Study On Forest Individual-Tree Segmentation Method Based On Airborne LiDAR Data

As the basic unit of the forest,individual trees,and its spatial structure,biophysical and biochemical components are the key factors needed for forest resource investigation and ecological environment modeling.Airborne LiDAR(light detection and ranging)is a kind of top-down scanning mode to observe ground objects.It has a strong penetration ability to forest canopy and can accurately describe the three-dimensional structure characteristics of the forest.Therefore,it has incomparable advantages in forest single tree segmentation and single tree factor extraction.In the multi-layer forest with high canopy density,the existing individual-tree segmentation methods have obvious over-or under-segmentation error for the underlying trees due to the overlapping of the canopy and the occlusion of branches and leaves,resulting in low segmentation accuracy.To solve this problem,271 regularly distributed circular plots(0.04 hectares in size)of Robinson forest(a multi-layer forest with complex torrain and vegetation conditions)of the University of Kentucky were selected as experimental samples.This paper studies the canopy stratification method,the relationship model between LiDAR point density and individual-tree segmentation accuracy,and the machine learning method of trunk information extraction.By extracting effective information from point cloud data,the project completes the construction of individual-tree segmentation models and the extraction of single wood factors.The specific research contents are as follows:Given the low accuracy of the individual-tree segmentation algorithm based on the digital surface model(DSM)in the multi-layer forest with complex forest structure and terrain,a individual-tree segmentation method based on a hierarchical strategy was proposed.Firstly,according to the structural distribution characteristics of the tree crown in the vertical direction,the crown is divided into canopies at different heights,and then the individual-tree segmentation algorithm is implemented for the middle canopy and understory canopy by layer stripping method,to improve the detection rate of intermediate trees and overtopped trees.The experimental results show that:compared with the DSM method,the detection rate and segmentation accuracy of the proposed method are improved by 20.2%and 10%respectively.For all trees,the detection rate and overall segmentation accuracy are improved by 10.3%and 4.2%respectively.To reduce the over-and under-segmentation errors of understory trees,a individual-tree segmentation method based on a multi-resolution segmentation threshold(MRST)was proposed.Firstly,the algorithm pre-processes the point cloud and encodes the LiDAR points to each canopy according to the number of LiDAR echoes.The algorithm uses a cross-validation method to obtain the optimal segmentation threshold of each canopy,and then implements individual-tree segmentation using the optimal threshold for the canopy,and then combines the crown segments across layers by combining criteria.Finally,the algorithm separates single trees from the point cloud and extracts the parameters of the single tree.The algorithm uses the multi-threshold segmentation to reduce the under segmentation error and improve the recognition rate;uses the cross-layer combination of tree segments to reduce the over-segmentation error and improve the accuracy.The experimental results show that,compared with the DSM method,the MRST based segmentation method significantly improves the detection rate(from 52.3%to 73.4%)and the overall segmentation accuracy(from 65.2%to 76.9%).The total accuracy of all trees increased from 75.1%to 82.6%.This paper studies the response mechanism of the leaf-on data and leaf-off point cloud data to the performance of the individual-tree segmentation model and proposes a single tree segmentation method combining trunk detection and canopy segmentation.Firstly,by analyzing the vertical histogram of the leaf-off point cloud,the algorithm detects and separates the potential trunk point clusters,and then uses the density-based spatial clustering of applications with noise(DBSCAN)method to cluster the trunk clusters and extract the trunk information.At the same time,the algorithm implements canopy segmentation based on the MRST method for leaf-on point cloud data and extracts tree crown information.Finally,the extracted trunk information and the segmented crown segments are used as cross-references to separate individual trees and extract single tree parameters.The experimental results show that the method can detect 84.0%of single trees,90.7%of detected trees is correct,and the total segmentation accuracy score is 87.2%.At the same time,the estimation accuracy of single tree parameters is improved correspondingly:compared with the DSM based method and MRST method,the linear regression fitting degree R2 of tree-height is increased by 1.0 and 6.0 percentage points respectively;for the position of the tree,the average-difference is reduced by 11cm and 34cm respectively.