基于流形学习的高分辨率SAR影像城市地物特征提取方法研究

发布时间：2018-08-02 21:14

【摘要】：合成孔径雷达(SAR)能够穿云透雾,具有全天时全天候监测的独特优势,能够弥补光学传感器在多云雨雾地区无法获取有效数据的缺陷,已成为遥感信息提取的重要手段。随着高分SAR卫星的陆续发射,城市典型地物提取已成为高分SAR的应用热点,可为城市规划、土地利用监测、人口密度调查等提供可持续的科学基础数据。由于高分辨率SAR影像城市地物复杂的散射特性,目前基于高分辨率SAR数据的典型地物提取的精度不高。同时,高分辨率SAR影像的高维非线性特点使城市典型地物自动提取的难度加大。流形学习作为一类机器学习的新方法,能够发现高维空间数据的内在本质特征。将善于处理非线性数据的流形学习方法应用于高分辨率SAR影像的特征提取,有利于提高目标识别精度。因此,为提高城市典型地物识别和提取精度,研究典型地物自动提取技术,论文研究基于流形学习方法的高分辨率SAR影像城市典型地物特征提取与识别方法。主要研究内容如下:(1)分析了高分辨率SAR影像的城市典型地物特征并构建了高维特征集。首先,从几何特征、辐射特征及各城市地物的图像特征对高分辨率SAR影像进行了详细的分析;其次,利用经典的二阶概率统计方法灰度共生矩阵(Gray Level Co-occurrence Matrix,GLCM)提取了图像8种纹理特征,与图像的灰度特征构建SAR影像的高维特征集;最后,通过实验分析,确定了纹理特征提取的最佳窗口参数。(2)分别选择拉普拉斯特征映射(Laplacian Eigenmap,LE)、局部线性嵌入(Locally Linear Embedding,LLE)、Hessian局部线性嵌入算法(Hessian Locally Linear Embedding,HLLE)、局部切空间排列(Local Tangent Space Alignment,LTSA)、局部保持投影(Locality Preserving Projections,LPP)5种典型的流形学习方法对三种城市典型地物类型(建筑区、水体、体育场)的高维特征集进行降维,最终提取出三种地物类型,并对5种方法提取的结果进行精度评价,综合分析了各种方法的优缺点。(3)对选择的局部切空间排列(Local Tangent Space Alignment,LTSA)方法,针对高分SAR数据分布不均的情况进行改进。综合考虑流形形态结构与样本邻域的欧式距离,对原方法的切空间估计进行加权改进,提出了一种基于距离与结构加权的局部切空间排列算法(Distance and Structure Weighting Local Tangent Space Alignment,DSWLTSA),并将算法应用到高分SAR图像高维特征集的维数约简中。以3种典型的地物为例,通过实验对DSWLTSA和LTSA算法进行对比分析,验证了DSWLTSA算法的有效性;通过实验深入的分析了DSWLTSA算法的适用性和应用价值。(4)针对局部线性嵌入算法(Locally Linear Embedding,LLE)对于样本采样稀疏效果不好的问题,提出一种基于均匀化样本距离的LLE算法(Distance Homogenization Locally Linear Embedding,DHLLE),新的方法通过对样本之间距离的重新计算,使问题得到改善。将算法应用到高分SAR图像高维特征集的维数约简中。验证了LLE算法与DHLLE算法对城市典型地物的识别能力,并对比分析了DHLLE算法与DSWLTSA算法的优缺点。
[Abstract]:Synthetic aperture radar (SAR) can wear cloud permeable fog, and has the unique advantage of all-weather all weather monitoring. It can make up for the defect that optical sensor can not obtain effective data in the cloudy and fog area. It has become an important means of remote sensing information extraction. With the launching of high SAR satellite, the city typical ground extraction has become a high grade SAR Hot spots provide sustainable scientific basic data for urban planning, land use monitoring, population density survey and so on. Due to the complex scattering characteristics of high resolution SAR images, the accuracy of typical terrain extraction based on high resolution SAR data is not high. At the same time, high dimensional nonlinear characteristics of high resolution SAR images make cities The difficulty of automatic extraction of typical objects is more difficult. Manifold learning, as a new method of machine learning, can discover the intrinsic characteristics of high dimensional spatial data. The manifold learning method, which is good at dealing with nonlinear data, is applied to the feature extraction of high resolution SAR images, which is helpful to improve the accuracy of target recognition. Therefore, to improve the city code The automatic extraction technology of typical ground objects is studied and the method of extracting and identifying the typical features of the city based on the manifold learning method is studied in this paper. The main research contents are as follows: (1) the characteristics of the city canonical features of high resolution SAR images are analyzed and the high vate collection is constructed. First, the SAR feature collection is constructed. The high resolution SAR images are analyzed in detail. Secondly, 8 texture features are extracted by the classical two order probability statistical method of Gray Level Co-occurrence Matrix (GLCM), and the high Vitter collection of the SAR image is constructed with the gray features of the image. Finally, through the experimental analysis, the optimum window parameters for texture feature extraction are determined. (2) select Laplasse Eigenmap (LE), local linear embedding (Locally Linear Embedding, LLE), Hessian local linear embedding algorithm (Hessian Locally Linear Embedding,), local tangent space arrangement Alignment, LTSA), Locality Preserving Projections (LPP), 5 typical manifold learning methods are used to reduce the dimension of the high Vette collection of three typical types of city objects (construction area, water body, stadium), and finally extract three types of ground objects, and evaluate the accuracy of the results extracted by the 5 methods. The advantages and disadvantages of the method. (3) to improve the Local Tangent Space Alignment (LTSA) method, to improve the uneven distribution of the high grade SAR data. Considering the Euclidean distance between the shape structure of the manifold and the neighborhood of the sample, a weighted improvement is made for the estimation of the tangent space of the original method, and a kind of distance and junction based on the distance and knot is proposed. The weighted local tangent spatial arrangement algorithm (Distance and Structure Weighting Local Tangent Space Alignment, DSWLTSA) is applied to the dimensionality reduction of high grade SAR image high vet collection. The validity of the algorithm is verified by comparison and analysis of DSWLTSA and LTSA calculations with 3 typical terrain objects. The applicability and application value of DSWLTSA algorithm are analyzed through experiments. (4) a LLE algorithm based on homogeneous sample distance (Distance Homogenization Locally Linear Embedding, DHLLE) is proposed for local linear embedding algorithm (Locally Linear Embedding, LLE), which is not good for sample sampling sparse results. The algorithm improves the problem by recalculating the distance between samples. The algorithm is applied to the dimensionality reduction of high SAR image high VAT collection. The recognition ability of LLE and DHLLE algorithm for typical urban terrain is verified, and the advantages and disadvantages of DHLLE algorithm and DSWLTSA algorithm are compared and analyzed.
【学位授予单位】：中国科学院大学(中国科学院遥感与数字地球研究所)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：P237

【参考文献】