Earth Observation Satellite Scheduling Based on DQN

The satellite mission planning problem for land resource census is highly challenging, characterized by a high-dimensional nonlinear solution space due to continuously adjustable satellite slewing angles and massive time windows, superimposed with strongly coupled resource constraints. This paper constructs a discrete decision-making model based on "observation opportunities," decoupling the original problem into two subproblems: observation sequencing and optimal imaging strip selection. To address the limitations of existing algorithms—specifically rule myopia and search inefficiency—when handling such coupled problems involving sequence optimization and parameter selection problems, this paper proposes a scheduling algorithm that integrates variable neighborhood search (VNS) with deep reinforcement learning. This method establishes a hierarchical scheduling model: at the macro level, it utilizes the multiple neighborhood structures of VNS to optimize the observation sequence and escape local optima; at the micro level, it introduces a deep Q-network (DQN) combined with a multi-dimensional state feature space to achieve adaptive evaluation of strip selection values, thereby replacing manually designed rules. Simulation experiments demonstrate that the proposed method exhibits both excellent convergence speed and solution quality, with the gap between the solution score and the theoretical total score being less than 15% for 99% of the test samples.