Uplink channel estimation is a classical and important issue for massive multiple-input multiple-output (MIMO) communication systems. Due to the existence of the pilot coupling effect, it is difficult to directly utilize the sparse characteristics of massive MIMO channels. The existing methods usually adopt a prior least squares (LS) estimator to decouple the pilot matrix, which could introduce significant modeling errors. Another solution is to utilize the expectation propagation approximation (EPA) instead to decouple the pilot matrix, but it always suffers from the performance loss caused the approximate message passing. In this regard, we try to propose a new sparse Bayesian learning (SBL) uplink channel estimation method, which decouples the sparse signals-of-interest automatically using the independent variational Bayesian inference (VB I) factorization, thereby avoiding the performance loss brought by the prior LS estimator and approximations. Numerical simulation results verify the superiority of our method.