With the development of space technology, the use of satellites for Earth observation is becoming more and more common. Among them, low-earth orbit (LEO) satellites have received a lot of attention because of their short return period and low cost. However, traditional satellite systems are unable to process data in orbit due to technical limitations and can only transmit massive amounts of data with useless information back to ground stations and then perform calculations, putting a tremendous strain on downstream bandwidth. Additionally, there are data walls between satellite affiliates, which makes it difficult to share and use data. To address the lack of data processing capabilities and data barriers in conventional satellite systems, as well as to effectively use satellite resources and enhance worldwide collaboration, we proposed a connection density-aware satellite-ground federated learning (FL) framework. According to this structure, satellites serve as workers in federated learning, analyzing data and transmitting the parameters to ground stations for aggregation. The ground station senses the satellite’s connection density during the aggregation process and implements a dynamic aggregation approach for the downstream parameters. Experiments based on real image datasets and on a simulated satellite network validate the effectiveness of the method in terms of execution efficiency and accuracy.