【作者】 陈杰；

【导师】 杨敏华； 邓敏；

【作者基本信息】 中南大学 ， 地图制图学与地理信息工程， 2010， 博士

【摘要】 遥感技术的巨大发展扩宽了对地观测的视野,给人们提供了极为丰富的地理信息。随着卫星传感器的空间分辨率不断提高,高分辨率遥感影像已经在城市规划、国土资源管理、地质调查、交通检测等区域性研究与相关的应用领域中扮演着重要角色。虽然遥感影像的分类技术取得了长足发展,但是已有的研究表明基于像元的高分辨率遥感影像分类存在明显的限制。为了克服基于像元的传统信息提取方法所面临的缺点,一种被称为面向对象图像分析(OBIA)的新方法应运而生。作为地理科学的一个分支,OBIA代表着遥感与地理信息系统学科发展的重要趋势。本文围绕着OBIA的主要研究内容与结论包括：(1)基于分水岭变换和小波变换提出多尺度分割方法用于波段融合后的高分辨率多光谱影像。利用该方法进行分割的过程包括多尺度图像生成、图像分割、区域合并和结果映射等四个方面。采用相位—致模型多尺度地提取各近似子图的梯度,并逐个尺度地进行梯度融合。进而分析不同尺度与不同地物的局部梯度方差,选择最佳的小波分解尺度。通过移动阈值与扩展最小变换多层次地标记纹理和灰度的均质区域。以空间相邻关系、面积、光谱与纹理等因素多约束地合并最初的分割区域。处理边界像元将最初的结果投影到更高的尺度直到原始图像。实验结果表明所提方法能够应用到高分辨率影像的分割且可取得较准确的分割效果。(2)基于对纹理频谱的分析提出一种高分辨率遥感影像最佳空间尺度的选择方法。分析了四种典型地物在傅里叶变换频域的频谱响应特性。采用点扩散函数对原始影像进行尺度扩展,进而根据不同尺度下影像纹理的径向与角向曲线变化情况选择最佳尺度。通过分析四种地物在6个尺度下的纹理特征可分性,说明本文方法能客观反映出地物的尺度效应,具备最佳尺度选择的可行性。进行了基于支持向量机的全色影像面向对象分类,实验结果表明在最佳尺度下可取得最佳精度。(3)提出一种基于粗糙集理论的面向对象分类方法以区分高分辨率遥感影像上的不同地物。利用了不可分辨关系、上下近似集和知识约简等方式发现隐含在Gabor纹理特征内的分类规则。在对象光谱特征的初步分类结果基础上,依据纹理分类规则得到最终结果。本文重点提出一种适用于面向对象分类的连续区间属性离散化方法。实验表明本文方法可取得较好分类结果与较高分类精度。(4)结合支持向量机技术与基于粗糙集的粒度计算,提出了一种新的高分辨率遥感影像面向对象分类方法。从多光谱波段数据中提取对象的光谱特征,并用Gabor滤波器组产生纹理特征,利用多核支持向量机进行初步的面向对象分类,对分类结果进行求交后生成信息颗粒。比较颗粒的特征均值与各样本中心的欧氏距离以区分颗粒类别,通过定量分析颗粒间的空间相邻关系判断待定类别的颗粒,利用少量人工交互的识别处理得到最终分类结果。与基于高斯径向基核函数的支持向量机和神经网络两种方法进行了对比,实验结果表明本文所提方法能够取得更好的分类效果。最后,总结了本文的研究成果。下一步需要深入研究的工作有：1)整合多种方法、从多角度进行分析以提高分割的效果；2)将遥感中的尺度因素和具体应用的尺度要求进行统一考虑；3)如何在遥感信息提取中充分利用智能方法。更多还原

【Abstract】 Great achievements of remote sensing technology having extended the visual field of Earth Observation, makes human obtain very abundant geographical information. With the progresses in spatial resolution of satellite sensors, high-resolution remotely sensed imagery has play an important role in regional studies of urban planning, territorial resources management, geological survey, vehicle detection, and concerned application fields. Though the technology of the remote sensing image classification has made considerable progress, available investigations have shown that the pixel-based approach has explicit limits in classification of such high-resolution remotely sensed imagery. For overcoming the drawbacks suffered by conventional information extraction method which based on pixel, a new method named Object-based Image Analysis has emerged as a reflection of the times. As a sub-discipline of GIScience, OBIA represents a significant trend in remote sensing and GIScience. The main contents and results of the study centred on OBIA were presented as follows.(1) An efficient multiscale approach based on watershed transform and wavelet transform is presented for segmentation of the pan-sharpened high-resolution multispectral remote sensing imagery. The procedure toward complete segmentation using the proposed method consists of pyramid representation, image segmentation, region merging and result projection. Multi-scale gradient images are obtained by applying phase congruency model to approximation coefficients, and gradient magnitudes of all bands are fused at each scale. The optimal scale of wavelet decomposition is chosen by analysis local gradient variance varying correspond to different scales and varieties of geo-objects. Multi-level marker location algorithm is subsequently used to locate significative regions that are homogeneous in terms of texture and intensity, by moving threshold and extended minima transform. A multi-constraint region merging strategy considering spatial adjacency relation, area, spectral and textural properties is proposed to merge the initial segments. Pixels at boundaries are assigned to refine object contours. The experimental results demonstrate that the developed method can be applied to the segmentation of high resolution remote sensing images and get the high accuracy segmentation.(2) An approach based on texture frequency analysis is proposed to determine the optimal spatial scale for high resolution imagery. Four typical geo-objects are used to analyze their frequency properties of the response to the Fourier transform domain. The original image is up-scaled to different spatial resolutions using point spread function. The adequate spatial scale is chosen from the up-scaled images according to the change patterns in the radius distribution and angle distribution curves. Separability among four types of objects at six scales is analyzed based on texture feature to approve the feasibility of the new method. Object-oriented classification of the panchromatic image by means of SVM is implemented, and results of experiment demonstrate that higher accuracy can be obtained at the optimal spatial scale.(3) An object-based algorithm based on rough set theory is proposed to classify different objects extracted from high-resolution remotely sensed imagery. The indiscernibility relation, upper or lower approximation, and knowledge reduction in the rough set theory are used to discover the connotative rules of classification from the Gabor texture feature. Based on the preliminary clustering result derived from spectral feature of objects, the ultimate classification is achieved referring to the rules. A new technique to discretize continuous interval-valued attributes is developed, which is very suitable for the object-based classification. The experiments demonstrate that the proposed method can achieve better results and better accuracies.(4) A new object-based method for classification of high-resolution remotely sensed imagery is proposed in the paper, which integrates support vector machine (SVM) technique with rough-set-based granular computing (RSBGC). Spectral characteristic is got from multi-spectral data and texture feature is extracted by Gabor filters. Multi-kernel SVM is used to present preparatory object-based classification, and information granularities are obtained through intersection of the classification results. Granularities are differentiated by means of comparing the Euclidean distance between average value of granularity and every sample central moment. Spatial adjacency relation among the granularities is quantitative analyzed in order to classify the uncertain granularities after the former clustering. The resulting classification is achieved by little artificial interaction identification. A comparative experiment is performed with both SVM and neural network methods based on RBF-kernel function. It is shown that the proposed method can obtain better classification results.Finally, after concluding all about research work in this dissertation, further work need be advanced study:1) the integration of multi-method and the multi-angle analysis will be conducive to improve the segmentation; 2) scale factors in remote sensing and scale requirement according the specific application should be considered together; 3) how to give full play to the advantages of intelligent methods in remote sensing information extraction.更多还原