图谱耦合的高分辨率城市精细不透水面提取方法研究

发布时间：2018-01-20 15:42

本文关键词： 高空间分辨率遥感图像不透水面图谱耦合多尺度云平台Hadoop　出处：《中国科学院大学(中国科学院遥感与数字地球研究所)》2016年博士论文　论文类型：学位论文

【摘要】：不透水面的覆盖度、类型组成、空间分布等指标在城市化进程及环境质量评估中具有重要意义,准确估算与提取不透水面信息可以为构建“海绵型”生态城市提供数据层面的支撑。随着成像技术的发展,成像空间分辨率得到不断提升,为提取精细的不透水面信息提供了可能。然而,由于城市背景复杂,不透水面组成多样,基于高分辨率影像的精确不透水面提取面临着严重的“同物异谱,异物同谱”的问题。高分影像不仅可以提供光谱信息,还含有丰富的形状、纹理、上下文信息等,这些特征的组合应用可以很大程度上提高不透水面的提取精度。在这种形势下,传统的基于光谱信息的中、低分影像不透水面信息提取方法明显不适用于高分影像。因此,本文以融合遥感影像的光谱和空间特征为出发点,充分挖掘影像的图谱特征,建立图谱协同利用的分类模型,进而实现城市尺度高分不透水面精确提取。本文研究的主要内容与创新点如下:(1)提出了先验知识引导下的城市地区多尺度多层次的遥感影像分割方法。为了解决传统的单尺度的分割方法无法满足复杂城市地物的多尺度分割需求,而产生严重的“过分割”、“欠分割”问题。本文提出了基于先验知识的多尺度多层次分割方法,将像素级的影像分类结果与第一层次大尺度的影像分割结果相融合,并转化为先验知识,辅助影像第二层次多尺度(不同地物)的分割过程,实现自适应的地物适宜分割尺度参数的应用,满足了不同地物分割尺度参数需求。实验结果表明,相比单尺度方法,本文方法可以获得更加均质而完整的地物斑块,有利于不同尺度地物对象形态和特征提取及后续属性的识别。(2)提出了基于距离度量学习的图谱耦合的高分不透水面分类方法。针对传统的矢量叠加的多特征融合方法在光谱异质度高的城市地表分类应用中的不足,本文以协同利用遥感影像的光谱和空间特征为出发点,将距离度量学习技术引入到不透水信息提取过程中,通过对当前影像已标签样本的学习,得到反映样本特征空间的距离度量函数,充分挖掘了有利于当前影像分类的图谱特征组合与应用模式,从而完成图谱耦合的高精度城市地表分类过程,进而提取精确的不透水面信息。实验结果表明,本文方法可以显著提升城市地表的分类精度,为后续高精度的不透水面提取打下坚实基础。(3)提出了一套基于Hadoop的不透水面分布式提取的流程。由于不透水面本身组成多样,城市背景复杂,高分精细不透水面提取算法复杂度比较高,无能满足城市尺度快速而精确地不透水面信息提取的应用需求,本文将计算机领域成熟的云平台Hadoop应用于高分遥感不透水面信息提取应用中,并针对遥感影像及高分不透水面提取算法的特点,设计并实现了基于Hadoop的高分影像自动分块与合并策略以及城市高分不透水面提取算法的并行机制。最终,通过遥感影像数据或数据集的分块处理,实现高分不透水面的分布式、快速提取。实验结果表明,本文方法在不影响高分不透水面提取精度的前提下,可以大幅提高其提取速率。
[Abstract]:Which type of coverage, impervious surface, spatial distribution index has important significance in the process of city and environmental quality assessment, accurate estimation and extraction of impervious surface information can provide data support for the building of the level of "sponge" eco city. With the development of imaging technology, imaging spatial resolution has been improved, providing the possibility for the fine extraction of impervious surface information. However, because the city complex background, impervious surface diversity, high resolution image accurate impervious surface extraction faces the serious "synonyms spectrum based on objects with the same spectrum". The image can not only provide high spectral information, but also rich in shape. Texture, context information, extraction accuracy of combination of these features can greatly improve impervious surface. In this situation, based on the traditional spectral information, low Images of impervious surface information extraction method was not suitable for high-resolution images. Therefore, based on the spectral and spatial features of remote sensing image fusion as a starting point, fully tap the image map features, the classification model of collaborative use, so as to realize the high city scale accurate extraction of impervious surface. The main research contents and innovation points are as follows: (1) the proposed segmentation method of remote sensing image multi scale city area under the guidance of prior knowledge. In order to solve the traditional single scale segmentation method can not meet the complex features of the city multi-scale segmentation requirements, and have serious over segmentation, less segmentation problem is put forward in this paper. The multi-level multi-scale segmentation method based on prior knowledge, the image classification results of pixel and the first level of large scale image segmentation and integration into prior knowledge, auxiliary To help the second level multi scale image segmentation process (different objects), the realization of adaptive features suitable segmentation scale parameters, to meet the demand of different objects segmentation scale parameters. The experimental results show that compared to the traditional method, this method can obtain a more homogeneous and complete surface patches, facilitates the identification of different scale morphology and characteristic extraction objects and subsequent properties. (2) proposed coupled map distance metric learning with high impervious surface classification method based on multi feature vector superposition for lack of traditional fusion methods in city surface classification using spectral heterogeneity in the high, based on the cooperative spatial and spectral features of remote sensing images that will be the distance metric learning technique is introduced to the impervious information extraction process, based on the current image labeled samples, reflect the sample characteristics The spatial distance measure function, fully tap the map feature combination and application mode of the image classification, thus completing the classification process of high precision surface City coupling, and then extract the accurate impervious surface information. The experimental results show that the classification accuracy of this method can significantly improve the city's surface, for the subsequent high precision is not through the water extraction and lay a solid foundation. (3) proposed a set of Hadoop impervious surface extraction process based on distributed. The impervious surface itself consists of diverse city background is complex, high precise impervious surface extraction algorithm complexity is relatively high, inability to meet the city scale rapidly and accurately through the application of water demand information this paper will extract the computer field mature cloud platform Hadoop is applied to the high resolution remote sensing of impervious surface information extraction applications, and according to the remote sensing image and high water Surface extraction algorithm, the design and implementation of the strategy and city block and with high impervious surface extraction algorithm of the parallel mechanism of automatic image score based on Hadoop. Finally, the block processing through remote sensing image data or data set, to achieve the distributed, high impervious surface extraction. The experimental results show that the the method without affecting the high impervious surface extraction accuracy, can greatly improve the extraction rate.

【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】：博士
【学位授予年份】：2016
【分类号】：P237

【相似文献】