Graph Learning for Multi-Satellite based Spectrum Sensing

Recently, Low Earth Orbit (LEO) satellite Internet has been deployed and provides access service. In the near future, with high-speed development and dense deployment of non-terrestrial and terrestrial infrastructures, limited spectrum resources will not be allocated enough. Consequently, dynamic spectrum access enables the coexistence of non-terrestrial and terrestrial networks in the same spectrum, and hence, efficient spectrum sensing technology plays a vital role. Unlike spectrum sensing in terrestrial networks, satellites in space are too far from the Earth, resulting in serious channel fading, and the spectrum sensing performance of any single satellite may be degraded significantly. To provide better spectrum sensing performance, multiple satellite collaboration can offer data diversity. However, that collaboration is a non-trivial task in LEO satellites, due to the heterogeneity of radio frequency channels between satellites and ground station. We leverage the graph model to represent the relationship of multiple satellites with different channel quality, and propose a graph attention neural network to fuse their signals for spectrum sensing. Extensive experiments demonstrate that our multiple satellites collaboration framework efficiently executes spectrum sensing tasks, and outperforms conventional deep learning methods.