Energy-efficient Federated Learning for Earth Observation in LEO Satellite Systems

Low Earth Orbit (LEO) satellites enable various Internet of Things (IoT) applications by providing plentiful observation data at different spatial scales. However, transmitting such a large volume of observation data to the ground for model training is challenging due to the intermittent connection between satellites and ground stations, leading to long training delays for task-oriented artificial intelligence models. Orbital federated learning has been investigated to enable on-board model training sharing model parameters instead of observation data, thus greatly reducing the communication overhead. However, these works underutilize the availability of inter-satellite links and ignore energy consumption. To fill this gap, we propose an energy-efficient federated learning scheme for earth observation applications in LEO satellite systems to minimize learning loss and energy consumption. Specifically, to reduce the effect of data heterogeneity on learning accuracy, we propose a satellite grouping scheme based on satellites' data distribution and communication delay to ground stations. Then, we optimize each satellite's transmission power and computing frequency under the constraint of training time. Experimental results based on popular datasets show the efficacy of the proposed scheme compared to benchmark methods.