Lightweight Deep Learning Models for Efficient Cucumber Disease Classification


ÖZKÖK F. Ö.

IEEE Access, cilt.14, ss.167576-167602, 2026 (SCI-Expanded, Scopus)

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 14
  • Basım Tarihi: 2026
  • Doi Numarası: 10.1109/access.2026.3747736
  • Dergi Adı: IEEE Access
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
  • Sayfa Sayıları: ss.167576-167602
  • Anahtar Kelimeler: computational efficiency, Cucumber disease classification, data augmentation, lightweight deep learning, MobileViT-XS
  • Erciyes Üniversitesi Adresli: Evet

Özet

Cucumber is an economically and nutritionally important vegetable crop. Cucumber diseases can significantly affect crop productivity and cause economic losses. Therefore, early and accurate disease detection is important for effective disease management. Deep learning models have been widely used for plant disease classification. However, their computational requirements can limit their use on resource-constrained devices. The main objective of this study is to compare lightweight deep learning models in terms of classification performance, computational efficiency, robustness, and generalization, and to examine how data augmentation and training data selection affect their performance. In this study, nine lightweight deep learning models were evaluated for cucumber disease classification using five-fold cross-validation. The models were evaluated using classification performance and computational efficiency metrics. Random, K-Means, and Diversity were evaluated using 25%, 50%, and 75% of the available training data within each fold. Robustness and external validation experiments were also conducted. RepViT-M1.0 achieved the highest mean accuracy of 94.65% and F1-score of 94.53%. ShuffleNetV2 had the lowest parameter count and estimated computational cost, with 1.26 million parameters and 0.152 G MACs. At the 75% training ratio, K-Means achieved the highest mean in-domain accuracy of 92.11% across the nine models, whereas Random achieved the highest mean external accuracy. In external validation using models trained with the full training data, EfficientNet-B0 achieved the highest accuracy of 86.00% in the eight-class evaluation and 91.50% in the restricted two-class evaluation on the external test partition. These results show that the models with the highest classification accuracy on the original dataset do not necessarily achieve the highest external accuracy or the lowest computational cost.