The Identification of Early Blight Disease on Tomato Leaves Utilizing DenseNet Based on Transfer Learning

Abstract

Early blight disease initiated by the fungus Alternaria solani causes reduced tomato harvests of up to 86%. Identification of these diseases manually was prone to identification errors. Thus, it was required to involve deep learning to reduce oversight. This study aimed to determine the performance of the CNN pre-trained model, namely DenseNet based on transfer learning, for identifying early blight disease in tomatoes. The transfer learning technique was carried out by changing the last layer in the model used. The total image data of as many as 3,000 datasets consisting of early blight, healthy, and other disease types were divided into training, validation, and testing data. The data was trained to employ eight different modeling scenarios based on the percentage of data sharing and a combination of hyperparameter tuning. The evaluation results obtained the A4 model as the best model, which uses 2,400 training data, 300 validation data, and 300 testing data. Using a dense layer of 8 neurons, as well as the Adam optimizer with a learning rate of 0.00001. The model succeeded in obtaining validation accuracy values of 91.33%, testing accuracy of 90%, average precision of 90%, average recall of 90%, and testing time of 1.75 seconds. In addition, during the model testing process, it was found that the model was less optimal if it identified new data with conditions and background images that blended with the identified leaf objects.