Existing on-board planning systems do not apply to small satellites with limited onboard computer capacity and on-board resources. This study aims to investigate the problem of autonomous task planning for small satellites. Based on the analysis of the problem and its constraints, a model of task autonomous planning was implemented. According to the long-cycle task planning requirements, a framework of rolling planning was proposed, including a rolling window and planning unit in the solution, and we proposed an improved genetic algorithm (IGA) for rolling planning. This algorithm categorized each individual based on the compliance of individuals with a time partial order constraint and resource constraint, and designed an appropriate crossover operator and mutation operator for each type of individual. The experimental result showed that the framework and algorithm can not only respond quickly to observation tasks, but can produce effective planning programs to ensure the successful completion of observation tasks.