Early Crop Disease Detection using Vision Transformers

Abstract

Crop diseases pose a significant threat to global food security, often resulting in substantial yield losses and economic instability for farmers. Traditional methods of disease identification, which rely on manual visual inspection, are labor-intensive, subjective, and frequently prone to error. While Convolutional Neural Networks (CNNs) have established a baseline for automated detection, they occasionally struggle with capturing global context within complex leaf patterns. This paper presents a robust image-based classifier utilizing Vision Transformers (ViT) to identify crop diseases from leaf images. Leveraging the self-attention mechanism, the proposed model effectively captures long-range dependencies in image data. The system is trained and validated on the PlantVillage dataset using transfer learning techniques. Experimental results demonstrate that the Vision Transformer architecture achieves a classification accuracy of 98.4%, outperforming traditional CNN architectures such as ResNet50 and VGG16. These findings suggest that transformer-based models offer a promising avenue for precision agriculture, enabling early intervention and reduced pesticide usage.

Code and data availability

The paper uses the public PlantVillage dataset and ViT-Base/16 fine-tuning, but provides no authors' public code, trained checkpoints, or paper-specific data deposit; no availability statements or URLs for their assets appear in the supplied blocks.

No evidence-backed public reproduction asset is currently recorded.