Abstract: Aiming at the autonomous scheduling problem of spacecraft remote sensing tasks in the field of deep space exploration, a sequence planning method based on deep reinforcement learning（DRL）is proposed. To improve the accuracy of physical modeling, a bucketized lookup table is used to compute maneuvering time, restoring the real dynamic process during attitude maneuvering. To address state drift caused by sequential decision-making, a time-nearest-neighbor state matching method based on historical decisions is proposed, which reversely retrieves the nearest executed node on the timeline as a physical baseline to recalibrate the current state of the spacecraft. A 1D-CNN and Pointer Network are used to build an end-to-end policy network, with tensor-based dynamic masking embedded at the decoder side to satisfy constraints on multi-frequency observation, time windows, energy, and storage. Simulation results demonstrate that the scheduling sequences produced by the algorithm satisfy all physical constraints and show advantages in computational efficiency and scheduling revenue, providing algorithmic support for autonomous mission planning of spacecraft.