【作者】 马利刚；

【导师】 王珂；

【作者基本信息】 浙江大学 ， 农业遥感与信息技术， 2014， 博士

【摘要】 近年来,工业园区中土地的浪费和低效利用已经成为加剧土地供需矛盾和制约社会经济发展的重要因素,其粗放利用方式亟需改变。目前管理部门已开始加强对工业园区用地效率的监管,但缺乏客观、详细和精确的基础数据如建筑覆盖率、建筑容积率。传统由用地单位提供的数据因缺乏真实性而广受质疑。随着航空遥感技术的发展,特别是以LiDAR (light detection and ranging,机载激光雷达)传感器平台为代表的最新遥感技术在城区建筑物识别方面表现出的明显优势,使得越来越多的学者将其认定为目前最有前景的城区建筑物提取方案。其客观、快速、低成本、实时、动态和人力投入低的特点十分适合开发区建筑密度的估算研究。到目前为止,尽管LiDAR数据在建筑物提取方面已有广泛应用,但在建筑类型的区分以及工业园区建筑密度估算方面的研究仍不多见。本研究探讨单独使用LiDAR数据和LiDAR/RGB影像结合两种情况下的建筑物分类和建筑密度估算方法。主要研究内容和结论如下：1、LiDAR点云是传感器对地面物体的三维采样,以其为基础计算出的空间自相关特征可以反映地物在高程上的自相关性和高值/低值集聚特点。在仅有LiDAR数据支撑的情况下,尝试充分挖掘点云的空间几何、自相关、高程异质性特征,结合随机森林变量选择、重要性度量和面向对象的分割、分类、基于规则集的分类后处理方法进行建筑物的识别与分类。精度评价结果表明,住宅/工业建筑物可以成功提取,两个区块的总体精度均达到90%以上,总体Kappa超过0.88,充分证明这一技术流程的可靠性和稳定性。2、工业和住宅是工业园区中的两类主体建筑,他们在空间中的分布具有明显差异：首先,住宅建筑单个图斑面积较小,绿化较多,在同样大小地块范围内,住宅建筑通常包含更多的空隙成分。其次,住宅建筑通常以小区为单位整体开发,在空间中的排列更加规则整齐,而工业建筑则杂乱无章。以上两种性质可以通过以建筑物为基础的二值化空隙度算法定量,本文对这一部分的内容做了理论和试验两方面探索。通过RGB影像与Gabor小波、共生矩阵纹理结合从高分辨率影像中提取出植被掩膜层,用其掩膜掉nDSM分割对象中的植被信息,得到建筑物。经过简单编辑,将建筑物提取结果二值化,从而提取出反映建筑物空隙含量和几何结构的空隙度特征并用于建筑类型的区分。对植被掩膜层的精度评价结果表明,两个区块总体分类精度均超过86%,总体Kappa超过0.7,说明光谱与纹理结合的大范围植被信息提取是成功的；对建筑物分类的分组试验结果表明,两个区块空隙度组的总体精度均超过了85%,显示出良好的总体识别率。Kappa分析结果表明,在建筑物类型的区分中,空隙度特征在95%显著水平上显著优于nDSM,证实了空间信息在建筑物类型识别中的重要价值。3、在建筑密度估算中,楼层高度的确定是影响其准确性的关键因素之一。传统的方法往往通过设置统一的层高值计算楼层数,少部分研究按照地域设置差异化的层高值。然而前者无可避免的会增大建筑密度估算误差,后者则不利于自动化的实现。考虑到影响层高的主要因素是建筑物类型,本文按照分类后的建筑类型设置相应的层高值,从而实现更加准确和自动化的建筑密度估算。4、针对不同楼层建筑面积不一致的复杂建筑物,提出了小尺度分割并以产生的分割单元为基本单位进行建筑密度的计算,有效降低了将整个建筑图斑作为整体计算时带来的高估误差。综合RMSE,相关分析和残差分析的建筑密度精度评价结果表明,总体误差在0.2以内,回归系数大于0.8,残差小于15%,结果可靠具有重复性。更多还原

【Abstract】 The waste and low efficiency of industrial land-use has become a critical factor of preventing social economic development and contradiction between land resource supply and demand. The extensive land-use type should be changed. The authenticity of data traditionally provided by land consuming companies becomes increasingly questioned under the background of enhancing monitoring and improving efficiency of land use. As the development of airborne remote sensing technology, the latest LiDAR sensor platform has been recognized as the most promising tools for building extraction in urban area. The property of fast, objective, real-time, low cost and labor saving make it extremely suitable for building density estimation in Economical and Technological Development Zone.So far, much effort has been made to the study of building extraction based on LiDAR data, but few attention has been paid to building type classification as well as the estimation of building density in Economic and technological Development Zone. This study focus on the methods of building type classification and building density estimation through two approaches:using LiDAR data alone and the combined use of LiDAR data and RGB imagery. The core points and conclusions are listed as follows:1LiDAR points are3D sampling of the earth surface, spatial autocorrelation features derived from this data can characterize the properties of autocorrelation and clustering of objects with high or low elevation. Various sources of features including geometric, autocorrelation and height heterogeneity were integrated on the recognition and classification of buildings through an object oriented approach and post classification technique based solely on LiDAR data. Variable selection and importance measure by random forest classifier were also participated in this process. Accuracy assessment indicated the successful extraction of industrial/residential buildings. Overall classification accuracy of the two blocks were both exceed90%with Kappa value more than0.88. The proposed method was demonstrated to be stable and reliable. 2In fact, residential and industrial buildings are the predominantly distributed anthropic structure types in Chinese urban region, and present obvious differences in spatial size and pattern which can be identified through building based lacunarity technique. First, the area of residential buildings is generally smaller than that of industrial buildings in Chinese ETDZs. In parcels of land of equal size, the number of buildings and the total area of gap are typically larger in residential areas. Second, residential buildings are often regularly distributed, whereas industrial buildings tend to be disorganized.Vegetation mask was generated firstly through classification by using a SVM classifier and an integration of Gabor, GLCM features and RGB imagery. Vegetation objects were then masked out with this vegetation mask. Buildings were thus left. A binary map containing building/non-building was created after some simple editing of the building layer. Lacunarity features that indicate percent and geometry of building gaps were extracted based on this binary map and then applied in building type classification. Vegetation mask was assessed to be high quality indicated by an overall accuracy of86%and a Kappa statistic of more than0.7. The integration of multispectral image and texture was illustrated to be successful for vegetation extraction. Overall classification accuracy of lacunarity group were both exceed85%. Kappa analysis result demonstrated the superiority of lacunarity to nDSM on building type classification in95%confidence level.3The floor height was one of the key factors to the accuracy of Building Density estimation. The number of floors were commonly determined by setting a unique threshold as floor height. This parameter was also defined with reference to the location of buildings in some few studies. Nevertheless, the former shall inevitably increase the residuals of building density while the latter suffers from human interaction. Because floor height is mostly correlated to building type, it was defined according to the building type in this study to estimate building density more accurately and automatically.4To those complicate structure buildings with multiple floor heights, we proposed a method that involves small scale segmentation and building density estimation based on those small segments. Overestimation error was thus significantly reduced when the whole building polygon was participated as basic unit for building density estimation.5The accuracy assessment on Building Density based on RMSE、Correlation Analysis and Residual Analysis illustrated a satisfactory result with the overall error below0.2, regression coefficient greater than0.8and residual error less than15%.更多还原