A Generalized Federated Learning Approach for Robust Crop Disease Classification Across Diverse Farms

In the current scenario, one of the major problems faced by developing countries like India is food security, as the world’s largest population lives here. Plant-based foods are the primary source of food, yet their susceptibility to diseases further compromises the requirements for food security. Although early detection of plant diseases might help in resolving the diseases in crops, farmers may not be able to identify them due to a lack of knowledge and facilities. Deep learning presents a potential solution that identifies diseases using crop images. Although deep learning is a promising solution, it may compromise data privacy because it may require image data from multiple farms. In this work, we produce a federated learning-based privacy-preserving crop disease classification solution where each farm classifies its own data using the Vision Transformer model and only shares model weights with the centralized server. We evaluated the experiment and found that the Vision Transformer, as a base model, provides $\mathbf{9 2 . 0 \%}$ accuracy in disease detection, which we further enhanced to $\mathbf{9 7 . 0 \%}$ with 20 clients. We find this solution promising for AI in agriculture, as it preserves farm data privacy and improves the accuracy of disease classification compared to existing work.