Since the segmentation results of SAR images by traditional transform domain hidden Markov tree (HMT) model were unsatisfactory in homogenous regions and exact edges, a new segmentation method based on second generation bandelet-domain HMT-3S model was proposed. The method was called BHMT-3Sseg shortly. HMT-3S is a special kind of HMT which combines the correlation of different subbands. It is more reasonable to characterize texture regions than HMT model. BHMT-3Sseg modeled the second generation Bandelet coefficients of an image by using HMT-3S, and the SAR image segmentation results were obtained by training the parameters of HMT-3S and computing the likelihood of each scale and multiscale fusion based on a contextual model. The segmentation results by BHMT-3Sseg not only have more exact and more continuous edges, but also retain better region information. The experiments show that BHMT-3Sseg is efficient and effective for SAR image segmentation.