Scholarlink

<p>Plant-parasitic nematodes (PPNs) cause significant yield losses in agricultural production, and the rapid and accurate identification of these organisms is crucial for the timely implementation of appropriate control strategies. This study aimed to classify the genera <i>Meloidogyne, Pratylenchus</i>, and <i>Helicotylenchus</i> using deep learning (DL), utilizing a&#xa0;dataset of 1629 microscopic images obtained from different ecosystems in Turkey. In this context, current Convolutional Neural Network (CNN)-based models including ResNet50, DenseNet121, EfficientNet-B0, and ConvNeXt-Tiny, along with Transformer-based Swin‑T and ViT-B/16 architectures, were trained for 30&#xa0;epochs with a&#xa0;batch size of&#xa0;32 using a&#xa0;transfer learning approach. Model performance was evaluated using accuracy, precision, recall, F1-score, ROC-AUC, and statistical significance analyses. The results showed that all models achieved high accuracy levels of approximately 98–99%. Swin‑T achieved the highest overall performance with 99.39% accuracy and an F1-score of 0.9936. Among the CNN-based architectures, DenseNet121 and ConvNeXt-Tiny also demonstrated competitive performance, each achieving 98.78% accuracy. Statistical significance analysis indicated that the observed performance differences among the evaluated models were not statistically significant. Grad-CAM++ analyses revealed that the models focused on genus-specific morphological regions during classification, confirming that the predictions were biologically meaningful. These findings demonstrate that DL offers strong potential as a&#xa0;fast, accurate, and interpretable decision-support tool for PPN diagnosis and may help reduce dependency on nematology experts.</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

Grad-CAM++ Guided Deep Learning-Based Genus-Level Identification of Plant-Parasitic Nematodes in Microscopic Images

  • 期刊论文

  • 发表日期：

    2026年8月2日

查看全文
  • Furkan Ulaş,
  • Kemal Adem,
  • Dilek Dinçer,
  • Muhammad Aasim

摘要

Plant-parasitic nematodes (PPNs) cause significant yield losses in agricultural production, and the rapid and accurate identification of these organisms is crucial for the timely implementation of appropriate control strategies. This study aimed to classify the genera Meloidogyne, Pratylenchus, and Helicotylenchus using deep learning (DL), utilizing a dataset of 1629 microscopic images obtained from different ecosystems in Turkey. In this context, current Convolutional Neural Network (CNN)-based models including ResNet50, DenseNet121, EfficientNet-B0, and ConvNeXt-Tiny, along with Transformer-based Swin‑T and ViT-B/16 architectures, were trained for 30 epochs with a batch size of 32 using a transfer learning approach. Model performance was evaluated using accuracy, precision, recall, F1-score, ROC-AUC, and statistical significance analyses. The results showed that all models achieved high accuracy levels of approximately 98–99%. Swin‑T achieved the highest overall performance with 99.39% accuracy and an F1-score of 0.9936. Among the CNN-based architectures, DenseNet121 and ConvNeXt-Tiny also demonstrated competitive performance, each achieving 98.78% accuracy. Statistical significance analysis indicated that the observed performance differences among the evaluated models were not statistically significant. Grad-CAM++ analyses revealed that the models focused on genus-specific morphological regions during classification, confirming that the predictions were biologically meaningful. These findings demonstrate that DL offers strong potential as a fast, accurate, and interpretable decision-support tool for PPN diagnosis and may help reduce dependency on nematology experts.