Addressing data scarcity in deep learning-based plant disease detection using gan-based synthetic data augmentation

Abstract

Early detection of plant diseases is critically important for preserving agricultural productivity and ensuring food security. Image processing and deep learning techniques can provide faster, more objective, and consistent results compared to human visual assessments in this field. However, the success of deep learning models largely depends on the availability of large and balanced datasets. In this study, a deep neural network based on the ResNet architecture was trained using an enriched image dataset, which was augmented via LeafGAN for the purpose of plant leaf disease detection. To overcome the limitations of the original dataset, such as data imbalance and insufficient sample size, the LeafGAN model was first trained to generate synthetic leaf images that represent plant diseases. These synthetic images were then combined with the original dataset to create a more balanced and diverse training set. Subsequently, a model was developed based on the ResNet architecture and trained using this dataset. The impact of the applied data augmentation method on model performance was thoroughly analyzed. The results show that the dataset augmented with LeafGAN significantly improves classification accuracy. This study demonstrates that GAN-based data generation provides a practical and effective solution for training deep learning models with agricultural images. It also offers a valuable methodology for enhancing model performance in other domains facing data scarcity challenges.