A forest vertical stratification technology based on near-ground 3D point clouds

Abstract

This invention provides a forest vertical stratification technique based on near-ground 3D point clouds, belonging to the research field of forest vertical structure stratification. The steps are as follows: data identification; ALS point cloud data height normalization; seed point filtering based on ALS data point density (volume filter); seed point filtering based on ALS point spatial distribution characteristics; ALS canopy segmentation; a two-level multi-scale voxel partitioning strategy based on ALS data; determination of stratification height within secondary voxels; TLS data height normalization and tree seed point extraction; trunk point tracking based on TLS data; initial canopy segmentation based on TLS data; overlapping canopy segmentation based on TLS data; and understory vegetation extraction based on TLS data. Compared with traditional methods, this invention considers different point cloud data characteristics and adopts different forest vertical structure stratification methods for different data types, achieving the segmentation of overlapping canopies and the separation and extraction of trees and understory vegetation in TLS data. Furthermore, it combines ALS data features and understory vegetation spatial variation characteristics to achieve the extraction of trees and understory vegetation from ALS data.

Description

I. Technical Field

[0001] This invention relates to a forest vertical stratification technology based on near-ground 3D point clouds. Specifically, it refers to an improved method for vertical forest structure stratification that identifies point cloud data types by identifying point cloud features and utilizes different types of point cloud data to account for spatial variations in overlapping canopies and understory vegetation (the process is attached). Figure 1 (As shown). II. Background Technology

[0002] Most forests can be vertically divided into two layers: the forest canopy composed of trees and the understory vegetation consisting of shrubs, herbs, and mosses. The forest canopy and understory differ significantly in species composition, structure, and function (such as photosynthesis). However, traditional optical remote sensing or lidar sensors mounted on different platforms acquire mixed signals that simultaneously include both trees and understory. Directly estimating forest structure parameters based on these mixed signals leads to significant discrepancies in the inversion results, such as leaf area index, canopy cover, and aboveground biomass. Therefore, quantitatively identifying and separating the forest canopy and understory vegetation is necessary and meaningful before using active or passive remote sensing data to invert forest structure parameters.

[0003] Traditional forest stratification methods primarily rely on visual identification of tree species and manual measurement of tree height and canopy width in the field. However, these field-based methods are time-consuming, labor-intensive, and limited by weather conditions, thus hindering their application to large spatial areas. The advent of remote sensing technology has made it possible to identify and separate forest canopy and understory vegetation with higher accuracy on a larger spatial scale. Forest stratification studies based on traditional optical remote sensing mainly rely on the inversion of forest background reflectivity from multi-angle remote sensing data. For example, in 2014, Jiao et al. extracted the seasonal background reflectivity of global forest regions in their article "Mapping global seasonal forestbackground reflectivity with Multi-angle Imaging Spectroradiometer data" published in Volume 119 of the *Journal of Geophysical Research: Biogeosciences*. In their 2017 paper, "Separating overstory and understory leafarea indices for global needleleaf and deciduous broadleaf forests by fusion of MODIS and MISR data," Liu et al. estimated the forest canopy and understory leafarea reflectance (LAI) based on Jiao's forest background reflectance results. However, separating forest canopy and understory reflectance from multi-angle remote sensing data suffers from limitations in acquiring forest vertical structure information, requires multiple satellite observation angles, and is susceptible to topographic and weather influences. The emergence of lidar technology has provided new ideas and methods for forest stratification research. Currently, many researchers use terrestrial laser scanning (TLS) and airborne laser scanning (ALS) to acquire forest structure information at the plot and landscape scales, respectively. TLS, based on the ground, acquires high-precision laser point clouds (millimeter-level), while ALS, primarily using manned aircraft, unmanned aerial vehicles, and other aerial vehicles, acquires data with lower point density and accuracy than TLS data. Therefore, TLS data and ALS data have different data characteristics such as point density and coverage area. However, no researchers have accurately identified the point cloud data type and combined it with point cloud features to extract forest structure parameters. Directly extracting forest structure parameters without considering the point cloud data type will result in significant errors in the extracted parameters.Therefore, based on the identified point cloud data type and corresponding point cloud features, processing algorithms for different point cloud types can be implemented, thereby improving the accuracy of point cloud data processing.

