The increasing prevalence of cotton leaf diseases poses substantial risks to crop productivity, highlighting the need for scalable and privacy-aware diagnostic frameworks in modern agriculture. Conventional centralized machine learning approaches require aggregation of field data from multiple institutions, raising concerns related to data ownership, privacy, security, and regulatory compliance. Federated Learning (FL) addresses these challenges by enabling collaborative model training while retaining data locally. In this study, each FL client is modelled as an independent agricultural entity, such as a cooperative, extension office, or research station, possessing region-specific image data that reflects agroclimatic variability. Building upon EfficientNetB0 as the base model, we propose a federated cotton leaf disease classification framework integrating the FedSCAFFOLD algorithm with progressive layer unfreezing and data augmentation strategies to mitigate client drift under heterogeneous data distributions. Experimental results show test accuracies of 97.01% with four clients and 95.92% with eight clients. These findings demonstrate competitive performance under the evaluated configurations and suggest robustness across heterogeneous client settings. Overall, the proposed framework indicates potential applicability for privacy-aware, collaborative agricultural disease monitoring in distributed environments.
Ram et al. (Mon,) studied this question.