A SAR image can be modeled as the multiplication of the noise-free image and speckles. So the noise-free image can be estimated from the observed image with the Bayesian technique. It’s crucial to choose a proper prior model for well matching the SAR images’ characteristics. In this article the Membrane MRF model is employed to model the prior information

which overcomes GMRF’s problem of sensitivity to parameters. And

pixels in homogeneous and non-homogeneous regions are processed separately by adjusting the model’s neighborhood adaptively. Experiments show that not SAR images can be despeckled efficiently while their structures are preserved well.