地物分析是指利用观测数据认识地球表面物体性质的过程，可广泛应用于资源勘查、土地规划、灾情防治诸多方面。数字高程模型（DEM）和多光谱图像是进行地物分析的两种重要数据源，它们分别提供了地物目标的地形信息与辐射信息。本文结合洪水防治的项目背景，分别利用 DEM 和多光谱图像对地形分析和土地利用分类问题进行了研究，同时还进一步研究了这两种数据的空间集成与信息集成。本文工作及贡献主要包括下面四个部分： （1）DEM 地形分析，针对地形特征表现出的多尺度特性，将小波变换应用于地形分类，综合考虑地形的绝对高程、整体起伏与局部复杂程度，给出了一种地形自动分类方法。另外，根据分洪区周围高程的分布特点，将两种交互式图分割技术应用于分洪区堤坝的检测，从而能够得到连续封闭的曲线对分洪区堤坝进行描述。 （2）多光谱图像的土地利用分类，提出了一种针对多光谱图像的非监督分类方法，其主要创新性在于：根据多光谱图像不同波段对于不同种类地物的不同辐射响应特性，采用了一种波段加权策略，能自动区别不同波段对于不同地物识别的贡献大小，能有效改善地物分类结果。同时，波段权重的选取是随不同地物区域而自动改变，因而具有自适应性。另外，在图像预处理中，拓展了普通均值漂移，提出一种空间均值漂移方法，能完成边缘及其它有用信息的提取。 （3）DEM 与光谱图像的几何配准，针对 DEM 与光谱图像之间由于成像方式不同在全局变换之外还存在局部形变的难点，提出了一种基于轮廓特征参数化描述的配准方法，其特点及创新在于：采用非均匀有理 B 样条（NURBS）曲线拟合轮廓特征，利用 NURBS 曲线的仿射变换不变性，通过 NURBS 控制点的对应估计图像变换参数；利用 NURBS 曲线的局部可控性，能有效避免由于对应特征局部形变造成的误匹配；迭代策略的使用能使配准不断精化。该方法适用于一般多源图像配准。 （4）融合 DEM 与多光谱图像的地物分类，在上述工作基础之上，为 DEM与多光谱图像这两种不同性质数据间的信息集成进行了初步探讨。综合利用DEM 与多光谱图像所提供的地形信息与光谱信息，融合各自数据所含的显著地物特征以完成地物分类。给出三种不同的融合 DEM 与多光谱图像的地物分类方法，并定性比较它们的优劣。
英文摘要
Land analysis is the process of identifying the properties of ground objects based on observed data, which can be used in a wide range of applications, such as resource survey, land planning, disaster management etc. Digital elevation model (DEM) and optical images are two kinds of important data for land analysis tasks, where DEM offer the terrain information and optical images provide the spectral information, respectively. With the supports of flood control and management projects, we pursue the studies about terrain analysis and land-use classification using DEM and multi-spectral images. Furthermore, we also concern on the integration of DEM and multi-spectral images, including their registration and fusion. The related work and main contributions are summarized as follows: 1. Terrain analysis using DEM. Based on the fact that terrain appears some multi-scale property, we apply wavelets transform in terrain analysis, and present an automatic method for terrain classification. In this method, terrain features are extracted by combing terrain elevation and roughness together. In addition, two interactive image segmentation techniques are applied in the task of detecting dike in DEM so that a continues and closed curve can be achieved as the dike description. 2. Land-use classification using multi-spectral images. We propose an unsupervised method for the land-use classification of multi-spectral images. The main contributions of this method lie in two aspects: First, we design a strategy to weight different image bands for their contributions to identify different ground objects, because different bands present different spectral characteristics for different ground objects. Moreover, such weighting strategy is adaptive to the local ground region to be processed. Second, a novel spatial mean shift procedure is proposed, from which some useful information can be extracted. 3. Registration between DEM and multi-spectral images. We propose a novel parametric registration method for multi-sensors data, including DEM and optical images. Control points of No-Uniform Rational B-Spline (NURBS) are used as the matching features due to their invariance under affine transform. Different acquisition times and imaging conditions may bring the local deformation between the corresponding features. By making use of the local controllability of NURBS, mismatching of control points with the local deformation between the corresponding curves can be avoided.
修改评论