Context-Aware Multi-UAV Fleet Coordination for Flood Response using Deep Reinforcement Learning

With the increasing frequency of severe flood events, the need for rapid and effective disaster response is critical. While autonomous fleets are important for situational awareness, conventional guidance systems utilize static trajectory planning, which proves inefficient in environments with heterogeneous infrastructure density. To address this limitation, we introduce a centralized coordination framework. We extend the standard Deep Reinforcement Learning (DRL) agents by fusing egocentric perception with static geospatial layers (OpenStreetMap, Digital Elevation Model) via a dual-stream Late-Fusion Architecture. This transforms the task from blind coverage to strategic navigation, allowing the fleet to proactively prioritize high-value residential zones. We evaluate the approach in a high-fidelity simulation of the 2021 Ahr Valley flood. Compared to a systematic Lawnmower baseline, the context-aware agents demonstrate superior efficiency. While maintaining comparable spatial coverage ( 64.3 % ), the system achieves a discovery rate of 50.5 %, representing a net semantic gain of 4.9 %. Crucially, the fleet exhibits emergent learned coordination and reactive deviation, diverting from linear paths to trace infrastructure corridors while maintaining safe inter-agent separation. These results validate that geospatial context awareness serves as a robust surrogate for global planning in time-critical environments.

Keywords: Multi-UAV Fleet, Deep Reinforcement Learning, Flood Response, Context-Aware Coordination