机器学习参与山区村落影像点云分类的研究
Study on image point cloud classification of mountain villages by machine learning
李霞 1杨正维 1黄俊伟 1杨亚复 1高莎2
作者信息
- 1. 云南省水利水电勘测设计研究院,昆明 650093,中国
- 2. 昆明理工大学国土资源工程学院,昆明 650500,中国
- 折叠
摘要
为了利用点云技术更好地获取地表信息,用无人机AA1300的内置光学镜头采集影像数据,构建2-D的数字正射影像地图(DOM),悬挂GS-1350N镜头采集3-D的激光雷达点云;通过k最近邻法(KNN)、支持向量机法(SVM)和随机森林法(RF)来实现DOM分类,用定量分析中精度高的方法分类3-D点云,并进行了 2-D和3-D的分类映射对比分析.结果表明,2-D的DOM分类中,相对于KNN和SVM,RF的kappa系数分别高3.74％和2.16％,全局精度分别高4.04％和2.88％;2-D的分类结果通过直接线性变换到3-D点云中,可实现2-D和3-D的点云分类,映射精度达94.15％;而在相同条件下,相对于2-D/3-D点云映射,直接3-D点云分类能更完整地呈现地物信息.3-D点云的精准分类对获取地表信息是有帮助的.
Abstract
In order to use point cloud technology to better obtain surface information,the built-in optical lens of unmanned aerial vehicle(UAV)AA1300 was used to collect image data and build a 2-D digital orthophoto map(DOM)and GS-1350N lens was hung to collect a 3-D light detection and ranging point cloud.DOM classification was realized by three methods,namely,the k-nearest neighbor(KNN)method,support vector machine(SVM)method,and random forest(RF)method.3-D point cloud was classified by the method with high accuracy in quantitative analysis.The comparative analysis of 2-D and 3-D classification mapping was carried out.The results show that,in 2-D DOM classification,kappa coefficients of RF are 3.74％and 2.16％higher,and the overall accuracy is 4.04％and 2.88％higher than those of KNN and SVM,respectively.The classification results of 2-D can be directly linearly transformed into 3-D point clouds,achieving 2-D and 3-D point cloud classification with a mapping accuracy of 94.15％.Under the same conditions,compared to 2-D/3-D point cloud mapping,direct 3-D point cloud classification can present more complete terrain information.This study indicates that the precise classification of 3-D point clouds can be helpful for better obtaining surface information.
关键词
激光技术/图像处理/机器学习/随机森林分类法/高原山区乡村Key words
laser technique/image processing/machine learning/random forest classification/highland mountain villages引用本文复制引用
基金项目
国家重点研发计划(2021YFC3000205-06)
出版年
2024
国家科技期刊平台
NSTL
万方数据