A Cross-Domain Transfer Learning Framework for Robust Plant Disease Identification from Leaf Imagery

Abstract

Harvest efficiency depends on the plant health and climate variations. Identification of leaf disease at right time, using leaf imagery, certainly helps in treating the crop and getting good yield. Recent advancements in transformer-based models, particularly Vision Transformers (ViT), have revolutionized agricultural image analysis by capturing complex, non-linear patterns. Despite their effectiveness, ViTs require large, labelled datasets, posing challenges in plant disease identification due to similar symptoms and limited data. This study explores data augmentation and transfer learning with two ViT variants, ViT- Base and ViT-Large, using training-from-scratch and feature extraction techniques. The Robo-Flow augmentation combined with ViT-GZSL feature extraction achieved 96.62% accuracy. The work uses transfer learning in the background. ViT-Base excelled in classifying corn, chili, tea, and tomato diseases, while ViT-Large performed best on peanut and potato crops, reaching up to 98.85 % accuracy.