Stratifying Leaf Disease Impact in Strawberry Agriculture with Federated Learning CNNs

Rapid diagnosis of crop diseases at the early stages is an essential instrument to build up agriculture adaptability, specifically for developing countries. The present work fails to describe a novel technique for identifying strawberry leaf diseases by implementing Convolutional Neural Network structures in the condition of Federated Learning architecture systems. The problem was divided into four levels of leaf affliction, and a model was trained on four different consumers. Customers and all the local data were used. The models illustrated the range of diseases and patients being accurate. The FL-CNN model manifests excellent results; the accuracy scores vary between 94% and 96%; such an approach is deeply differentiated on disease level. We combined all these individual data studies into a sound model through aggregation by the federated averaging technique. The macro F1-score, accuracy, and recall ranges were between 87.42% and 91.91%, which can be interpreted as a well-balanced model with relatively even performance across the different severity classes that were the case of class imbalances. Through weighted averages, which considered class majority (56%) to class minority (31%), the model performance was exhibited, ranging from 87.38% to 91.84%. The micro averages, an expression of the model's overall accuracy, were composed of many indicators ranging from 87.38% to 91.8%, substantially in line with the weighted averages. Another crucial factor affecting our ability to act on KPIs from the company depended on the federated learning approach, during which we guaranteed the privacy of every customer data at a local level and enabled the integration of local knowledge of our customers to the global scope of the company. Achieving the objectives indicates that the FL-CNN models work for detecting strawberry diseases and are, therefore, scalable in agriculture as they are privacy-protecting technologies.