Bottom-up method for single-tree segmentation of rubber plantation using UAV LiDAR point cloud.

Abstract

This invention discloses a bottom-up UAV LiDAR point cloud segmentation method for individual trees in rubber plantations, comprising: collecting and processing initial point cloud data of the rubber plantation to obtain target point cloud data; stratifying the target point cloud data vertically and using a density-based point cloud clustering algorithm to calculate the centroids of clusters from bottom to top as the hierarchical initial skeleton points of the rubber plantation; extracting trunk skeleton points using density differences between branch points; identifying branch points based on the directional consistency of the trunk skeleton points and the density differences between the trunk and canopy, using the branch points as the starting seed points for canopy segmentation to separate the trunk from the canopy; fusing multiple features and using different feature combinations to process the skeleton points in different regions of the canopy, capturing the feature differences between individual trees, and constructing a multi-feature segmentation method for individual trees. This invention effectively improves segmentation accuracy and noise resistance, and has good segmentation results for canopies with general, medium, and complex canopy closures.

Description

Technical Field

[0001] This invention belongs to the field of point cloud data segmentation and classification technology, and in particular relates to a bottom-up method for segmenting individual trees in rubber forest UAV LiDAR point clouds. Background Technology

[0002] Rubber plantations are important artificial forest ecosystems in tropical regions. Extracting and mapping structural parameters of rubber plantations plays a crucial role in ecosystem assessment, tree species cultivation and breeding, refined management, and carbon sequestration. Traditional methods for acquiring structural information mainly rely on field measurements, which are time-consuming and labor-intensive. Unmanned aerial vehicle (UAV) LiDAR has become an important method for extracting high-precision individual tree structural parameters in tropical rubber plantations due to its low cost, high accuracy, high point density, and detailed representation of forest 3D structure. In recent years, LiDAR technology has received widespread attention for its advantages in rapidly and accurately acquiring forest vertical structure information. In particular, the use of low-altitude UAV platforms can acquire data with densities of up to thousands of points per square meter, offering strong operability, high geometric resolution, and highly detailed characterization of rubber plantation tree structure. How to segment high-precision individual tree point clouds from massive point cloud data is key to subsequent individual tree structural parameter extraction and is also an important foundation for the construction of 3D rubber tree models and accurate biomass inversion.

[0003] LiDAR point cloud segmentation for individual trees is a key technology in forestry surveying and mapping. Currently, LiDAR point cloud segmentation methods are mainly divided into two categories: segmentation based on projection images and segmentation directly based on point clouds. Segmentation based on projection images generates raster images through canopy height model (CHM) interpolation, identifies tree apexes using local extremum algorithms, and then uses clustering algorithms to estimate the canopy extent of each tree apex for individual tree segmentation. Using CHM to detect treetop locations yields good results in coniferous forests and other stands with clearly defined treetops. However, because some forest structure information is lost during the interpolation process to generate CHM from the original point cloud, the application of projection image-based segmentation is limited in tropical rubber plantations with complex canopy structures.

[0004] Point cloud-based segmentation can be categorized into top-down and bottom-up methods depending on the search strategy. Top-down methods detect individual trees by using the tree vertex as the initial seed point, reducing data loss during the generation of raster surfaces and models from the canopy height model. They can utilize adaptive thresholds and additional classification rules to improve canopy segmentation accuracy and mitigate the impact of overlapping branches on segmentation precision to some extent. However, starting from the treetop makes it difficult to accurately locate the treetop in forests with overlapping branches, especially when multiple canopies overlap, hindering accurate seed point localization and reducing accuracy.

[0005] Bottom-up tree segmentation methods start from the trunk, using trunk points as seed points to more accurately associate the canopy with its corresponding trunk, thus segmenting individual trees more precisely. However, existing bottom-up segmentation strategies are somewhat dependent on the density of the point cloud data. Low data density may lead to a decrease in trunk detection rate, thereby reducing the accuracy of tree segmentation. Furthermore, relying on single feature constraints is insufficient for identifying canopies with severe branch overlap, resulting in oversegmentation and undersegmentation. Although this method improves upon traditional methods in handling complex forest structures, accurate segmentation of canopy edges remains challenging in plots with severe branch overlap, requiring further research and technological advancements to overcome these challenges.

[0006] Mature rubber plantations often have high canopy density, leading to intense competition for branch and leaf growth among the canopies and significant branch overlap. Although previous studies have proposed many methods for individual tree segmentation, accurate segmentation of the canopy edges remains challenging in rubber plantations with overlapping canopies. Therefore, there is an urgent need to develop a robust and universal individual tree segmentation algorithm capable of accurately segmenting rubber tree canopies, providing a necessary methodological foundation for the precise extraction of individual tree structural parameters and subsequent 3D model construction and accurate biomass inversion. Summary of the Invention

[0007] To address the problems of indistinct tree crown apex, severe branch overlap, and difficulty in accurate segmentation of single-tree point clouds in rubber plantations, this invention proposes a bottom-up UAV LiDAR point cloud segmentation method for single trees in rubber plantations, including:

[0008] Initial point cloud data of the rubber forest was collected and processed to obtain target point cloud data;

[0009] The target point cloud data is layered vertically, and a density-based point cloud clustering algorithm is used to calculate the centroids of the clusters from bottom to top as the initial hierarchical skeleton points of the rubber forest.

[0010] The main trunk skeleton points are extracted by utilizing the density difference between the upper and lower branches. Branching points are identified based on the directional consistency of the main trunk skeleton points and the density difference between the main trunk and the crown. The branching points are used as the starting seed points for crown segmentation to separate the main trunk from the crown.

[0011] By comprehensively considering multiple feature information, fusing multiple features and using different feature combinations to process the skeleton points of different regions of the tree canopy, we can capture the feature differences between individual trees and construct multiple features to segment individual trees.

[0012] Preferably, the process of collecting and processing initial point cloud data of the rubber forest to obtain target point cloud data includes:

[0013] The initial point cloud data of the rubber plantation includes ground point and noise point information;

[0014] The process of processing the initial point cloud data includes,

[0015] Outlier detection is used for point cloud denoising, and a cloth-simulated filtering algorithm is used to separate ground points from non-ground points.

[0016] Then, the vegetation points are subjected to elevation normalization to reflect the true height of the trees. In the elevation normalization process, inverse distance weighted interpolation of ground points is used to generate a digital elevation model. The difference between the elevation of the vegetation points and the elevation of the DEM is calculated to generate a normalized point cloud, which is the target point cloud data.

[0017] Preferably, the process of stratifying the target point cloud data vertically and using a density-based point cloud clustering algorithm to calculate the centroids of clusters from bottom to top as the hierarchical initial framework points for the rubber forest includes:

[0018] Based on the vertical height of the rubber plantation, the plantation is divided into sections from bottom to top according to height intervals. A density-based point cloud clustering algorithm is used to cluster the point cloud of each layer. The center point of each cluster is then determined, which is the cluster center point of the local branch point cloud of each individual rubber tree in each layer. The branch point clouds of different layers are marked according to the cluster center points to obtain the initial skeleton points of the rubber trees in the entire plantation.

[0019] Preferably, the formula for obtaining the initial skeleton points of rubber trees in the entire forest based on the branch point clouds of different layers marked by the cluster center points is as follows:

[0020]

[0021] Among them, S i,j Let x represent the centroid of the j-th cluster in the i-th layer, and let Ni,j represent the number of points in the j-th cluster in the i-th layer. i，j，k y i，j，k and z i，j，k Let x, y, and z represent the x, y, and z coordinates of the k-th point in the j-th cluster of the i-th layer, respectively.