[0004] Based on the different lidar sensor mounting platforms, forest stratification methods based on lidar data can be divided into two categories: the first category is forest stratification based on TLS data, and the second category is stratification based on ALS data. (1) Forest stratification based on TLS data. Forest stratification based on TLS data mainly uses the single tree segmentation algorithm to extract the height information of a single tree to achieve vertical stratification of the forest structure. For example, Tao et al. in 2015 and Xi et al. in 2020 extracted point clouds of a single tree based on TLS data. However, they ignored the influence of understory vegetation in TLS data on segmentation, and few researchers considered the existence of understory vegetation to estimate the forest leaf area index based on TLS point clouds. In addition, the accurate segmentation of overlapping canopies affects the accuracy of forest aboveground biomass and canopy volume estimation. How to better segment canopies with overlapping areas is still an unsolved problem. (2) Forest stratification based on ALS data. Segmenting canopies with overlapping areas using ALS data is still a challenge. Overlapping canopies have a significant impact on the accurate estimation of canopy width, canopy volume and biomass. In addition, there are some problems with forest stratification. Most methods use a fixed stratification height threshold for any area of ​​study. However, due to differences in topography, canopy closure, tree species, and light conditions, the stratification height will vary. Therefore, it is unreasonable to use a fixed height stratification method within the study area. III. Summary of the Invention

[0005] The purpose of this invention is:

[0006] Based on the density characteristics of the input point cloud data, the data type is identified, and different forest vertical structure stratification algorithms are used to extract point cloud data of trees and understory vegetation for different types of point cloud data. For TLS data, the influence of understory vegetation and overlapping canopies is taken into account when performing forest vertical structure stratification. For ALS data, the spatial variation characteristics of overlapping canopies and understory vegetation are taken into account when performing forest vertical structure stratification.

[0007] The technical solution of this invention mainly includes the following steps:

[0008] (1) Input data recognition. Use the KD-tree data structure to organize the input point cloud data. Set the neighborhood search radius of the KD-tree algorithm to 0.1 meters, and count the number of points in the neighborhood of each point, as well as its average value mean and standard deviation std. Set the average value threshold mean_thd and standard deviation threshold std_thd for the number of points in the neighborhood. If mean < mean_thd and std < std_thd, it is determined that the current input data is airborne lidar (ALS) point cloud data; otherwise, it is terrestrial lidar (TLS) point cloud data. According to the recognized type of point cloud data, different forest vertical structure stratification algorithms are used to separate and extract the point clouds of trees and understory vegetation. If the recognized input data type is ALS data, execute steps (2) to (7). If the recognized input data type is TLS data, execute steps (8) to (12).

[0009] (2) ALS point cloud data height normalization. If the detection result in step (2) is ALS point cloud, use the cloth simulation filtering (CSF) algorithm to filter out the ground points from the original point cloud, and perform interpolation on the filtered ground points and generate DTM based on the ordinary kriging (OK) interpolation algorithm. Subtract the DTM from the z coordinate value of the point cloud to obtain the relative height of the point cloud, that is, the height normalization of the ALS point cloud data.

[0010] (3) Seed point filtering based on the point density (volume filter) of ALS data. First, use a local search sphere to find the local maximum, and set the radius r to 3 - 5 times the nearest point distance (NPD). Second, determine the reference value of the volume filter based on the three-dimensional voxel. The length, width, and height of the voxel are set to be more than twice the NPD to include at least one point.

[0011]

[0012] Among them, the length, width, and height of a single voxel are all l, point_num is the total number of points in the point cloud; voxel_num is the number of non-empty voxels, and volume_density is the number of points per unit volume, that is, the point density. v_threshold is the upper threshold of the volume filter. Therefore, use v_threshold and a lower threshold (usually set to 1 or 2) to filter out some false seed points.

[0013] (4) Seed point filtering based on the spatial distribution characteristics of ALS points. Construct a vector from the centroid of the neighborhood points to the local highest point Rotate the point cloud from the direction to the positive direction of the Z axis The rotation axis is and The cross product of, and the calculation formula for the rotation angle θ is:

[0014]

[0015] The minimum and maximum height difference of the rotated point cloud is considered as the branch length. The maximum horizontal distance from the rotated point cloud to the centroid is taken as the point cloud width. A shape ratio threshold t is set; if the aspect ratio of the rotated point cloud is greater than t, the current seed point is determined to be a branch vertex; otherwise, it is a true seed point.

