Scholarlink

<p>Apple leaf diseases such as Apple Scab, Cedar Rust, and Black Rot create significant threats to global fruit production. Traditional disease identification methods rely heavily on manual visual inspection, which is time-consuming, error-prone, and impractical for large-scale agricultural deployment. Deep learning has shown promise in automating plant disease recognition, yet current models often lack the combined strengths of segmentation accuracy, classification robustness, and interpretability. This study aims to develop a&#xa0;unified and explainable deep learning framework capable of performing accurate classification, fine-grained segmentation, and visual interpretability of apple leaf diseases under real-world conditions. This study proposes Quantum-Enhanced TransUNet (QTransLeafNet), an integrated model that combines a Transformer-based U‑Net (TransUNet) for lesion segmentation, a&#xa0;Quantum Variational Classifier (QVC) for enhanced class separation, and Gradient-weighted Class Activation Mapping Plus Plus (Grad-CAM++) for visual interpretability. The model was trained and validated on the Apple Leaf Disease dataset with custom preprocessing and five-fold cross-validation. Performance was benchmarked against Convolutional Neural Network with Long Short-Term Memory (CNN + LSTM), EfficientNet-B0, ResNet-50, and classical TransUNet models. QTransLeafNet achieved an accuracy of 96.21%, F1-score of 96.10%, and an area under the receiver operating characteristic curve (AUC-ROC) of 98.35%, significantly outperforming all baseline models. Segmentation yielded a&#xa0;Dice coefficient score of 0.923 and Intersection over Union (IoU) of 0.887, while Grad-CAM++ achieved an IoU of 0.861 with annotated disease regions. Paired t‑tests (<i>p</i> &lt; 0.01) confirmed the statistical significance of the improvements. Future work will focus on extending QTransLeafNet to field-acquired, multiseason datasets, integrating real quantum hardware for classifier acceleration, and enhancing multimodal interpretability using SHapley Additive exPlanations (SHAP) or Local Interpretable Model-agnostic Explanations (LIME techniques).</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

QTransLeafNet: A Quantum-Enhanced and Explainable Deep Learning Framework for Segmentation and Classification of Apple Leaf Diseases

  • 期刊论文

  • 发表日期：

    2025年9月3日

查看全文
  • Krishan Kumar,
  • Shiva Mehta,
  • Alka Singh,
  • Tejinder Pal Singh Brar,
  • Harpal Singh,
  • Vinay Kukreja

摘要

Apple leaf diseases such as Apple Scab, Cedar Rust, and Black Rot create significant threats to global fruit production. Traditional disease identification methods rely heavily on manual visual inspection, which is time-consuming, error-prone, and impractical for large-scale agricultural deployment. Deep learning has shown promise in automating plant disease recognition, yet current models often lack the combined strengths of segmentation accuracy, classification robustness, and interpretability. This study aims to develop a unified and explainable deep learning framework capable of performing accurate classification, fine-grained segmentation, and visual interpretability of apple leaf diseases under real-world conditions. This study proposes Quantum-Enhanced TransUNet (QTransLeafNet), an integrated model that combines a Transformer-based U‑Net (TransUNet) for lesion segmentation, a Quantum Variational Classifier (QVC) for enhanced class separation, and Gradient-weighted Class Activation Mapping Plus Plus (Grad-CAM++) for visual interpretability. The model was trained and validated on the Apple Leaf Disease dataset with custom preprocessing and five-fold cross-validation. Performance was benchmarked against Convolutional Neural Network with Long Short-Term Memory (CNN + LSTM), EfficientNet-B0, ResNet-50, and classical TransUNet models. QTransLeafNet achieved an accuracy of 96.21%, F1-score of 96.10%, and an area under the receiver operating characteristic curve (AUC-ROC) of 98.35%, significantly outperforming all baseline models. Segmentation yielded a Dice coefficient score of 0.923 and Intersection over Union (IoU) of 0.887, while Grad-CAM++ achieved an IoU of 0.861 with annotated disease regions. Paired t‑tests (p < 0.01) confirmed the statistical significance of the improvements. Future work will focus on extending QTransLeafNet to field-acquired, multiseason datasets, integrating real quantum hardware for classifier acceleration, and enhancing multimodal interpretability using SHapley Additive exPlanations (SHAP) or Local Interpretable Model-agnostic Explanations (LIME techniques).