A biologically structured hierarchical vision transformer-CNN framework for robust tomato leaf disease classification.

Abstract

Precise and reliable diagnosis of leaf diseases in tomato is essential for enhancing crop cultivation and minimizing agricultural losses. While deep learning models have performed well on benchmark datasets, the majority of present techniques rely on flat multi-class classification, which predicts all disease categories simultaneously. Such formulations promotes inter-class confusion, particularly when biologically different diseases with similar visual symptoms are learned within a same model. To overcome this constraint, we propose a biologically structured hierarchical deep learning framework in this study. Instead of directly classifying 10 disease classes, the proposed method first classifies leaf images into meaningful biological groups such as bacterial, fungal, pest-associated and healthy using a vision transformer (ViT) model. Then, specialized convolutional neural network (CNN) experts perform fine-grained classification within each category. The proposed hierarchical model shows an overall accuracy of 97.8%, when validated on PlantVillage tomato dataset. A flat ViT model trained with class-weighted loss obtained 96.3% accuracy, whereas a flat CNN model reached 99.3% under clean conditions but decreased sharply to 41% under Gaussian perturbation ( σ = 0.05). On the other hand, the hierarchical model performed steadily under noise with 97.4% accuracy at the same perturbation level. These results indicate that adding biological structure to model design reduces confusion, helps prevent imbalance effects and increases robustness, providing a more trustworthy and interpretable solution for real-world agricultural disease diagnosis.

Plant phenotyping relevance

トマト葉画像から病害状態を分類する深層学習手法を開発し、既存モデルとの比較およびノイズ下での技術検証を行っており、植物表現型（病害状態）の取得・推定が中心である。

Code and data availability

The paper's phenotyping/classification measurements are based on the public PlantVillage tomato leaf image dataset, which the authors explicitly state was analyzed and provide a public Kaggle URL. No author analysis code, trained models, or paper-specific supplementary assets are described with availability language.