Citation

Iftee, Iftehad Kamal and Ahmed, Farhana and Jobaier, Mahmudul Hasan and Uddin Osama, Amad and Abdul Aziz, Nor Hidayati (2025) Smart Agriculture: Cucurbit Leaf Disease Detection Using Swin Transformer. In: 2025 International Conference on Engineering and Emerging Technologies (ICEET), 22-23 October 2025, Kuala Lumpur, Malaysia.

Text 5.pdf - Published Version Restricted to Repository staff only Download (1MB)

Abstract

Diseases such as downy mildew, anthracnose, and mosaic virus result in extensive 30–70% yield loss in cucurbit crops, threatening food security in countries like Bangladesh. Recent advances in deep learning have improved the identification of plant diseases; however, traditional CNN-based methods typically lack robustness to heterogeneity in the field and are yet to be practical for scalable use. To address these limitations, this work compares five state-of-the-art models, CNN with self-attention, ResNet, DenseNet, MobileNetV2, Vision Transformer, and Swin Transformer, on a newly developed cucurbit leaf dataset of 12,206 field-harvested images across five balanced classes. From the CNN-based models, the self-attention CNN achieved 88.10% accuracy but performed poorly for visually similar classes, misclassifying downy mildew and healthy leaves. ResNet, DenseNet, and MobileNetV2 also offered competitive performance (87–88% accuracy) but showed lower generalization in noisy field settings. The Vision Transformer improved overall performance (89.17% accuracy), but the Swin Transformer was the best-performing model and achieved peak accuracy (94.49%), macro F1-score (0.922), and improved per-class robustness, such as +9.1% F1 gain on downy mildew versus CNN. These results demonstrate hierarchical vision transformer capability for fine-grained disease classification and lay a foundation for real-time deployment on drone and mobile platforms to enable scalable smart agriculture in remote resource-limited regions.

Item Type:	Conference or Workshop Item (Paper)
Uncontrolled Keywords:	Deep learning, Swin Transformer, Vision Transformer (ViT), Resnet, resource-limited agriculture, CNN with self-attention model, Smart agriculture, curcubit leaf diseases detection
Subjects:	S Agriculture > SB Plant culture
Divisions:	Faculty of Engineering and Technology (FET)
Depositing User:	Ms Suzilawati Abu Samah
Date Deposited:	20 Apr 2026 01:36
Last Modified:	20 Apr 2026 01:36
URII:	http://shdl.mmu.edu.my/id/eprint/15743

Downloads

Edit (login required)