[0022] Preferably, the process of extracting the main trunk framework points by utilizing the density difference between the upper and lower branches, identifying the branches, and separating the main trunk from the crown includes:

[0023] Remove non-trunk objects around the trunk, and then use two constraints—the number of neighborhood points and the main clustering direction—to extract the individual trunks.

[0024] Using the point with the maximum local main elevation as the seed point, a radius neighborhood search is performed, and skeleton points that meet the neighborhood number limit are merged to obtain the main skeleton points;

[0025] By combining the straight trunk characteristics of rubber trees and the density variation pattern of skeletal points near the branching point, the branching point is determined. All skeletal points below the branching point are regarded as the main trunk, thus completing the division of the main trunk.

[0026] Preferably, the process of using the point with the maximum local main elevation as the seed point to perform a radius neighborhood search, and merging the skeleton points that meet the neighborhood number limit to obtain the main skeleton points includes:

[0027] The local backbone skeleton points are sorted by height. The point with the highest local backbone skeleton elevation is used as the seed point to search for its neighboring points. ρ1 is set as the density parameter threshold for identifying branch points. If the number of skeleton points to be classified within the search range is less than ρ1, it is added to the skeleton point set q. i A covariance matrix is ​​constructed with the local backbone skeleton points, and the eigenvalues ​​and corresponding eigenvectors of each point are obtained through principal component analysis. The skeleton points in the local backbone clustering are S = {s...} i |i=1,2,…,N s} and the skeleton point to be classified Q={q i |i=1,2,…,N q The formula for the local covariance matrix M of} is:

[0028]

[0029] In the formula, N represents the number of skeleton points and unclassified skeleton points within the local backbone cluster. The center point between the local main skeleton point and the skeleton point to be classified is calculated using the following formula: The eigenvector corresponding to the largest eigenvalue represents the principal direction. The angle θ between the principal direction and the skeleton point to be classified and the local center point is obtained by the following formula:

[0030] S·V=S x V x +S y V y +S z V z

[0031]

[0032] In the formula, S(S x ,S y ,S z V(V) is the normalized vector of the skeleton point to be classified and the local center point. x V y V z ) is the normalized vector with the main direction as the main direction.

[0033] Preferably, constructing multiple features to segment a single tree includes segmenting the skeleton points of the crown core point;

[0034] The process of segmenting the skeletal points of the tree crown core includes,

[0035] The crown core point is identified based on the horizontal distance from the crown core point to the branch point and the local density. The set of unsegmented skeleton points is U = {u i |i=1,2,…,N u}, N u To determine the number of unsegmented skeleton points, a density parameter threshold ρ3 is set to identify canopy points, prioritizing those with a horizontal distance of less than T from branch points. HD1 Furthermore, the skeleton points with local density ρ≥p3 are assigned to the tree to which the branch point belongs;

[0036] The skeleton point is defined as the core skeleton point of the tree crown, and its expression is:

[0037] H∈Core point of tree crown, if d HA ∈(0,T) HD1 )&ρ≥ρ3

[0038] The remaining unsegmented skeleton point set is the non-core points of the tree crown C = {c i |i=1,2,…,N c}, N c It represents the number of remaining undivided skeleton points, and all crown core points and crown non-core points are obtained by classifying them accordingly.

[0039] Preferably, constructing multiple features to segment a single tree includes segmenting the skeleton points of the crown boundary points;

[0040] The process of segmenting the skeleton points of the tree canopy boundary includes,

[0041] The canopy boundary is identified by analyzing the local density of each point within the boundary. The neighborhood search radius is set to Rs, and neighborhoods with fewer than ρ3 and within a horizontal distance (T) are selected. HD1 T HD2 Points within the range () are added as boundary points to the boundary point set B = {b} i |i=1,2,…,N b}, perform a 3D shortest distance search on the boundary point and assign the point to the tree containing the nearest split point. The formula is as follows:

[0042] H∈Boundary points, if d HA ∈(T HD1 ,T HD2 )&ρ<ρ3

[0043] In the formula, H is the skeleton point attribute to be classified, and T HD1 It is the minimum crown radius, T HD2ρ is the maximum crown radius, and ρ3 is the number of skeleton point constraints within the radius's search range.

[0044] Preferably, constructing a multi-feature segmentation method for a single tree includes segmenting the skeleton points of disputed points at the intersection of branches and trunks;

[0045] The process of segmenting the skeletal points of disputed points at branch intersections includes:

[0046] The horizontal distance from the branch point is (T) HD1 T HD2 Points with local density ρ > ρ2 are considered disputed points where branches overlap severely, and the formula expression is:

[0047] H∈Disputed point if d HA ∈(T HD1 ,T HD2 )&ρ<ρ2

[0048] In the formula, H is the skeleton point attribute to be classified, and T HD1 It is the minimum crown radius, T HD2 ρ2 is the maximum crown radius, and ρ2 is the density threshold of the disputed skeleton points dividing the boundary.

[0049] Based on the characteristic that the growth trend of rubber tree crowns and branches extends obliquely upwards at a certain angle, the growth angle between branches and leaves is within a certain range. Therefore, the set of disputed points Z = {z...} is used to remove the core points and boundary points. i |i=1,2,…,N z Search for each point in its neighborhood, select the nearest skeleton point with attributes, calculate the angle between the two and compare their sizes. If the calculated angles both conform to the direction of branch growth, then the point belongs to the tree of the skeleton point with the smaller angle. If there is no skeleton point with existing attributes within the neighborhood search range of the disputed point's radius, then it is assigned to the tree of the nearest skeleton point based on the shortest 3D distance.

[0050] Preferably, the vector V of the nearest segmented points C(Cx,Cy,Cz), D(Dx,Dy,Dz), and disputed point E(Ex,Ey,Ez) within the radius neighborhood search range of the disputed point is calculated. CE V DE The angles θ1 and θ2 between the vector and the Z-axis unit vector P = (0,0,1) are given by the following formulas:

[0051]

[0052] If θ1 is less than θ2, then assign E to Tree#1; otherwise, assign E to Tree#2.

[0053] Compared with the prior art, the present invention has the following advantages and technical effects:

[0054] This invention identifies trunk skeleton points by analyzing the local density of skeleton points at branching points and the changes in the angle between trunk skeleton points. A bottom-up segmentation strategy, starting from trunk identification, is then used to segment individual tree point clouds. Multi-feature fusion is used instead of a single threshold constraint as a constraint for canopy segmentation in the individual tree segmentation algorithm, and angular feature metrics are used instead of Euclidean spatial distance to determine the attribution of disputed points at branch overlaps. This strategy effectively improves segmentation accuracy and noise resistance, demonstrating good segmentation results for canopies with general, medium, and complex canopy closures. Attached Figure Description

[0055] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0056] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0057] Figure 2 This is an example diagram of hierarchical DBSCAN clustering extraction of initial skeleton points according to an embodiment of the present invention;

[0058] Figure 3 This is a schematic diagram of the main skeleton point extraction process according to an embodiment of the present invention;

[0059] Figure 4 This is a rule diagram of the fusion of horizontal distance and density features to segment skeleton points according to an embodiment of the present invention;

[0060] Figure 5 This is a diagram illustrating the attribution of disputed points based on the angle threshold in an embodiment of the present invention. Detailed Implementation

[0061] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0062] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0063] like Figure 1 As shown, the present invention provides a bottom-up method for single-tree segmentation of rubber forest UAV LiDAR point cloud, which includes the following steps:

[0064] Step 1: Collect initial point cloud data of the rubber forest and process it to obtain target point cloud data.

