To take full use of context information to learn the spatial relationship of the image, a novel model based on Generative Model and Region Connection Calculus (RCC) is proposed in this paper. We name it GM-RCC model, which is used for SAR image classification. Firstly, a SAR image is over-segmented into multi-scale super pixels via adopting the image pyramid. Then the hierarchical RCC model is utilized to describe the spatial relationships among these super pixels. All hierarchical RCC relationships are learned and reasoned under the Generative Model reasoning framework. The experiments are carried out on SAR image datasets and polarimetric features are selected with other typical features. The results reveal the efficient performances and superiorities of the proposed algorithm.