Faced with the ever-increasing number of satellites and diversified mission requirements, the difficulty of task scheduling for on-orbit imaging satellite constellations is constantly escalating. Most existing methods are mainly aimed at small-scale satellites or single-type tasks, and are thus difficult to meet the collaborative scheduling demands of multi-type tasks under large-scale satellite resources. Therefore, the paper presents a scheduling model and algorithm for imaging satellite constellations that takes into comprehensive account multiple types of tasks. Regarding model construction, a generalized observation benefit function is designed, which is capable of uniformly evaluating the benefits of different task types, including common point targets, periodic point targets, and regional targets. Simultaneously, the model also takes into consideration the limitations of satellite resources and the specific requirements of tasks to ensure that the generated scheduling scheme is reasonable and efficient. In terms of algorithm design, an adaptive large neighborhood search algorithm based on deep reinforcement learning (DRL-ALNS) is proposed. This algorithm makes use of the autonomous learning ability of DRL to intelligently select operators and configure parameters within the framework of the adaptive large neighborhood search, thereby effectively coping with the challenges of large-scale search spaces. To verify the effectiveness of the proposed method, experiments are conducted by comparing it with multiple comparison algorithms. The results show that the DRL-ALNS achieves an average increase of 7.6% in task benefit value, demonstrating its effectiveness and superiority in addressing the scheduling problem of multi-type task imaging satellite constellations.