Study On TLS Point Cloud Classification Of Forest By Using Geometric Features

Terrestrial Laser Scanning can obtain the geometric data with high precision and high levels and it has been applied to the research of forest parameters extracting?stem extracting?Individual tree segmentation and modeling,biomass estimating.Labeling the TLS point cloud to ground?stem and leaf can be help to all above research,so it is meaningful to study on forest point cloud classification.We design classification algorithms by calculating fixed nearby features?dynamic nearby features and multi levels segmentation features based on feature selection.The fixed-neighbor feature is obtained by calculating 160 neighboring point features point by point;considering that the three-dimensional structure of the point set with small feature entropy of the local point cloud is relatively stable,this study searches the neighbors to calculate the feature entropy in the set number of neighbor search to determine the optimal neighbor number;in the smaller voxel space,the point clouds most belong to the same category,the small scale can separate the point cloud with close distribution in the space,and the large scale can divide the connected domain to the same voxel,this study introduces multi-scale segmentation to calculate point features,which can effectively improve the computational efficiency of features.The point cloud classification tasks are all based on feature selection.So we use feature selection technique by xgboost to select 6 features which is good for classifier's performance from 19 features.The test precision of classification based on nearby features?dynamic nearby features and multi levels segmentation features are 95.51%,95.19%and 95.94%.it shows our methods are suitable for TLS point cloud classification.