Unsupervised branch and leaf separation method for LiDAR forest based on super voxel segmentation

Abstract

The application discloses a LiDAR tree unsupervised branch and leaf separation method based on super voxel segmentation, belongs to the field of fine tree breeding, stand management and forest carbon sink measurement, and comprises the following steps: obtaining ground LiDAR point cloud data by using a ground laser scanner and pre-processing the ground LiDAR point cloud data; performing super voxel segmentation on the point cloud based on a global energy optimization boundary keeping super voxel segmentation algorithm; constructing a connected network topological structure based on point cloud super voxel representative points; detecting wood nodes by calculating the access frequency of nodes through path frequency detection; expanding nodes in the neighborhood of the wood nodes based on curvature and perpendicularity; detecting branch nodes by judging the saliency features of nodes in the same path during path backtracking; and optimizing the branch and leaf separation result based on DBSCAN. The method can improve the situation that branch and leaf nodes are insufficiently detected in path backtracking, so that accurate detection of tree branches and accurate separation of branch and leaf nodes are realized.

Description

Technical Field

[0001] This invention belongs to the fields of superior tree selection, forest stand management, and forest carbon sequestration measurement, and specifically relates to a method for unsupervised branch and leaf separation of trees based on hypervoxel segmentation using LiDAR. Background Technology

[0002] The structure of tree branches and leaves directly affects their photosynthesis and transpiration, thus influencing their relative competitive advantage and growth, and ultimately impacting forest carbon, water, and nutrient cycles. Accurate extraction of tree branch and leaf structure parameters is crucial for the selection of superior tree species, forest stand management, and forest carbon sequestration. Furthermore, understanding the three-dimensional spatial distribution patterns of forest canopy materials and further grasping the changes in branch and leaf growth within the forest and their interactions with the environment is essential for the selection of superior tree varieties, sustainable forest management, and ecosystem quality improvement. Conventional forest resource surveys do not consider detailed tree branch and leaf structures, and their parameter extraction relies primarily on field surveys and statistical analysis, resulting in low efficiency and often limited precision, hindering large-scale practical application. Ground-based LiDAR data can obtain detailed three-dimensional information on forest canopy structure, recording the spatial distribution of branches and leaves within the canopy and their complex combinations. Simultaneously, its superhighpoint density facilitates the accurate identification and separation of tree twigs and leaves. Precise separation of tree branches and leaves helps to achieve high-precision extraction of tree branch and leaf structural parameters.

[0003] In existing technologies, the paper "Improved salient feature-based approach for automatically separating photographic and nonphotosynthetic components within terrestrial LiDAR point cloud data of forest canopies," published in Volume 54 of *IEEE Transactions on Geoscience and Remote Sensing*, is a study on forest branch and leaf separation based on ground-based LiDAR. This study calculates a set of geometric features for each point in the point cloud and uses a Gaussian mixture model (GMM) to classify them to achieve branch and leaf separation. The paper "Single Tree Branch and Leaf Separation Based on Ground Laser Intensity Correction Data," published in Volume 48 of *Chinese Journal of Lasers*, uses a multinomial model to correct laser intensity data and combines color information and a random forest algorithm for branch and leaf separation. The paper "LeWoS: Universal leaf-wood classification method to facilitate the 3D modelling of large tropical trees using terrestrial LiDAR," published in Volume 11 of *Methods in Ecology and Evolution*, uses graph segmentation recursively to separate branch and leaf points, while also introducing class regularization to obtain spatially smooth branch-leaf separation results. However, these methods generally suffer from difficulties in intensity data correction, model parameter determination, and misclassification of fine branches, failing to achieve stable and accurate branch-leaf separation precision, thus hindering the accurate identification and separation of forest tree branches and leaves. Furthermore, no comprehensive and in-depth analysis of the three-dimensional spatial distribution characteristics of branch and leaf points, or a method that comprehensively considers the structural differences between branches and leaves for branch-leaf separation, has been found.

[0004] Furthermore, existing methods are based on statistical features or simple connections between leaves and branches, without employing a multi-level identification strategy that integrates node expansion and path merging, which can easily lead to missed detections of branches and leaves. Summary of the Invention

[0005] In response to the problems mentioned in the background art, this invention proposes a LiDAR unsupervised branch and leaf separation method for trees based on supervoxel segmentation, which achieves accurate detection of tree branches and trunks and accurate separation of branch and leaf nodes.

[0006] Technical Solution: To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] A method for unsupervised branch and leaf separation of trees using LiDAR based on supervoxel segmentation includes the following steps:

[0008] S1: Use a ground-based laser scanner to acquire ground-based LiDAR point cloud data;

[0009] S2: Preprocessing of point cloud data from ground-based lidar;

[0010] S3: A random discrete noise identification method based on hierarchical adaptive voxels is used to remove noise from the point cloud;

[0011] S4: A boundary-preserving supervoxel segmentation algorithm based on global energy optimization is used to perform supervoxel segmentation on the point cloud;

[0012] S5: Constructing a connected network topology based on point cloud hypervoxel representation points;

[0013] S6: Detect wood nodes by calculating the access frequency of the nodes through path frequency detection, and expand the neighborhood of the wood node based on curvature and verticality.

[0014] S7: Detect branch nodes by judging the saliency features of nodes that merge the same path during path backtracking;

[0015] S8: Optimize the branch and leaf separation results based on the DBSCAN clustering algorithm, and evaluate the accuracy of the separation results.

[0016] As a preferred option, the specific content of the preprocessing of ground-based lidar point cloud data in S2 is as follows:

[0017] The original ground-based lidar point cloud data is preprocessed by denoising, ground point filtering, height normalization, and single-tree segmentation to obtain a complete three-dimensional point cloud of a single tree.

[0018] As a preferred option, the specific process in S3 is as follows:

[0019] S31: Slice the 3D point cloud of a single tree into points and calculate the average point spacing in each point cloud slice;

[0020] S32: Voxelize the point cloud slice based on the average point spacing in each slice;

[0021] S33: Traverse all voxels in each point cloud slice to identify random discrete noise points.

[0022] Preferably, in S4, the algorithm process includes merge and swap operations, specifically as follows:

[0023] S41: Merging operation: The boundary-preserving supervoxel segmentation algorithm iteratively merges supervoxels, using the feature distance of the supervoxel representative point to represent the feature distance of the supervoxel, and prioritizes merging points at the trunk position;

[0024] S42: Exchange operation: After determining the representative point, perform an exchange operation on the boundary points, assigning the boundary point to the representative point with the smallest feature distance.

[0025] As a preferred option, the specific process for constructing the point cloud network structure diagram in S5 is as follows:

[0026] S51: Input the center point Q of the supervoxel and the point d with the lowest elevation among them;

[0027] S52: Create a blank graph G;

[0028] S53: Use the KNN algorithm to search for n neighboring points around each point in Q, and store the index and the corresponding point spacing r;

[0029] S54: Add all points in Q to graph G. Starting from d, use r as the weight and selectively add edges and weights to each point in Q and the n neighboring points searched in S53 according to the constraints.

[0030] S55: Output a weighted topology graph G;

[0031] After constructing the undirected graph, Dijkstra's single-source shortest path algorithm is used to calculate the shortest path information from the lowest point in graph G to all other nodes, and the shortest path tree is generated.

[0032] Preferably, in S6, by calculating the frequency of each node visited in all shortest paths of the shortest path tree, nodes whose visit frequency is greater than half of the logarithm of the node with the highest visit frequency are identified as wood nodes, specifically:

[0033]

[0034] in, Represents node P i The number of times a node is visited, max(log(f)) represents the number of nodes with the highest number of visits; Wood represents a wood node; n represents the number of nodes.

[0035] As a preferred approach, the k neighboring points of the wood node are optimized and detected. Neighboring nodes with similar perpendicularity and curvature to the wood node are expanded into wood nodes. Specifically:

[0036]

[0037] Where Verticality() represents the verticality of a node, Curvature() represents the curvature of a node, and δ represents the threshold; P i and P j Both represent nodes.

