Automated Leaf Disease Classification via Three-Way Information-Guided Knowledge Distillation with Feature Interaction Maximization
IEEE Access, cilt.14, ss.131546-131573, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 14
- Basım Tarihi: 2026
- Doi Numarası: 10.1109/access.2026.3722572
- Dergi Adı: IEEE Access
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
- Sayfa Sayıları: ss.131546-131573
- Anahtar Kelimeler: convolutional neural networks, EfficientNet-B7, knowledge distillation, Leaf diseases, MobileNet-V3
- Gazi Üniversitesi Adresli: Evet
Özet
Accurate and efficient leaf disease classification is essential for intelligent crop health monitoring and precision agriculture. Although deep convolutional neural networks achieve remarkable performance, their high computational complexity limits deployment on resource-constrained devices. Moreover, conventional knowledge distillation methods primarily focus on output-level supervision while overlooking informative intermediate feature interactions. To address these challenges, this paper proposes Three-Way Information-Guided Knowledge Distillation (TWIG-KD), a lightweight framework that integrates an EfficientNet-B7 teacher and an improved MobileNetV3 student through a three-way information-guided distillation strategy. Furthermore, a Feature Interaction Maximization (FIM) module is introduced to enhance feature-level knowledge transfer by maximizing class-relevant information while suppressing redundant feature interactions. The proposed framework is evaluated on the PlantVillage and Rose Leaf Disease datasets. Experimental results achieve 99.81% accuracy and 99.85% F1-score on PlantVillage, and 92.58% accuracy and 92.36% F1-score on the Rose Leaf Disease dataset, outperforming several state-of-the-art methods. Ablation and interpretability analyses further demonstrate that TWIG-KD and FIM learn compact, discriminative, and semantically consistent feature representations while maintaining computational efficiency. These findings indicate that the proposed framework is well suited for practical plant disease diagnosis on resource-constrained agricultural platforms.