一种联合特征值信息的全极化SAR图像监督分类方法

摘要

基于H/a平面的分类器对于具有相似散射类型的地物的分类能力很差，为此该文直接使用特征值特征来进行分类。首先提取特征值特征，并使用一种自适应调整高斯分量个数的高斯混合模型对特征值分布进行较为准确地拟合，然后采用朴素贝叶斯分类器进行初步分类。针对可能存在特征值分布较为相近导致错分的问题，计算每两类地物的特征值分布的相似度，将相似度大于给定阈值的类别对组成相似性表，对于这些相似对再用基于Wishart距离的K近邻分类器进行细分。综合分析机载和星载SAR数据上的实验结果，表明这种方法能够克服基于H/a的非监督分类方法对于特征值利用的一些不足，且与基于SVM的分类方法效果相当。%Since classification methods based onH/a space have the drawback of yielding poor classification results for terrains with similar scattering features, in this study, we propose a polarimetric Synthetic Aperture Radar (SAR) image classification method based on eigenvalues. First, we extract eigenvalues and fit their distribution with an adaptive Gaussian mixture model. Then, using the naive Bayesian classifier, we obtain preliminary classification results. The distribution of eigenvalues in two kinds of terrains may be similar, leading to incorrect classification in the preliminary step. So, we calculate the similarity of every terrain pair, and add them to the similarity table if their similarity is greater than a given threshold. We then apply the Wishart distance-based KNN classifier to these similar pairs to obtain further classification results. We used the proposed method on both airborne and spaceborne SAR datasets, and the results show that our method can overcome the shortcoming of theH/a-based unsupervised classification method for eigenvalues usage, and produces comparable results with the Support Vector Machine (SVM)-based classification method.