Automatic detection of grape varieties with the newly proposed CNN model using ampelographic characteristics

• Published in: Scientia horticulturae 2024-08, Vol.334, p.113340, Article 113340
• Main Authors: Terzi, Ismail, Ozguven, Mehmet Metin, Yagci, Adem
• Format: Article
• Language: English
• Subjects: AlexNet; Ampelography; automatic detection; class; CNN; data collection; Deep learning; fruits; Googlenet; Grape; grapes; leaves
• ISSN: 0304-4238
• DOI: 10.1016/j.scienta.2024.113340

•In this study, 50 grape varieties were classified with deep learning techniques using ampelographic features on leaf, bunch and fruit images.•A total of 27,320 images, including 9854 leaf, 8745 bunch and 8721 fruit images, were used as the data set in the study.•The new and uniquely developed CNN model was used in classification processes with deep learning techniques, and the performance of this model was evaluated with pre-trained models such as GoogleNet and AlexNet.•In the study, it was determined that the newly developed model gave better results than other models.•These results suggest that deep learning architectures can be used in the automatic classification of vine varieties, that this approach can help producers and the wine industry, and that integrated systems with vehicles such as autonomous robots and drones can be developed in the future. In this study, leaves, bunches and fruits of fifty grape varieties were classified using deep learning techniques using ampelographic features. A new and unique CNN model has been proposed for the classification process. In addition, GoogleNet and AlexNet models adapted to the data set with the transfer learning method were also used. The dataset was divided into two groups: leaf and cluster/fruit. A total of 27,320 images of 227 × 227 × 3 size, including 9854 leaves, 8745 bunches and 8721 fruits, were used in the data set. The dataset was randomly divided 80 % (21,856 images) for training and 20 % (5464 images) for testing. Grape varieties were classified in a total of nine different categories: five different categories in the leaf group and four different categories in the cluster/fruit group. Each class in the categories represents a grape variety. Each category of the leaf group consists of ten classes, and each category of the cluster/fruit group consists of eleven classes. The results were obtained by calculating the Accuracy, Sensitivity, Recall and F1 Score values of the categories separately for the three models. In the newly developed CNN model, the highest accuracies were determined as 94.10 % in the leaf group and 97.20 % in the cluster/fruit group in Category 4. The accuracies of GoogleNet and AlexNet models were determined as 84.39 % and 92.31 %, respectively. According to the experimental results obtained, it was determined that the proposed model showed successful performance in the classification of grape varieties. Thus, it was demonstrated that deep learning models can be used successf