With the increasing scarcity of spectrum resources in satellite and terrestrial communications, spectrum sharing becomes a promising option. In this paper, we investigate the spectrum access and power control problem in multibeam-based cognitive satellite communication network. Differ from the most existing spectrum access problems in terrestrial networks, we consider not only the interference between cognitive users, but also the co-channel interference from multi-beam satellite communication system to cognitive users. We formulate a spectrum access and power control game, and it is proved to be an ordinal potential game. The sufficient conditions for cognitive users not to interfere with each other are given, and the upper bound of the aggregation interference experienced by all cognitive users is deduced theoretically. Then, based on the trial and error (TE) algorithm, we propose a learning-based distributed spectrum access algorithm, which statistically converges to the best Nash equilibrium(NE). Furthermore, to simplify the coefficients design and improve the convergence performance, we propose an improved learning-based distributed spectrum access algorithm. Simulation results show that the average network throughput of the proposed game is close to the best, which validates the game-theoretic solution. Simulation results also show that the improved learning-based distributed spectrum access algorithm is superior to the original algorithm in terms of convergence speed, throughput performance, and practicability, which confirms the effectiveness of the improved algorithm.