Performance-Guided Aggregation for Federated Crop Disease Detection Across Heterogeneous Farmland Regions

A region-aware federated learning framework (RAFL) is proposed to address the non-IID heterogeneity in multi-regional crop disease recognition while reducing communication and computation costs. RAFL integrates three complementary modules: a region embedding module that captures region-specific representations, a cross-region feature alignment module that aligns semantic distributions across regions on the server, and an attention-based aggregation module that dynamically weights client updates based on performance through Transformer attention. Without sharing raw images, RAFL achieves efficient and privacy-preserving collaboration among heterogeneous farmlands. Experiments on datasets from Bayan Nur, Zhungeer, and Tangshan demonstrate substantial improvements: a classification accuracy of 89.4%, an F1-score of 88.5%, an AUC of 0.948, while the detection performance reaches mAP@50=62.5. Compared with FedAvg, RAFL improves accuracy and F1 by over 5%, and converges faster with reduced communication overhead (total 2822 MB over 95 rounds). Ablation studies verify that the three modules act synergistically—regional embeddings enhance local discriminability, feature alignment mitigates cross-domain drift, and attention-based aggregation stabilizes training—resulting in a robust and deployable solution for large-scale, privacy-preserving agricultural monitoring. Furthermore, the framework enables regional-level economic analysis by correlating disease incidence with yield reduction and estimating potential economic losses, providing a data-driven reference for agricultural policy and resource allocation.