The Application Of The Transfer Learning Method In The Classification Of Eggplant Diseases Using The Densenet-201 Architecture

Abstract

Disease affecting eggplant fruit (Solanum melongena L.) is one of the main obstacles that reduce crop productivity, while conventional detection methods are considered inefficient as they require expert knowledge and a considerable amount of time. This study aims to apply the transfer learning method with the DenseNet201 architecture to classify eggplant fruit diseases into three categories, namely healthy fruit, rotten fruit, and fruit borer, based on digital images. The research dataset consisted of 2,417 images, divided into training, validation, and testing data with a 70:15:15 ratio, which were then processed through resizing, normalization, denoising, and data augmentation on the training set so that each class contained 1,000 images. The model was built using pre-trained ImageNet weights and trained for 20 epochs with the Adam optimizer, achieving a training accuracy of 90.0% and a validation accuracy of 87.5% with no significant indication of overfitting. During testing, the model achieved a test accuracy of 81.69% with a weighted F1-score of 81.68%. The healthy fruit class showed the best performance with an F1-score of 90.77%, while the rotten fruit and fruit borer classes achieved F1-scores of 78.48% and 74.11%, respectively, due to the visual similarity between the two classes. System testing on sample images from each class resulted in entirely accurate classifications, with confidence values ranging from 78.35% to 99.69%. These results indicate that the DenseNet201 model based on transfer learning performs well and has the potential to be used as an early detection support system for eggplant fruit disease.