Determination of grape varieties with deep learning techniques

In viticulture, the ampelographic characteristics of the shoot, leaf, inflorescence and fruit of the varieties are used to determine the grape varieties. After the ampelographic features are determined, they are expressed numerically or verbally. In this study, classification of fifty grape cultivars with deep learning techniques was done by using ampelographic features of leaves, bunch and fruits. In the study, a new and originally developed CNN model is proposed. The proposed CNN model was developed on the MATLAB 2021a platform, using a total of 27 320 dataset images, 227x227x3 in size, 9 854 leaves, 8 745 bunch and 8 721 fruits. 80% of the data set is reserved for training and 20% for validation. Classification was carried out in two stages with leaf and bunch /fruit images. Since the characteristics of grape varieties are very similar to each other, in order to avoid memorization in the model, a total of nine different categories were created, five different categories in the leaf group and four different categories in the bunch /fruit group. There are ten classes in each of the five different categories with leaf images, and eleven classes in each of the four different categories with bunch /fruit images. Model performance was determined by calculating the Accuracy, Sensitivity, Sensitivity and F1 Score values of each category separately. In addition, GoogleNet and AlexNet models were also trained with the newly created dataset to compare the proposed model performance. Accuracy success rates in the leaf group for the new model; 82.50% for category 1, 90.03% for category 2, 87.44% for category 3, 94.10% for category 4 and 92.40% for category 5. Accuracy and success rates in the bunch /fruit group; 85.88% for category 1, 84.90% for category 2, 96.40% for category 3 and 97.20% for category 4. The success rate was 84.39% in the GoogleNet model and 92.31% in the AlexNet model. As a result of the study, the highest success rate among the three models was calculated with 97.20% in category 4 in the bunch/fruit group, and the lowest success rate in category 1 in the leaf group was calculated with the new model with an accuracy of 82.50%. When these results are examined, it has been revealed that the models work successfully in the classification of grape varieties, can learn the ampelographic properties of grape varieties and can be used in studies in this field.