UAV Swarm Coordination for Flood Area Coverage in Populated Regions using Reinforcement Learning

Flooding is one of the most prevalent natural disasters worldwide and is increasingly recognized as a consequence of climate change. Floods cause substantial economic damage and, moreover, endanger human lives. We present a Deep Reinforcement Learning-based approach, using a centralized Proximal Policy Optimization (PPO)-based agent to coordinate a UAV swarm for the systematic identification of locations with a high likelihood of human endangerment. The agent acts adaptively based on the real-time coverage state, which is crucial for effective inspections of affected areas under a time constraint. We incorporate flood locations and areas of interest-defined by damaged infrastructure-into the decision-making process. We also present a method for extracting relevant data from satellite imagery, based on a previous flood event in the Ahr Valley in Germany in 2021. Our results demonstrate increasing effectiveness in the coverage of diverse flood scenarios. Further advancements are needed before real-world deployment, but the collected data could ultimately be crucial for planning rescue operations and mitigating human risks, especially during the initial disaster response phase. In addition to optimizing coverage efficiency, we highlight key operational risk factors in UAV swarms, such as unpredictable environmental conditions, communication disruptions, and energy constraints, which are essential considerations for ensuring reliable UAV swarm performance in real-world flood scenarios.

Keywords: Unmanned Aerial Vehicles (UAV), Reinforcement Learning (RL), UAV swarm operations, Flood response, Aerial coverage, Autonomous navigation, Emergency management.