The spectral occupancy model of satellite spectrum sensing data in the temporal dimension is analyzed. Because the bottom noises in satellite receiving data are undulated in different time and space, the traditional modeling methods with fixed threshold are defective. Therefore, the adaptive threshold method is introduced to determine the noise threshold and preprocess the satellite spectrum data to obtain the satellite spectrum occupancy length sequence. In order to make an effective statistical analysis on the situation of the satellite spectrum, the probability density curve of the spectral occupation time length series is fitted by using the Poisson and exponential distribution methods, and a probability distribution model suitable for the satellite spectrum occupation time series is obtained. Based on the obtained satellite spectral occupancy state model, the state transfer matrix at a certain frequency point of the satellite channel is calculated by two-state Markov chains to predict the probability of outgoing channel occupancy and idle. In addition, the Back-Propagation(BP) neural network is trained through the data set constructed by satellite spectrum sensing data to predict the occupancy length of a certain frequency point. By calculating the Root Mean Square Error(RMSE) of the BP neural network and the conventional Long Short-Term Memory(LSTM) neural network prediction methods, 0.172 8 and 2.208 1 are obtained respectively. The evaluation results show that the BP neural network bears the advantage of high accuracy.