[0065] Specifically, the process involves: First, using DBSCAN clustering to generate hierarchical initial framework points for the rubber forest, facilitating further individual tree segmentation; second, extracting trunk framework points using density differences between branch points to provide trunk information for optimizing canopy segmentation; and finally, comprehensively considering multiple feature information to fully capture feature differences between individual trees, constructing multi-feature segmentation for individual trees to improve accuracy.

[0066] To verify the effectiveness and superiority of the algorithm of this invention, three experimental plots of rubber trees of the same variety but with different canopy closure were selected to verify the method, and the crown width of individual trees and the crown projection area parameters were extracted to further evaluate the accuracy of individual tree segmentation.

[0067] The data collected in this invention comes from a rubber tree experimental base in Danzhou City, northwestern Hainan Province. Danzhou City is located between 19°11′ and 19°52′ north latitude and 108°56′ and 109°46′ east longitude. It is the largest rubber production base on Hainan Island, with a wide variety of rubber tree strains and extremely abundant rubber resources. Danzhou City has a tropical monsoon climate, receiving approximately 100 to 130 kilocalories of solar radiation per square meter annually. Summers are not excessively hot, winters are not severely cold, and there is abundant sunshine. The average annual temperature is 23.2℃, and the average annual rainfall is 1815 mm, with most areas receiving 1500 mm or more, indicating ample water resources. To verify the single-tree segmentation performance of the algorithm in this embodiment, point clouds of rubber trees from three sample plots were obtained. The terrain of the three sample plots was relatively flat, and the rubber trees were of the same variety. The canopy closure of the three sample plots was different. The canopy closure of sample plots A, B, and C was general, medium, and complex, respectively. The effectiveness of the single-tree segmentation method in this study was verified by using three experimental sample plots with different canopy closures. The data information of the three experimental sample plots is shown in Table 1.

[0068] Table 1

[0069]

[0070] The experiment used a DJI M300RTK drone equipped with a CBI-Lite lidar measurement system to collect data from a rubber plantation during its deciduous period, obtaining LiDAR point cloud data. The drone flew at an altitude of 80m, 40-50m from the canopy top, with an average speed of 4m / s and a lateral overlap of 65%. The system was equipped with a PandarXT customized lidar sensor, emitting a wavelength of 905nm, rotating 360° at a frequency of 10Hz per second, and detecting the scattered laser pulses from objects via a receiver. It can generate approximately 1.92 million points per second, with an effective detection range of 120m.

[0071] Furthermore, the bottom-up UAV LiDAR point cloud segmentation method for rubber forests provided by this invention mainly consists of three sub-processes: initial skeleton point generation, single-tree trunk skeleton extraction, and single-tree segmentation through multi-feature fusion. The first step involves vertically stratifying the sample point cloud and using bottom-up clustering to obtain the centroids of the clusters as the initial skeleton points for the rubber forest. These initial skeleton points reflect the true morphology of the trees, and their generation reduces the impact of uneven dataset density distribution on the clustering results. Then, branching points are identified based on the directional consistency of the rubber tree trunk skeleton points and the density differences between the trunk and crown, using these branching points as the starting seed points for crown segmentation. Branching point identification achieves the separation of the trunk and crown, avoiding missegmentation of the trunk and crown by using a uniform feature threshold. Finally, to improve the segmentation accuracy of single trees in complex and overlapping crown structures, multi-feature fusion is used, employing different feature combinations to process the skeleton points of different regions of the crown, which can more effectively segment single trees.

[0072] Raw LiDAR data of rubber plantations includes ground point and noise point information, requiring preprocessing such as point cloud denoising and ground filtering to obtain a clean rubber tree point cloud. This invention employs outlier detection for point cloud denoising and uses the cloth-based simulated filtering (CSF) algorithm to separate ground and non-ground points. Furthermore, to reflect the true height of the trees, vegetation points undergo elevation normalization. In the preprocessing, inverse distance weighted interpolation (IDW) is used to generate a digital elevation model (DEM) from the ground points, and the difference between the vegetation point elevation and the DEM elevation is calculated to generate a normalized point cloud.

[0073] Step 2: The target point cloud data is layered vertically, and the centroids of the clusters are calculated from bottom to top using a density-based point cloud clustering algorithm as the initial hierarchical skeleton points of the rubber forest.

[0074] Specifically, bottom-up single-tree segmentation employs point-by-point analysis and calculation, which is inefficient for processing massive point cloud data. To improve computational efficiency, existing technologies use voxelization to reduce the number of point clouds; however, establishing a voxel grid using a uniform standard in dense rubber plantations may result in the loss of true tree morphology information. This embodiment utilizes a density-based point cloud clustering algorithm, DBSCAN (Density-Based Spatial Clustering of Applications with Noise), to effectively eliminate noise influence and cluster trees into arbitrary shapes while preserving their true morphology.

[0075] Based on the vertical height of the rubber forest, the rubber forest is divided into sections from bottom to top according to the height interval. The point cloud of each layer is clustered using the DBSCAN algorithm. Then, the center point of each cluster is obtained, which is the cluster center point of the local branch point cloud of each single rubber tree in each layer. By marking the branch point cloud of different layers according to these center points, the initial skeleton points of the rubber trees in the entire forest can be obtained, as shown in formula (1).

[0076] Taking a section of a single rubber tree branch as an example, the effect of extracting layered framework points is as follows: Figure 2 As shown.

[0077]

[0078] Among them, S i,j N represents the centroid (skeleton point) of the j-th cluster in the i-th layer. i,j x represents the number of points in the j-th cluster of the i-th layer. i，j，k y i，j，k and z i，j，k Let x, y, and z represent the x, y, and z coordinates of the k-th point in the j-th cluster of the i-th layer, respectively.

[0079] Step 3: Extract the main trunk skeleton points by utilizing the density difference between the upper and lower branches. Identify the branches based on the directional consistency of the main trunk skeleton points and the density difference between the main trunk and the crown. Use the branches as the starting seed points for crown segmentation to separate the main trunk from the crown.

[0080] Specifically, due to the differences in density characteristics between the trunk and the crown, they cannot be effectively segmented using a uniform density threshold. To accurately segment a single tree, the key step is to first segment the trunk, a process that relies on accurately identifying the branching points, i.e., the junction between the crown and the trunk.

[0081] The trunk may contain some non-trunk objects, which need to be removed. This invention improves the accuracy of single-tree trunk extraction by utilizing two constraints: the number of neighborhood points and the main clustering direction. The point with the maximum local trunk elevation is used as the seed point for radius neighborhood search. Skeleton points that meet the neighborhood number constraint are merged as trunk skeleton points. Possible branching points are determined by combining the straight trunk characteristic of rubber trees and the density variation pattern of skeleton points near branching points. Once a branching point is determined, all skeleton points below it are considered part of the trunk, at which point the trunk segmentation is complete. The segmentation process is as follows: Figure 3 As shown, the skeleton points in the L1 layer are used as local trunks. Skeleton points that meet the conditions are identified from bottom to top as trunks. Abnormal points or noise points that are inconsistent with the main direction of the local trunk are not part of the trunk and are removed. When the number of points is greater than ρ1, growth stops and the location of the point is a branch point.

