【作者】 肖志强；

【导师】 鲍光淑；

【作者基本信息】 中南大学 ， 地球探测与信息技术， 2004， 博士

【摘要】 空间数据的更新已成为目前世界各国的地理信息系统的数据库面临的突出问题。随着遥感技术的不断发展，特别是各种SAR传感器的相继上天，遥感作为空间数据获取的一种手段，其在空间数据更新中的作用显得越来越重要。当前，从遥感图像中，尤其是高分辨率SAR图像中提取道路网络已成为遥感技术应用研究中的热点之一，其目的就是利用自动或半自动提取技术为道路中心线的描述和GIS空间数据库的更新提供一种行之有效的方法。论文在总结分析国内外已有研究成果的基础上，对高分辨率SAR图像中道路网络提取方法及地理信息系统空间数据更新方法进行了深入研究。 高分辨率SAR图像中细节丰富，目标背景异常复杂，同时SAR图像受其固有的斑点噪声的影响，很难直接从原始图像中提取道路网络。针对这一特性，论文对高分辨率SAR图像进行聚类分析，将道路类象素从图像中分离出来，使问题得到简化。然后分别建立不同的数学模型，利用遗传算法和Snakes方法提取道路中心线。提出了两种从高分辨率SAR图像中提取道路网络的半自动提取方法。实验结果表明基于遗传算法的道路网络提取算法具有较高的计算精度，但计算时间稍长，而基于Snakes的提取方法计算速度较快，但计算精度有所下降。 对于合成孔径雷达这种侧视遥感器图像，SAR图像中道路与周围背景对比度强弱与雷达的侧视方向及道路的延伸方向密切相关，针对这一问题，提出了一种多阈值局部检测算子。局部检测的各方向的道路象素经过曲线拟合形成相应方向的道路线段。以提取的道路线段为图的顶点建立道路网络的数据结构，将这些线段有效地组织起来。针对局部检测的道路线段之间“相邻”等关系具有不确定性，提出了一种利用模糊推理系统进行道路线段全局连接的方法。利用该方法连接道路线段的具体步骤包括共线连接，线段整合，交叉点连接等，每一步包含一个Mamdani类型的模糊推理系统。该方法具有较好的适应性，只要合理选择模糊规则，就可以利用模糊系统将道路线段连接起来形成道路网络。 同一区域各种不同传感器数据之间存在一定的冗余和互补信息，融合多种传感器数据可以减少系统总的不确定性，增加提取特征的精确性。为充分利用各种传感器数据更准确地提取道路网络，论文提出了一种融合多传感器数据提取道路网络的方法。在局部检测时，基于DRO线性特征检测算子的基本思想，针对雷达图像和光学图像特性，分别设计了一种改进的局部检测算子，该算子可以检测任意方向的道路象素。利用D一S证据理论融合两算子分别在SAR图像和TM图像中的检测结果，融合检测结果经细化和线性拟合后，利用模糊推理系统融合先验知识对道路线段进行连接。试验结果表明融合检测的网络较单独从SAR图像中提取的结果无论是完整性还是正确性都有明显提高。同时，论文还对文中提出的各种道路网络提取算法的定量评价方法进行了探讨。 最后，论文对利用遥感图像更新地理信息系统道路网络数据的方法进行了研究，提出了一种线性特征变化检测的方法。该方法以地理信息系统的道路网络图层中单个道路特征为变化检测单元，对道路特征中两坐标点之间线段与遥感图像中融合检测的相应道路线段进行匹配，判断是否应对原有数据进行校正或增加新的道路特征。该方法克服了缓冲区检测算子对特征局部变化不敏感的局限性。更多还原

【Abstract】 The update of spatial data has become a serious problem confronted by the database of geographical information system all over the world. With the development of the technique of remote sensing, especially the launch of the SAR remote sensors, remote sensing, as a tool to update the spatial data, is becoming more and more important. Currently, the road network extraction from the remote sensing images, especially from the high-resolution SAR images, has become a hotspot in remote sensing application research. And the goal is to offer viable and cost-effective approaches for road centerline delineation and for revision of spatial databases using automated or semi-automated extraction techniques. Based on the analysis and summarization of the research home and abroad, the dissertation makes deep research on the methods to extract road network from SAR images and to update the spatial data of GIS.In high-resolution SAR images, rich details and the complicated background of objectives, along with the intrinsic speckle, make it difficult to extract road network directly from original SAR images. Aiming at this problem, fuzzy C means is first used to classify the SAR images to extract road pixels, and then GA or Snakes to extract road centerline according to different mathematical models. The dissertation presents two semi-automated methods to extract road network from high-resolution SAR images. The experimental results show that the method based on GA bears higher accuracy, but the computing time is a little longer, while the speed of the method based on Snakes is faster, but the accuracy falls.The contrast of road in SAR images with the background is tightly related with the side-looking direction of radar and the extension direction of roads. Aiming at this problem, a multi-threshold detector is presented. The road pixels detected by the local detectors form into road segments of the corresponding direction by using curve fitting. And a data structure of graph whose vertices are the extracted road segments is constructed for road network to organize the segments effectively.Because of the uncertainty existing in the relations (such as proximity) between the road segments, a method to connect road segments using fuzzy inference system is presented. In this method, there are three steps ?Collinearity connection, segment mergence and junction connection. And each step includes a fuzzy inference system of Mamdani type. The method possesses very preferable adaptability. If only fuzzy rules selected rationally, the fuzzy inference system can connect segments into road network.Because of the redundant and complementary information existing in different sensor data from the same areas, it can reduce the uncertainty and improve the accuracy of the extracted features by fusing multi-sensor data. In order to make full use of all kinds of sensor data to extract road network more exactly, the dissertation presents a method to extract road network by fusing multi-sensor data. During the local detection, aiming at the features of radar images and optical images, the respective line detectors are devised based on the thought of Dudo Road Operator (DRO), and they can detect road pixels at any direction. The Dempster-Shafer evidence theory is used to fuse the detected results of different sensor images such as SAR images and TM images. After the road pixels obtained by fusion being thinned and fitted linearly, the road segments are connected conditionally by using the fuzzy reasoning system that fuses some prior knowledge. The result shows that both the completeness and correctness of the road network obtained by the fusion method are sharply higher than that of the road network extracted only from SAR images. And the quality of the road network extraction algorithm presented in this dissertation is evaluated.Finally, the dissertation makes research on the method to update road network data of GIS by using remote sensing images, and presents an algorithm for linear features to detect changes. Taking each road feature in the更多还原