[0016] (5) ALS Canopy Segmentation. Starting from the seed point, sort the points by height from highest to lowest. Calculate the distance between each point and the nearest point in the tree to be segmented, and the nearest point outside the current tree. Assign the point to be segmented to the nearest tree. For example, if the distance to the nearest point in the tree is less than the distance to the nearest point outside the current tree, then the point to be segmented is assigned to a tree, thus obtaining the initial result of canopy segmentation.

[0017] Fine-grained canopy segmentation. The initial segmentation results of individual trees are projected onto a horizontal plane, and their projected area and perimeter are calculated using a 2D convex hull algorithm. The ratio of area to perimeter is used to identify whether the canopies overlap. For trees with overlapping canopies, polynomial fitting (PFT) is used to detect the presence of understory vegetation in the initial segmentation results. In PFT, a height-point curve is fitted using a ninth-order polynomial. If the first derivative is zero and the second derivative is greater than zero, it is identified as a trough point, and the elevation corresponding to the trough point is V0. If V0 is not detected in the curve, it indicates that the overlapping canopies are caused by trees. and These are two local maxima of the two trees, which will wrap around the point cloud. Rotate until the line connecting the two local maxima is parallel to the x-axis, and slice along the y-axis at fixed intervals. Detect the minimum point in each slice using the PFT technique. For two minimum points P1 and P2 on adjacent slices detected by PFT, we construct a vector from P2 to P1, which is then compared with the projected vector... ( Together they construct plane α (on a plane) Figure 2 a) The normal vector of plane α It can be determined by the following formula:

[0018]

[0019] in, It is parallel to The vector, This is the normal vector of plane α. Similarly, more planes can be determined through the remaining connection points. Multiple planes can separate overlapping canopies. If PFT detects V0, it indicates that canopy overlap is caused by understory vegetation. We created a slice with a thickness greater than NPD at the V0 height and applied the DBSCAN algorithm based on density noise to cluster the points. Clusters with fewer points in the clustering results were considered understory vegetation points. Based on the centroids B1 and B2 of the clusters, we constructed a cluster boundary point and... The dividing plane is the normal vector, and overlapping tree crowns are separated based on this dividing plane.

[0020] (6) A two-level, multi-scale voxel partitioning strategy based on ALS data. The first level of partitioning divides the entire study area into voxels of only one level, i.e., it is divided into multiple plots (e.g., Figure 2 In plots D1, D2, D3, and D4 of layer b, the voxel dimensions are: length: l1, width: l1, height: the global maximum z-value. The reference value for l1 is set to a commonly used forest plot size, for example, ranging from 30 meters to 200 meters. For the second layer, considering the spatial variation in the proportion of trees and understory vegetation, the plots obtained in the first layer are further subdivided into smaller voxels of different sizes. Each plot has one layer of voxels with dimensions: length: l2, width: l2, height: the maximum z-value in the plot. For example, D3 and D4 are subdivided again. To determine the reference value for l2, we calculate the standard deviation of tree height (SDTH) based on the tree subdivision results in each plot (D1, D2, D3, and D4). A higher SDTH indicates a more complex forest structure. After trial and error, the value of l2 in this study is determined by an equation.

[0021]

[0022] Each voxel obtained in the second layer partitioning is used to determine the layer height. To avoid the influence of the tree crown being segmented by the voxel edges, trees whose centroids fall within the voxels in the horizontal direction are collected as the inlier set of that voxel.

[0023] (7) Determination of stratification height within secondary voxels. All points within a secondary voxel are divided into height intervals with a gap greater than NPD, and the number of points in each height interval is recorded to generate a height-point count histogram. PFT technology is used to identify troughs in the curve; multiple troughs in the curve indicate multiple vegetation layers, so the maximum height of the trough is considered the stratification height. For secondary voxels where no trough is detected, the stratification height within the surrounding voxels is used as its stratification height. The stratification height in each voxel is repeatedly calculated, and the stratification heights at different locations within the entire area are combined. The stratification surface can be obtained by adding different stratification heights to the ground in all voxels. Finally, if the tree height is lower than the stratification surface, it is determined to be understory vegetation; otherwise, it is a tree.

