Scholarlink

<p>Accurate and timely identification of plant diseases along with their severity is critical for effective crop management and minimizing agricultural losses. While recent advances in deep learning have demonstrated high performance in plant disease classification, limited attention has been given to quantifying disease severity, which is essential for informed agronomic decision-making. To address this gap, this study proposes TomatoMTL, a unified multi-task learning framework for simultaneous disease classification and severity estimation of tomato leaf diseases from a single image. The proposed architecture employs a shared ResNet50-based convolutional backbone augmented with CBAM-based feature refinement, followed by task-specific branches for disease classification and severity prediction. Furthermore, a cross-task attention mechanism is introduced to enable interaction between disease-specific and severity-related features, thereby enhancing the robustness of severity estimation. To effectively leverage partially labeled data, a masking strategy is incorporated during training. Experimental evaluation on a publicly available tomato leaf disease severity dataset demonstrates that the proposed model achieves 97.85% disease classification accuracy and 77.66% severity prediction accuracy, outperforming state-of-the-art single-task classifiers including EfficientNetV2-S, ViT-B/16, and ConvNeXt-Tiny as well as existing multi-task learning baselines including Cross-Stitch Networks and MTAN. Comprehensive ablation studies confirm the individual contributions of CBAM, MixUp and CutMix augmentation, and the cross-task attention mechanism. Statistical significance analysis across five independent runs yields p-values less than 0.001 and Cohen’s d greater than 14, establishing the reliability of the reported improvements. Quantitative localization analysis reveals that the model achieves 89.4% Pointing Game accuracy, confirming that attention maps focus on biologically meaningful disease regions. The proposed framework represents a complete and effective approach for integrated plant disease analysis with strong potential for real-world precision agriculture applications.</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

Tomato leaf disease and severity prediction using multi-task learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月18日

查看全文
  • Anusri Kadam,
  • Parnika Jain,
  • Srishti Tripathi,
  • Tushar Garg,
  • Aniket K. Shahade,
  • Shruti Sunnad,
  • Priyanka V. Deshmukh,
  • Ketan Kotecha

摘要

Accurate and timely identification of plant diseases along with their severity is critical for effective crop management and minimizing agricultural losses. While recent advances in deep learning have demonstrated high performance in plant disease classification, limited attention has been given to quantifying disease severity, which is essential for informed agronomic decision-making. To address this gap, this study proposes TomatoMTL, a unified multi-task learning framework for simultaneous disease classification and severity estimation of tomato leaf diseases from a single image. The proposed architecture employs a shared ResNet50-based convolutional backbone augmented with CBAM-based feature refinement, followed by task-specific branches for disease classification and severity prediction. Furthermore, a cross-task attention mechanism is introduced to enable interaction between disease-specific and severity-related features, thereby enhancing the robustness of severity estimation. To effectively leverage partially labeled data, a masking strategy is incorporated during training. Experimental evaluation on a publicly available tomato leaf disease severity dataset demonstrates that the proposed model achieves 97.85% disease classification accuracy and 77.66% severity prediction accuracy, outperforming state-of-the-art single-task classifiers including EfficientNetV2-S, ViT-B/16, and ConvNeXt-Tiny as well as existing multi-task learning baselines including Cross-Stitch Networks and MTAN. Comprehensive ablation studies confirm the individual contributions of CBAM, MixUp and CutMix augmentation, and the cross-task attention mechanism. Statistical significance analysis across five independent runs yields p-values less than 0.001 and Cohen’s d greater than 14, establishing the reliability of the reported improvements. Quantitative localization analysis reveals that the model achieves 89.4% Pointing Game accuracy, confirming that attention maps focus on biologically meaningful disease regions. The proposed framework represents a complete and effective approach for integrated plant disease analysis with strong potential for real-world precision agriculture applications.