测绘学报 ›› 2018, Vol. 47 ›› Issue (4): 508-518.doi: 10.11947/j.AGCS.2018.20170417

• 摄影测量学与遥感 • 上一篇    下一篇

机载LiDAR点云数据降维与分类的随机森林方法

熊艳1, 高仁强2, 徐战亚1   

  1. 1. 中国地质大学(武汉)信息工程学院, 湖北 武汉 430074;
    2. 北京大学遥感与地理信息系统研究所, 北京 100871
  • 收稿日期:2017-07-21 修回日期:2017-11-28 出版日期:2018-04-20 发布日期:2018-05-02
  • 通讯作者: 徐战亚 E-mail:zhanyaxu@163.com
  • 作者简介:熊艳(1992-),女,硕士生,研究方向为点云数据处理、数据可视化。E-mail:1453258599@qq.com

Random Forest Method for Dimension Reduction and Point Cloud Classification Based on Airborne LiDAR

XIONG Yan1, GAO Renqiang2, XU Zhanya1   

  1. 1. Faculty of Information Engineering, China University of Geosciences(Wuhan), Wuhan 430074, China;
    2. Institute of Remote Sensing and GIS, Peking University, Beijing 100871, China
  • Received:2017-07-21 Revised:2017-11-28 Online:2018-04-20 Published:2018-05-02

摘要: 探索自动化的激光点云分类方法对于三维建模、城市土地分类、DEM制图等应用具有重要作用。考虑到现有的点云分类算法在提取依赖邻域结构的特征参数时面临邻域尺度的选择难、数据维度高、计算复杂,并且缺乏对分类特征参数的重要性评估和选择等问题,本文提出了基于随机森林的机载LiDAR点云数据降维与分类方法。在分析点云数据的高程、回波、强度等属性特征的基础上,提取归一化高度、高度统计量、表面特征、空间分布特征、回波特征及强度特征6大类特征参数,并构建多尺度特征参数,运用随机森林的特征选择算法对分类特征集进行优化,然后进行点云分类。试验结果表明,基于随机森林的特征选择方法可以有效地降低特征维度,并且使得总体分类精度达到94.3%(Kappa系数为0.922),相比于使用全部特征分类和SVM分类方法而言,该方法的总体分类精度均有一定程度的提高;特征的重要性度量结果表明,归一化高度特征在点云分类中所起的作用最大。

关键词: 机载激光雷达, 特征选择, 点云分类, 随机森林

Abstract: Exploring automatic point cloud classification method is of great importance to 3D modeling,city land classification,DEM mapping and etc.To overcome the problem that extracting geometric feature for point cloud classification involved neighbor structure meets the challenge that the optimal neighbor scale parameter,high data dimension and complex computation,lacking efficient feature importance analysis and feature selection strategy,this paper proposed a point cloud classification and dimension reduction method based on random forest.After analyzing the characteristic of elevation,intensity and echo of laser points,this paper extracted a total of 6 feature types like normalized height feature,height statistic feature,surface metric feature,spatial distribution feature,echo feature,intensity feature,then built a multi-scale feature parameter from them.Finally,a supervised classification was conducted using a random forest algorithm to optimal the feature set and choose the best feature set to classify the point cloud.Results indicate that,the overall accuracy of the proposed method is 94.3% (Kappa coefficient is 0.922).The proposed method got an improvement in the overall accuracy when compared with no feature selection strategy and SVM classification strategy; The feature importance analysis indicates that the normalized height is the most important feature for the classification.

Key words: LiDAR, feature selection, point cloud classification, random forest

中图分类号: