Mitigating the Impact of Guava Leaf Diseases Using CNNs and Federated Learning

Abstract

Farmers may suffer considerable financial losses due to guava leaf infections, notably in India, where guava planting is common. This study uses federated learning and convolutional neural networks (CNNs) to categorize and diagnose guava leaf diseases at five severity levels. A framework comprising five clients has been used to guarantee a varied and inclusive data representation. Each client represents a distinct severity degree of guava leaf disease. Utilizing local data from each customer, the performance of the created model is assessed. The model demonstrated exceptional precision, recall, F1-score, and accuracy outcomes. The accuracy measures continuously remained over 95%, with precision values ranging from 91.25% to 96.79%, recall from 90.83% to 96.87%, and F1-scores from 86.90% to 96.83%. Additionally, the model's overall performance, as determined by federated averaging, showed outstanding values for accuracy, recall, and F1-score, ranging from 89.43% to 94.09%, 89.74% to 94.18%, and 89.56% to 94.13%, respectively. These numbers demonstrate the model's ability to generalize well from various datasets, with astounding accuracy exceeding 96%. Additionally, macro, weighted, and micro averages were used to examine the overall performance. The weighted averages varied from 90.59% to 94.38%, the micro averages from 90.57% to 94.38%, and the macro averages for all customers fell between 89.57% and 94.13%. The promise of federated learning and CNNs as effective methods for controlling and treating agricultural diseases is highlighted in this research. This will help farmers increase crop productivity and ensure food security.