0.42 seconds), and low inference latency (2.1 milliseconds per sample), making it highly suitable for resource-constrained IoT deployments.

Contributions

• First systematic benchmarking of five regression and ensemble learning algorithms (SVR, AdaBoost, XGBoost, Random Forest, CatBoost) specifically for sweet potato irrigation forecasting within an actual IoT-based deployment.
• Implementation of a multi-day (43-hour) continuous IoT deployment, capturing two complete diurnal cycles, which significantly enhances dataset robustness and addresses limitations of single-session monitoring in previous studies.
• Demonstration that Random Forest's bagging approach generalizes more effectively than boosting methods (XGBoost) under limited sample conditions (n < 300), a crucial insight for resource-constrained IoT agricultural deployments.
• Rigorous methodological approach including chronological train-test splitting, explicit exclusion of relay status, and proper lag-feature engineering to prevent data leakage and ensure valid time-series forecasting.

Funding

The author(s) received no financial support for this research.

Citation

@article{Rahmah2025Comparative,
  author = {Rahmah, Muthia and Maulana, Indra},
  title = {Comparative Performance of Regression and Ensemble Learning Algorithms in Precision Irrigation Forecasting of Sweet Potato},
  journal = {Jurnal Elektronika dan Telekomunikasi},
  year = {2025},
  doi = {10.55981/jet.799},
  url = {https://doi.org/10.55981/jet.799}
}