A segmentation algorithm based on wavelet domain hidden Markov tree model was improved. The pixel level segmentation result can not be obtained because of choosing wavelet coefficients as training feature directly in traditional methods. At the same time, traditional methods ignore the feature of labeling maps at different scales by using one single context to all scales in fusion phase. Hence, this study dealt with the initial parameters set problem and chose better feature for training. In this way, the fine pixel level segmentation can be acquired directly in the raw segmentation step, and in multiscale fusion phase, the characteristics of labeling maps at different scales are used sufficiently. Among them, both the information from coarse-scale segmentation and the one from fine-scale segmentation were considered. Experiments show that the visual effects of our algorithm are the best compared with the HMTseg method proposed by Choi and the WD-HMTseg algorithm of remote sensing image segmentation presented by Sun Q.