针对谱聚类方法应用于合成孔径雷达（SAR）图像分割时Laplace矩阵的特征值和特征向量难以计算的问题，结合SAR图像在多个尺度的统计信息，给出了一个包含顶点凝聚、初始分割和分割细化3个步骤的SAR图像多尺度分割方法。首先，用一个顶点数不断减少的凝聚图序列来逼近从SAR图像得到的图；然后应用谱聚类方法对最粗尺度的凝聚图进行分割得到初始分割结果；最后根据SAR图像的统计性质，利用基于混合模型估计的分类后验概率将初始分割结果逐尺度进行细化得到SAR图像的最终分割。实验结果表明了方法的有效性。

To solve the computational demands of spectral clustering approach when applied to SAR image segmentation，a multiscale method is proposed.It consists of three steps，i.e.coarsening，initial segmentation and refining.First，a sequence of smaller graphs，each with fewer vertices，is constructed from the SAR image.Second，spectral clustering is applied on the smallest graph to obtain initial segmentation.Third，the initial segmentation is refined scale by scale to get the final segmentation of the SAR image based on the posterior probability of classification which is estimated by the mixture model.Finally，experimental results demonstrate the effectiveness of the proposed method.