Revolutionizing in Agriculture: Federated CNN Models for Sunflower Leaf Diseases

• Main Authors: Joshi, Jyoti, Aeri, Manisha, Kukreja, Vinay, Mehta, Shiva
• Format: Conference Proceeding
• Language: English
• Subjects: (CNN)_(FL); Adaptation models; Augmentation image; Convolutional neural networks; Data models; data privacy; data security; Disease; Federated learning; Leaf diseases; Sunflower leaves; Training; Transforms
• ISSN: 2769-2884
• DOI: 10.1109/ICRITO61523.2024.10522441

With the advent of an age when the reason for adopting precision agriculture to ensure essential food security and economic sustainability is foremost, the precise recognition and tattooing of sunflower leaf disease is an inauspicious part of agricultural study. In a learning process characterized by interconnectedness, collaboration, and data protection and confidentiality maintenance, the five participants (xc_1 to xc_5) have applied different data representations to the model. The summary explains our approach to analytical matters, citing the influence of global and local data analyzed via the trained AI model using federated averaging. The most considerable average accuracy was observed for all clients at xc_5, 95.71%, whereas the first version was much lower at xc_1, only 88.76 % . It demonstrated clear progress from the eventual to the initial stages of the model learning without bias towards certain classes. Weighted averages are essential to the analysis of the datasets where there are imbalances. With xc_1, the weighted average starts at 89.15 and reaches the top with xc_5 of 95.72. The fact that it is more accurate and adaptable when categorizing common diseases is demonstrated by the improved precision that is illustrated in general. The micro average is the resultant average that depicts the general model accuracy, while the weight averages are the averages based on the weight. To be specific, the accuracy for sample 1 reached 89.13%, and sample 5 overviewed the accuracy level and reached 95.72%. The core of the work is federated averaging, which transforms the world's local neighborhood analytics into a universal understanding of the world. It refers to the mix of local training for separate models inside and outside the globe.