Federated Semantic Segmentation of Synthetic UAV Imagery under Non-IID Conditions

Unmanned Aerial Vehicles (UAVs) rely on semantic segmentation for monitoring, inspection, and disaster response tasks. Traditional centralized training requires the transfer of large data volumes, which is often infeasible because of privacy, bandwidth, and decentralization constraints. Federated Learning (FL) offers a promising alternative; however, non-IID data hinder its effectiveness because UAVs operate in diverse environments with heterogeneous distributions. This study evaluated the adaptation of a semantic segmentation model to a federated setting using synthetic UAV imagery. We investigated four training configurations by varying the communication rounds, local epochs, and learning rates, and analyzed their effects on convergence and class-level accuracy. The results suggest that moderate communication with local training achieves balanced performance, whereas high learning rates or excessive synchronization cause instability and class-specific biases. These findings open avenues for designing robust federated UAV segmentation strategies under non-IID conditions.