A lightweight privacy-aware federated vision transformer for tomato leaf disease classification

Smart farming increasingly depends on intelligent disease diagnosis systems to improve crop productivity while protecting the privacy of agricultural data distributed across farms, research institutions, and edge devices. Conventional centralized deep learning approaches require collecting large volumes of data at a central location, raising concerns about data privacy, ownership, and communication overhead. To address these challenges, this study proposes FedVRT, a lightweight privacy-aware federated learning framework for tomato leaf disease classification. The proposed framework combines a Vision Transformer (ViT) with a lightweight ResNet26 architecture within the Federated Averaging (FedAvg) framework, enabling collaborative model training without exchanging raw agricultural images. The Vision Transformer captures global contextual relationships among disease symptoms, while the ResNet26 branch extracts fine-grained local lesion and texture features. Their complementary feature representations are fused to improve disease classification under heterogeneous non-IID data distributions commonly encountered in real-world agricultural environments. The proposed framework was evaluated on a dataset of 11,108 tomato leaf images across six disease categories in federated settings with 3, 12, 24, and 36 distributed clients. Performance was assessed using accuracy, precision, recall, F1-score, area under the receiver operating characteristic curve (AUC), confusion matrix analysis, confidence intervals, and statistical significance tests. FedVRT achieved a classification accuracy of 98.50%, with precision 98.42%, recall 98.35%, F1-score 98.38%, and an AUC of 0.997 across 36 participating clients. Comparative experiments further demonstrate that the proposed framework consistently outperforms established federated optimization methods, including FedSGD, FedProx, FedNova, SCAFFOLD, FedOpt, and FedDyn, while maintaining lower communication overhead and faster convergence. Unlike existing federated Vision Transformer-based approaches that primarily focus on classification accuracy, FedVRT integrates lightweight hybrid feature extraction, comprehensive statistical validation, scalability analysis, and communication efficiency evaluation within a unified framework. These characteristics demonstrate the effectiveness of FedVRT as a scalable, privacy-aware, and practically deployable solution for intelligent tomato leaf disease diagnosis in next-generation smart farming environments.