Scholarlink

<p>Timely disease detection is essential for sustainable table grape production in tropical regions where fast-spreading pathogens and restricted expert access delay timely treatment. This paper presents an interpretable deep learning architecture combining Progressive Generative Adversarial Network (ProGAN) augmentation, multi-scale Vision Transformer ensemble, and a&#xa0;Hierarchical Capsule Attention Network (HCANet) for reliable grape leaf disease diagnosis. Quality-insured ProGAN augmentation verified via Fréchet Inception Distance (FID: 11.4–14.2) and expert agreement (Fleiss’ κ = 0.87) addresses class imbalance while ensuring biological realism. The transformer ensemble of local-global features leverages adaptive fusion, and HCANet incorporates dynamic routing and self-attention to enable interpretable spatial understanding. On the dataset of 5847 infield images across three districts of Tamil Nadu, we achieved 98.73% accuracy, 98.46% macro F1 score, and 0.992 area under the receiver operating characteristic curve (AUC-ROC), which outclass state-of-the-art baselines by 1.6–2.0&#xa0;percentage points (<i>p</i> &lt; 0.001). Cross-district validation substantiated 91.2% generalization accuracy with just 7.5% drop compared to conventional models’ typical 18–35% declines. Gradient-Weighted Capsule Activation Mapping obtained 94.3% Intersection over Union (IoU) agreement with pathologist labeling, outperforming Convolutional Neural Network Gradient-weighted Class Activation Mapping (CNN Grad-CAM) (78.4%) and transformer attention (84.6%). The optimized 49-MB architecture runs in 47 ms on smartphones.</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

An Explainable Deep Learning Framework Integrating Progressive GAN Augmentation and Capsule-Transformer Networks for Early Grape Leaf Disease Detection in Indian Vineyards

  • 期刊论文

  • 发表日期：

    2026年3月18日

查看全文
  • S. Priya,
  • A. Subashini,
  • Elamurugan Balasundaram

摘要

Timely disease detection is essential for sustainable table grape production in tropical regions where fast-spreading pathogens and restricted expert access delay timely treatment. This paper presents an interpretable deep learning architecture combining Progressive Generative Adversarial Network (ProGAN) augmentation, multi-scale Vision Transformer ensemble, and a Hierarchical Capsule Attention Network (HCANet) for reliable grape leaf disease diagnosis. Quality-insured ProGAN augmentation verified via Fréchet Inception Distance (FID: 11.4–14.2) and expert agreement (Fleiss’ κ = 0.87) addresses class imbalance while ensuring biological realism. The transformer ensemble of local-global features leverages adaptive fusion, and HCANet incorporates dynamic routing and self-attention to enable interpretable spatial understanding. On the dataset of 5847 infield images across three districts of Tamil Nadu, we achieved 98.73% accuracy, 98.46% macro F1 score, and 0.992 area under the receiver operating characteristic curve (AUC-ROC), which outclass state-of-the-art baselines by 1.6–2.0 percentage points (p < 0.001). Cross-district validation substantiated 91.2% generalization accuracy with just 7.5% drop compared to conventional models’ typical 18–35% declines. Gradient-Weighted Capsule Activation Mapping obtained 94.3% Intersection over Union (IoU) agreement with pathologist labeling, outperforming Convolutional Neural Network Gradient-weighted Class Activation Mapping (CNN Grad-CAM) (78.4%) and transformer attention (84.6%). The optimized 49-MB architecture runs in 47 ms on smartphones.