This paper introduces Bayesian supervised and unsupervised segmentation
algorithms aimed at oceanic segmentation of SAR images. The data term,
\emph{i.e}., the density of the observed backscattered signal given the region,
is modeled by a finite mixture of Gamma densities with a given predefined
number of components. To estimate the parameters of the class conditional
densities, a new expectation maximization algorithm was developed. The prior is
a multi-level logistic Markov random field enforcing local continuity in a
statistical sense. The smoothness parameter controlling the degree of
homogeneity imposed on the scene is automatically estimated, by computing the
evidence with loopy belief propagation; the classical coding and least squares
fit methods are also considered. The maximum a posteriori segmentation is
computed efficiently by means of recent graph-cut techniques, namely the
$\alpha$-Expansion algorithm that extends the methodology to an optional number
of classes. The effectiveness of the proposed approaches is illustrated with
simulated images and real ERS and Envisat scenes containing oil spills.