[0082] Further optimization of the scheme involves using the point with the maximum local main elevation as the seed point to perform a radius neighborhood search, and merging the skeleton points that meet the neighborhood number limit to obtain the main skeleton points. The process includes:

[0083] The skeleton points are sorted from bottom to top by height. After multiple experiments and statistical analyses, the first 5 layers of skeleton points were found to be the most accurate and effective local trunks. The traversal layer index retains the skeleton points of L1. The local trunk L1 is obtained through DBSCAN clustering and numbered. This number also serves as the tree number, and subsequent skeleton points belonging to the same tree will be assigned the same number.

[0084] The local backbone skeleton points are sorted by height. The point with the highest local backbone skeleton elevation is used as the seed point to search for its neighboring points. ρ1 is set as the density parameter threshold for identifying branch points. If the number of skeleton points to be classified within the search range is less than ρ1, it is added to the skeleton point set q. i A covariance matrix is ​​constructed with the local backbone skeleton points, and the eigenvalues ​​and corresponding eigenvectors of each point are obtained through principal component analysis. The skeleton points in the local backbone clustering are S = {s...} i |i=1,2,…,N s} and the skeleton point to be classified Q={q i |i=1,2,…,N q The local covariance matrix of} is shown in Equation (2).

[0085]

[0086] In the formula, N represents the number of skeleton points and unclassified skeleton points within the local backbone cluster. The center point between the local main skeleton point and the skeleton point to be classified is calculated using the following formula: The eigenvector corresponding to the largest eigenvalue represents the principal direction, and the angle θ between the principal direction and the skeleton point to be classified to the local center point is obtained by equations (3) and (4).

[0087] S·V=S x V x +S y V y +S z V z (3)

[0088]

[0089] In the formula, S(S x ,S y ,S z V(V) is the normalized vector of the skeleton point to be classified and the local center point. x V y V z) is the normalized vector with the main direction as the main direction.

[0090] This invention eliminates θ greater than θ th skeleton points, θ th The angle between the skeleton point to be classified and the main direction of the trunk is used to determine whether the skeleton point is a linearly perpendicular object. Only when the direction of the skeleton point to be classified is approximately the same as the main direction of the local trunk is the point considered a linearly perpendicular object and added to S. However, absolute perpendicularity is generally not possible. This invention uses θ... th =20° boundary. Select the point with the highest elevation from S and repeat the above steps until the number of neighborhood points of the skeleton point to be classified exceeds the set range ρ1, indicating that the point is a branch point and the main trunk segmentation is completed.

[0091] Step 4: Taking into account multiple feature information, the skeleton points of different regions of the tree crown are processed by combining multiple features and different feature combinations to capture the feature differences between individual trees and construct multiple features to segment individual trees.

[0092] Specifically, it includes:

[0093] (1) Constructing multiple features to segment a single tree includes segmenting the skeleton points of the crown core point;

[0094] (2) Constructing multiple features to segment a single tree includes segmenting the skeleton points of the crown boundary points;

[0095] (3) Constructing multiple features to segment a single tree includes segmenting the skeleton points of disputed points at the intersection of branches and trunks;

[0096] Further optimization of the scheme, (1) constructing multi-feature segmentation of individual trees, including the specific process of segmenting the skeleton points of the crown core point, includes:

[0097] Segmentation using a single feature presents challenges for rubber plantations with severe branch overlap, as different branches and trees within the plantation may share similar values ​​for certain features, making accurate segmentation difficult with a single threshold. To address this issue, this invention proposes a multi-feature fusion-based single-tree segmentation method. By comprehensively considering multiple feature information (distance, local density, and angular features), it more comprehensively captures the feature differences between individual trees, improving segmentation accuracy. The most challenging task in single-tree segmentation is segmenting areas with severe branch overlap. Different segmentation methods are proposed for skeletal points at different locations (crown core point, crown boundary point, and disputed points at branch overlap), as detailed below.

[0098] Generally, the crown core point closest to the branching point is more likely to belong to the tree. However, in plots with high canopy closure and complex canopy structures, considering only the horizontal distance between the branching point and the crown point may assign crown points with larger canopies to trees with smaller canopies. To address this challenge, crown core point identification is based on its horizontal distance from the branching point and local density, such as... Figure 4 As shown at point A. The set of undivided skeleton points is U = {u i |i=1,2,…,N u}, Nu represents the number of unsegmented skeleton points. A density parameter threshold ρ3 is set to identify canopy points, selecting those with a horizontal distance less than T from the branch point. HD1 Furthermore, skeleton points with local density ρ≥p3 are assigned to the tree to which the branch point belongs; a skeleton point is defined as a crown core skeleton point, and its expression is:

[0099] H∈Core point of tree crown, if d HA ∈(0,T) HD1 )&ρ≥ρ3

[0100] The remaining unsegmented skeleton point set (non-core points) C = {c i |i=1,2,…,N c}, N c It represents the number of remaining unsegmented skeleton points. Based on this method, all core points and non-core points can be distinguished.

[0101] Further optimization of the scheme, (2) constructing multi-feature segmentation of individual trees, including the specific process of segmenting the skeleton points of the crown boundary points, includes:

[0102] The local density of skeletal points at the center of the canopy should be higher than that at the canopy boundary. For skeletal points near the canopy boundary, the canopy boundary can be clearly distinguished by analyzing the local density of each point within the boundary. Figure 4 Point B is shown in the diagram. This invention sets the neighborhood search radius to Rs, and considers neighborhoods with fewer than ρ3 and within a horizontal distance (T). HD1 T HD2 Points within the range () are added as boundary points to the boundary point set B = {b} i |i=1,2,…,N b}, perform a 3D shortest distance search on the boundary point and assign the point to the tree containing the nearest split point. The formula is as follows:

[0103] H∈Boundary points, if d HA ∈(T HD1 ,T HD2 )&ρ<ρ3 (5)

[0104] In equation (5), H is the skeleton point attribute to be classified, and T HD1 It is the minimum crown radius, T HD2 ρ is the maximum crown radius, and ρ3 is the number of skeleton point constraints within the radius's search range.

[0105] The horizontal distance from the branch point is (T) HD1 T HD2 Points with local density ρ > ρ2 are considered disputed points where branches overlap severely, and the formula expression is:

[0106] H∈Disputed point if d HA ∈(T HD1 ,T HD2 )&ρ<ρ2

[0107] In the formula, H is the skeleton point attribute to be classified, and T HD1 It is the minimum crown radius, T HD2 ρ2 is the maximum crown radius, and ρ2 is the density threshold of the disputed skeleton points dividing the boundary.

[0108] like Figure 4 As shown, Figure 4 The leftmost point represents the initial skeleton points of the two trees before they were split. The middle point represents the rules for splitting the skeleton points by fusing horizontal distance and density features. Point A has a local density (ρ≥p3) and is within the horizontal distance (0, T) of the branch point. HD1 Within the range of ), it belongs to the core skeleton point; the local density of point B is ρ < ρ3 and it is within the horizontal distance (T) of the branch point. HD1 T HD2 Within the range of ), it belongs to the crown boundary point; the local density at point C is ρ < ρ2, and it is within the horizontal distance (T) of the branching point. HD1 T HD2 Within the range of ), points are considered boundary dispute points; the rightmost points are the remaining points after the core skeleton points and boundary skeleton points have been separated, and these are disputed skeleton points.

[0109] If only a single horizontal distance is used for comparison to determine the ownership of disputed skeleton points, such as Figure 5 As shown on the right, point E will be assigned to Tree #2, which is closer in horizontal distance. In fact, point E belongs to Tree #1, so it is necessary to compare the angle threshold θ to determine its affiliation.

