Classification of Tea Leaf Clone Using Custom Convolutional Neural Network Based on VGG-16

Each type of tea leaf has unique characteristics which may affect the taste and aroma of tea beverage. However, manual recognition of the tea leaf clone only can be carried out by experts in tea picking because the similarities in shape and color between several types of tea leaves. Therefore, this study proposed an automatic model to classify tea leaf clone from its image by using a custom convolutional neural network (CNN) based on VGG-16. The 13 convolutional layers of VGG-16 is used for feature extraction, then to reduce the size of model the first and the second hidden layer are removed, replaced with the use of batch normalization and dropout layer. The image dataset to train the model was captured directly using a digital camera at tea plantations of PT. Pagilaran. This study focused on the three types of tea leaf clones, namely Gambung 7, Gambung 9, and TRI 24. The dataset is divided into 70% training data, $15 \%$ validation data, and $15 \%$ test data. The training process implements transfer learning to initialize weights and speed up model training time. The hyperparameters adjusted for model training are batch size, learning rate, and dropout. The best hyperparameter values are used to build the final model. The evaluation of the final model on testing data is able to reach the highest test accuracy of ${9 3. 2 0 \%}$. In addition, the size of the proposed model is almost ten times lower than original VGG network.