Agile Earth Observation Satellite Scheduling: A Graph-Based Learning Approach

With the rapid expansion of Internet of Things (IoT) applications in remote and uncovered areas, Agile Earth Observation Satellites (AEOS) serve as nodes in space-borne sensor networks, playing an increasingly critical role in acquiring remote sensing data from ground targets. However, the complexity of AEOS scheduling poses significant challenges. This paper proposes a graph-based learning approach for AEOS scheduling to efficiently enable IoT data acquisition. First, we build a window conflict graph to transform the scheduling of feasible Observation Windows (OWs) into a conflict-free vertex set construction problem. On this basis, we propose Interactive Conflict-Free Adaptive Learning (InterCFAL), coupling a conflict-free vertex set construction module with an adaptive learning module driven by Reinforcement Learning (RL). The construction module iteratively explores feasible solutions on the graph, while the adaptive learning module encodes graph structure into the RL state via graph embedding and trains a policy network to guide vertex selection. To reduce computation and enable effective training on large-scale conflict graphs, the sampling operation is incorporated to reduce the problem scale, and the dynamic weighted aggregation reward adaptively balances conflicting objectives. Finally, simulation results demonstrate significant advantages in observation profit, energy consumption, load balancing, convergence, and generalizability, confirming the scalability of InterCFAL to large-scale AEOS scheduling scenarios.