FedPome: federated deep learning for real-time pomegranate disease classification

Pomegranate cultivation faces significant productivity losses due to fungal and bacterial diseases, yet existing automated detection systems rely on centralized deep learning pipelines that require raw image data aggregation, raising data privacy concerns in distributed agricultural settings. This paper proposes a federated learning (FL) framework for raw-data-decentralized pomegranate disease classification, systematically evaluating six architecturally diverse deep learning models such as Custom CNN, ResNet50, ConvNeXt_V2, ViT-B/16, EfficientNetV2-S, and MobileViT-S (under the FedAvg aggregation protocol across five simulated clients and 10 communication rounds). All six architectures are initialized with publicly available ImageNetpretrained weights to ensure a fair architectural comparison. Experiments are conducted on a working set of 6,823 images derived from the Mendeley Pomegranate Fruit Diseases Dataset (5,099 original images) spanning five disease classes-Alternaria, anthracnose, bacterial blight, Cercospora fruit spot, and healthy-using a split-before-augmentation pipeline: the train/validation/test split is performed prior to any augmentation; augmentation is applied only to the training partition (5,800 images), while the validation (511) and test (512) sets contain only original, unaugmented images. EfficientNetV2-S achieves the single highest test accuracy of 99.22%, with ViT-B/16, ResNet50, and ConvNeXt_V2 tied at 98.83% and MobileViT-S close behind at 98.63%, offering the strongest efficiency-accuracy tradeoff. We further report results under non-IID (Dirichlet-partitioned) client distributions, centralized and local-only training baselines, and multi-seed statistical evaluation. All six models are exported to ONNX format and deployed as a real-time web application on Hugging Face Spaces, demonstrating practical deployment readiness. The results confirm the viability of federated learning for high-accuracy, raw-data-decentralized crop disease classification while also characterizing its performance degradation under realistic non-IID conditions.