Deep Reinforcement Learning-Based Attention Decision Network for Agile Earth Observation Satellite Scheduling

Agile Earth observation satellite scheduling is crucial for space-based remote-sensing services. The sharply rising demands and explosion of the solution space pose significant challenges to the optimization of observation task scheduling. To address this issue, we propose a deep reinforcement learning-based attention decision network (ADN) to determine the task scheduling sequence. We also construct a Markov decision process model in which the original and direct attributes are defined to describe the environment and used as the input of the ADN. Moreover, a start-time-shift-based local search is proposed to improve the observation plan generated by the ADN model. A comprehensive experiment was conducted, and the results proved that the attention mechanism in our ADN was beneficial for the training process to converge to better strategies. Compared with other advanced algorithms, the proposed method obtained a better total profit in the test sets. Furthermore, our methods exhibit considerable time efficiency, even for large-scale problems.