Identification of Plant Leaf Disease Using Synthetic Data Augmentation ProGAN to Improve the Performance of Deep Learning Models

Abstract

Detecting plant diseases at an early stage is a critical task in agriculture. Plant disease detection investigations have been conducted over a period of several years. Existing approaches still face challenges in achieving high accuracy and generalization. One key limitation lies in the requisite for a significant amount of labeled data to properly train deep learning models. Moreover, imbalanced and limited datasets often lead to poor predictions and over fitting. Data augmentation strategies such as rotation can be used to get around these obstacles, scaling, and horizontal flipping have been commonly employed. However, recent advancements in synthetic data augmentation, particularly through Generative Adversarial Networks (GANs) have shown promising development in generating high-quality synthetic images for training. This work proposes utilization of Progressive GAN (ProGAN) to generate realistic infected leaf images of Grape for effective data augmentation. By combining the original dataset with these created images, VGG16 deep-learning model, and through extensive analysis, it has been demonstrated that the method greatly improves plant disease identification accuracy. With a notable 4.27% improvement achieved through data augmentation with ProGAN. This work provides a novel solution to deal with the limitations in plant disease detection and underscores the potential of synthetic data augmentation using ProGAN in enhancing the effectiveness of deep learning models for precise diagnosis.