Scholarlink

<p>Reliable detection of potato leaf diseases is vital for crop protection and sustainable agriculture. This study introduces PotatoXViT, a lightweight and explainable Vision Transformer (ViT)–based model designed to improve generalization under real-world field conditions. The model includes a customized patch embedding module, a two-layer Transformer encoder for long-range feature modeling, a global average pooling for compact representation, and a classification head that enhances inter-class separability. To enhance transparency, Grad-CAM–based explainability is employed to visualize disease-discriminative regions and guide optimizer selection, providing interpretable insights into the model’s decision-making while ensuring stable convergence. Experimental results demonstrate that PotatoXViT achieves high and consistent classification performance, with accuracies of 99.16% on the PlantVillage potato leaf subset and 99.29% on the Irish potato dataset, indicating strong robustness to varying data distributions. Compared to existing studies, the proposed model provides an improved balance between accuracy, computational efficiency, and interpretability. Indeed, it provides a compact, interpretable, and effective solution for potato leaf disease classification, contributing to more reliable, trustworthy, and practical agricultural AI systems.</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

PotatoXViT: XAI-Enhanced Lightweight Vision Transformers for Potato Leaf Disease Classification

  • 期刊论文

  • 发表日期：

    2026年6月15日

查看全文
  • Amal Jlassi,
  • Amira Soudani,
  • Walid Barhoumi

摘要

Reliable detection of potato leaf diseases is vital for crop protection and sustainable agriculture. This study introduces PotatoXViT, a lightweight and explainable Vision Transformer (ViT)–based model designed to improve generalization under real-world field conditions. The model includes a customized patch embedding module, a two-layer Transformer encoder for long-range feature modeling, a global average pooling for compact representation, and a classification head that enhances inter-class separability. To enhance transparency, Grad-CAM–based explainability is employed to visualize disease-discriminative regions and guide optimizer selection, providing interpretable insights into the model’s decision-making while ensuring stable convergence. Experimental results demonstrate that PotatoXViT achieves high and consistent classification performance, with accuracies of 99.16% on the PlantVillage potato leaf subset and 99.29% on the Irish potato dataset, indicating strong robustness to varying data distributions. Compared to existing studies, the proposed model provides an improved balance between accuracy, computational efficiency, and interpretability. Indeed, it provides a compact, interpretable, and effective solution for potato leaf disease classification, contributing to more reliable, trustworthy, and practical agricultural AI systems.