[0110] Further optimization of the scheme, (3) constructing multi-feature segmentation of individual trees, including the specific process of segmenting the skeleton points of disputed points at the intersection of branches and trunks, includes:

[0111] The rubber tree crown grows in an "inverted broom" shape, and the branches extend obliquely upwards at a certain angle. Therefore, the growth angle between the branches and leaves is within a certain range. Based on this characteristic, the set of disputed points Z = {z} is removed from the core points and boundary points.i |i=1,2,…,N z Search for each point in its neighborhood, select the nearest skeleton point with attributes, calculate the angle between the two and compare their sizes. If the calculated angles both conform to the direction of branch growth, then the point belongs to the tree of the skeleton point with the smaller angle. If there is no skeleton point with existing attributes within the radius of the disputed point, then it is assigned to the tree of the nearest skeleton point according to the shortest 3D distance.

[0112] Calculate the vector V of the nearest split points C(Cx,Cy,Cz), D(Dx,Dy,Dz), and the disputed point E(Ex,Ey,Ez) within the radius neighborhood search range of the disputed point. CE V DE The angles θ1 and θ2 between the vector P = (0,0,1) and the Z-axis are given by formula (6).

[0113]

[0114] If θ1 is less than θ2, then assign E to Tree#1; otherwise, assign E to Tree#2. Figure 5 As shown, Tree#1 and Tree#2 on the left have similar growth patterns, and the horizontal distances from point E to seed points A and B are the same. The assignment of point E is determined by the angle between the skeleton points C and D with the smallest 3D distance within the radius neighborhood of point E and which have tree attributes. This point belongs to Tree#1. On the right, when the growth patterns of Tree#1 and Tree#2 are significantly different, point E tends to be assigned to Tree#1, which better conforms to the branch growth pattern.

[0115] The skeleton points are labeled using the above processing steps. Therefore, tree crown skeleton points and trunk skeleton points with the same label are combined to obtain a single tree skeleton. Skeleton points are obtained by calculating the centroids of the point clouds in the cluster. After skeleton point segmentation is complete, point clouds with index relationships to the skeleton points are called and assigned to the corresponding skeleton point tree attributes, thus completing the single tree segmentation.

[0116] The experimental results show that this method achieves high accuracy in segmenting three plots with varying canopy closure from low to high, with F (harmonicity) values ​​of 97.3%, 98.2%, and 95.5%, respectively. The R² (coefficient of determination) between the measured and estimated values ​​of crown width and canopy projection area ranges from 0.9744 to 0.9976, demonstrating good correlation and consistency. This study presents an algorithm that improves the accuracy of individual tree information segmentation by combining the geometric and topological relationships of skeleton points with a multi-feature bottom-up growth approach. This is of great significance for the refined management of tropical rubber forests, tree competition, and monitoring.

[0117] Further optimization of the scheme involved sensitivity analysis of individual tree segmentation parameters: The density difference between the skeletal points on the trunk and crown is significant, and the main directions of the skeletal points from bottom to top are relatively consistent. If the density change exceeds ρ1 and the angle between the direction vectors of the upper and lower skeletal points and the main direction does not exceed θ... th If the value is 0, then the point is considered a branch point. The density threshold used to identify the main framework points is denoted as . Sensitivity analysis is performed to verify the approximation of manually extracted branch point locations by comparing them with those extracted manually. The density of the main framework points is relatively consistent across the three sample plots; a density threshold of 3 yields good results.

[0118] Based on the canopy width of the rubber trees in both row and column directions, two horizontal distance thresholds T are set. HD1 and T HD2 To avoid misassigning the core framework of trees with larger canopies to trees with smaller canopies in the sample plot, T HD1 The setting should be close to the minimum canopy radius of the sample plot. It is more reasonable to use the smaller canopy radius as the horizontal distance segmentation constraint to distinguish between core points and non-core points. HD2 The radius of the canopy was set to be similar to the maximum canopy radius of the sample plot. ρ2 was used as the second condition to distinguish between core and non-core points. Core points are usually located at the center of the canopy and have a higher local density than non-core points. By performing a 0.4m radius neighborhood search on the skeleton points in the sample plot, the local density of each point was distinguished and displayed, including the local density of skeleton points in the core point area and the local density of skeleton points at the canopy boundary. The local density (ρ≥p2, p3<ρ<p2) and the horizontal distance from the branch point (0, T) were considered. HD1 The core point is the skeleton point within the range of ) where the local density ρ < ρ3 and the horizontal distance from the branch point is (T). HD1 T HD2 The skeleton points within the range are designated as boundary points. Among them, the canopy branches in plot C are densely grown and heavily overlapped, with higher densities of core points and boundary points than in plots A and B. Based on the branch growth characteristics, an angular relationship θ is established between existing attribute skeleton points and unclassified skeleton points (disputed points). This angular characteristic conforms to the growth trend between branch skeleton points or leaf skeleton points.

[0119] Further optimization schemes and accuracy evaluation of single-wood segmentation;

[0120] To evaluate the accuracy of the tree segmentation algorithm in this embodiment, the crown outline of each tree in each experimental plot was measured. Based on the measured results, the point cloud of the plot was manually segmented into individual trees. This data was used as the measured reference value for tree segmentation, and the accuracy of the algorithm's segmentation theoretical value was evaluated. The ratio of the number of points in the manually segmented tree point cloud to the number of points in the algorithm-segmented tree point cloud (ρalgorithm / reference) was used to evaluate the segmentation accuracy.

[0121] Regarding the detection rate of individual rubber trees, this embodiment uses recall (R), precision (P), and the harmonic value F, which comprehensively considers R and P, for evaluation. Recall represents the proportion of validly detected individual rubber trees to the actual number of trees, while precision represents the proportion of validly detected individual rubber trees to the total extraction result. The individual tree segmentation algorithm can segment point clouds into multiple categories, each of which can be considered a single rubber tree. The segmentation results can be roughly divided into three cases: (1) Correct segmentation (TP): the number of correctly segmented rubber trees; (2) Undersegmentation (FN): the number of rubber trees that were not segmented because they were considered to belong to other types of rubber trees; (3) Oversegmentation (FP): trees that were originally one tree but were mistakenly segmented into multiple rubber trees.

[0122] Therefore, R, P, and F can be described as shown in equation (7-9).

[0123]

[0124]

[0125] In the formula, the values ​​of F, R, and P all range from 0 to 1. To obtain a higher F value, the value of F is equal to 1 only when the values ​​of R and P both reach 1. The accuracy evaluation criteria based on the point count criterion are shown in Table 2.

[0126] Table 2

[0127]

[0128] By identifying the canopy boundary, a convex polygon is used as the canopy boundary to calculate its area. The average canopy width in both the east-west and north-south directions is measured as the canopy width. The same method is used to evaluate the accuracy of the manually segmented original single-tree point cloud parameters and the extracted parameters. The coefficient of determination (R²), root mean square error (RMSE), and normal root mean square error (NRMSE) are used to evaluate the accuracy of the canopy width parameter extraction in this embodiment, thereby verifying the accuracy of the single-tree segmentation in this embodiment. R² reflects the degree of fit between the extracted values ​​and the measured values. RMSE quantifies the extraction error of the single-tree structural parameters. The closer R² is to 1 and the smaller the RMSE, the higher the accuracy of the parameter extraction. NRMSE reflects the consistency between the extracted and measured values. The smaller the NRMSE value, the higher the consistency between the extracted and measured values.

[0129]

[0130] In the formula, y i These are the measured values ​​of the parameters for a single timber structure; These are the parameter values ​​of a single timber structure extracted using an algorithm; is the measured average value of the structural parameters of a single tree; n is the number of rubber tree samples.