[0038] Preferably, in S7, path merging is performed using saliency features. The formula for calculating saliency features is as follows:

[0039]

[0040] Where Linearity, Planarity, and Sphericity represent linear characteristics, planar characteristics, and scattering characteristics, respectively; λ1, λ2, and λ3 all represent eigenvalues.

[0041] As a preferred method, in S8, the DBSCAN clustering algorithm is used to segment the identified tree point cloud and leaf point cloud respectively, calculate the ratio of the maximum feature value of each cluster, recover the misclassified tree points and leaf points, and optimize the branch and leaf separation results.

[0042] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0043] (1) Conventional forest resource surveys do not consider the detailed structure of tree branches and leaves. Furthermore, their parameter extraction mainly relies on field surveys and statistical analysis, which is inefficient and often lacks precision, making it difficult to apply and promote in large areas. The ground-based LiDAR data of this invention can obtain detailed three-dimensional structural information of the forest canopy, recording the spatial distribution of branches and leaves within the canopy and their complex combinations. Simultaneously, its superhighpoint density characteristic facilitates the accurate identification and separation of tree twigs, leaves, etc. Through precise separation of tree branches and leaves, high-precision extraction of tree branch and leaf structural parameters can be achieved.

[0044] (2) Previous methods were based on the statistical characteristics or simple connections between leaf and branch points, while this method comprehensively and deeply analyzes the three-dimensional spatial distribution characteristics of branch and leaf points, and then identifies and separates them based on the morphological structural characteristics of branches and leaves. Because this method takes into account the morphological structural differences of tree branches and leaves, it enhances the ability and accuracy of separating tree branches and leaves.

[0045] (3) Previous methods did not employ a multi-level identification strategy that integrates node expansion and path merging, resulting in missed detections of branch and leaf points. This method can fundamentally reduce the probability of misclassification and incorrect classification, thereby improving the accuracy of tree branch and leaf separation.

[0046] (4) The method of this invention considers the impact of random noise on the branch and leaf separation effect and proposes a hierarchical adaptive voxel denoising method. This method removes noise by judging the relationship between neighboring voxels based on the spatial arrangement of noise points, thereby improving the accuracy of branch and leaf separation. Furthermore, this invention not only achieves morphological separation of tree branches and leaves, which is beneficial for extracting fine structural parameters of trees, but also facilitates method transfer (i.e., it can be applied to different forest types in different regions). Verification results show that, compared with other similar remote sensing methods, the overall accuracy of unsupervised branch and leaf separation using this invention is improved by more than 10%. Attached Figure Description

[0047] Figure 1 This is a flowchart of the unsupervised branch and leaf separation of trees based on hypervoxel segmentation according to the present invention;

[0048] Figure 2 This is a schematic diagram of the principle of random discrete noise removal based on hierarchical adaptive voxels, where (a) is a point cloud map of a single tree, (b) is a schematic diagram of noise distribution of a single tree, (c) is a schematic diagram of noise recognition based on hierarchical adaptive voxels, and (d) is a comparison diagram before and after noise removal.

[0049] Figure 3 This is a schematic diagram of the point cloud supervoxel segmentation principle of the present invention, wherein (a) is a schematic diagram of neighboring point merging, and (b) is a schematic diagram of the result based on minimum distance exchange;

[0050] Figure 4 The diagram shows the results of the branch and leaf separation according to the present invention, wherein (a) is the overall effect and a magnified view of the parts after the branch and leaf separation, (b) is the point cloud diagram of the separated leaves, and (c) is the point cloud diagram of the separated branches. Detailed Implementation

[0051] The present invention will be further illustrated below with reference to specific embodiments. These embodiments are implemented based on the technical solutions of the present invention, and it should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0052] The method for unsupervised branch and leaf separation of LiDAR trees based on supervoxel segmentation provided in this embodiment refers to constructing supervoxels from a ground-based LiDAR point cloud, and using the supervoxels to represent points to construct an undirected graph of individual trees for shortest path analysis to obtain a shortest path tree; then, detecting tree nodes by calculating the access frequency of nodes through path frequency detection, and performing "node expansion" on the neighborhood of tree nodes based on the two constraints of curvature and verticality; finally, detecting branch nodes by judging the salient features of nodes merging the same path during path backtracking, thereby improving the insufficient detection of branch nodes in path backtracking, so as to achieve accurate detection of tree branches and accurate separation of branch and leaf nodes.

