Satellite edge computing (SatEC) is an emerging computing paradigm for remote and disaster-affected areas with scarce computing resources. It is crucial to make the appropriate offloading strategy to maximize quality of service (QoS) due to the scarcity of satellite computing resources. However, existing researches mostly focus on the offloading of independent tasks without considering tasks with dependencies. To address this problem, we first design a satellite-aerial task offloading (SATO) framework. In the SATO, unmanned aerial vehicles (UAVs) collect ground tasks and upload them to high-altitude platform (HAP), while HAP provides local computing, and low earth orbit (LEO) satellites provide edge computing. Furthermore, we propose SatEC directed acyclic graph (DAG) task offloading (SECDTO) algorithm to investigate the task offloading problem. Specifically, we model tasks with dependencies as a DAG, then we propose an encoder-decoder model with attention based on deep reinforcement learning (DRL) to translate graph vector information into offloading strategies. In order to maximize QoS, we use proximal policy optimization (PPO) to train the model. Simulation results show that our method converges quickly and stably during the training process and outperforms several baseline algorithms in terms of QoS@.