Machine Learning for Agricultural Variable Prediction in Wireless Sensor Networks and IoT Environments: A Comparative Study with Statistical Validation
Year: 2026
Type: conference-paper
Abstract: Precision agriculture depends on reliable sensor data for efficient crop and resource management. Wireless Sensor Networks transmit environmental measurements continuously, imposing a high energy cost on battery-powered nodes. This paper evaluates three machine learning models, namely Random Forest, XGBoost, and Long Short-Term Memory (LSTM) networks, for one-step-ahead prediction of agricultural variables collected from two distinct sensor datasets: a greenhouse WSN dataset with four Bluetooth Low Energy and Wi-Fi scenarios (192 observations each), and a large IoT dataset with 37,920 timestamped records. A persistence baseline is included in every comparison as a scientifically rigorous calibration reference. Feature engineering follows a strict anti-leakage protocol using lag, rolling-window, and difference features derived exclusively from past observations. Hyperparameters are selected through exhaustive grid search with TimeSeriesSplit cross-validation for tree-based models and Bayesian optimisation for the LSTM. Statistical significance is assessed with a Friedman test followed by pairwise Wilcoxon signed-rank tests with Bonferroni correction. The numerical analysis reveals that while Random Forest achieved the lowest RMSE among the trained models in 9 of the 12 WSN tasks and XGBoost did so for all three IoT targets, the naive persistence baseline remained the overall best predictor in 9 of the 12 WSN tasks and 2 of the 3 IoT tasks due to the low high-frequency variance of the micro-climate signals. The Friedman test confirms that the structural differences among the trained candidates are significant (χ 2 = 14.00, p = 0.0009). Under an operational dualprediction paradigm, offline transmission-suppression analysis indicates that data-driven virtual sensing can eliminate an upper bound of approximately 68% of communications on nontrivial WSN tasks, offering substantial potential transmission reductions and energy-saving opportunities. Temperature and humidity models transfer across communication protocols without retraining (R 2 ≥ 0.83), facilitating flexible cross-protocol edge deployment scenarios.
Source: HAL
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7224974284