Beyond boundaries: synergizing SAR, UAV and federated learning for flood mapping in Australia

Flooding has become increasingly severe in recent years, causing major displacement and infrastructure damage. While synthetic aperture radar (SAR) satellites and UAVs aid flood monitoring, SAR is limited by resolution and dynamic flood boundaries, and UAVs struggle with large-area coverage. To address this, we propose a novel framework that fuses multiscale SAR and unmanned aerial vehicles (UAVs) data, enhanced by a federated learning (FL) scheme and feedback loop to build a vertical space-air sensing structure. FL effectively handles the data heterogeneity and decentralized nature of real-world sensing platforms, avoiding the challenges of central aggregation. Tested in flood-prone Moama, NSW, Australia, the approach significantly improves flood assessment. Compared to centralized training, the FL framework boosts average kappa and F1 scores by 1.7%–4%, with the best vector (Vector 4) improving from 82.30% to 85.21% in kappa and from 83.01% to 85.87% in F1, demonstrating strong scalability, accuracy, and robustness.