Abstract: Aiming at the problems of time frequency aggregation and cross term interference in traditional time-frequency analysis methods. The received frequency hopping signals are segmented to constructs a time-frequency sparse model. Then the block sparse Bayesian learning algorithm is used to reconstruct the time-frequency representation of the frequency hopping signals by using the statistical characteristics and structural characteristics of the model. High precision time-frequency representation can be obtained without prior knowledge of sparseness and noise intensity. However, this algorithm has high computational complexity due to the parameter estimation in the high-dimensional parameter space. The approximate substitution method is used in this paper to improve the performance, and the high-dimensional parameter space is converted to the original parameter space. Simulation results show that the improved algorithm can obtain the high precision time-frequency representation of the frequency-hopping signals effectively and reduce the complexity greatly.