Scholarlink

<p>Nitrogen is essential for plant growth, yet its deficiency often remains undetected until visible damage occurs. Conventional detection methods are typically time-consuming, costly, and destructive. This study proposes a lightweight deep learning framework for non-destructive detection of nitrogen deficiency in rice crops using RGB leaf images. A publicly available dataset containing 4,890 rice leaf images is used at four levels of nitrogen deficiency. To enhance model robustness, we introduce a novel augmentation technique, RiceLeafAUG, which simulates real-world image variations through controlled object rotation. This method expands the dataset to 41,936 images, improving generalization. A compact convolutional neural network, called Res-CNN, is developed. It is trained on the augmented dataset and achieves 97.91% classification accuracy and 93.80% Intersection over Union (IoU). It performs better than several existing state-of-the-art models. Model interpretability is supported through t-SNE visualizations and calibration analysis, while heatmaps highlight areas of misclassification. The lightweight architecture and high accuracy of Res-CNN make it suitable for mobile agricultural applications, offering a practical solution for monitoring large-scale nitrogen deficiency.</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

Lightweight Deep Learning Approach for Monitoring Plant Stress Caused by Nitrogen Deficiency

  • 期刊论文

  • 发表日期：

    2025年9月10日

查看全文
  • Chiranjit Pal,
  • Imon Mukherjee,
  • Sanjoy Pratihar,
  • Sanjay Chatterji

摘要

Nitrogen is essential for plant growth, yet its deficiency often remains undetected until visible damage occurs. Conventional detection methods are typically time-consuming, costly, and destructive. This study proposes a lightweight deep learning framework for non-destructive detection of nitrogen deficiency in rice crops using RGB leaf images. A publicly available dataset containing 4,890 rice leaf images is used at four levels of nitrogen deficiency. To enhance model robustness, we introduce a novel augmentation technique, RiceLeafAUG, which simulates real-world image variations through controlled object rotation. This method expands the dataset to 41,936 images, improving generalization. A compact convolutional neural network, called Res-CNN, is developed. It is trained on the augmented dataset and achieves 97.91% classification accuracy and 93.80% Intersection over Union (IoU). It performs better than several existing state-of-the-art models. Model interpretability is supported through t-SNE visualizations and calibration analysis, while heatmaps highlight areas of misclassification. The lightweight architecture and high accuracy of Res-CNN make it suitable for mobile agricultural applications, offering a practical solution for monitoring large-scale nitrogen deficiency.