This paper proposes a novel,robust and multiscale filtering algorithm,which robustly filters speckle from Synthetic aperture radar(SAR) image using robust multiscale scale-recursive estimation algorithm of multiscale autoregressive(MAR) model on dyadic tree.Firstly,a multiscale sequence of SAR images is constructed based on the statistic character of SAR image at successive scales.Then,a suitable MAR model,which provides a powerful framework for describing random process and fields that evolve in scale,is selected to model SAR image.And,filtering for the image is given based on the robust scale-recursive M estimation algorithm for MAR graphical model.The proposed algoirhtm is not only computationally efficient,but also capable of obtaining better filtering results due to the merging of SAR image at different resolution.Experiments over SAR images confirm the robustness of the filtering algorithm we proposed.