Enhanced Vision Transformer Model with Multi-Scale Attention for Robust Plant Leaf Disease Identification and Classification

Abstract

Plant leaf diseases pose a major risk to global food security, requiring accurate and robust detection systems. Traditional deep learning models struggle with diverse plant species, multiple disease types, and complex visual variations, affecting their generalization in real agricultural settings. To address this, the AgriViT-MSCA model is proposed, combining a Vision Transformer architecture with Multi-Scale Channel-Spatial Attention to enhance both global and local feature learning for improved disease classification. Evaluations on the PlantVillage dataset show high performance with 99.28% prediction accuracy and class-specific accuracies above 99.9% for Tomato, Apple, Maize, and multi-class datasets. This demonstrates AgriViT-MSCA’s strong generalization ability, making it a promising approach for intelligent plant disease monitoring.