3D Reconstruction And Parameter Extraction Of Single Tree Based On Semantic Segmentation

Tree parameter is an important support for applications in ecosystem productivity simulation,carbon cycle research,and forestry management.Traditional manual measurement is time consuming,laborious,and destructive.The extraction method based on 2D optical remote sensing image can only extracts tree top surface parameter.Terrestrial Laser Scanning(TLS)provides a fast and eff-icient technique for extracting high-accuracy tree parameter.In order to improve the accuracy of tree parameter and reduce the computational complexity,it is necessary to segment the leaf and branch(semantic segmentation)and perform 3D reconstruction of the branch.In this study,three aspects of leaf and branch segmentation,branch 3D reconstruction and branch parameter extraction are studied.(1)Extending the single optimal scale segmentation method to the multi-optimal scale segmentation method.First,multiple optimal scales are selected for each point in many predefined candidate scales.Second,12 frequently-used 3D and 2D geometric features are calculated on these optimal scales.Finally,the extracted geometric features are used to train a Linear Discriminant Analysis(LDA)model,and then segments leaf and branch.Compared with the single optimal scale segmentation method,the multi-optimal scale segmentation method has higher accuracy.Compared with the stochastic multi-scale method,the multi-optimal scale method has better stability with fewer scales.(2)Improving the 3D reconstruction method of the branch.First,K-Nearest Neighbors(KNN)region growth is combined with density clustering to cut the branch TLS data into cylindrical data blocks.Second,the geometric centroid of these cylindrical data blocks are taken as the branch skeleton points.Next,constructing the adjacency map of each branch skeleton point to the 10 nearest neighbors and extracting the branch skeleton map by the Minimum Spanning Tree(MST)algorithm.Most importantly,the Principal Component Analysis(PCA)is used to transform a given cylindrical data block into another orthogonal space.The eigenvalues of the covariance matrix are used to calculate the branch radius.Finally,the 3D model of the branch is reconstructed by the cylindrical fitting method.The experimental results show that the improved 3D reconstruction method successfully reconstructs the 3D structure of the branch.Compared with the 3D model of the branch reconstructed by the SimpleTree software,the improved 3D reconstruction method is more anti-noise and accurate(especially twigs).(3)The tree height,DBH,branch volume and branch surface area of each tree are extracted based on the reconstructed 3D model of the branch.Compared with the measured values,the average error of the surface area of the branch extracted by our method is 28.72%,and the average error of other parameters are less than 8%.The average error of tree height and DBH extracted by the SimpleTree software are less than 6%,and the average error of extracted branch volume and branch surface area are as high as 246.35%and 185.24%,respectively.Compared with the values extracted by the SimpleTree software,the DBH,branch volume and branch surface area extracted by our method are more accurate.Data integrity and noise have a large impact on the accuracy of the parameter which extracted by the SimpleTree software and our method.