【雷达与对抗】【2017.03】半城市环境下谱激光雷达分析与地形分类

本文为美国海军研究生院（作者：Charles A. McIver）的硕士论文，共181页。

对海军研究生院校园进行了遥感分析，包括建筑物、不透水表面（沥青和混凝土）、天然地面和植被。数据来自Optech Titan扫描仪，以10-15点/平方米的速度提供三波长激光数据（532、1064和1550纳米）。激光扫描仪（LiDAR）数据的分析技术传统上只使用X、Y、Z坐标信息。传统方法用于将分类过程初始化为广泛的空间类别（未分类、地面、植被和建筑物）。谱分析为分类过程提供了一种独特的方法。

设计用于多光谱图像的工具和技术以适于本文中的激光雷达分析。ENVI的N维可视化工具被用来开发监督分类技术的训练集，主要是最大似然方法。对组合空间/谱数据的无监督分类使用K均值分类器进行比较。校园被划分为10 - 16个类别，相比之下，传统的方法有4个。增加光谱成分可以改善不透水表面、其他地面元素和建筑材料之间的区别。最大似然度显示总体分类准确度为75%，包括草地（99.9%）、草皮（95%）、沥青瓦（94%）、轻质建筑混凝土（89%）、沙子（88%）、灌木（85%）、沥青（84%）、树木（80%）和粘土瓦（77%）。通过“返回数”进行后处理过滤，总体精度可以提高到82%。

Remote-sensing analysis is conducted forthe Naval Postgraduate School campus, containing buildings, impervious surfaces(asphalt and concrete), natural ground, and vegetation. Data are from the OptechTitan scanner, providing three-wavelength laser data (532, 1064, and 1550 nm)at 10-15 points/m2 . Analysis techniques for laser-scanner (LiDAR) datatraditionally use only x, y, z coordinate information. The traditional approachis used to initialize the classification process into broad-spatial classes(unclassified, ground, vegetation, and buildings). Spectral analysiscontributes a unique approach to the classification process. Tools andtechniques designed for multispectral imagery are adapted to the LiDAR analysisherein. ENVI’s N-Dimensional Visualizer is employed to develop training setsfor supervised classification techniques, primarily Maximum Likelihood.Unsupervised classification for the combined spatial/spectral data isaccomplished using a K-means classifier for comparison. The campus isclassified into 10-16 classes, compared to the four from traditional methods.Addition of the spectral component improves the discrimination among impervioussurfaces, other ground elements, and building materials. Maximum Likelihooddemonstrates 75% overall classification accuracy, with grass (99.9%), turf(95%), asphalt shingles (94%), light-building concrete (89%), sand (88%),shrubs (85%), asphalt (84%), trees (80%), and clay-tile shingles (77%). Post-processfiltering by ‘number of returns’ increases overall accuracy to 82%.

1 引言
2 星载激光雷达系统的历史与文献综述
3 数据集与准备工作
4 处理方法
5 分类分析与结果
6 结论
附录 基础数据处理：NASA数据集衍生图、点云和数字高程模型

更多精彩文章请关注公众号：