DeepSMOTE for Tea Leaf Disease Classification Using DenseNet-121 Transfer Learning

Abstract

Data imbalance is a common problem in image classification that causes models to tend to be biased towards the majority class, so that the performance of the minority class decreases. To overcome this problem, this study applies the DeepSMOTE approach as a deep learning-based oversampling technique that produces synthetic images through interpolation in the latent space of the autoencoder. The balanced training data images were used to train the DenseNet-121 model through a transfer learning approach. Evaluations were carried out on three approaches, DeepSMOTE as the main approach, as well as two comparators, namely no augmentation and conventional augmentation. The results showed that DeepSMOTE had the highest accuracy, which was 97.73%, higher than the 8.99% non-augmentation approach and 4.49% from the conventional augmentation. F1-score value on minority classes such as Anthracnose increased by 0.2833 in the non-augmentation approach and 0.1318 in the conventional augmentation approach.These findings confirm that data balance plays an important role in improving model performance and that DeepSMOTE is able to provide a more even distribution of data thus supporting better generalization of image classification tasks.