Earth Observation Satellite Scheduling for Dynamic Tasks with Deep Reinforcement Learning

Existing deep reinforcement learning (DRL)-based methods for scheduling Earth observation satellites (EOSs), which use DRL to accept or reject dynamic tasks, are limited in their ability to adaptively adjust satellite schedules, leading to suboptimal performance and inadequate handling of dynamically arriving tasks. To overcome this limitation, this study introduces a novel DRL-based method for the EOS scheduling problem in scenarios with dynamic tasks, called dynamic satellite scheduling (DSS). Unlike existing DRL-based methods, the proposed DSS method allows for modifying satellite schedules in response to dynamic tasks, transforming the EOS scheduling problem into a decision-making problem that determines which tasks the satellites will perform, thus enabling better handling of dynamic tasks. The proposed DSS method incorporates the long short-term memory network into the decision-making network to achieve dynamic updates of decision results. Experimental results show that our proposed method significantly enhances scheduling performance compared with existing methods, achieving higher rewards in scenarios with dynamic tasks.