Learning-based constellation scheduling for time-sensitive space multi-target collaborative observation

The space time-sensitive multi-target collaborative observation scheduling problem extends the traditional Earth Observation Satellite (EOS) scheduling by incorporating the time dimension. Multiple effective and collaborative observations over time are required to collect relevant information on time-sensitive space targets in high-speed motion. Furthermore, changes in target spatial relationships cause conflicts in resource allocation. Due to the urgency of the time-sensitive targets, the observation system must respond promptly and efficiently upon receiving a task. This study proposes a mathematical model and a Time-Sensitive Observation Scheduling (TSOS) algorithm to address these problem characteristics. Firstly, a mathematical model for scheduling is established by introducing the completion quality index and resource switching frequency index, given the requirement for collaborative multiple satellites in this problem. The TSOS algorithm consists of two parts: task pre-processing and a circular scheduling strategy. In the task pre-processing module, the entire scheduling is divided into multiple observation decision intervals. This enables the candidate satellite set corresponding to each target to remain stable, which reduces the complexity of the problem. An action mapping method is used in the circular scheduling module, and a decision network based on Deep Reinforcement Learning (DRL) called the Observation Pair Decision Network is proposed. Experimental results demonstrate that the trained TSOS algorithm can effectively schedule resources in scenarios with different conflict degrees and exhibits strong robustness.