高分辨率遥感影像中城市道路网的提取方法研究

In the large amount of related papers, there are some methods which are partially effective for road extraction. Combining them may help us to find a new way for extracting roads more accurately and more efficiently. To extract road networks in urban area from high resolution remotely sensed images will inevitably involve solving the following problems: segmenting roads, generating road segments, smoothing road segments, linking road segments and vectorizing roads. The most critical problem is segmenting roads. Ideal segmentation results output continuous and smooth road ribbons which contain no holes and are not attached by irregular blocks, meanwhile roads are segmented sequentially. All these mean that no post-segmentation processes are needed, because the segmentation results are the very vectorized road networks we want. We hope a new segmentation method can segment roads as ideal as possible..In this project we attempt to segment roads from the background one by one, by the integrated use of Markov random field theory, mixture model theory, target tracking theory and shape prior technique. Multispectral data and texture values consist of the feature vector .Our previous studies showed that smooth roads may be obtained in the segmentation step with the aid of Markov random field theory and shape prior knowledge, and an obtained road segment may be used to help the segmentation of its following part and to fill fractures by the aid of target tracking principle. The difficulty of this project in how to express the probability density functions of road points and background points in terms of powerful mixture models. Because of the integration of advantages of the several excellent theories and techniques the segmentation scheme may output smooth road ribbons which are very similar to the ideal segmentation results mentioned above, and relatively few post-segmentation processes are needed. .This new scheme can be embedded in both an automatic road extraction strategy and a semi-automatic road extraction strategy. In both cases, initial road bands are necessary. Usually parts of a complete road are clear, and traditional segmentation methods combined with some simple post-segmentation processes may help us to find one or more road bands for each road. These road bands naturally serve as initial bands based on which the new segmentation scheme start working. If no band of a road is detected automatically, we put a point on the road surface manually. Then the new segmentation scheme is applied to the window around the point to get an initial road band. In such a case the scheme is embedded in a semi-automatic road extraction strategy. .The probability density functions expressed by mixture models will further be combined with hidden Markov model theory to test the existence of an old road in a new satellite image. The implementation of this project will exert a positive influence on the improvement of urban road extraction performance.

现有的道路提取和图像处理理论中，存在着一些可部分解决道路提取不同方面问题的有效思想，可惜少有人对其进行总结和归纳。由前期研究得知：从一幅普通的高分辨率影像中提取城市道路网必然涉及到道路点分割、道路段形成和修整、道路段连接和道路矢量化几个环节，其中最关键的步骤是道路分割。本项目试图综合运用马尔科夫随机场、混合模型、形状先验以及轨迹跟踪等理论和技术将道路从背景中逐条分割出来。分割所用的特征值包括原始的多光谱数据和道路纹理信息。申请者的前期研究表明：利用随机场理论和形状先验知识可获取连续、光滑的道路；而借鉴轨迹跟踪思想可依据已获取道路段指导后续路面的分割，并能填补断裂。利用混合模型描述路面和背景的概率密度是本项目的难点。此路面分割方案综合了几种理论的优点，可应用于道路的自动和半自动提取，还可结合已有道路网信息，检测旧道路在新影像中是否存在，整个项目的实施对提高城市道路提取的水平将产生积极影响。

自项目生效之日起，项目组全组人员竭尽全力地执行计划书中的每一项任务。. 根据项目计划书的要求，设计了提取一维Gabor纹理特征的方案；利用一维Gabor纹理特征生成路面点与非路面点的高斯混合模型概率密度函数；基于形状先验模型及MRF理论构建分类图；在跟踪外推原理的引导之下，实现对道路的逐段识别；探索利用HMM模型和线划地形图判定道路条带变化的具体流程。此部分成果实现了项目计划书中相关条目的主要研究目标，对实验图像取得了良好的效果。. 在按照项目计划书展开研究的过程中，出现了一些有碍研究进展的新问题：一是难以在遥感影像中标识大量实验用样本；二是形状先验图割法耗时较多；三是分类特征未经优化；四是识别道路子图像的准确性不是很高；五是缺少简单有效的道路条带中心线提取方法。为了获得解决这些新问题所需要的理论和技术，项目组拓展了研究范围。. 拓展研究内容包括：利用半监督学习手段减少人工标记样本的数量；尝试并行实现形状先验图割算法；引入概率增进树方法及CNN卷积网络获取道路初始条带；借助机器学习环境筛选最优分类特征值；基于纹理特征和稀疏编码技术对场景进行分类；提出了基于测地线搜索和优化的道路条带中心线提取新方法。. 为减小建筑物对城市道路提取的影响，项目组研究了新的建筑物轮廓提取方法。作为建立半自动道路提取系统的基础，项目组研究了一种同时顾及像点空间邻域相似性及差异性的主动学习新算法。. 总而言之，项目组对项目计划书中所规定的每一项要求都展开了研究，取得了良好的效果。对实施研究过程中所衍生出的新问题，均能及时地从相关前沿研究成果中寻找应对之策略。项目的实施为提高城市道路提取的水平奠定了扎实的基础。