[0024] (8) TLS Data Height Normalization and Tree Seed Point Extraction. Initial ground points were filtered using the Fabric Simulation Filter (CSF) algorithm, and more accurate ground points were extracted from the filtered results based on the Local Minimum Method. Ground points were interpolated using the ordinary kriging method provided by ArcGIS software to generate a Digital Terrain Model (DTM) with a resolution of 0.5 meters. Height-normalized points were obtained by subtracting the corresponding terrain elevation from the z-values ​​of the original point cloud in the DTM model, and ground points were deleted to obtain height-normalized vegetation points. A sixth-order polynomial curve function was fitted to the vertical profile of the point density in each height box to determine the layered surface. The sixth-order polynomial curve function used was:

[0025]

[0026] B is the intercept of the polynomial. A i (i = 1...6) represents the coefficients corresponding to xi, whose values ​​are determined by the least squares method. The initial segmentation height (H0) is defined as the position where the first derivative is zero and the second derivative is greater than zero. If multiple H0 values ​​are detected, values ​​near the treetop are considered invalid. Layered surfaces can be obtained by adding height H0 to the DTM. Tree trunk detection is performed based on the point cloud located between H0 and H0+0.2. Assuming the tree trunk is a cylindrical geometric object, the tree trunk can be calculated by fitting the data. A cylinder in 3D space can be expressed by the following formula:

[0027]

[0028] Where p i0 (x i0 y i0 ,z i0 (i = 1, 2, ..., Q) is the i-th geometric center point of the cylinder, and the central axis is... Q represents the total number of tree trunks detected within the height interval from H0 to H0+0.2m; r is the cross-sectional radius of the cylinder. The cylinder is detected within the extracted height interval using the RANdom SAmpleConsensus (RANSAC) algorithm and total least squares method. The points constituting the cylinder are marked as tree trunk points, and point p... i0 (x i0 ,y i0 ,z i0 This is used as the first seed point for each trunk.

[0029] (9) Trunk point tracing based on TLS data. Starting from the first seed point p i0 (x i0 ,y i0 ,z i0Starting from the center point, two search spheres with different radii are used to identify the trunk points in the upward and downward directions respectively (R2>R1 and R2>r). R1 ​​is defined as the radius of the inner search sphere used for trunk point identification. The value of R2 can be set with reference to the trunk diameter. R2 determines the thickness of the horizontal slice and the level of detail in the trunk curvature. For the upward direction, the value is calculated from z... i0 +R1 to z i0 The geometric center point p of the horizontal slice of +R2 i1 (x i1 ,y i1 ,z i1 (In the next step, the new center of the ball can replace p.) i0 Point). Then vector The growth direction of the tree trunk in local three-dimensional space ( Figure 3 a). Point p i1 (x i1 y i1 , z i1 ) will be the center point of the two new search balls and will be used to continuously identify trunk points. Simultaneously, we also calculated the maximum distance D between a point within the horizontal slice and its geometric center. i1_max If the maximum distance within a horizontal slice exceeds a reference threshold, it exceeds the r value. The upward search process typically stops at the bottom of the canopy. The same search sphere algorithm is also used with p... i0 (x i0 ,y i0 ,z i0 () represents the downward direction of the starting seed point. The height within the horizontal slice is calculated in the z-axis. i0 -R2 to z i0 - Obtain the geometric center point of all points within the R1 range, and repeat this process until the ground is reached. Finally, identify and extract the trunk points from the ground surface to the base of the tree canopy at the plot level and store them separately.

[0030] (10) Initial canopy segmentation based on TLS data. First, the remaining points in the forest plot are sorted in ascending order of height. The minimum two-dimensional distance between the point to be classified and the known point set is compared with the two-dimensional distance between the point to be classified and the point not belonging to the current tree. The point's affiliation is determined according to a preset threshold. This process is repeated until all points are classified. The combination of the identified trunk points and the trunk points is the preliminary segmentation result.

[0031] (11) Overlapping Canopy Segmentation Based on TLS Data. For overlapping areas of trees in the initial segmentation results, horizontal point cloud slices are first performed at height intervals greater than the nearest neighbor distance (NPD) (e.g., 0.3m-0.5m). The density-based noise spatial clustering algorithm (DBSCAN) is used to detect the number of classes in each horizontal slice. The DBSCAN algorithm is applied from two vertical directions (i.e., from bottom to top and from top to bottom) to identify all possible points of change in the number of points in the overlapping area.

