【作者】 程晓光；

【导师】 龚健雅；

【作者基本信息】 武汉大学 ， 摄影测量与遥感， 2014， 博士

【摘要】 作为一种主动遥感方式,极化合成孔径雷达(PolSAR)具有全天时全天候工作能力,其分辨率一般高于普通的真实孔径雷达。最近几年,它开始在军事,测绘,农业,林业,地质等领域得到广泛应用。作为一种从PolSAR中提取信息的重要方法,极化分解,尤其是基于模型的非相干极化分解,是最近几年PolSAR领域内最活跃的方向之一。它可以获得不同散射机制的功率和其它参数,进而用于PolSAR影像分类,干涉SAR,极化相干斑滤波,土壤粗糙度和湿度估计等。自从Freeman和Durden提出三分量分解法后到现在,研究人员已经提出了二十多种基于模型的非相干分解法。这些方法虽然得到了不少成功的应用,但是普遍存在诸如不满足特征值非负约束,出现负功率值,高估体散射功率,对极化信息利用不充分,对地面散射采用相干散射模型进行模拟,不能描述去极化,以及难以有效区分森林和分布方向不平行于SAR方位向的建筑物等问题。一般采用真实数据对分解法进行验证,缺乏和真值的比较,难以定量评估分解法对各分量功率及其它参数估计的准确度。针对上述问题,本文首先创建了一个基于极化分解的模拟框架,模拟不同分量的参数,计算它们不基于反射对称的散射模型,得到功率加权后的相干矩阵。通过对模拟得到的相干矩阵利用不同方法进行分解,可以将分解结果与模拟参数进行定量比较。作者还挑选了不同散射机制主导的模拟数据,以更好地模拟真实情况。本文提出了两种高度自适应的分解法。这两种方法都进行去方位处理,应用特征值非负约束到螺旋散射和体散射参数的计算,采用Neumann自适应散射模型和双极子来描述体散射,选择能让体散射解释最多交叉极化功率的参数作为最优的体散射参数。但是第一种分解法不基于反射对称计算体散射参数(简称为RAVD),导致在一般情况下,螺旋散射和体散射不能解释所有的交叉极化功率。为此,采用Neumann模型描述主导地面散射以解释剩余的交叉极化功率,采用相干模型描述次要地面散射。而第二种分解法计算体散射参数时基于反射对称假设(简称为RSVD),使得在大部分以表面或双次散射为主的区域,体散射和螺旋散射能解释全部的交叉极化功率,再由van Zyl分解即可获得表面散射和双次散射的参数。但是在部分森林地区,少部分交叉极化功率不能由体散射和螺旋散射解释。在这种情况下,对观测到的相干矩阵进行三分量分解,其中体散射和主导地面散射均由反射对称的Neumann模型描述。如果上述分解不能取得合理结果,则进行三分量反射不对称分解。利用模拟数据和UAVSAR数据所做的实验表明,在绝大多数情况下,这两种方法可以匹配除T13外其它观测到的相干矩阵中的元素。如果进行三分量反射不对称分解,则有可能匹配除T13虚部之外的其它相干矩阵元素。RSVD避免了负功率值的出现,而RAVD的结果中,负功率值的比例也低于0.070%。两种分解法明显降低了对体散射功率的高估,估计各分量功率的准确度高于几种最新的特征值非负分解法。在大多数情况下,RAVD估计不同分量的方位角随机度和复散射系数的准确度优于RSVD,而且在它的结果中,森林和延伸方向不平行于SAR方位向的建筑物可以较为容易地区分开。但是在以表面散射或双次散射为主的区域,RSVD估计各分量功率的效果优于RAVD.本文还提出了一种基于功率的非监督散射机制分类法。散射机制类被定义为不同主导和次要散射机制的组合。通过分析不同散射机制的特征以及两种散射机制混合时的特征,作者给出了一种基于极化特征和特征域分割的散射机制分类法。由于该分类法基本不依赖于极化分解,所以避免了对体散射功率的高估和特征值分解。该分类法的效率大大高于Wishart-H/alpha法和模糊H/alpha法,而且能够提供次要散射机制的类别。该方法可以用于PolSAR影像的快速分类,其分类结果可以作为更复杂的分类器的初始分类。它还可能用于简化基于模型的非相干分解。在利用模拟数据的实验中,该方法给出的Kappa系数为0.864。该方法识别主导散射机制的效果显著优于H/alpha法,Wishart-H/alpha法和模糊H/alpha法。UAVSAR数据的实验表明,该方法能够有效识别森林和城区的主导和次要散射机制。更多还原

