Decentralized Coconut Leaf Disease Detection with Privacy-Preserving Federated CNNs

Several ailments are known to affect coconuts, which pose a great threat to the production of coconuts and, in return, threaten farmers' livelihood and the economy. Traditional approaches to disease diagnosis involve centrally located data repositories, leading to privacy and spatial distribution problems. To this end, this research proposes an FL approach involving CNN for the decentralized detection of coconut leaf diseases across different clients. This framework helps maintain data privacy while using less bandwidth than the previous method. We used datasets from four different areas to train our model: there are 3,190 images in total, which are healthy and affected coconut leaves. The federated CNN model, as suggested, got the global accuracy to about 27 variables out of 87. 5%, and the recall of 85.3%, recall of 84. 1%, the recall 80 %, and the F-measure 84.7%. Nonindependency of the data was dealt with in the aspect of the IID by applying a weighted Federated Averaging (FedAvg). This technique helped ensure clients with more diversified datasets were given more control over the global model. Client 2, employing a wider data set, obtained a local accuracy rate of 89.1%, while Client 2’s accuracy was at 71.4% due to their employing specialized data. The federated strategy decreases bandwidth requirements by approximately 75% compared to traditional centralization methods, making it suitable for implementation in rural or remote areas with limited access.