Mango leaf disease detection and classification using a federated learning based lightweight vision transformer model

The detection and management of horticultural crop diseases is crucial for enhancing productivity, reducing costs, and promoting environmentally sustainable practices. Recently, various techniques have been developed for managing crop-related issues, including leaf diseases and pest control. A centralized control approach has been implemented for disease detection. This centralized system faced significant challenges, including concerns over data privacy and the costs associated with data migration. This study addressed the problems with centralized solutions through a federated learning approach enhanced by the attention mechanism of the vision transformer (ViT). We investigated the key features of the federated learning method and propose a deep learning model based on federated learning, utilizing the attention mechanism of the ViT to detect diseases in mango plant leaves. The study utilized an openaccess dataset of mango leaf diseases sourced from Kaggle. The performance of the federated model depended heavily on several factors, including data quality, the number of participating agents, their communication efficiency, and the number of local iterations. Although various pretrained models were tested to address these challenges, the lightweight ViT was the most effective, achieving an impressive accuracy rate of 98.99% in federated learning scenarios.