针对常规极化SAR影像监督分类在分类时由于极化信息利用不完全, 导致分类结果不可信的缺陷, 在分析地物极化散射特性的基础上, 提出一种结合复Wishart分类器思想的极化SAR监督分类方法, 在监督分类的基础上, 利用几乎包含所有极化信息的相干矩阵进行后续分类, 利用地物所有极化散射信息完成类别的划分。 首先, 利用Cloude&Pottier极化特征组合对影像进行监督分类, 获得初始类别划分结果； 为纠正监督分类时仅利用特征矢量的空间分布特性进行类别划分导致的错误, 对监督分类结果精度评价及研究区地物散射相似性分析, 综合确定初始分类结果中分类效果较差、 待后续调整类别的像素集； 然后以各类别的相干矩阵均值为初始聚类中心, 利用核模糊C均值算法, 结合相干矩阵的复Wishart分布特性, 对待修正的像素集进行后续类别迭代调整, 获得精细分类结果。 采用国产X-SAR获取的海南陵水地区的全极化SAR影像进行分类试验, 结果表明: 提出的对极化SAR影像监督分类后利用地物散射特性相似性对错分像素重新进行聚类调整类别的方法, 与仅进行监督分类相比, 分类结果总体精度更高, 也更满足地物的散射特性。

The present paper, on the basis of analyzing the terrain scattering characteristics, proposed a novel supervised classification method combined with complex Wishart classifier ideology. This method used coherent matrix which almost contains all the polarization information to make subsequent adjustments for the supervised classification result to achieve higher accuracy division categories. For the first beginning, supervised classification was carried out on the Cloude & Pottier polarimetric characteristics combination to get the initial classification result．Then, in order to achieve the purpose to correct the mistakes resulting from just using the spatial distribution of feature vectors in supervised classification, we did some analysis as follow. The accuracy analysis of the classification results and the analysis of study area feature scattering similarity play an important role in our study to help us make the determination that the pixels need to be adjusted. Furthermore, taking the mean value of each category coherence matrix as the initial cluster centers of subsequent iterations, and using Kernel Fuzzy C-Means algorithm to adjust the fixed pixel set categories by subsequent category iterative correction, the fine and high-accuracy classification results were obtained, combined with complex Wishart distribution of coherence matrix. The domestic X-band full polarization SAR data of Lingshui area in Hainan province was applied in this classification experiment. The experiment results demonstrate that the proposed method can obtain a favorable classification accuracy polarization SAR image classification results, and better meet the scattering characteristics of the surface objects compared to the original method.