Design and evaluation of strategies to minimize operating costs and support in-orbit Distributed Learning within satellite networks for Earth Observation

Earth Observation (EO) satellites currently acquire images to be then stored in their on-board memory, until they fly over a ground station, when they shall transmit a high amount of data in a short time. Thus, a high data rate is requested, but this is constrained by the power available on-board, in turn limited by the dimensions and masses of solar panels and batteries. However, not all data transmitted to ground are actually useful to the application. A solution can be obtained by endowing satellites with on-board processing capacity and Inter-Satellite Links (ISLs) to make them able to offload data processing to other satellites whenever they have not enough resources to accomplish the processing task. This would allow for an on-board extraction of the useful information from acquired images, leading to an increased efficiency in bandwidth usage and to a reduction of both the time needed to deliver information to the ground station and of the energy to be used by ground stations to process information. However, transmission, storage, and computational capacity available for in-orbit processing are valuable resources and could be not always available. For this reason, it is necessary to design strategies to appropriately allocate bandwidth and processing resources on satellites and to leverage the possibilities opened by the network of satellites made possible by ISLs, while optimizing a desired metric. In this thesis, I propose strategies to minimize operating costs of EO satellite networks, to save energy due to image processing on ground stations, and to support in-orbit training of machine learning models in a distributed manner to allow for faster accuracy convergence while reducing both the bandwidth and on-ground energy consumption with respect to centralized learning solutions. In particular, I first introduce and solve an optimization problem to allocate transmission, memory and processing resources to minimize the total operating cost to be paid for transferring, elaborating and storing EO data. Furthermore, since the proposed optimal strategy is NP-hard, I also define and evaluate two heuristics to be applied in real orbital scenarios, proving that they outperform benchmark solutions. Second, I introduce two optimal strategies to maximize energy saving on ground stations by leveraging at most in-orbit EO image processing. In particular, in the first strategy I do not take into account any constraint on the level of usage of on-board batteries to optimize operative life. Since this strategy results to be NP-complete, I also introduce a heuristic to be applied in real orbital scenarios. Instead, the second optimal strategy aims to maximize ground station energy saving while also optimizing satellite operative life by assuring that batteries are not discharged under a certain threshold. Results obtained with all the proposed strategies also provide useful insights on how the on-board CPU clock frequency has to be chosen to obtain optimal results, given limitations on energy available on satellites. Finally, I propose a communication strategy to support in-orbit distributed training of deep learning models by leveraging the satellite network made of both intra-orbital and inter-orbital ISLs. In particular, the proposed distributed learning strategy provides for satellites exchanging locally trained models within themselves, without having to lean on a central parameter server as it happens in federated learning schemes available in literature. Obtained results show that such strategy allows for reaching model convergence in a shorter time if compared to federated learning-based schemes.