AHE-FNUQ: An Advanced Hierarchical Ensemble Framework with Neural Network Fusion and Uncertainty Quantification for Outlier Detection in Agri-IoT

Agricultural Internet of Things (Agri-IoT) systems need strong anomaly detection to monitor crops effectively. However, current approaches lack accuracy and efficiency. To mitigate this problem, we proposed an advanced hierarchical ensemble framework with neural network fusion and uncertainty quantification (AHE-FNUQ). This framework combines six detection algorithms: Isolation Forest, ECOD (empirical cumulative distribution-based outlier detection), COPOD (copula-based outlier detection), HBOS (histogram-based outlier score), OC-SVM (one-class support vector machine), and KNN (k-nearest neighbors). It uses a three-level decision process: (1) selecting models with good performance (ROC AUC > 0.75), (2) applying recall-weighted ensemble fusion, and (3) using a fusion neural network (FusionNN) to improve uncertain predictions in the confidence range [0.75,0.9]. The framework was tested on three agricultural datasets with contamination levels between 10% and 50%. The result showed strong performance: ROC AUC between 0.93 and 0.99, PR AUC between 0.90 and 0.98, and F1-scores between 0.85 and 0.90. Moreover, we have conducted a statistical test (Friedman test, χ2=63.02, p<0.0001) and confirmed that AHE-FNUQ is significantly better than common methods such as COPOD, ECOD, HBOS, Isolation Forest, and KNN.