Abstract
This paper introduces an efficient IoT-based framework for predicting crop damage within smart agricultural systems, focusing on the integration of Internet of Things (IoT) sensor data with advanced machine learning (ML) and ensemble learning (EL) techniques. The primary objective is to develop a reliable decision support system capable of forecasting crop health status classifying crops as healthy, pesticide-damaged, or affected by other stressors while addressing a critical challenge: the presence of missing data in real-time agricultural datasets. To overcome this limitation, the proposed approach incorporates robust data imputation strategies using both traditional ML methods and powerful EL models. Techniques such as K-Nearest Neighbors, linear regression, and ensemble-based imputers are evaluated for their effectiveness in reconstructing incomplete data. Furthermore, Bayesian Optimization is applied to fine-tune EL classifiers including XGBoost, CatBoost, and LightGBM (LGBM), enhancing their predictive performance. Extensive experiments demonstrate that XGBoost outperforms all other models, achieving an average sensitivity of 88.1%, accuracy of 89.56%, precision of 83.4%, and F1-score of 84.8%. CatBoost and LGBM also deliver competitive results, with CatBoost achieving 90.50% accuracy and LGBM reaching 90.23%. In addition, the imputation capability of the XGBoost model is validated through a low Mean Squared Error (MSE) of 0.0213 and a high R-squared (R(2)) value of 0.99, confirming its effectiveness for both prediction and data recovery tasks. The key contributions of this innovative work include the design of a low-cost, power-efficient, and scalable crop damage prediction system, the integration of real-time IoT data with optimized ensemble learning, and a comprehensive evaluation of imputation techniques to enhance model robustness. This framework is particularly suited for deployment in resource-constrained agricultural environments, advancing the field of smart farming through intelligent, data-driven solutions.