Hierarchical Semi-Supervised Federated Learning for UAV-Enabled Fire Monitoring

Wildfire detection and monitoring increasingly rely on distributed sensing platforms capable of operating in dynamic and hazardous environments. While Unmanned Aerial Vehicles (UAVs) are playing an increasingly key role in fire-monitoring applications, their deployment is limited by technical and environmental constraints. Despite advances in detection and segmentation algorithms, several challenges persist due to the limited availability of labelled data, infrastructural and technological limitations, and unique characteristics of each wildfire. In this paper, we present Semi-DHFL, a new framework for fire and smoke segmentation that combines the advantages of Hierarchical Federated Learning with semi-supervised learning techniques on UAV-enabled distributed edge computing systems. By leveraging the limited amount of labelled data together with the large volume of unlabelled data continuously collected in real time by UAVs, we define a semi-supervised local training procedure for each client, and implement a pixel-wise FedAvg aggregation both at the edge server level and at the global cloud level, enabling effective model updates. The experimental results show that the proposed approach outperforms state-of-the-art semi-supervised federated learning baselines, achieving improvements of up to 26.37% in mIoU, 12.51% in mean Dice, and 6.15% in accuracy. Furthermore, comparison with a fully supervised scenario demonstrates that Semi-DHFL can achieve comparable performance, confirming its effectiveness in realistic, partially labelled wildfire monitoring settings. In addition, Semi-DHFL achieves lower computational energy by 6.78% and total training time by 6.82%, demonstrating higher efficiency for resource-constrained UAV deployments.