The acquisition of channel state information (CSI) has a significant impact on the performance of orthogonal time frequency space (OTFS) modulated communication systems. By exploiting the sparsity in the delay-Doppler domain, the channel estimation problem of OTFS can be transformed into a two-dimensional sparse representation problem. However, the computational complexity of solving the two-dimensional sparse signal is extremely high, and existing methods generally cannot effectively against errors caused by fractional Doppler/delay. To address these issues, we propose a non-uniform grid-based sparse representation method to model the OTFS channel estimation as a lower-dimensional sparse signal recovery problem. Sparse Bayesian learning is then used to infer the optimal solution in the sense of maximum a posteriori, achieving two-dimensional sparse signal recovery with low computational complexity. To counteract the fractional Doppler/delay errors, we treat the coordinates of each grid point as adjustable variables and iterate them adaptively. Simulation results illustrate that the proposed method can effectively improve the performance of OTFS channel estimation.