Scholarlink

<p>Potato (<i>Solanum tuberosum </i>L<i>.</i>) crop is highly susceptible to early blight and late blight, which necessitates accurate and prompt disease management. Although recent deep learning methods have shown excellent results in automated detection, most of them are computationally expensive, which restricts their use in resource-constrained agricultural environments. This research introduces the recurrent vision transformer (R-ViT) framework for classifying blight diseases in potato leaf—healthy, early blight, and late blight. This method adapts the recurrent mechanism typically used for temporal sequence modeling to progressively enhance spatial feature extraction from static leaf images. The model resulted in a test accuracy of 99.00%, with precision, recall, and F1 scores of 0.99. These results indicate that the proposed network is on par with convolutional neural networks (CNNs) and standard vision transformers, with better parameter efficiency. To biologically interpret the model, score-weighted class activation mapping (Score-CAM) was employed, whose activation maps emphasized symptomatic necrotic lesions as identified by the model, instead of background noise. By focusing on computational efficiency, training stability, and interpretability, this study illustrates the potential of R-ViT in precision agriculture. Future research will aim at deployment on edge devices and validation with noisy, diverse field data.</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

A Lightweight and Interpretable Recurrent Vision Transformer for Robust Potato Blight Classification

  • 期刊论文

  • 发表日期：

    2026年4月28日

查看全文
  • Prakash Sandhya,
  • B Venkataramana

摘要

Potato (Solanum tuberosum L.) crop is highly susceptible to early blight and late blight, which necessitates accurate and prompt disease management. Although recent deep learning methods have shown excellent results in automated detection, most of them are computationally expensive, which restricts their use in resource-constrained agricultural environments. This research introduces the recurrent vision transformer (R-ViT) framework for classifying blight diseases in potato leaf—healthy, early blight, and late blight. This method adapts the recurrent mechanism typically used for temporal sequence modeling to progressively enhance spatial feature extraction from static leaf images. The model resulted in a test accuracy of 99.00%, with precision, recall, and F1 scores of 0.99. These results indicate that the proposed network is on par with convolutional neural networks (CNNs) and standard vision transformers, with better parameter efficiency. To biologically interpret the model, score-weighted class activation mapping (Score-CAM) was employed, whose activation maps emphasized symptomatic necrotic lesions as identified by the model, instead of background noise. By focusing on computational efficiency, training stability, and interpretability, this study illustrates the potential of R-ViT in precision agriculture. Future research will aim at deployment on edge devices and validation with noisy, diverse field data.