Abstract: Second generation bandelet can make full use of intrinsic geometric regularity and provide optimal image representation adaptively. In this paper, we modeled the second generation bandelet coefficients of an image using hidden Markov tree (HMT) model and obtained the image segmentation results using multiscale parameter training and context-based maximum posterior probability. In order to evaluate the performance of the proposed method, we made experiments on synthetic mosaic images, aerial images, and SAR images. The segmentation results were compared with the wavelet-domain HMTseg (WD-HMTseg) one and contourlet-domain HMTseg (CHMTseg) one. Experiment results showed that our method not only obtained more exact boundary and uniform regions, but also reduced the misclassification rate of texture images.