Miftahul Walid, Muhammad Ashar, Heru Wahyu Herwanto
The reliability of Internet of Things (IoT)-based irrigation systems depends heavily on the quality of data acquired from various sensors. In real-world applications, missing values and faulty data, caused by transmission faults or sensor failures, can severely compromise the application of Machine Learning (ML), particularly in critical tasks such as detecting pump failures. This study investigates the impact of missing data imputation on the performance of an Extreme Gradient Boosting (XGBoost) classifier under a controlled setting. Missing data were simulated at rates ranging from 5% to 30%, and the imputation techniques applied included Multiple Imputation by Chained Equations (MICE), k-Nearest Neighbors (KNN), Random Forest (RF), and iterative gradient-boosting methods. The quality of the imputed data was evaluated using Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), whereas classifier performance was assessed using accuracy, precision, recall, F1-score, and Area Under the Receiver Operating Characteristic Curve (AUC). For 5% missing data, MICE achieved the lowest errors (MAE = 0.018, RMSE = 0.026), whereas at 30% missing data, errors increased (MAE = 0.072, RMSE = 0.094). Classifier performance declined with increasing missing data: accuracy decreased from 0.972 to 0.818 and recall from 0.961 to 0.742. The application of MICE mitigated this decline, maintaining an accuracy of 0.982 and recall of 0.975 at 5% missing data, and an accuracy of 0.863 and recall of 0.812 at 30% missing data, with AUC remaining above 0.880 in both cases. Notably, the imputation method with the lowest MAE and RMSE did not always produce the best classification performance, indicating that numerical precision does not necessarily translate into improved classification skill. These results highlight the importance of selecting imputation techniques based on the specific nature of the problem to preserve feature integrity and sustain the performance of Artificial Intelligence (AI)-based smart irrigation systems in real-world sensor environments. Licensed under a CC-BY 4.0 license | Copyright (c) by the authors
Department of Electrical Engineering and Informatics, Universitas Negeri Malang, Malang, Indonesia; Department of Informatics Engineering, Universitas Islam Madura, Pemekasan, Indonesia
