Research On Typical Tree Species Classification In Laser Point Cloud Based On Deep Learning

Accurate tree species classification is important basic of managing forest resource and protecting ecological system.Remote sensing technology can large-scale and regularly investigate forest resource.Light Detection and Ranging(Li DAR)is an active remote sensing technology.Li DAR can fast and accurate acquisition of 3-dimensional structure information of single tree in forest,which is the significance for increasing accuracy of tree species classification in forest.In the past years,with the computer hardware improved,it has driven what the deep learning technology is from theoretical design to practical application.The point cloud acquired by Li DAR has the characterize of randomness and rotation invariance.At present,there are three ideas for the classification of tree species of forest point cloud based on deep learning,converting point cloud to multi-view images,converting point cloud to voxels and directly using point cloud.This paper focuses on the classification method of terrestrial and Unmanned Aerial Vehicle(UAV)Li DAR point clouds based on point cloud deep learning for typical tree species in the north.To improve the robustness of the model,two types of trees,white birch and larch,from two forest sample plots(natural forest and plantation)with large variability(growth environment and tree structure)were selected to conduct single tree extraction based on point cloud and canopy height model.The complete individual tree point cloud was selected by the location of single tree,file size of point cloud,manual visual interpretation to construct the dataset in this paper.The point cloud tree classification is achieved by training the deep learning tree classification model,and the analysis of the results is studied in detail.The experimental results show that the point cloud tree species classification method proposed in this paper can achieve high-precision tree species classification for both terrestrial and UAV Li DAR acquired forest point clouds,providing a new idea for forest point cloud tree species classification.The main conclusion of this paper are follows:(1)A method based on point density analysis is proposed to extract individual tree diameter at breast height(DBH)from terrestrial Li DAR point cloud.The entire forest point cloud of normalization height is sliced.The section point cloud is removed outliers by improving K-means algorithm.Then,the section point cloud is conducted point density analysis.The central point of tree trunk is extracted.The DBH can be automatically extracted according to the relationship between the center points of tree trunk sections and Euclidean distance,which reducing the workload of DBH extraction.(2)A tree species classification network model based on point cloud deep learning is proposed.The network uses geometric sampling and improved farthest point sampling for downsampling to preserve more feature points.The network can extract the coordinate information,normal vector and intensity of feature points simultaneously,and up-dimension the spatial features(coordinates and normal vector)and attribute features(intensity)respectively to extract their deep features.The tree species classification model is trained to achieve high accuracy classification of tree species.(3)This paper also conducts experiments on machine learning(support vector machines and random forests)and other deep learning(Res Net,Vox Net,Point Net family)tree classification methods.When the quantity of sample is same,the experimental result shows that the tree species classification average accuracy of this paper method is higher 20.3% and6.6%,respectively.The experimental results show that the number of feature points and the highest mapping dimension have a great influence on the classification accuracy of tree species for both terrestrial and UAV Li DAR acquired single tree point clouds,and the highest classification accuracy of tree species can reach up to 96%.