A Bilevel Evolutionary Algorithm for Large-Scale Multiobjective Task Scheduling in Multiagile Earth Observation Satellite Systems

This article studies a multiagile earth observation satellite system, in which a group of satellites provides observation services to acquire images of targets on the earth's surface. In this system, the large-scale multiobjective task scheduling problem is studied by jointly optimizing the task assignment scheme and the observation window allocation to maximize the total profit of all executed tasks on all satellites and the loading balance among satellites. Note, however, that it is challenging to solve this problem since the task assignment scheme and the observation window allocation are tightly coupled. Therefore, a bilevel optimization problem is formulated, where the tasks are assigned at the upper level and the observation windows are allocated at the lower level. In this way, the observation windows are allocated based on the given task assignment scheme, thus decoupling the task assignment scheme and the observation window allocation. Furthermore, the observation windows can be allocated in parallel on different satellites to improve computational efficiency. Subsequently, a bilevel evolutionary algorithm is proposed. Specifically, at the upper level, an initialization strategy is devised to efficiently generate feasible task assignment schemes by constructing the candidate satellite set for each task, and then a constrained multiobjective evolutionary algorithm is adopted to optimize the task assignment schemes. In addition, at the lower level, for each task assignment scheme, a greedy strategy is proposed to allocate the observation windows to as many tasks as possible on each satellite and a local search method is suggested to further improve the observation window allocation. Experiments on a diverse set of instances involving up to 1000 tasks demonstrate that the proposed algorithm exhibits better or at least competitive performance against other compared algorithms on each instance.