Abstract: To address the issue that the dictionary matrix is too large in sparse Bayesian learning channel estimation for estimating two-dimensional delay-Doppler parameters, an improved method is proposed in this paper. Firstly, the ambiguity function of the received signal is obtained by using matching filter. Secondly, a threshold is set to retain larger value channel responses, corresponding index values and atoms. Then, a new dictionary matrix and a hyper-parameter matrix are constructed at same scale using these retained atoms. At last, the retained atoms are substituted into sparse Bayesian learning algorithm one by one to iterate and converge the hyper-parameters, and the channel impulse response is finally obtained. This method enhances parameter estimation resolution without requiring prior knowledge of channel sparsity. Compared to traditional sparse Bayesian learning, it significantly reduces calculation quantity and enables estimating two-dimensional time-delay Doppler parameters. Simulation results demonstrate that the mean squared error(MSE) estimated with this improved method is at least one order of magnitude lower than that with original sparse Bayesian learning algorithm, while the run time of new method is reduced significantly.