This paper presents a new method for unsupervised classification of terrain types using polarimetric synthetic aperture radar data. This unsupervised classification combines the target decomposition theory and the support vector machines. The initial cluster centers are firstly determined by target decomposition advanced by Cloude and Pottier. Then the pixels near to the cluster centers are selected to train the support vector machines using Wishart distribution. The classified results are then used to define training sets for the next iteration if necessary. Finally

by the optimal separating hyperplanes and the kernel method this method obtains extraordinary classification results and neednot much iteration. And the effects of feature vectors consisted of several polarimetric parameters are discussed in detail.