[0131] Results and Analysis

[0132] Single-wood segmentation results

[0133] Extracting the skeleton points of a single tree can be seen as a derivative of the local minimum clustering algorithm based on point clouds. First, the lowest point is found as the seed point. Then, starting from the seed point, the tree trunk and crown are segmented sequentially from bottom to top. The boundary between the trunk and crown lies in the identification of branch points. In this embodiment, the point with the maximum local trunk elevation is used as the initial seed point. Starting upwards, the location of abrupt changes in the local density of skeleton points and the degree of offset between the skeleton points and the main direction of the trunk are used to identify branch points. The portion below the branch point is considered the trunk of the tree, facilitating subsequent segmentation of the crown using trunk information.

[0134] After the single-wood skeleton point segmentation algorithm is completed, there are still some unclassified disputed points. The angular features and 3D shortest distance algorithm are used to determine the ownership of the skeleton points and complete the single-wood skeleton point segmentation.

[0135] After segmenting individual trees into plots with canopy closure levels of general, moderate, and complex, the resulting 3D point clouds of individual trees were displayed from both top and side views, with adjacent tree point clouds rendered in different colors. It can be seen that the algorithm segments the point cloud into reasonable results that conform to the morphology of individual trees.

[0136] Plots A and B showed high accuracy with relatively less oversegmentation. Although the canopy closure of plot A was lower than that of plot B, some trunk point cloud data was missing, causing the algorithm to terminate at trunk breaks, resulting in lower overall accuracy than plot B. While there was some branch overlap among the rubber trees in plot B, the data completeness was high, leading to better algorithm recognition than in plot A. The overall accuracy of plot C was relatively low, mainly due to severe branch overlap, making it difficult to accurately identify the canopy boundaries of all rubber trees in plot C using a uniform density threshold. This method can provide a reference for 3D modeling and refined management of forest scenes.

[0137] Since visual interpretation cannot determine whether individual trees have been misclassified or missed, this embodiment uses a ratio method to verify the segmentation of individual trees. When the ratio exceeds a certain threshold, it is considered an incorrect segmentation. Experimental results show that the detection rate of the algorithm decreases as the canopy closure of the sample plot increases. This is because in high-canopy-closure sample plots, the overlap between tree canopies is more severe, increasing the difficulty of algorithm recognition and thus leading to misclassification and missed segmentation.

[0138] To further verify the segmentation performance of the algorithm, the identification of individual tree density in three sample plots with different canopy closures was considered. Regarding individual tree density, as the canopy closure of the sample plots increased, the segmentation accuracy ρ of individual trees in the sample plots was either below 95% or above 105%, indicating that missegmentation or omissions occurred in the sample plots.

[0139] In plot C, severe branch overlap made it difficult to accurately identify all canopy boundaries using a uniform density threshold, leading to misclassification and omissions during the algorithm's identification process. In plot B with medium canopy closure, one instance of omission occurred, primarily due to the tree's low point density and missing point cloud data for branch tips, causing the algorithm to terminate. This meant that even small changes in individual tree density had a significant impact on the overall ratio. No misclassification or omissions occurred in the low canopy closure plots because the larger spacing between trees made it easier for the algorithm to distinguish and segment individual trees, reducing canopy overlap and interference, thus improving the accuracy and reliability of segmentation. Overall, the algorithm performed best in the low canopy closure plots, achieving a comprehensive accuracy of 97.4%. In the medium and high canopy closure plots, the comprehensive accuracies were 96.9% and 90.6%, respectively.

[0140] Comparison of results from various single-tree segmentation algorithms

[0141] To further verify the single-tree segmentation performance of the algorithm in this embodiment, the bottom-up skeleton segmentation algorithm proposed in this embodiment is compared with other algorithms. The hierarchical DBSCAN algorithm, the threshold-based region growing algorithm (PCS), the layer stacking algorithm, the K-means algorithm, and the watershed algorithm are compared and analyzed. The threshold-based region growing algorithm (PCS), the layer stacking algorithm, the K-means algorithm, and the watershed algorithm are commonly used clustering algorithms in the single-tree segmentation problem of UAV LiDAR point cloud data. Comparing them verifies the performance of the algorithm in this embodiment in single-tree segmentation based on clustering methods. The hierarchical DBSCAN algorithm, as an intermediate process of the algorithm in this embodiment, is compared to it to examine the effectiveness of the improvements made to the hierarchical DBSCAN clustering algorithm in this embodiment. The accuracy evaluation of the single-tree segmentation comparison experiment is shown in Table 3.

[0142] Table 3

[0143]

[0144]

[0145] Table 3 compares the segmentation results of different segmentation algorithms with the original single-tree point cloud data for rubber plantation plots with different tree densities. After obtaining the correct, under-segmented, and over-segmented data for each plot, a quantitative accuracy evaluation was conducted. For the three plots with different canopy closures, the five segmentation methods achieved the highest accuracy (F-score) in plot A with lower canopy closure: 90%, 85%, 75%, 64%, and 45%, respectively. Furthermore, the segmentation accuracy and single-tree recognition rate of the bottom-up segmentation algorithm (layered DBSCAN and region growing) were higher than those of the top-down segmentation algorithm (K-means, layer stacking, and watershed). This indicates that the accuracy of obtaining the initial seed point affects the segmentation of individual trees. The indistinct apex of rubber plantation trees makes it difficult for the top-down segmentation method to obtain the actual initial seed point at the tree top, leading to multiple trees sharing a single tree apex as the seed point, resulting in missegmentation and omissions.

[0146] To more intuitively compare the performance of several algorithms across multiple sample plots, this embodiment uses a bar chart to statistically compare the accuracy of the algorithms, showing that the accuracy gradually decreases with increasing canopy closure. For the hierarchical DBSCAN algorithm, since its theory is similar to the algorithm in this embodiment, the same parameters are used for parameter settings, resulting in a similar single-tree detection rate within the same sample plot. It can be seen that the hierarchical DBSCAN algorithm has the second-highest single-tree detection rate among all sample plots. Meanwhile, the single-tree segmentation accuracy of the other five algorithms decreases to varying degrees with increasing canopy closure. Among the five segmentation methods, the hierarchical DBSCAN algorithm achieves the best accuracy in the matching results of all three sample plots, with overall segmentation accuracies of 90%, 81%, and 64%, respectively. The hierarchical DBSCAN algorithm can effectively identify the lowest point of a tree and successfully classify it as a tree. However, when dealing with challenges of different heights and point cloud densities, the hierarchical DBSCAN algorithm has some limitations, which may lead to an increase in undersegmentation or oversegmentation. The region growing algorithm showed a 5%, 6%, and 13% decrease in overall segmentation accuracy compared to the hierarchical DBSCAN algorithm in three plots with varying canopy closure from low to high. This is primarily because region growing algorithms are typically based on local neighborhood information, which can lead to inaccurate segmentation results in point clouds with complex structures or significant occlusion. Hierarchical DBSCAN is not subject to these limitations as it considers global point density. The K-means algorithm, layer stacking algorithm, and watershed algorithm, which employ a top-down segmentation approach, performed poorly in rubber plantations where branches overlap significantly at the canopy level and tree apexes are not clearly defined. These algorithms rely on tree apexes as initial seed points; inaccurate initial seed point acquisition can lead to severe oversegmentation and undersegmentation. The algorithm in this embodiment uses a bottom-up approach, acquiring the lowest point of the tree trunk as the initial seed point, thus avoiding the difficulty in obtaining initial seed points due to complex canopy structures. As the height increases, based on the straight growth characteristics of rubber tree trunks and the geometric topological relationship between the skeleton points, thresholds such as the angle, distance, and density between skeleton points are used as constraints to solve the problem of reduced segmentation quality caused by a single threshold, and it is more consistent with the morphology of individual trees. The segmentation accuracy of individual trees in high, medium, and low density plots is 97%, 98%, and 95%, respectively.

