Deep reinforcement learning–driven multi-satellite collaborative observation planning for emergency and disaster monitoring

Satellite remote sensing plays a crucial role in emergency disaster monitoring. However, variations in orbital parameters and inconsistent coverage frequently lead to inefficient observation and underutilization of valuable satellite resources. Therefore, we propose a coordination framework based on deep reinforcement learning to optimize multi-satellite observation scheduling, enabling more rapid and accurate disaster sensing. Specifically, the approach consists of three key components. First, it involves compiling an on-orbit satellite resource database by aggregating various orbital parameters. Second, it refines satellite imaging modes based on the geometric principles of on-orbit observation. Third, it introduces a Multi-Satellite Planning algorithm based on Proximal Policy Optimization (PPO) deep reinforcement learning, which dynamically schedules off-nadir angles and imaging durations to achieve optimal observation configurations. Moreover, experiments conducted with historical satellite imagery demonstrate that the proposed method significantly outperforms comparable algorithms in terms of coordination efficiency and timeliness of disaster information acquisition. The results validate the feasibility of this theoretical approach for multi-satellite emergency sensing. Importantly, the method offers considerable value for enhancing satellite-based disaster response, command and dispatch efficiency, and life-saving operations.