Enhancing Vision Transformers with Kolmogorov–Arnold Networks for Plant Leaf Disease Classification

Creative Commons License

Tekin B., Zontul M., ADEM K.

Potato Research, cilt.69, sa.5, 2026 (SCI-Expanded, Scopus)

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 69 Sayı: 5
  • Basım Tarihi: 2026
  • Doi Numarası: 10.1007/s11540-026-10095-y
  • Dergi Adı: Potato Research
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Ultimate (EBSCO), Natural Science Collection (ProQuest)
  • Anahtar Kelimeler: Kolmogorov–Arnold Networks (KAN), Plant leaf disease classification, PlantVillage dataset, Potato and maize disease classification, Vision Transformer (ViT), ViT-KAN
  • Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
  • Sivas Cumhuriyet Üniversitesi Adresli: Evet

Özet

Early and accurate detection of plant diseases is vital for global food security and sustainable agriculture. While deep learning offers promising solutions, there is a continuous need for architectures that enhance learning capacity and efficiency. This study introduces ViT-KAN, an innovative hybrid model merging the powerful feature extraction of Vision Transformers (ViT) with the flexible, learnable activation functions of Kolmogorov-Arnold Networks (KAN). By replacing the standard Multilayer Perceptron (MLP) classification head of ViT with a KAN module, the proposed architecture aims to better capture nonlinear patterns in agricultural images. Evaluated on the PlantVillage dataset for potato and maize leaf diseases using standard fivefold cross-validation, with final results reported as mean ± standard deviation across the five folds, the model was trained entirely from scratch. ViT-KAN achieved 99.49 ± 0.13% accuracy on the maize dataset and 98.28 ± 0.51% on the potato dataset, compared with 98.92 ± 0.40% and 97.77 ± 0.88%, respectively, for the standard ViT model. Beyond mean accuracy, ViT-KAN showed lower standard deviation across folds, while representative fold curves suggested smoother early training trajectories under the shared training configuration. These findings suggest that ViT-KAN is a promising alternative to conventional ViT-based classification models for plant disease diagnosis.