[0032] Then, a new vector is constructed by connecting two adjacent points of change, and a dividing plane is generated by combining the vector determined by the geometric center points of the two point sets (e.g., ...). Figure 3 (as shown in b). For example, given two trees with overlapping regions, in the direction from bottom to top, for a horizontal slice of height i (i.e., the bottom of the overlapping region, see...) Figure 3 b) The two-point set generated by the DBSCAN algorithm can utilize their geometric center C. 1i and C 2i Generate a new vector By connecting the geometric center points of these two point sets, boundary points (P1) can be detected along this vector direction. At this height, the number of point sets generated by the DBSCAN algorithm changes from two to one. In this way, change points P1 and P3 can be detected by slicing layer by layer from the bottom to the top of the tree. When performing horizontal slicing from the top to the bottom of the tree, change points P2 and P4 can be identified using the DBSCAN algorithm. and A new vector can be obtained. The dividing plane is defined by the normal vector of this vector. Together with a point on the dividing plane, it is constructed ( Figure 3 c).

[0033]

[0034] A(x-x0)+B(y-y0)+C(z-z0)=0 (10)

[0035] Where (A, B, C) is the normal vector. Let P1 be the coordinates of point P1, and (x, y, z) be the coordinates of the plane. The coordinates of the previous point. Substitute the coordinate values ​​from the point cloud into formula (10) one by one, and compare the calculation result with 0. Points that satisfy the equation and are greater than or less than 0 are located on the two sides of the plane, while points equal to 0 are on the plane. In this way, the overlapping parts of the tree canopy in the forest can be separated from each other. A schematic diagram of tree segmentation can be found in [reference needed]. Figure 3 Furthermore, we can obtain continuous segmentation planes. and dividing plane Based on these three segmentation planes, we can effectively divide the overlapping area between two trees. For cases of overlapping canopies of multiple trees, we repeat the process designed for two canopies, processing each pair of canopies sequentially. After processing all possible combinations of multiple canopies and grouping points belonging to the same canopy, we can obtain the final canopy segmentation result of the trees.

[0036] (12) Understory vegetation extraction based on TLS data. A horizontal slice is set at 30% of the maximum height, and the DBSCAN algorithm is used to detect seed points for each understory canopy. Each obtained cluster is considered as an understory seed point. Once all seed points are found, we will follow a similar segmentation process as the tree canopy to assign the understory points to each canopy.

[0037] Compared with existing technologies, the advantages of this invention are as follows:

[0038] By leveraging the features of point cloud data to identify data types, and targeting the characteristics of different data types, this invention successfully achieves precise stratification of forest vertical structure. For airborne lidar data, it simultaneously considers the influence of understory vegetation and overlapping canopies to achieve effective stratification of individual trees and forest vertical structure. For terrestrial lidar data, this invention takes into account the spatial variation characteristics of overlapping canopies and understory vegetation, achieving even more accurate forest vertical structure stratification. This method, while ensuring accuracy, fully utilizes the characteristics of different lidar data, exhibiting stronger adaptability and universality.

[0039] This invention utilizes near-ground 3D point cloud data to perform forest vertical structure stratification based on point cloud data features. It provides a method for forest vertical structure stratification that considers data types, overlapping canopies, and the spatial distribution variations of understory vegetation. Firstly, this invention uses the density features of 3D point cloud data acquired from different platforms to identify point cloud data types. This method is simple and efficient, providing a foundation and prerequisite for subsequent targeted algorithm development that combines data features at a deeper level. For data identified as ground-based lidar point clouds, this invention uses statistical methods to remove understory interference based on the uneven vertical distribution of trees and understory vegetation. Furthermore, it uses the variation points formed between overlapping canopies to generate segmentation planes for overlapping canopies, which improves the accuracy of individual tree segmentation and lays the foundation for further forest vertical structure stratification in ground-based lidar point cloud data. This invention targets data identified as airborne lidar point clouds. It utilizes valley lines formed between tree canopies to segment overlapping areas and divides sample plots into voxels of different levels and scales. This enables the vertical stratification of forest structure by combining the distribution characteristics of understory vegetation. This not only improves the separation accuracy and speed of tree and understory vegetation point clouds but also lays the foundation for further forest structure parameter inversion that takes into account the spatial distribution characteristics of forests.

[0040] Practical applications demonstrate that this invention provides an effective method for forest vertical structure stratification, capable of accurately processing different types of point cloud data and applicable to forest vertical structure stratification of 3D point cloud data on various platforms. This method overcomes the low stratification accuracy caused by overlapping canopies and understory vegetation interference in traditional methods, avoiding the waste of time, manpower, and resources associated with manual observation. By combining point cloud data features with spatial distribution changes in understory vegetation, this invention significantly improves the accuracy and efficiency of forest vertical structure stratification, enhancing the application value and widespread adoption of lidar technology in forest resource surveys and vegetation ecology remote sensing. This not only helps improve the efficiency of forest resource management and monitoring but also provides more reliable technical support for research in vegetation ecology and environmental protection. IV. Description of the attached drawings

[0041] Figure 1 This is a schematic diagram of the process of the present invention.

