FedRefiner: Robust Federated Learning with Noisy Label Refinement for Remote Sensing

Remote sensing applications rely on cross-regional, multi-sensor data collected by distributed institutions, where privacy and bandwidth constraints prevent centralized aggregation. Data privacy concerns, limited bandwidth, and transmission constraints restrict centralized training and motivate distributed learning paradigms. Federated learning (FL) emerges as a promising distributed training approach, only requiring the model parameter transmission without sharing raw data. However, obtaining clean supervision for remote sensing imagery is costly, and low-cost alternatives such as crowd-sourcing and auto-labeling inevitably introduce label noise. To tackle the label noise challenge in distributed remote sensing applications, we propose FedRefiner, a framework that cautiously refines unreliable samples with noisy labels. It separates reliable and noisy samples through a dual-view two-component Gaussian mixture model that jointly considers per-sample loss and representation-to-centroid consistency, and then refines high-confidence noisy samples with model predictions. In addition, we introduce learning-to-cluster representation optimization to compact intra-class representations and separate inter-class representations. All refinement operations run locally on each client without transmitting additional sensitive information. Extensive experiments on two real-world remote sensing datasets demonstrate that FedRefiner outperforms 21 baselines across diverse noise settings. To facilitate communities, our code will be open-sourced.