超像素在多极化SAR数据分类中的应用--以ALOS PALSAR为例

针对已提出的极化合成孔径雷达数据地物分类方法较难同时获得地物边界及相邻信息的问题,并为了减少图像处理的消耗时间,本文引入一种超像素生成算法--线性迭代聚类方法,对日本先进对地观测卫星多极化SAR数据进行地物分类研究。本文以四川省彭州市与什邡市交界地区为研究区,先采用Pauli分解生成RGB假彩色图像并进行滤波,再以此为基础使用线性迭代聚类方法生成超像素,最后用支持向量机分类方法,合理选取极化熵、各向异性度及平均散射角等极化特征组合在一起作为分类参数,对基于像素超像素的极化SAR图像的分类结果进行对比分析。使用超像素比其他基于像素的分类方法能够获得更好的结果,基于超像素分类的总体精度为95.23%,Kappa系数为92.58%。

Application of superpixels in multipolar SAR data classification: taking ALOS PALSAR as an example

In view of the proposed polarization synthetic aperture radar data classification method is difficult to obtain both the boundary and adjacent information of the ground, and in order to reduce the consumption time of image processing, a super-pixel generation algorithm-inear iterative clustering method is introduced, and the geoclassification of the advanced earth observation satellite SAR multipolar data in Japan is studied. Based on the border area of Pengzhou and Shifang city in Sichuan Province, the paper uses Pauli decomposition to generate RGB false color images and filter them, and then uses linear iterative clustering method to generate superpixels on this basis, and finally uses the support vector machine classification method to select polarization entropy reasonably, The polarization features such as anisotropy and average scattering angle are combined as classification parameters to compare and analyze the classification results of pixel-based and hyper-pixel-based polarization SAR images. Experiments show that the use of superpixels is better than other pixel-based classification methods, the overall accuracy of superpixel classification is 95.23% and the Kappa coefficient is 92.58%.