John Michael D. Villagomez, Matthieu Blue D. Montecillo, Christian Maderazo
This paper proposes an explainable hybrid deep learning ensemble for multi-crop leaf disease classification across the three crops: coconut, banana, and sugarcane. Three architectures: ResNet50, EfficientNet-B0, and ViT-Small are trained independently and combined using rank-based exponential weighted averaging voting, achieving 99.79% overall accuracy, macro precision of 99.88%, recall of 99.90%, and F1-score of 99.89%. across 18 disease classes, including an outstanding perfect 100% accuracy on sugarcane. A three-layer explainability framework integrates Gradient-weighted Class Activation Mapping (Grad-CAM) for Convolutional Neural Networks (CNN) heatmaps, Attention Rollout for Vision Transformer attention visualization, and Shapley Additive Explanations (SHAP) for unified feature attribution. Experimental results demonstrate that the hybrid ensemble outperforms all individual models and other related works models while producing interpretable, biologically meaningful visual explanations suitable for adoption by agricultural technicians. © 2026 IEEE.
Dept. of Computer, Information Science and Mathematics, University of San Carlos, Cebu City, Philippines
