Performance comparison of deep and transfer learning models for smart soil texture classification

ÇETİN N., KAVUNCUOĞLU E., BUZPINAR M. A., Gunaydin S., KAPLAN S.

Computers and Electronics in Agriculture, cilt.237, 2025 (SCI-Expanded) identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 237
  • Basım Tarihi: 2025
  • Doi Numarası: 10.1016/j.compag.2025.110722
  • Dergi Adı: Computers and Electronics in Agriculture
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, PASCAL, Aerospace Database, Applied Science & Technology Source, BIOSIS, CAB Abstracts, Communication Abstracts, Compendex, Computer & Applied Sciences, Environment Index, Food Science & Technology Abstracts, INSPEC, Metadex, Veterinary Science Database, Civil Engineering Abstracts
  • Anahtar Kelimeler: Convolutional Neural Networks, Deep Learning, Soil Texture Classification, Transfer Learning, YOLOv5
  • Sivas Cumhuriyet Üniversitesi Adresli: Evet

Özet

Classification of soil is crucial for implementing precision agriculture practices to achieve better crop planning and resource utilization. The method of this study involved utilizing a deep learning-based transfer learning model for soil texture classification using image data. For background removal to precisely extract soil features, we processed a dataset of 720 soil images from six different soil types using YOLOv5. Nine state-of-the-art transfer learning models, namely, DenseNet121, Xception, MobileNetV2, and VGG19, were evaluated in terms of classification accuracy, computational efficiency, and memory usage. Experimental results showed that the test accuracy for DenseNet121 was the best with 97.22 %, Xception was 93.52 %, and MobileNetV2 was 72.22 %. The computational efficiency analysis showed that MobileNetV2 converged the fastest (769.53 sec) and used the smallest memory (0.74 GB). On the contrary, DenseNet121 and Xception, though consuming more memory, showed a better reliability of classification. Future research may focus on improving lightweight architecture or optimizing facility extraction techniques to enhance classification accuracy and reduce computational costs. These results indicate the promise of deep learning models for soil texture classification, which can be applied in accurate agriculture and sustainable land management.