In this work we detail a tree-structured MRF (TS-MRF) prior model useful for segmentation of multispectral satellite images. This model allows a hierarchical representation of a 2-D field by the use of a sequence of binary MRFs, each corresponding to a node in the tree. In order to get good performances, one can fit the intrinsic structure of the data to the TS-MRF model, thereby defining a multi-parameter, flexible, MRF. Although a global MRF model is defined on the whole tree, optimization as well estimation can be carried out by working on a single node at a time, from the root down to the leaves, with a significant reduction in complexity. Indeed the overall algorithm is proved experimentally to be much faster than a comparable algorithm based on a conventional Ising MRF model, especially when the number of bands becomes very large. Thanks to the sequential optimization procedure, this model also addresses the cluster validation problem of unsupervised segmentation, through the use of a stopping condition local to each node. Experiments on a SPOT image of the Lannion Bay, a ground-truth of which is available, prove the superior performance of the algorithm w.r.t. some other MRF based algorithms for supervised segmentation, as well as w.r.t. some variational methods.