全极化SAR地物分类与极化方位角补偿

发布时间：2018-11-27 09:23

【摘要】：全极化合成孔径雷达（Synthetic Aperture Radar，，SAR）作为一种先进的遥感信息获取手段，更加完整的记录了目标的回波散射信息。全极化SAR分类结果既可为目标检测、边缘提取等进一步的分析或解译提供辅助信息,也可作为最终结果。与普通遥感图像相比，全极化SAR分类技术对于揭示地物极化散射信息更具有研究价值。本文以提高全极化SAR数据的分类精度为主要目的,研究了全极化SAR图像分类方法。目前使用全极化SAR进行地物分类的方法主要有两种：非监督分类和监督分类。大部分非监督分类方法的优点在于提供了用于指派最终地物类型的辅助信息；但每个聚类对应某种单一散射机制，并不能代表实际地物，因此在对大规模的遥感数据处理时必须依靠人工专家的参与解译。监督分类方法通常基于像素或小区域得到底层特征，利用这些底层特征对具有单一散射机制的地物进行分类非常有效，但对复杂地物进行分类时会遇到困难。针对上述两种问题，本文提出一种使用中间层特征MLF(Middle-Level-Feature)的监督分类方法。即统计以某像素为中心的一定区域（矩形窗口）内各“中间成分（基于底层极化特征得到的非监督聚类类别）”的占比作为该像素的MLF，依此计算所有位置像素的MLF，然后利用支持向量机进行监督分类。本文在覆盖武汉地区的Radarsat-2全极化数据上，与基于经典全极化特征的SVM（Support Vector Machine）监督分类方法进行了对比，研究了不同中间成分获取方法以及特征支持窗口对于分类性能的影响，结果显示本文方法有很好的性能并有进一步提升的空间。对于全极化SAR数据中众多的极化信息，极化方位角反映了散射目标相对于雷达视线的旋转角度，即方位向坡度。在极化SAR图像分类中，同一类别地物目标所处方位向坡度的差异，体现在极化SAR数据中是极化特性的不同，将导致被分为不同的类别。为了消除这种由于地形因素造成的分类误差，本文进行了极化方位角的补偿，以改善极化SAR数据的分类结果。本文利用DEM（Digital Elevation Model）估计出极化方位角并做了极化方位角补偿。实验发现，极化方位角补偿后，经过极化补偿之后，Freeman分解中体散射功率会减小、二次散射功率均减小、绝大多数像素的面散射分量也会减小，但是减小值大部分都在0附近。对于Cloude分解，极化方位角补偿后，极化数据区分两个相对较弱的散射分量的能力增强，同时散射介质的随机性增强，代表散射过程物理机制的alpha值约60%的像素值减小。
[Abstract]:Fully polarized synthetic aperture radar (Synthetic Aperture Radar,SAR) as an advanced remote sensing information acquisition method, more complete recording of the target echo scattering information. The results of fully polarized SAR classification can not only provide auxiliary information for further analysis or interpretation, such as target detection, edge extraction and so on, but also can be used as final results. Compared with conventional remote sensing images, the fully polarized SAR classification technique is more valuable to reveal the polarimetric scattering information of ground objects. In order to improve the classification accuracy of fully polarized SAR data, a method of full polarization SAR image classification is studied in this paper. At present, there are two main methods for ground object classification using fully polarized SAR: unsupervised classification and supervised classification. The advantage of most unsupervised classification methods is that they provide auxiliary information for assigning final feature types. However, each cluster corresponds to a single scattering mechanism, which can not represent the real objects. Therefore, the interpretation of large-scale remote sensing data must rely on the participation of artificial experts. Supervised classification methods are usually based on pixels or small regions to obtain bottom features. Using these underlying features to classify objects with a single scattering mechanism is very effective, but it will be difficult to classify complex objects. In order to solve the above two problems, a supervised classification method using MLF (Middle-Level-Feature) is proposed. That is, the percentage of the "intermediate components (unsupervised clustering categories based on the underlying polarization feature) in a certain region (rectangular window) centered on a pixel is counted as the MLF, of the pixel. The MLF, of all the pixels is calculated accordingly." Then support vector machine is used for supervised classification. In this paper, the Radarsat-2 full polarization data covering Wuhan area are compared with the SVM (Support Vector Machine) supervised classification method based on classical full polarization features. The effects of different intermediate component acquisition methods and feature support windows on classification performance are studied. The results show that the proposed method has good performance and further improvement. For all polarimetric SAR data, the polarization azimuth reflects the rotation angle of the scattering target relative to the radar line of sight, that is, the azimuth slope. In the classification of polarimetric SAR images, the difference of azimuth gradient of the same ground object is reflected in the polarization characteristics of the polarimetric SAR data, which will lead to the classification of different categories. In order to eliminate the classification error caused by topographic factors, the polarization azimuth compensation is carried out to improve the classification results of polarized SAR data. In this paper, the polarization azimuth angle is estimated by DEM (Digital Elevation Model) and the polarization azimuth compensation is made. The experimental results show that after polarization compensation and polarization compensation, the volume scattering power and secondary scattering power in Freeman decomposition will decrease, and the surface scattering components of most pixels will also decrease, but most of the decreases are near zero. For Cloude decomposition, after polarization azimuth compensation, the ability of polarization data to distinguish two relatively weak scattering components is enhanced, and the randomness of scattering medium is enhanced, and the pixel value of alpha representing the physical mechanism of scattering process is reduced by about 60%.
【学位授予单位】：贵州师范大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN957.52

【参考文献】