Effective plant disease diagnosis using Vision Transformer trained with leafy-generative adversarial network-generated images

Abstract

Agriculture, as the foundation of human civilization, is critical to the global economy, providing food for billions. Plant diseases, caused by factors such as bacteria, fungi, viruses, and others, loom large over crop yields, jeopardizing farmers’ livelihoods worldwide. Rapid and accurate identification of these diseases is critical for agricultural productivity protection and to date, several automated plant disease diagnosis methods have been developed by researchers worldwide. However, the issue of having limited labeled datasets for certain plant leaf diseases poses a significant challenge in training classification models effectively. This scarcity often results in class imbalance which adversely affects a model's ability to accurately predict all the disease classes. It appears there is a need to explore synthetic data generation techniques to train the model for making a better prediction. Further, the disease prediction model should be lightweight so that it can be conveniently integrated with low-end devices with less computational power that farmers can afford to purchase. In this work, we aim to develop an effective neural augmentation model that can render synthetic disease patterns on uninfected leaf images thereby enhancing the leaf disease dataset by adding artificial samples corresponding to those disease classes for which only minor ground truth information is available. Our work extends the state-of-the-art by introducing a new model for leaf disease augmentation, termed “LeafyGAN”, that comprises two key elements: a segmentation model and a disease translation model, both of which are GAN-based. The segmentation model is a pix2pix GAN that is trained to separate foreground leaf images from the background and is trained using a combination of L1 loss and standard GAN loss. The disease translation model is a CycleGAN which is trained using a combination of adversarial loss and cycle consistency loss, which uses the generated segmented mask to render synthetic disease patterns to the extracted leaf regions. A lightweight MobileViT model trained using this augmented data has been seen to perform disease diagnosis with a remarkable accuracy of 99.92% on the PlantVillage dataset and 75.72% on the PlantDoc dataset. Notably, our model achieves an accuracy that is comparable with the recent CNN and Transformer-based models with a significantly lesser number of parameters. © 2024 Elsevier Ltd