Hierarchical Federated Learning for Natural Disaster Management

The increasing impact and frequency of natural disasters due to climate change highlight the need for improved disaster management strategies. Centralized systems are often too slow and inefficient for near real-time responses. To address this, the Horizon Europe-funded TEMA project aims to develop a disaster management solution that utilizes distributed architectures and Federated Learning in constrained environments. This paper presents the Drones Hierarchical Federated Learning (DHFL) algorithm, designed to train neural networks through the dynamic aggregation of variable unmanned aerial vehicles (UAVs) clusters. We evaluate DHFL’s performance in terms of accuracy, training time, energy consumption, and network communications, and compare it against traditional Hierarchical Federated Learning (HierFL) and FedAVG. Our results, based on tests using the MNIST dataset, show that DHFL achieves a 92.39% accuracy—comparable to HierFL—while operating in highly variable conditions.