A Hierarchical Transformer–CNN Fusion Model for Classification of Mustard Leaves: Distinguishing Healthy, Abiotic Disorder, and Alternaria Leaf Blight Disease

作者：Debapriya Dhar, Pinki Debnath, Tanoy Paul · 年份：2026 · DOI：10.1109/qpain69676.2026.11545956 · 研究领域：Fungal Plant Pathogen Control、Smart Agriculture and AI、Plant Physiology and Cultivation Studies

One of the most devastating mustard diseases is Alternaria leaf blight, which causes huge yield loss if it is not identified accurately. Moreover, diseases and disorders are often misunderstood. Misdiagnosis and the improper application of pesticides have adverse effects on both the food chain and the environment. Therefore, to enable early prediction and precise identification of Alternaria leaf blight disease and abiotic disorders in mustard, this study employed a Transformer-based Vision Transformer (ViT) and a CNN-based EfficientNet-B0 model. A total of 1122 mustard leaf photos were captured from several farmers' fields in the Sylhet region. We chose 853 photos and categorized them into three classes: Healthy, Disorder, and Alternaria. The ViT and EfficientNet models were fine-tuned using the collected dataset to fit the specific task of classification. Using 5-fold cross-validation, both models were evaluated, achieving mean accuracies of 95.08 % for ViT and$\mathbf{9 5. 5 4 \%}$for EfficientNet on the three-class dataset. Two additional binary sub-datasets were prepared: (i) Healthy-Unhealthy (ii) Alternaria-Disorder. The study then compared the performance of ViT and EfficientNet on the sub-datasets (i) and (ii) using 5-fold cross-validation. ViT outperformed EfficientNet on sub-dataset (i), achieving a mean accuracy of 98.59 % compared to 97.42 % of EfficientNet. In contrast, EfficientNet outperformed ViT, attaining a mean accuracy of 97.50 % versus 96.79 % on subdata...