A Unified Framework for Chili Leaf Disease Classification Using Federated Learning CNN

The increasing complexities in agricultural practices, namely in identifying chili leaf diseases, need intelligent systems that can efficiently analyze large volumes of data. This research presents a unique methodology that utilizes the combination of Federated Learning and Convolutional Neural Networks (CNNs) to facilitate the categorization of five distinct classes of chili leaf illnesses among five separate clients. The architectural design incorporates four convolutional layers, four max-pooling layers, and two fully connected layers. This configuration supports a dimensionality of 234×234 pixels, resulting in ten output nodes representing twelve distinct classes. The technique used in this study yields a distinctive global model via the synthesis of local data models using Federated Averaging methods. While doing the Result Analysis, a comprehensive assessment was conducted on three distinct forms of averages: Macro, Micro, and Weighted. The objective of this evaluation was to get a thorough understanding of the overall performance of the global model. For example, one of the five customers, ES-l, exhibited a Macro average of 82.77%, Weighted average of 83.81%, and Micro average of 83.84%. At the upper range of the spectrum, ES-5 recorded percentages of 91.44%, 91.60%, and 91.60% for Macro, Weighted, and Micro averages, respectively. The figures above demonstrate the model's resilience and ability to effectively manage the trade-off between accuracy and recall across several classes, as seen by an Fl-Score spanning 82.75
^%
to 91.44%. For example, ES-l had a worldwide accuracy rate of 94%, but ES-5 achieved a peak performance with an accuracy rate of 97%.