Reinforcement learning-enhanced routing for multiple unmanned aerial vehicles in satellite-assisted geohazard monitoring with emergency prioritization

Satellite remote sensing and unmanned aerial vehicles (UAVs) provide complementary capabilities for large-scale geohazard monitoring. Satellites enable rapid wide-area surveillance to identify potential geohazards (PGs), while UAVs perform high-resolution inspections to verify identified PGs and routinely monitor recorded geohazard sites (GSs). However, coordinating multiple UAVs under limited flight endurance and dynamic emergency-priority updates remains a challenging optimization problem. In this paper, we formulate a multi-UAV routing problem with dynamic emergency prioritization under a rolling-horizon monitoring framework. In the proposed system, satellite-identified PGs must be verified by UAVs, after which confirmed sites are dynamically integrated into the routine monitoring process with updated emergency levels. The objective is to jointly optimize mission efficiency and risk coverage by prioritizing inspections of high-risk sites within limited mission time. To solve this highly constrained problem, we develop an Adaptive Large Neighborhood Search algorithm enhanced with Reinforcement Learning (ALNS-RL), which utilizes a deep Q-learning mechanism for state-aware operator selection. Finally, a real-world case study in southern Shaanxi Province, China demonstrates that the proposed ALNS-RL significantly outperforms conventional heuristics in both solution quality and convergence performance.