In order to guarantee the long-term stable communication of underwater devices, the IFM-SBL(sparse Bayesian learning based on improved fast marginal likelihood maximization) algorithm was proposed to estimate underwater acoustic channels with low complexity and high performance. Especially in the case of low SNR(signal-to-noise ratio), the performance of proposed algorithm can be further improved by threshold denoising and discrete Fourier transform denoising. Simulation and sea trial results show the output bite error rate after channel estimation of IFM-SBL is similar to that of EM-SBL(sparse Bayesian learning based on expectation maximization), and it has good robustness in both low SNR and fast or slow time-varying channels. The running speed of FM-SBL and IFM-SBL algorithm is 90% better than that of EM-SBL algorithm, which greatly reduces the estimation time.