Scholarlink

<p>This paper introduces the Hybrid deep learning model that fuses a MobilenetV2 CNN as base network, in pipeline followed by a convolutional block attention mechanism and a transformer working in parallel for local and global feature extraction respectively. The objective to use this hybrid model is fine-grained four-class severity classification of yellow rust in wheat. The publicly available YellowRust-19 dataset, treated as gold standard for yellow rust, has been restructured into four disease progression stages (Healthy, Early, Middle, and Advanced) aligned with plant stages followed by agronomists and scaled as per Cobb’s scale. The dataset comprising of 15&#xa0;K RGB wheat leaf images, the model is trained with a class-imbalance-aware strategy employing balanced class-weighted cross-entropy loss with label smoothing and a two-phase frozen/unfrozen backbone curriculum. The training accuracy of 99.9% and the overall accuracy of 93.13% for testing dataset samples has been achieved. Notably high per-class precision is significant for early detection contribution of the proposed algorithm. The cross validation and statistical pair t-test establish the model stability and robustness against the baseline CNN. Computational complexity analysis helps understand the strength and efficiency of the proposed architectural pipeline. Comparison of the proposed approach against recent state-of-the-art methods has been presented. The findings highlight the promise of attention-augmented CNN architectures for precision disease staging and open practical pathways for real-time field diagnostic applications in wheat-growing regions. To further validate generalizability, the proposed has been evaluated on an independent real-world augmented dataset of 87 wheat field images collected from the fields across Punjab by experts of Punjab Agricultural University (PAU), Ludhiana, India from Jan, 2024 to Feb, 2026. The model successfully classified all images into three rust progression stages—early, middle and advanced stages. The results are significant with respect to performance gap in inoculated studies used predominantly in the existing published work and real fields with little control on environmental conditions. To understand the misclassifications, explainable AI(XAI) Grad-CAM has been applied and heatmaps analysed which endorse the capabilities of the model.</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

HYRNet- hybrid yellow rust stage-aware severity classification XAI enabled deep learning framework

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月28日

查看全文
  • Gaganpreet Kaur,
  • Jasmine Panesar,
  • Jaspal Kaur

摘要

This paper introduces the Hybrid deep learning model that fuses a MobilenetV2 CNN as base network, in pipeline followed by a convolutional block attention mechanism and a transformer working in parallel for local and global feature extraction respectively. The objective to use this hybrid model is fine-grained four-class severity classification of yellow rust in wheat. The publicly available YellowRust-19 dataset, treated as gold standard for yellow rust, has been restructured into four disease progression stages (Healthy, Early, Middle, and Advanced) aligned with plant stages followed by agronomists and scaled as per Cobb’s scale. The dataset comprising of 15 K RGB wheat leaf images, the model is trained with a class-imbalance-aware strategy employing balanced class-weighted cross-entropy loss with label smoothing and a two-phase frozen/unfrozen backbone curriculum. The training accuracy of 99.9% and the overall accuracy of 93.13% for testing dataset samples has been achieved. Notably high per-class precision is significant for early detection contribution of the proposed algorithm. The cross validation and statistical pair t-test establish the model stability and robustness against the baseline CNN. Computational complexity analysis helps understand the strength and efficiency of the proposed architectural pipeline. Comparison of the proposed approach against recent state-of-the-art methods has been presented. The findings highlight the promise of attention-augmented CNN architectures for precision disease staging and open practical pathways for real-time field diagnostic applications in wheat-growing regions. To further validate generalizability, the proposed has been evaluated on an independent real-world augmented dataset of 87 wheat field images collected from the fields across Punjab by experts of Punjab Agricultural University (PAU), Ludhiana, India from Jan, 2024 to Feb, 2026. The model successfully classified all images into three rust progression stages—early, middle and advanced stages. The results are significant with respect to performance gap in inoculated studies used predominantly in the existing published work and real fields with little control on environmental conditions. To understand the misclassifications, explainable AI(XAI) Grad-CAM has been applied and heatmaps analysed which endorse the capabilities of the model.