Quantified Explainability and Uncertainty-Aware Apple Leaf Disease Detection: Uncertainty-Aware Enhanced EfficientNet
Keywords:
Abstract
In information technology, accurately detecting apple leaf diseases is crucial to minimise losses and support sustainable agriculture. This paper proposes a reliable, efficient four-class apple disease classification system based on EfficientNet-B0 with CBAM (Convolutional Block Attention Module). The system uses Grad-CAM for interpretability, Monte Carlo Dropout for uncertainty estimation, and evaluation metrics to assess the reliability of confidence estimates. Grad-CAM explanations achieved a mean IoU of 0.25 and a Dice score of 0.783 against ground-truth lesion masks, while Monte Carlo Dropout produced a mean entropy of 0.0098. The model achieved 99.58% test accuracy, 0.0040 ECE, and a Brier Score of 0.006. During deployment, ONNX Runtime reduced CPU inference time from 36.11 ms to 5.94 ms, achieving 168.2 frames per second and a model size of 18.56 MB.
Downloads
Published
How to Cite
Issue
Section
License

This work is licensed under a Creative Commons Attribution-NoDerivatives 4.0 International License.