Iot-Enabled Evidence-Fusing Graph Neural Network With Attention-Based Feature Engineering For Uncertainty-Aware Crop Recommendation

Abstract

Precision agriculture demands not only accurate crop recommendations based on heterogeneously measured soil and environment, but also confidence information for the decisions being made in the field. In this study, a novel IoT-Enabled Attention Multiscale Evidence-Fusing Graph Neural Network Framework (IoT-AM-EFGNN) for intelligent and uncertainty-aware crop recommendation is developed. The framework combines the IoT sensing, AMTNet-inspired multiscale attention feature engineering, Evidence Fusing Graph Neural Network (EFGNN), Enzyme Action Optimizer (EAO), and SHapley Additive exPlanations (SHAP). Experimental evaluation has been performed on Kaggle Crop Recommendation Dataset which consists of data points comprising 22 crop classes and 7 agricultural parameters of 2200 samples. Missing value imputation, Z-score outlier detection, Min-Max normalization, Label encoding, train-test split with ratio 80:20 stratified was applied. We then created Soil Health, Climate, and Moisture indices, which were further enhanced using attention-based feature fusion. The EFGNN learned relationships between agricultural samples by propagating multi-level graphs and fused class-specific Dirichlet evidence to achieve crop probabilities and uncertainty measures. EAO fine-tuned the model, and SHAP offered feature-level interpretability. Experimental results demonstrated 99.18% accuracy, 99.12% precision, 99.08% recall, 99.10% F1-score, and 99.42% ROC-AUC. The cross-validation results show that the method achieves high accuracy with a mean of 99.18% and a standard deviation of 0.16%, which indicates that the method is stable. The proposed framework's accuracy was 1.72% higher than the baseline MLP and 0.44% higher than a conventional GNN. Confidence distribution was between about 0.95 and 1.00 on every prediction. The deployment of IoT-cloud is also a facility for real-time agricultural decision making.