Scholarlink

<p>Timely and precise identification of rice leaf ailments is crucial for maintaining the well-being of the crop and achieving optimal output. This study assesses advanced machine learning structures for the automatic recognition of rice leaf diseases by examining two standard datasets with XAI incorporation. The initial dataset comprises 4,684 pictures depicting three categories of diseases: bacterial blight, brown spot, and leaf smut. The other one contains 10,165 images comprising bacterial leaf blight, healthy leaf, rice blast, tungro, and rice. An organized evaluation of ResNet50 with Vision Transformer (ViT) and Hybrid ConvNeXt models is performed under exactly identical preprocessing and training conditions. The three-class dataset gets to, accuracy from both ViT and Hybrid ConvNeXt models was 100%, while ResNet50 reached 99.72%. The highest accuracy for the five-class dataset was 90.69% with ResNet50, followed by Hybrid ConvNeXt at 87.94% and ViT at 86.76%. We use Grad-CAM, LIME, SHAP, and ViT attention maps to interpret model decisions, which show the regions of the leaf which greatly influence the predictions of disease. Our results offer an elaborate and open structure for precision agriculture in facilitating a timely yet very accurate diagnosis of rice diseases.</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

Enhancing green AI through explainable deep learning-based multi-model for automated rice leaf disease classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月17日

查看全文
  • Ruaa A. Al-Falluji,
  • Marwan Ali Albahar

摘要

Timely and precise identification of rice leaf ailments is crucial for maintaining the well-being of the crop and achieving optimal output. This study assesses advanced machine learning structures for the automatic recognition of rice leaf diseases by examining two standard datasets with XAI incorporation. The initial dataset comprises 4,684 pictures depicting three categories of diseases: bacterial blight, brown spot, and leaf smut. The other one contains 10,165 images comprising bacterial leaf blight, healthy leaf, rice blast, tungro, and rice. An organized evaluation of ResNet50 with Vision Transformer (ViT) and Hybrid ConvNeXt models is performed under exactly identical preprocessing and training conditions. The three-class dataset gets to, accuracy from both ViT and Hybrid ConvNeXt models was 100%, while ResNet50 reached 99.72%. The highest accuracy for the five-class dataset was 90.69% with ResNet50, followed by Hybrid ConvNeXt at 87.94% and ViT at 86.76%. We use Grad-CAM, LIME, SHAP, and ViT attention maps to interpret model decisions, which show the regions of the leaf which greatly influence the predictions of disease. Our results offer an elaborate and open structure for precision agriculture in facilitating a timely yet very accurate diagnosis of rice diseases.