Good performance has been obtained by extending traditional image denoising algorithm from local computation model to non-local one with non-local means algorithm. For synthesis aperture radar (SAR) image, however, the similarity measured by observations and isotropic window is not robust and without direction, which is bad for capturing the structure of image. In this paper, Non-subsample Shearlet feature and directional neighborhood based non-local means algorithm are proposed. Experimental results demonstrated that the improved non-local means algorithm can not only remove the speckle, but also preserve the geometrical structure information which is essential for understanding and interpretation of SAR image.