ViT-PlantNet: Enhancing Plant Disease Classification with Encoder-Decoder Architecture

Agriculture is essential to human civilization, and numerous diseases often affect crops. Early and accurate classification of these diseases can help prevent significant damage. Traditional methods employed by farmers are usually slow and inaccurate, leading to potential crop loss. This study uses a Vision Transformer (ViT) Encoder-Decoder model to categorize and identify plant diseases. The model is trained using the rich PlantVillage dataset with various plant types and diseases. The proposed ViT model's architecture separates the input image into uniformly sized patches and uses positional embedding to preserve spatial information. The transformer self-attention mecha-nism captures intricate details, leading to accurate and reliable disease classification. Advanced data augmentation techniques such as random rotations, resizing, and color jittering, combined with dynamic callbacks like EarlyStopping, ModelCheckpoint, and ReduceLROnPlateau, enhanced generalization and robustness. The ViT self-attention mechanism further boosted accuracy by effectively capturing local and global dependencies compared to existing state-of-the-art methods, achieving a superior accuracy of 98.73%, highlighting its possibilities for real-world applications and its contribution to the field.