Scholarlink

This research paper advances an automated technique to detect guava diseases using image processing and deep learning algorithms. In addition, it distinguishes between the performances of Convolutional Neural Network (CNN) and DenseNet201 models in classifying guava diseases as well as the need for effective and accurate disease detection in agriculture. There are five classes of data in the dataset; 527 images of guava leaves and fruits including: Disease Free, Phytopthora, Red Rust, Scab and Styler Root. To appreciate their decision-making process, this study makes use of Explainable AI techniques specific to Gradient-weighted Class Activation Mapping (Grad-CAM). In terms of overall accuracy and consistency across all disease categories, DenseNet201 performed better than CNN especially with respect to identifying Phytopthora as well as Styler Root diseases. DunsetNet201 had higher micro-average and macro-average ROC areas compared to CNN. More specifically, Grad-CAM visualizations indicate that DenseNet201’s heatmaps are more spatially detailed vis-a-vis visible disease symptoms hence indicative of better feature detection. Therefore, through this research a more efficient, accurate and explainable approach can be provided for guava disease detection which would help in early management against these diseases in guava growing.

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

Explainable DenseNet201 and CNN for Automated Guava Disease Classification: A Comparative Analysis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nidhi Garg,
  • Gifty Gupta,
  • Vikas Khullar,
  • Isha Kansal,
  • Preeti Sharma,
  • Aditya Singh

摘要

This research paper advances an automated technique to detect guava diseases using image processing and deep learning algorithms. In addition, it distinguishes between the performances of Convolutional Neural Network (CNN) and DenseNet201 models in classifying guava diseases as well as the need for effective and accurate disease detection in agriculture. There are five classes of data in the dataset; 527 images of guava leaves and fruits including: Disease Free, Phytopthora, Red Rust, Scab and Styler Root. To appreciate their decision-making process, this study makes use of Explainable AI techniques specific to Gradient-weighted Class Activation Mapping (Grad-CAM). In terms of overall accuracy and consistency across all disease categories, DenseNet201 performed better than CNN especially with respect to identifying Phytopthora as well as Styler Root diseases. DunsetNet201 had higher micro-average and macro-average ROC areas compared to CNN. More specifically, Grad-CAM visualizations indicate that DenseNet201’s heatmaps are more spatially detailed vis-a-vis visible disease symptoms hence indicative of better feature detection. Therefore, through this research a more efficient, accurate and explainable approach can be provided for guava disease detection which would help in early management against these diseases in guava growing.