An effective unsupervised multiscale segmentation of Synthetic Aperture Radar(SAR) imagery based on Expectation Maximization (EM) and Genetic Algorithm(GA) was proposed. The statistical variations between pixels of scale-to-scale and same scale in SAR imagery were described due to radar speckle for the Mixture Multiscale AutoRegressive(MMAR) model, then the estimation of parameters in MMAR model was given by combining GA with EM algorithm. The number of components of the model was selected by using the Minimum Description Length (MDL) criterion and the segmentation of SAR imagery was implemented. This approach benefits from the properties of GA and the EM algorithm by combining of both into a single procedure. The local optimal solutions were avoided with less sensitivity to its initialization. The experiments on SAR images show that the GA-EM outperforms the EM method.