A New Federated Learning Framework for Plant Leaf Disease Diagnosis

Abstract

Several Deep Learning-based approaches have recently been proposed for plant leaf disease detection. Traditional learning-based models necessitate centralizing large image datasets, requiring huge volumes of storage space. It appears that a better predictive model can be developed if the available data at multiple data sites can be effectively utilized using a collaborative learning strategy. Federated Learning (FL) offers a decentralized alternative, enabling collaborative model training by allowing multiple participants to update a global model using their local data without sharing high volumes of raw image datasets. Our work presents an FL framework to train a Vision Transformer-based model with an improved Federated Averaging (FedLossAvg) algorithm. Experimental results demonstrate our model's capability to handle large, varied datasets. Using four clients, our global model achieved an F1-score of 95.62% and an accuracy of 99.89% on the PlantVillage data. © 2024 IEEE.