Abstract: Constructing a measuring matrix with training sequence and then using sparse recovery algorithms is a usual approach to multipath sparse channel estimation. In this paper, an improved Bayesian matching pursuit is proposed and applied to underwater multipath sparse channel estimation. We illustrate the method theoretically and test it on two models of underwater multipath sparse channel. Performance of this algorithm is shown in comparison with conven-tional estimating methods. Numerical simulations demonstrate that estimated result of this method converges faster than that of BMP, thus it estimates multipath sparse channel more efficiently. What's more, the proposed method provides better performance than conventional ones in low-SNR conditions and in the channels with many close paths.