[0147] Single-wood structure parameter extraction accuracy

[0148] Based on the extraction results of individual tree segmentation, the crown width and crown projected area are calculated from the point cloud of the crown portion. Crown width is a crucial characteristic factor representing the crown structure, typically obtained by averaging the crown width in both the east-west and north-south directions. This embodiment uses a direct method to extract the crown width from the LiDAR point cloud of a rubber tree, calculating it based on the two-dimensional projection points of the crown portion. The crown projected area is a tree parameter derived from the crown width, providing a more intuitive representation of the crown size. While the convex hull method effectively considers the overall crown morphology, using convex polygons as the crown boundary to calculate its area may overestimate the crown projected area. Since the outline of the crown point cloud projection is regular, a k-nearest neighbor concave hull algorithm is used to more closely approximate the crown projection outline. This algorithm calculates the boundary of the crown projection points. As the number of neighboring points (k) increases, if sequentially connecting the vertices forms a closed polygon, a concave polygon that meets the requirements is obtained. The area of ​​this polygon is then used as the crown projected area.

[0149] By obtaining the fitted line graphs of the extracted crown width and crown projection area and the measured crown width in the three study areas, it can be concluded that the correlation and consistency between the extracted crown width values ​​and the measured values ​​are high. Specifically, there is a certain deviation between the extracted crown width values ​​and the measured values. The reason for the error between the extracted values ​​and the measured values ​​is that the crowns of rubber trees are severely intersecting, and there is a certain deviation between the results of the manually extracted crown boundary on both sides and the results extracted based on single tree segmentation.

[0150] The linear regression equations and accuracy evaluation results between the measured and estimated crown width values ​​of the three study areas are shown in Tables 4 and 5. Table 4 shows the crown width regression equation and evaluation results, and Table 5 shows the crown projection area regression equation and evaluation results. As can be seen from the tables, the R² of study area A is 0.9865, RMSE is 0.173m, and NRMSE is 2.65%; the R² of study area B is 0.9882, RMSE is 0.082m, and NRMSE is 2.34%; and the R² of study area C is 0.9762, RMSE is 0.174m, and NRMSE is 3.17%. The crown width extraction results and measured values ​​of the three study areas all show good correlation and consistency, fully demonstrating the effectiveness of the algorithm in this embodiment. The RMSE of study area B is slightly lower than that of study areas A and C, mainly because the numerical quality of study area B is higher, there are fewer missing data, and the visually identified boundaries are more consistent with the single-tree segmentation results.

[0151] The linear regression equations and accuracy evaluation results between the measured and estimated values ​​of the canopy projection area in the three study areas are shown in the table. As can be seen from the table, the R² of study area A is 0.9976, RMSE is 0.5431 m², and NRMSE is 3.31%; the R² of study area B is 0.9876, RMSE is 0.70 m², and NRMSE is 2.58%; and the R² of study area C is 0.9744, RMSE is 1.527 m², and NRMSE is 4.33%. The extracted canopy projection area results and the measured values ​​in the three study areas also show good correlation and consistency, fully demonstrating the effectiveness of the algorithm in this embodiment. The RMSE of study area B is slightly lower than that of study areas A and C, mainly because the numerical quality of study area B is higher, there are fewer missing data, and the visually identified boundaries are more consistent with the single-tree segmentation results.

[0152] Table 4

[0153]

[0154] Table 5

[0155]

[0156] UAV LiDAR point clouds, providing both horizontal and vertical information, have been successfully used to estimate rubber plantation structural parameters such as canopy height, canopy width, and biomass, offering new avenues for better modeling and management. However, further scale improvements are needed for more accurate rubber plantation modeling and management. Since individual trees are the core component of rubber plantations, successful tree segmentation is crucial for management; therefore, increasing efforts are focused on developing tree segmentation methods suitable for UAV LiDAR point cloud data. However, current tree segmentation methods based on UAV LiDAR data are not ideal in forest scenarios with significant branch overlap, including the specific scenario of the rubber plantation described in this invention. In some studies, the accuracy of these algorithms ranges from 40% to 80%, which is insufficient for individual tree modeling applications. A recently developed method using a top-down region growing approach for coniferous forest tree segmentation achieved an accuracy of 90% in some study plots, but its performance was less than ideal when applied to rubber plantations where tree apexes are not clearly defined. The reason is that, unlike conifers, the crown shape of rubber trees in this invention is irregular, with branches overlapping severely at the crown and the top of the tree not being obvious. Therefore, in rubber forests, the local maximum method is difficult to avoid misidentification of the crown.

[0157] To address the above issues, this embodiment's algorithm starts from the base of the tree trunk. In artificial rubber plantations with sparse low-lying vegetation, the point clouds of the underlying tree trunks are clearer and easier to detect than those of the canopy layer. This algorithm can detect trees covered by larger trees. The study plots showed first branch heights of 3–5 m and tree heights of 16–21 m, indicating that this method is applicable to trees with different stem densities, first branch heights, and canopy sizes. This embodiment's algorithm considers the geometric and topological relationships of the skeleton points. For complex canopy structures, the multi-feature fusion single-tree segmentation algorithm comprehensively considers multiple feature information, capturing more comprehensive feature differences between individual trees and improving segmentation accuracy. This algorithm is suitable for forests with indistinct tree apexes and relatively sparse understory vegetation.

[0158] The second advantage of this invention is the application of the three-dimensional structure of UAV LiDAR. In methods utilizing point cloud mesh transformation, uncertainties introduced by the interpretation method can affect the accuracy of the segmentation results. The algorithm in this embodiment operates directly on the point cloud, thus reducing uncertainty. While the segmentation results from the bottom up may be affected by shrubs in certain areas, the study plot in this embodiment is an artificial rubber plantation with relatively few understory shrubs, therefore the impact on the individual tree segmentation accuracy is minimal.

[0159] This embodiment proposes a bottom-up skeletal growth-based individual tree segmentation method. Three rubber plantation plots of the same variety but with different canopy closures were selected to verify the accuracy of individual tree segmentation in rubber plantation plots with indistinct tree apexes and severe branch overlap. This method benefits from the effectiveness of extracting trunks based on local trunk direction and density differences, the accuracy of locating branch points, and the robustness of segmentation strategies for different canopy points. Experimental results from three plots at the rubber plantation breeding base in Danzhou City, Hainan Province, show that this method can effectively extract individual trees. Compared with other clustering methods, this method significantly reduces the impact of canopy overlap and undersegmentation. Overall, the proposed method achieves comprehensive accuracies of 97.4%, 96.9%, and 90.6% in the three plots, respectively.

