The tree-structured Markov Random Field (TS-MRF) model defined on a single spatial resolution, which is capable of expressing the hierarchical structure implied in the image to be segmented, fails to describe its non-stationary property. In order to solve this problem, a new image modeling method in Wavelet domain—WTS-MRF was proposed. In this model, a sequence of MRFs were hierarchically defined in the format of the classification tree structure. Each node was associated with a set of MRFs defined on different resolutions, wherein the correlation between neighbor MRFs with different resolutions was considered in the form of conditional probability. The child MRF was nested in the region of the parent one on the same resolution. Based on the WTS-MRF model, a supervised recursive segmentation algorithm was proposed. The classification hierarchical tree was manually set as the priori information, and the corresponding statistics for each leaf node were obtained by the training data on each resolution. The implementation of this algorithm was both on the inner-scale and inter-scale level. The inner-scale recursion was executed on the lowest resolution, where the MRF corresponding to each node was sequentially and recursively estimated by the ICM algorithm from the root to leaves. The inter-scale recursion was implemented on the next finer resolution, in which the estimation of MRFs was sequentially initialized by the direct projection from the next lower resolution and recursively refined by the ICM algorithm. The final segmentation was obtained when the MRFs were estimated on the primary resolution. Two experiments verify the validity of the proposed method in terms of both visual quality and quantitative indicators (e.g. overall accuracy and Kappa coefficient).