Federated Learning with Dynamic Aggregation Based on Connection Density at Satellites and Ground Stations

Abstract

Earth observation images collected by low earth orbit (LEO) satellites can enhance the capability of machine learning to solve some global problems: e.g., climate change, disease transmission, etc. However, downloading all data from LEO satellites for modeling on the ground is not feasible. This is mainly due to the downlink bandwidth of ground stations and the limited connection time between satellites and ground stations. In addition, the images collected by satellites are often of high resolution, which further exacerbates the difficulty of the data transfer task. To address this problem, we propose an efficient synchronous federated learning (FL) optimization method applied to LEO satellite constellations, where the ground stations and the satellites collaborate to train a global model and the satellites do not need to transmit raw data to the ground stations. In this approach, the ground station can dynamically aggregate the local model of the satellites based on the density of connections. When connections are sparse, ground stations perform aggregation after a predefined time period, ignoring stragglers who fail to return in time for updates. When connections are dense, ground stations can perform aggregation based on a buffering strategy to increase the frequency of global model updates. Extensive experiments on both satellite-based simulated networks and real-world image datasets demonstrate the effectiveness of our approach. Compared to the state-of-the-art FL optimization method, the proposed method accelerates the speed of convergence by 4.0 times in the IID setting, and 1.9 times in the Non-IID setting.