基于特征选择的TLS蒙古栎人工林点云分类研究  

Study on TLS point cloud classification of Quercus mongolica forest by feature selection

作　　者：邢涛[1] 汪献义 邢艳秋[1] XING Tao;WANG Xianyi;XING Yanqiu(Center for Research Institute of Forest Operations and Environment,Northeast Forestry University,Harbin 150040,Heilongjiang,China)

机构地区：[1]东北林业大学森林作业与环境研究中心

出　　处：《中南林业科技大学学报》2020年第3期1-7,共7页Journal of Central South University of Forestry & Technology

基　　金：国家重点研发计划项目（2017YFD060090402）

摘　　要：【目的】针对点云分类过程中多依据经验盲目构造特征的问题,本研究提出使用基于xgboost的特征选择弥补上述不足。【方法】本研究的数据为地面激光雷达扫描获得的蒙古栎人工林数据。本研究考虑构造适当的特征训练分类器将TLS点云快速分离为地面、树干与枝叶3个类别。在分类过程中,先在训练集中逐点搜索100个近邻构造19个特征,然后使用这些特征训练xgboost分类器,并依据控制分类器节点分裂的特征频率获得特征重要性。获得特征重要性之后将特征按重要性做降序排列,并依据该序列依次增加特征数量训练xgboost。因为构造了19个特征,所以在上述训练分类器的过程中可获得19个关于特征重要性的分类器模型。依次将上述模型应用于测试集的分类,在保证分类器性能的情况下,依据测试集的表现选择了前6个特征,从而实现了TLS点云分类的特征选择。【结果】使用基于特征选择获得的6个特征与依据经验构造的19个特征训练分类器的测试准确率分别为0.954 8与0.956 2。相较于使用19个特征,使用6个特征的分类器性能仅降低了0.001 4。在训练集与测试集中计算6个特征用时分别占计算19个特征用时的53.13%与54.33%。【结论】结果表明特征选择策略可有效提高特征计算效率,而且在保证分类器性能的前提下可以避免特征构造的盲目性。【Objective】Most researchers designing features in point cloud classification are based on experience.The xgboost feature selection technique was introduced to solve the problem.【Method】The data in our research is Quercus mongolica artificial forest point cloud scanned by terrestrial laser scanning.We want to design some features to train classifier which can be used to classify the TLS point cloud into ground,stem and leaf.In the classifying process,we first searched 100 nearby points of every point in train sets to calculate 19 experienced features,and they were all used to train xgboost.After that,we could get the features importance through the node split frequency in xgboost.When we gotten the features importance,we rank the them into descending order.And then,we train the xgboost again by adding the features from the rank one by one.Because we designed 19 features,we could get 19 classifiers relevant to feature importance.In the last,the feature selection was done.On the basis of ensuring the accuracy of point cloud classification,we used the 19 classifiers to label the test point cloud and we chosen 6 top features according to classifiers performance.【Result】The test accuracy were 0.9548 and 0.9562 between the classifiers trained by selected 6 features and original 19 features.When compared to classifier trained by 19 features,the performance of classifier trained by 6 features descended 0.0014.The computing time’s ratios between 6 features and 19 features in train sets and test sets were 53.13%and 54.33%.【Conclusion】The result shows that feature selection can increase feature computing efficiency,and on the basis of ensuring classifier performance,feature selection can avoid the blindness of designing features.

关 键 词：地面激光雷达 特征选择 点云分类 蒙古栎 

分 类 号：S771.8[农业科学—森林工程]