Planning the activities of an Earth observation satellite is a highly combinatorial task. It consists in regularly computing the sequence of observations to be performed by a satellite to collect images of candidate points of interest (POIs), while taking into account the time-dependent maneuvers required to point the satellite to the successive POIs. To solve such a recurrent optimization problem, we propose a novel approach that exploits offline learning techniques to approximate scheduling feasibility for sets of observation tasks. For this, we build a 0/1 neural network classifier whose inputs are related to the geographical positions of the POIs to be observed over the satellite orbit. We also learn hard capacity constraints limiting the number of observable POIs within orbit portions of various sizes. Finally, we introduce a hybrid algorithm, called HySSEO, optimizing the observation schedules based on a two-step process. The latter first searches for an optimal selection of POIs given the learned constraints, and then exploits this selection to bootstrap the search for an optimal schedule satisfying detailed time-dependent transition constraints. This hybrid optimization approach significantly improves the solution quality when compared to a standard scheduling approach.