Scalable and Privacy-Severity Analysis of Pomegranate Leaf Diseases: Federated Learning with CNNs

Abstract

Pomegranate leaf ailments represent a danger to the production and yield of this widely produced fruit. Convolutional neural networks (CNNs), a machine learning methodology that combines computer vision techniques, have shown promise in identifying and classifying various disorders. However, the difficulty of data privacy and ownership often prevents these models from being trained effectively. This paper suggests using CNNs as part of a federated learning technique to overcome this obstacle and maintain the privacy of the training data. According to the approach, five client datasets representing various degrees of pomegranate leaf disease severity were used to train local models. The parameters of the local models were then combined into a global model. A comparison of the local customers showed impressive precision, recall, F1-Scores, and accuracy rates, ranging from 93.74% to 97.71%, demonstrating the model's effectiveness in identifying the disease's development in various regions. With precision rates ranging from 93.74% to 96.32% and equivalent rates for recall, F1-Score, and accuracy, the global model, created by weighting the local models, likewise showed strong performance across all customers. Macro, weighted, and micro averages were used to further analyse the performance of the global model, with findings ranging from 93.79% to 96.52%. These results support the model's balanced performance across various severity levels and its capacity to address class inequality. As a result, the work demonstrates the possibility of a federated learning strategy using CNNs to provide a reliable, scalable, and privacy-preserving method for identifying pomegranate leaf diseases. The results widen the use of federated learning in other fields where data privacy is crucial and provide opportunities for improving agricultural practices.