Proposal and Investigation of a Distributed Learning Strategy for training of Neural Networks in Earth Observation Application Scenarios

One of the key enabling solutions to in-orbit extract information from Earth Observation images is given by deep learning techniques. However, the accuracy of these algorithms is strictly related to the availability of large datasets of satellite images for training purposes. Limitations on the available transmission bandwidth in the orbital context may prevent the possibility to downlink all acquired images to a node where centralized training happens. For this reason, we propose a communication strategy to support a completely distributed learning (DL) technique to train a deep learning model in-orbit, by leveraging the fact that satellites may form a network thanks to the potential availability of Inter-Satellite Links (ISLs) within and between orbital planes. Our proposal is different from a FL approach since we provide for each satellite to receive all the information needed to calculate an updated global model by itself, without leaning on a central parameter server. Numerical results show that DL outperforms FL in number of learning rounds completed in the unit time, allowing for reaching validation accuracy convergence in a shorter time.