[0160] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A bottom-up method for single-tree segmentation of rubber plantation UAV LiDAR point cloud, characterized in that, include: Initial point cloud data of the rubber forest was collected and processed to obtain target point cloud data; The target point cloud data is layered vertically, and the centroids of the clusters are calculated from bottom to top using a density-based point cloud clustering algorithm as the initial hierarchical skeleton points of the rubber forest. The main trunk skeleton points are extracted by utilizing the density difference between the upper and lower branches. Branching points are identified based on the directional consistency of the main trunk skeleton points and the density difference between the main trunk and the crown. The branching points are used as the starting seed points for crown segmentation to separate the main trunk from the crown. By comprehensively considering multiple feature information, fusing multiple features and using different feature combinations to process the skeleton points of different regions of the tree canopy, capturing the feature differences between individual trees, and constructing multiple features to segment individual trees; The process of extracting the main trunk framework points by utilizing the density difference between the upper and lower branches, identifying branch points, and separating the main trunk from the crown includes: Remove non-trunk objects around the trunk, and then use two constraints—the number of neighborhood points and the main clustering direction—to extract the individual trunks. Using the point with the maximum local main elevation as the seed point, a radius neighborhood search is performed, and skeleton points that meet the neighborhood number limit are merged to obtain the main skeleton points; Based on the straight trunk characteristic of rubber trees and the density variation pattern of skeletal points near the branching point, the branching point is determined, and all skeletal points below the branching point are regarded as the main trunk, thus completing the division of the main trunk. The process of using the point with the maximum local main elevation as the seed point to perform a radius neighborhood search, and merging the skeleton points that meet the neighborhood number limit to obtain the main skeleton points includes the following steps: The local backbone skeleton points are sorted by height. The point with the highest elevation in the local backbone skeleton is used as the seed point to search for its neighboring points. ρ1 is set as the density parameter threshold for identifying branch points. When the number of skeleton points to be classified within the search range is less than ρ1, it is added to the skeleton point set qi and a covariance matrix is ​​constructed with the local backbone skeleton points. Principal component analysis is used to obtain the eigenvalues ​​and corresponding eigenvectors of each point. The skeleton points in the local backbone cluster are then analyzed. skeleton points to be classified The formula for the local covariance matrix M is: In the formula, N represents the number of skeleton points and unclassified skeleton points within the local backbone cluster. The center point between the local main skeleton point and the skeleton point to be classified is calculated using the following formula: The eigenvector corresponding to the largest eigenvalue represents the principal direction, and the angle between the principal direction and the skeleton point to be classified to the local center point is... It is obtained through the following formula: In the formula, This is the normalized vector of the skeleton points to be classified and their local center points. The normalized vector in the main direction.

2. The bottom-up rubber forest UAV LiDAR point cloud single-tree segmentation method according to claim 1, characterized in that, The process of collecting and processing initial point cloud data of rubber plantations to obtain target point cloud data includes: The initial point cloud data of the rubber plantation includes ground point and noise point information; The process of processing the initial point cloud data includes, Outlier detection is used for point cloud denoising, and a cloth-simulated filtering algorithm is used to separate ground points from non-ground points. Then, the vegetation points are subjected to elevation normalization to reflect the true height of the trees. In the elevation normalization process, inverse distance weighted interpolation of ground points is used to generate a digital elevation model. The difference between the elevation of the vegetation points and the elevation of the DEM is calculated to generate a normalized point cloud, which is the target point cloud data.

3. The bottom-up rubber forest UAV LiDAR point cloud single-tree segmentation method according to claim 1, characterized in that, The process of stratifying the target point cloud data vertically and using a density-based point cloud clustering algorithm to calculate the centroids of clusters from bottom to top as the hierarchical initial framework points for the rubber forest includes the following steps: Based on the vertical height of the rubber plantation, the plantation is divided into sections from bottom to top according to height intervals. A density-based point cloud clustering algorithm is used to cluster the point cloud of each layer. The center point of each cluster is then determined, which is the cluster center point of the local branch point cloud of each individual rubber tree in each layer. The branch point clouds of different layers are marked according to the cluster center points to obtain the initial skeleton points of the rubber trees in the entire plantation.

4. The bottom-up rubber forest UAV LiDAR point cloud single-tree segmentation method according to claim 3, characterized in that, The formula for obtaining the initial skeleton points of rubber trees in the entire forest by marking the branch point clouds of different layers based on the cluster center points is as follows: in, Let represent the centroid of the j-th cluster in the i-th layer, and let Ni,j represent the number of points in the j-th cluster in the i-th layer. , and Let x, y, and z represent the x, y, and z coordinates of the k-th point in the j-th cluster of the i-th layer, respectively.

5. The bottom-up rubber forest UAV LiDAR point cloud single-tree segmentation method according to claim 1, characterized in that, Constructing multi-feature segmentation for individual trees includes segmenting the skeleton points of the crown core point; The process of segmenting the skeletal points of the tree crown core includes, The crown core point is identified based on the horizontal distance between the crown core point and the branch point, as well as the local density. The set of unsegmented skeleton points is... Nu represents the number of unsegmented skeleton points. A density parameter threshold ρ3 is set to identify tree crown points. Skeleton points whose horizontal distance from the branch point is less than THD1 and whose local density ρ≥p3 are assigned to the tree to which the branch point belongs. The skeleton point is defined as the core skeleton point of the tree crown, and its expression is: The remaining unsegmented skeleton point set consists of non-core points of the tree canopy. Nc is the number of remaining undivided skeleton points, and all crown core points and crown non-core points are obtained by classifying them in this way.

6. The bottom-up rubber forest UAV LiDAR point cloud single-tree segmentation method according to claim 1, characterized in that, Constructing multi-feature segmentation for individual trees includes segmenting the skeleton points of the crown boundary points; The process of segmenting the skeleton points of the tree canopy boundary includes, The canopy boundary is identified by analyzing the local density of each point in the boundary. The neighborhood search radius is set to Rs, and points with fewer than ρ3 neighbors and within a horizontal distance (THD1, THD2) are added to the boundary point set as boundary points. Perform a 3D shortest distance search on the boundary point and assign the point to the tree containing the nearest split point. The formula is as follows: In the formula, H is the skeleton point attribute to be classified, THD1 is the minimum crown radius, THD2 is the maximum crown radius, and ρ3 is the number of skeleton point constraints within the radius neighborhood search range.

7. The bottom-up rubber forest UAV LiDAR point cloud single-tree segmentation method according to claim 1, characterized in that, Constructing multi-feature segmentation for individual trees includes segmenting the skeleton points of disputed points at branch-trunk intersections; The process of segmenting the skeletal points of disputed points at branch intersections includes: Points with a horizontal distance of (THD1, THD2) from the branching point and a local density ρ > ρ2 are considered disputed points where branches overlap severely. The formula expression is: In the formula, H is the attribute of the skeleton point to be classified, THD1 is the minimum crown radius, THD2 is the maximum crown radius, and ρ2 is the density threshold of the disputed skeleton point at the dividing boundary. Based on the characteristic that the branches and trunks of rubber trees extend obliquely upwards at a certain angle, the growth angles between branches and leaves are within a certain range. Therefore, the set of disputed points, including core points and boundary points, is removed. Search for each point in its neighborhood, select the nearest skeleton point with attributes, calculate the angle between the two and compare their sizes. If the calculated angles both conform to the direction of branch growth, then the point belongs to the tree of the skeleton point with the smaller angle. If there is no skeleton point with existing attributes within the neighborhood search range of the disputed point's radius, then it is assigned to the tree of the nearest skeleton point based on the shortest 3D distance.

8. The bottom-up rubber forest UAV LiDAR point cloud single-tree segmentation method according to claim 7, characterized in that, Calculate the angles between the vectors VCE and VDE of the nearest segmented points C(Cx,Cy,Cz), D(Dx,Dy,Dz), and the disputed point E(Ex,Ey,Ez) within the radius neighborhood search range of the disputed point and the Z-axis unit vector P = (0,0,1). and , and The formula expression is: if Less than If the condition is met, then E will be assigned to Tree#1; otherwise, E will be assigned to Tree#2.

Images