The near-field steering vector contains the angle and range parameters. They are coupled with each other and difficult to separate. Most existing methods adopt the approximate decoupling model to estimate the angle and range parameters step by step. However， such an approximate decoupling model will inevitably introduce a systematic model error， which could lead to severe localization performance degradation. To address the above challenges，this paper proposes a near-field sources localization method for sparse representation via a non-uniform grid. It directly models the complex near-field sources localization as a lower-dimensional sparse signal recovery problem and adopts sparse Bayesian learning （SBL） to adaptively recover the sparse signal， avoiding the approximate error and improving the parameters estimation accuracy.In the proposed method， the non-uniform grid only contains a few points， reducing the computational complexity greatly. The nearby points neither share the same direction of arrival （DOA） nor the range value， effectively overcoming the high correlation basis. And the grid refinement trick is additionally introduced to further solve the mismatch problem caused by the coarse grid. The numerical simulation results confirm the superiority of the proposed method.