Innovative Approaches to Java Plum Leaf Disease Identification: Federated Learning meets Convolutional Neural Networks

The spread of plant diseases negatively impacts agricultural output considerably, demanding quick, accurate, and reliable detection techniques. This paper presents a revolutionary method for classifying four types of Java plum leaf illnesses using a federated learning convolutional neural network (FL-CNN) model. The study achieves the combined goals of improving illness diagnosis accuracy and protecting data privacy across distant datasets using federated learning capabilities. The FL-CNN model was tested on a testbed consisting of five clients, each with its unique local dataset. The local performance indicators, such as precision, recall, F1-score, and accuracy, consistently showed great values of over 90% for each client across all illness classes, highlighting the model's effectiveness. Client 4 achieved the most incredible local precision for Class A illness (96.35%). The global model showed similarly encouraging results once federated averaging was applied to the model parameters, with precision, recall, F1-score, and accuracy values consistently close to 95%, 88%, 91%, and 0.96 across the five clients, respectively. Three global aggregation techniques were also studied in this study: macro, weighted, and micro averages. For the five customers, the macro average showed performance values ranging from 89.52% to 91.67%. A similar pattern was seen in the weighted average, with values ranging between 89.47% and 91.72%. Finally, the micro average showed that Client 1 performed better than the other clients (91.82%), with values for the other clients hovering around 91.41%. These consistently strong performance measures, locally and internationally, show how the FL-CNN model is resilient and flexible regarding various assessment approaches. The study demonstrates how federated learning may improve illness diagnosis while maintaining data privacy, opening the door for more research in this exciting area.