The satellite observation task planning has the characteristics of complex constraints, large solution space, and unfixed length of input task sequence. The deep reinforcement learning (DRL) method was used to solve the problems. The satellite observation task planning problem was modeled by taking into account the constraints of time windows, transfer time between tasks, and satellite power and memory constraints. A sequence decision algorithm model was established based on the operating mechanism of pointer networks (PN), Mask vector was used to consider various constraints in the satellite observation task planning problem, and the model was trained by Actor Critic reinforcement learning algorithm to obtain the maximum reward. The PN was improved by referring to the multi-head attention (MHA) mechanism, and the multi-head attention pointer networks (MHA-PN) algorithm was proposed. Experimental results show that the MHA-PN algorithm significantly improves the training speed and the generalization performance of the model, and the trained MHA-PN algorithm model can carry out end-to-end reasoning on the input sequence, avoiding the iterative solution process of traditional heuristic algorithm, has a high efficiency of solution.