Federated learning meets few-shot learning: A voting ensemble based combined approach to cauliflower leaf disease classification across Non-IID data distributions

In recent years, the classification of cauliflower (Brassica oleracea var. botrytis) leaf diseases has gained significant attention within agricultural research, particularly for optimizing yield and enhancing disease management. Traditional machine learning methods often rely on large, well-distributed datasets, which are difficult to obtain due to diverse data characteristics across regions and farms. To address this challenge, we proposed a novel framework that integrates Federated Learning (FL) with Few-Shot Learning (FSL) for robust cauliflower leaf disease detection, even in environments with non-identical data distributions. Our hybrid method generalizes at the task level without needing large unlabeled datasets, unlike traditional FL methods that rely heavily on data augmentation or semi-supervised techniques to make up for a lack of data. Few-Shot Learning makes it easier for each client to adapt to new disease classes with only a few samples. This makes the hybrid framework more flexible and efficient in situations where the data is not independent and identically distributed (IID). In our study, we utilized five clients, each having their own support and query sets in an n-way k-shot configuration, where each client trains on a small set of labeled data and evaluates the model on unseen data. Federated Learning facilitates collaborative, decentralized training among these clients, while Few-Shot Learning, implemented through the Reptile meta-learning algorithm, allows each client to efficiently adapt to new classes with limited samples. Moreover, we enhance prediction accuracy through ensemble voting, where predictions from multiple pre-trained deep learning models (VGG16, ResNet50V2, Xception, DenseNet169, and MobileNetV2) are combined. The ensemble voting aggregates the individual model predictions, and the most common class prediction across all models is selected, improving the overall classification performance. This ensemble mechanism is implemented after each federated round, where the model weights are aggregated and updated based on local client training, followed by a final ensemble decision for disease classification. Additionally, we have performed an ablation study on our framework to evaluate the contribution of each component (FL, FSL, and voting ensemble). Experimental results show that the ensemble model achieves an test accuracy of 95% for 2-shot, 97% for 3-shot and 4-shot, and 100% for 5-shot configurations, demonstrating the effectiveness of the framework. The results demonstrate that our Fed-FSL hybrid framework, combined with ensemble voting, provides accurate disease classification even in heterogeneous environments, offering a scalable and adaptable solution for precision agriculture and smart farming systems.