FedMLKD: Communication-Efficient Federated Learning for UAV Swarm Flash Flood Segmentation
TL;DR
Federated Multi-Level Knowledge Distillation (FedMLKD), a novel framework that enables UAVs to exchange both logits and adaptively pooled intermediate features, is proposed, showing its ability to balance accuracy and efficiency, making it a practical candidate for real-time disaster mapping in bandwidth-constrained UAV swarms.
Abstract
Rapid and precise semantic segmentation of flash floods, such as landslides and debris flows, is critical for emergency response and post-disaster assessment. In recent years, Uncrewed Aerial Vehicle (UAV) swarms offer flexible large-area coverage but face two major challenges for collaborative model training: (i) highly non-Independent and Identically Distributed (non-IID) data distributions that degrade global accuracy, and (ii) prohibitive communication costs in mobile ad hoc networks. To address these challenges, we propose Federated Multi-Level Knowledge Distillation (FedMLKD), a novel framework that enables UAVs to exchange both logits and adaptively pooled intermediate features. This design preserves critical spatial structure while drastically reducing transmission overhead. A client quality–aware aggregation strategy further ensures robust global knowledge construction. Experiments on the Landslide segmentation dataset demonstrate that compared to state-of-the-art baselines, FedMLKD improves mIoU by 9.19% and the average Boundary F1-score by 85.54%, while reducing communication cost by nearly 200 times. These results highlight FedMLKD’s ability to balance accuracy and efficiency, making it a practical candidate for real-time disaster mapping in bandwidth-constrained UAV swarms.
