CONVOLUTIONAL NEURAL NETWORKS FOR MULTI-CLASS WHEAT FUNGAL DISEASE DIAGNOSIS: A COMPARATIVE STUDY OF TRANSFER LEARNING MODELS

---

Gedik C., Sayıncı B., Yıldız T.

Black Sea Journal of Agriculture, cilt.9, sa.3, ss.332-344, 2026 (TRDizin)

• Yayın Türü: Makale / Tam Makale
• Cilt numarası: 9 Sayı: 3
• Basım Tarihi: 2026
• Doi Numarası: 10.47115/bsagriculture.1836017
• Dergi Adı: Black Sea Journal of Agriculture
• Derginin Tarandığı İndeksler: TR DİZİN (ULAKBİM)
• Sayfa Sayıları: ss.332-344
• Bilecik Şeyh Edebali Üniversitesi Adresli: Evet

Özet

 Wheat (Triticum aestivum L.) is a staple food for billions worldwide and plays a critical role in global food security. However, fungal pathogens cause yield losses and significant economic damage, posing a serious threat to wheat production. Therefore, early and accurate detection is strategically important to mitigate these losses. In this study, a dataset comprising 5,325 images of seven major wheat fungal diseases Blast, Brown Rust, Stripe Rust, Fusarium Head Blight, Loose Smut, Powdery Mildew, and Septoria along with healthy samples was used. The images were split into training (70%), validation (15%), and test (15%) subsets, and four pretrained convolutional neural network (CNN) models, MobileNet, NASNetMobile, DenseNet121 and DenseNet169 were fine-tuned using transfer learning. Results showed that DenseNet169 achieved the highest classification accuracy (86.87%), followed by DenseNet121 (84.16%) and MobileNet (81.07%). NASNetMobile, however, demonstrated the lowest performance during validation. The findings highlight the strong potential of DenseNet169 in achieving high accuracy for wheat disease classification and its applicability in developing early detection systems that support sustainable agricultural production.