Large-Scale Complex Correlated Earth Observation Task Planning Based on Graph Neural Networks and Reinforcement Learning

In light of increasingly complex demands for Earth observation tasks, this paper proposes a satellite resource planning approach that integrates Graph Neural Networks (GNN) and Reinforcement Learning (RL). By mapping multiple observation modes (e.g., multi-frequency observations, collaborative guidance, on-board autonomous tasks, and multi-strip mosaicking) into a graph structure, the GNN can perform unified modeling and conflict resolution concerning spatial coverage, resource conflicts, and temporal dependencies among tasks and strips. On this basis, we introduce a reinforcement learning framework based on the Markov Decision Process, incorporating multiple constraints (e.g., time, payload, data transmission, and onboard storage) into the reward function design. This approach enables adaptive learning of near-optimal resource allocation strategies in dynamic environments. Experimental results show that, compared with conventional heuristic algorithms and integer programming methods, the proposed approach exhibits significant advantages in task completion rate, resource utilization, and adaptability to on-the-fly task modifications. Our research provides a novel technical avenue for large-scale, complexly correlated satellite Earth observation mission planning and holds importance for further enhancing the scheduling efficiency and intelligence level of satellite constellations under demands for high temporal and spatial resolution, real-time capabilities, and diversity.