A GAN-Based Vision Transformer Model for Explainable Plant Disease Identification

Early identification of plant diseases is essential to avoid yield losses and support sustainable agriculture. Deep learning has shown remarkable results in identifying plant diseases, but the issue of little intra-class variation and great interclass variance of visually similar disease categories remains difficult. We show a new deep structure for potato and tomato leaf disease detection and classification. We demonstrate in the current study that training generative adversarial networks (GANs) can improve database images to high quality samples, but this does not affect feature characteristics that convolutional neural networks can exploit. To explain the image feature in training and classification, we provide a new deep architecture depending on the Deep Vision Transformer (DeepViT) model for training and classification and the XAI model based on Gradient-weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP). Under lab-controlled imaging circumstances, the model trained and tested on the PlantVillage dataset of nine tomato and potato diseases classes. With macro ROC-AUC of$0.995,92.11 \%$test accuracy, great class-wise balance as shown by macro precision of 89.70% and recall of 91.30%, we reach 92.11% test accuracy. The visual explanations show the XAI emphasizing the leaf lesion regions generating disease rather than background noise. More clarity and comprehension.