【Abstract】 As an active remote sensing tool, polarimetric Synthetic Aperture Radar (PolSAR) is with "all weather", day and night imaging capabilities. Its resolution is generally higher than that of real aperture radar. In recent years, it began to be widely used in the field of military, surveying and mapping, agriculture, forestry, geology et al. As an important information extracting method, polarimetric decomposition, especially incoherent model-based decomposition, is one of the most active research directions in PolSAR field. It can get the powers and parameters of different scattering mechanisms, hence being used for PolSAR image classification, SAR interferometry, speckle filtering, soil moisture and surface roughness estimation, and so on.Since Freeman and Durden proposed the three-component decomposition, more than20incoherent model-based decompositions have been published. Although many successful applications have been achieved, several severe problems exist, including violating non-negative eigenvalue constraint (NNEC), occurrence of negative powers, overestimation of volume scattering power, insufficient utilization of polarimetric information, employing coherent models for ground scattering hence being incapable of describing depolarizing effect, the confusion between forests and buildings whose orientation is unparallel to SAR azimuth direction, etc. In general, real data is used to verify the decomposition results in experiment, so ground truth is lack and quantitative evaluations of the accuracy of decomposed powers and other parameters are difficult.To solve these problems, a simulation framework on the basis of model-based incoherent decomposition was first established in this paper. By simulating parameters of all components, computing corresponding reflection asymmetric scattering models, the power-weighted coherency matrix [T] were obtained. With simulated data, we are able to quantitatively compare the results given by diverse decompositions with simulation parameters. Simulated [T] with different dominant mechanisms were specially selected to better model the real conditions.Two highly adaptive decompositions were proposed in this paper. Both methods perform deorientation processing, apply NNEC to the calculation of the parameters of helix scattering and volume scattering, utilize Neumann’s scattering model and dipole to describe volume scattering, and select the parameters that let volume scattering explain the most cross-polarized power. But the first decomposition computes volume scattering parameters without reflection symmetry assumption (denote this decomposition as RAVD). As a result, generally, helix scattering and volume scattering cannot explain all cross-polarized power. Therefore, apply Neumann’s depolarizing model to describe the dominant ground scattering to explain the remaining cross-polarized power and coheren model to describe the secondary ground scattering. The second decomposition computes volume scattering parameters with reflection symmetry assumption (denote this decomposition as RSVD). In most areas dominated by surface scattering and double-bounce scattering, cross-polarized power is explained by helix scattering and volume scattering, then we can get the parameters of surface scattering and double-bounce via van Zyl decomposition. However, in some forests, a small proportion of cross-polarized power cannot be explained by volume scattering and helix scattering. In this case, perform a three-component decomposition to the observed coherency matrix, during which volume scattering and the dominant ground scattering are both modeled by Neumann’s reflection symmetric model. If desirable results cannot be achieved, then perform a three-component reflection asymmetric decomposition to the observed coherency matrix.Experiments using simulated data and UAVSAR data revealed that, in most cases, all elements in the observed coherency matrix except T13can be well fitted by both decompositions. If three-component reflection asymmetric decomposition is implemented, it is likely to fit all the elements in the observed coherency matrix except the imaginary part of T13. RSVD completely avoids the occurrence of negative powers, meanwhile, in the results of RAVD, the proportion of negative powers is lower than0.070%. Compared with several latest Nonnegative Eigenvalue decompositions, the two proposed decompositions significantly lower the estimation of volume scattering power and better estimate component powers. In most instances, RAVD performs better than RSVD in estimating orientation angle randomness and complex scattering coefficients of different components. It is also worth noticing that in the decomposition results of RAVD, forests and the buildings whose orientation direction is unparallel to SAR azimuth direction can be easily separated. But in the areas dominated by surface scattering or double-bounce, RSVD better estimate component powers than RAVD.An unsupervised scattering mechanism classifier was advanced based on powers of scattering mechanisms. Scattering mechanism classes are defined as the combinations of dominant and secondary mechanisms. Through analysis of the characteristics of several typical mechanisms alone and mixture of two mechanisms, a classifier on the basis of characteristics and segmentation of characteristic space, was given. Since the proposed classifier is nearly free of polarimetric decomposition, it avoids the overestimation of volume scattering power and eigenvalue decomposition. The proposed classifier has much higher efficiency than Wishart-H/alpha classifier and fuzzy H/alpha classifier, and could provide secondary mechanism. The proposed classifier can be used for a fast classification of PolSAR images, or as a pre-classification step of sophisticated classifier. It also has the potential to simplify model-based incoherent decomposition.In experiments with simulated data, the Kappa coefficient by the proposed method was0.864. It performed much better than H/alpha classifier, Wishart-H/alpha classifier as well as fuzzy H/alpha classifier in the identification of dominant scattering mechanism. It was demonstrated by UAVSAR data that the proposed classifier was able to identify dominant and secondary mechanism in forests and urban areas.更多还原