[0042] Figure 2 Forest vertical structure hierarchies identified as ALS data type.

[0043] Figure 3 Forest vertical structure hierarchies identified as TLS data type. Figure 4 A schematic diagram illustrating the results of forest vertical structure stratification based on different types of point cloud data. V. Detailed Implementation Methods

[0044] The invention will be further explained by the following examples:

[0045] Point cloud data acquired from different platforms are placed in the same folder. The algorithm reads the point cloud files, automatically identifies the data type based on data characteristics, and implements a forest vertical structure layering, as shown in the attached diagram. Figure 4 As shown, Figure 4 'a' represents the hierarchical forest vertical structure of TLS data. Figure 4 b represents the forest vertical structure stratification result from ALS data. As can be seen from the results, this invention can also identify different types of point cloud data based on data features and employ different methods to extract tree and understory vegetation data from forest point clouds, which is difficult to match by other methods.

Claims

1. A method for vertical forest stratification based on near-ground 3D point clouds, characterized in that, The method includes the following steps: (1) Pre-identification of point cloud data types. The point density characteristics of the acquired data are used to pre-identify the input data type (ALS or TLS data). Specifically, if the average number of points in the neighborhood and the standard deviation are greater than a set threshold, the data is identified as TLS data. (2) Based on the different data types identified, algorithms targeting the point cloud data characteristics of different platforms are used to perform forest vertical structure stratification.

2. The method according to claim 1, characterized in that, When identified as ALS data, the forest vertical structure stratification includes the following steps: (1) Seed point filtering: Set the upper and lower thresholds of the number of points in the search ball to initially filter seed points, and further use the larger aspect ratio of the tree to remove some error-prone seed points. (2) Canopy Segmentation: Coarse segmentation is performed based on the distance interval method. For coarse segmentation results of overlapping canopies caused by trees, the segments are rotated to be parallel to the X-axis and sliced ​​along the Y-axis. Valley points of each slice in the overlapping canopy are extracted using polynomial fitting and connected to form valley lines between canopies. For coarse segmentation results of overlapping canopies caused by understory vegetation, vegetation height is extracted using height point curve fitting of the coarse segmentation results and sliced. A segmentation plane passing through the boundary points is constructed based on the two center points of the clustering results to achieve segmentation of trees and understory vegetation with overlapping canopies. (3) Forest stratification based on two-level, multi-scale voxel space: First, the study area is divided into first-level voxels of plot size. The point cloud data within each first-level voxel is divided into grids, and its height standard deviation (STDH) is calculated. The larger the STDH, the more complex the vertical structure. The side length of the second-level voxels is dynamically set according to the STDH to further divide the first-level voxels into second-level voxels. Point cloud slices are generated for points within each second-level voxel to produce height point curves. The trough points are fitted and extracted as the stratification heights of trees and understory vegetation. Points above the stratification height are trees, and those below are understory vegetation.

3. The method according to claim 1, characterized in that, When identified as TLS data, the forest vertical structure layering includes the following steps: (1) Extraction of tree trunk seed points before segmentation: The TLS point cloud with ground points removed is divided into height layers, and a height point count curve is generated. The layer height is extracted using a ninth-order polynomial fitting. A layered surface is generated by combining ground points and layer heights. The RANSAC algorithm is used to fit and extract cylindrical points within a certain range above the layered surface as tree trunk seed points. Tree trunk points are identified and extracted iteratively based on upward and downward search spheres and diameter thresholds. The tree crown is initially segmented using the distance interval method. The DBSCAN algorithm is used to detect the change points between different height layers in the overlapping area. The tree crown segmentation plane is constructed using two change points, the centroid of the cluster, and the boundary points of the clustering results to achieve the segmentation of the overlapping area of ​​the tree crown. The tree point cloud is obtained by combining it with the extracted tree trunk points. (2) Segmentation of understory vegetation after tree extraction: For the remaining point cloud after tree point cloud extraction, the result of clustering after slicing at 30% maximum height is used as seed point, and understory vegetation is extracted using the distance interval method.

4. The method according to claim 1, characterized in that, By combining the point density features of point cloud data to identify the input data type, different forest vertical structure layering algorithms are adopted for different data types to separate and extract point clouds of trees and understory vegetation.

Images