测绘学报 ›› 2023, Vol. 52 ›› Issue (3): 419-431.doi: 10.11947/j.AGCS.2023.20210153

• 摄影测量学与遥感 • 上一篇    下一篇

多维GMM与邻域约束的多光谱机载LiDAR数据城市地物分类

王丽英1, 吴际2, 有泽1, 李玉1, CAMARA Mahamadou1   

  1. 1. 辽宁工程技术大学测绘与地理科学学院, 辽宁 阜新 123000;
    2. 黑龙江地理信息工程院, 黑龙江 哈尔滨 150081
  • 收稿日期:2021-03-25 修回日期:2022-09-25 发布日期:2023-04-07
  • 作者简介:王丽英(1982-),女,博士,教授,博士生导师,研究方向为激光雷达数据处理及应用。E-mail:wangliyinglntu@163.com
  • 基金资助:
    国家自然科学基金(42071343; 42201482)

Urban object classification of multispectral airborne LiDAR data with multidimensional Gauss mixture model and neighborhood constraints

WANG Liying1, WU Ji2, YOU Ze1, LI Yu1, CAMARA Mahamadou1   

  1. 1. School of Geomatics, Liaoning Technical University, Fuxin 123000, China;
    2. Heilongjiang Institute of Geomatics Engineering, Harbin 150081, China
  • Received:2021-03-25 Revised:2022-09-25 Published:2023-04-07
  • Supported by:
    The National Natural Science Foundation of China (Nos. 42071343;42201482)

摘要: 如何充分利用空间位置及多光谱信息完整、准确地区分各类地物是多光谱机载激光雷达点云应用的重要前提。传统的基于点的分类算法受点云无法明晰表达拓扑及邻域信息的局限导致算法设计困难、执行效率低,而将点云插值为图像的分类算法则受图像存在信息及精度损失的局限导致分类精度较低。另外,点云和图像两种结构均无法直接表达地物的三维几何形体,不利于地物三维建模及分析。为此,本文提出了一种多维高斯混合模型(Gauss mixture model,GMM)与邻域约束的多光谱机载LiDAR城区地物分类算法。该算法首先以无损且明晰表达邻域信息为原则将多光谱LiDAR数据体素化为多值虚拟体素结构,其中,虚拟体素为体素与其内激光点的联合体,虚拟体素值是由体素内激光点的多波段光谱、高程、局部法向量分布及点密度等特征构成的特征矢量。然后,构建模糊聚类模型对多值虚拟体素结构进行分割,获得各虚拟体素的模糊隶属度矩阵。其中,特征空间地物呈现的多峰分布用多维GMM拟合,从而建立标号场,并将多维GMM的概率分布作为非相似性测度;标号场中相邻体素类别的空间相关性用隐马尔可夫随机场模型建模,从而建立邻域约束下的先验概率,并将其作为控制聚类尺度的参数;采用附有规则化项的目标函数解决聚类尺度敏感问题。最后,对隶属度矩阵进行反模糊化确定分类结果。采用Optech Titan实测的不同场景的、不同数据量的多光谱机载LiDAR数据定量评价本文算法的有效性和可行性。试验结果表明,本文算法的平均总体精度可达91.32%、Kappa系数可达0.872,可有效实现对城市各类地物的分类;本文算法综合利用了地物的辐射、空间及几何一致性信息,扩大了信息利用种类,为多光谱机载LiDAR数据的空间位置及多光谱信息的综合利用提供了可行方案。

关键词: 多光谱机载激光雷达, 虚拟体素, 体素模型, 隐马尔可夫随机场, 多维高斯混合模型, 模糊聚类

Abstract: Making full use of the spatial location and multispectral information of multispectral airborne LiDAR point cloud to completely and accurately segment objects is an important prerequisite for the subsequent application of the data.However, the traditional point-based classification algorithms are difficult to design and inefficient to execute due to the limitations of point cloud that cannot clearly represent topology and neighborhood information.Moreover, the classification algorithms which interpolate point cloud into image is limited by the information and precision loss of image, which leads to low classification accuracy. In addition, neither point cloud nor image can directly represent the three-dimensional geometry of objects, which is not conducive to the subsequent three-dimensional modeling and analysis of objects.Therefore, this paper proposes a classification algorithm based on multi-dimensional Gauss mixture model (GMM) and neighborhood constraints for multispectral airborne LiDAR data. The multispectral airborne LiDAR data are firstly voxelized into a multi-valued virtual voxel structure based on the principle of no information loss and representing neighborhood information explicitly, in which the virtual voxel is the combination of the voxel and its internal laser points, and the virtual voxel value is a feature vector composed of the spectrum, elevation, local normal vector distribution and point density features of the laser points in the voxel.Then, a fuzzy clustering segmentation model is constructed to segment the virtual voxel structure, and the fuzzy membership matrix of each virtual voxel is obtained.Among them, the multi-peak distribution of objects in feature space is fitted by multi-dimensional GMM to establish the label field, and the probability distribution of multi-dimensional GMM is used as the dissimilarity metric.The spatial correlation of classes for the adjacent voxels in the label field is modeled by the hidden Markov random field (HMARF) model, and the prior probability under the neighborhood constraint is established, which is used as the parameter to control the clustering scale.An objective function with a regularization term is used to solve the scaling sensitivity problem of clustering.Finally, the membership matrix is defuzzified to determine the classification result.The effectiveness and feasibility of the proposed algorithm are quantitatively evaluated based on Optech Titan multispectral airborne LiDAR data of different complexity and different data volume.The experimental results show that the average overall accuracy and Kappa coefficient of the algorithm is 91.32% and 0.872, respectively, which can effectively realize the classification of various urban objects.The proposed algorithm makes comprehensive use of the spectral, geometric and neighborhood consistency information of objects, expands the types of information utilization, and provides a feasible scheme for the comprehensive utilization of the spatial location and multispectral information of multispectral airborne LiDAR data.

Key words: multispectral airborne LiDAR, virtual voxel, voxel model, hidden Markov random field, multi-dimensional Gaussian mixture model, fuzzy clustering

中图分类号: