Comparative Analysis of Outlier Detection Algorithms in Agricultural IoT Datasets

This paper evaluates the efficiency of eleven outlier detection algorithms applied to three agricultural IoT datasets: Weather in Szeged, Greenhouse, and Crop Recommendation. It focuses on how data quality affects IoT systems' reliability in agriculture. The analyzed algorithms include statistical, distance-based, density-based, ensemble, and hybrid methods, and their ability to detect outliers is investigated using recall, precision, and F1-score. The results demonstrate that algorithm performance is influenced by dataset characteristics such as dimensionality, nonlinearity, and temporal dependencies. Isolation Forest (iForest) consistently outperformed other algorithms due to its non-parametric design and recursive partitioning. In contrast, algorithms like Cluster-Based Local Outlier Factor (CBLOF), K-Nearest Neighbors (KNN), and Gaussian Mixture Model (GMM) showed variable effectiveness depending on dataset features. The study emphasizes the need for selecting the appropriate outlier detection method tailored to specific agricultural datasets and recommends future research on hybrid models or deep learning approaches to enhance anomaly detection in agricultural IoT systems.