Reinforcement Learning Guided UAV Swarm Navigation for Disaster Surveillance via Multimodal Satellite Image Intelligence
Аннотация
Disaster surveillance using UAVs is often limited by static flight planning, weak swarm coordination, and poor integration with real-time satellite intelligence, resulting in low coverage and high energy consumption. This paper proposes RL-SwarmNav, a reinforcement learning-based UAV swarm navigation framework that integrates multimodal satellite image intelligence for priority-driven disaster monitoring. The system utilizes Proximal Policy Optimization (PPO) to train UAV agents for energy-efficient, collision-free navigation, while disaster hotspots extracted from the FloodNet dataset guide adaptive path planning. UAV learning is performed using the SimDrone dataset to model realistic flight behavior. Experimental results show that RL-SwarmNav achieves a 94.7% coverage rate and 12.3 s hotspot response time, outperforming conventional PPO and DQN methods. The framework demonstrates scalable, intelligent, and real-time UAV coordination, offering an effective solution for large-scale disaster management and emergency response.