With LEO mega satellites constellation coming into operation, the available spectrum resources are more overcrowded. To improve spectrum utilization, cognitive satellite communication technology composed of GEO relay satellites and LEO satellites has become an important choice. The most critical step in the cognitive satellite communication scenario is the spectrum sensing technology used to quickly determine the presence or absence of the primary user.
Since most current spectrum sensing algorithms are model-driven, they rely heavily on the assumed statistical model for their detection performance, which makes it more difficult to model and deploy in variable satellite communication scenarios. In this paper, we firstly analyze the signal-to-noise ratio fluctuations during LEO satellite transit, and secondly propose a spectrum sensing algorithm combining bidirectional long short-term memory network and Bayesian likelihood ratio test for this variable channel environment. The algorithm does not require any a priori knowledge of PU signals and can automatically learn features from PU signal data and make decisions. Simulation results show that the proposed algorithm still achieves 83% detection performance at a signal-to-noise ratio of -14 dB and consistently outperforms convolutional neural networks, multi-layer perceptron, and model-driven energy detection-based algorithms.