Abstract:
In precision agriculture, machine learning models are increasingly revolutionizing
disease detection and management for crops. Recent advances in computer vision,
particularly Vision Transformer (VIT) architectures, have shown promising results
in accurately identifying plant diseases. This study explores three VIT models.
variants: Pre-trained VIT, Mobile VIT, and Scratch VIT applied to Okra leaf disease
detection, an area where effective early identification can significantly improve crop
yield. The dataset, containing 3,775 Okra leaf images includes four disease classes
and represents diverse environmental conditions to ensure robust model
performance. Among the models, the Pre-trained VIT demonstrated the highest
performance, achieving 96% validation and test accuracy with an AUC score of 0.985,
indicating strong generalizability and minimal overfitting. Scratch VIT followed
closely with 93% validation accuracy, 94% test accuracy, and a 0.98 AUC score,
showcasing reliable classification despite being trained from scratch. Mobile VIT
achieved 83% validation and 85% test accuracy with an AUC of 0.962, suggesting
some limitations in handling complex features. The study highlights the Pre-trained
VIT’s potential as a reliable and efficient solution for okra disease detection,
providing an effective tool for farmers to make informed, proactive decisions in crop
management. This approach emphasizes the value of VIT models in advancing
precision agriculture through accurate, disease classification.
