TransLeaf Intelligence: A Hybrid Vision Transformer Framework for Next-Gen Phytopathology

Abstract

Precision agriculture increasingly demands intelligent and scalable solutions for early plant disease detection to minimize crop losses and enhance productivity. This study presents an optimized plant disease classification framework based on Vision Transformers (ViT) integrated with transfer learning and lightweight machine learning classifiers. The proposed system combines deep feature extraction using Vision Transformer (ViT) and DenseNet121 (DN121), followed by classification through Perceptron (PC), Nearest Centroid (NC), and K-Nearest Neighbors with Reduced Nearest Centroid (KNN-RNC) models. A Tkinter-based graphical interface is developed to support secure role-based access, dataset management, model training, and real-time prediction. The ViT model effectively captures global contextual dependencies, while DN121 extracts fine-grained local features, enabling robust feature representation. An Explainable AI (XAI) module further enhances system reliability by validating input images and providing semantic interpretations. Experimental evaluation demonstrates that the proposed ViT–Perceptron (ViT-PC) model achieves superior performance with 99.88% accuracy, precision, recall, and F1-score, significantly outperforming DN121-PC, DN121-NC, and DN121-KNN-RNC models. The results highlight the effectiveness of transformer-based architectures in handling complex visual patterns and inter-class similarities in plant diseases. The proposed framework offers a reliable, interpretable, and high-performance solution for real-world phytopathology applications, supporting timely decision-making and sustainable agricultural practices.