[0053] The specific steps are as follows:

[0054] S1: Acquire high-density lidar data via a ground-based platform;

[0055] A plantation of *Liriodendron tulipifera* was selected as the research object. Ground-based LiDAR point cloud data was acquired using a ground-based laser scanner, RIEGL VZ-400i (RIEGLLaser Measurement Systems GmbH, Horn, Austria). The scanning angle was 100° (vertical) × 360° (horizontal), the scanning rate was 500,000 points / second, and the scanning accuracy was 5 mm / 100m. A multi-station, multi-angle scanning method was employed during data acquisition, with 19 scans conducted at 9 locations.

[0056] S2: Preprocessing of point cloud data from ground-based lidar;

[0057] Preprocessing is performed on the original ground-based lidar point cloud data, including denoising, ground point filtering, height normalization, and single-tree segmentation.

[0058] First, noise points are removed using a noise point identification method based on neighborhood distance, and ground point filtering is performed using a cloth simulation filtering algorithm.

[0059] Then, by calculating the average height of ground points within each pixel, a digital elevation model (DEM) with a resolution of 0.1m is generated, and it is used to normalize the height of the denoised point cloud.

[0060] Finally, the normalized point cloud was segmented using the Shortest Path Comparison (CSP) algorithm combined with visual interpretation to obtain a complete 3D point cloud of a single tree. The point cloud information of a single tree is shown in Table 1 below.

[0061] Table 1 Summary Table of Single Tree Point Cloud Information

[0062] serial number Branch type Points Tree height (m) Diameter at breast height (cm) Average point spacing / m 1 Simple 597399 21.52 17.8 0.0059 2 Simple 632257 23.355 20.8 0.0054 3 Simple 664178 25.142 25.5 0.0067 4 Simple 717921 22.135 32.1 0.0039 5 Simple 766925 24.741 27.4 0.0074 6 Simple 952533 25.601 26.7 0.0065 7 complex 723175 25.559 34.6 0.006 8 complex 586990 24.962 33.4 0.0079 9 complex 963192 25.054 27.9 0.0066 10 complex 1495627 26.183 37.1 0.005 11 complex 1633618 26.512 32.6 0.0049 12 complex 1911356 26.347 30.7 0.0051

[0063] S3: A random discrete noise identification method based on hierarchical adaptive voxels is used to remove noise from the point cloud;

[0064] S31: First, the single-tree 3D point cloud is vertically divided into point cloud slices with a thickness of 0.5m from bottom to top, and the average point spacing in each point cloud slice is calculated.

[0065] S32: Then, the point cloud slice is voxelized using the average point spacing in each point cloud slice;

[0066] S33: Finally, iterate through all voxels in each point cloud slice to identify random discrete noise points;

[0067] If a voxel is not empty, and the total number of point clouds contained in the neighboring voxels (i.e., the surrounding 26 voxels) of the voxel is less than or equal to 1, then the points contained in the voxel are considered as random discrete noise points.

[0068] S4: Point cloud supervoxel segmentation;

[0069] The boundary-preserving supervoxel segmentation (BPSS) algorithm based on global energy optimization is used to perform supervoxel segmentation on the point cloud.

[0070] The algorithm uses energy descent to transform the supervoxel segmentation problem into a subset selection problem. Specifically, N points are represented by K representative points, and the subsets corresponding to these representative points are the resulting supervoxels. The algorithm consists of merging and swapping operations, and does not require initializing seed points; it operates directly on the point cloud rather than on individual voxels. The objective function is:

[0071]

[0072] The objective function E(Z) takes the form of an energy function, with the first term representing the feature distance constraint and the second term representing the number constraint; C(Z) represents the number of representative points; K represents the desired number of representative points, which can be calculated using the supervoxel resolution R; λ is a regularization parameter that balances the two terms, and this value is automatically evaluated using an adaptive strategy that accelerates optimization using local information. The initial value is set to the median of the minimum feature distances between each point and its neighbors, and then doubled in each iteration; N represents the number of points; D(P) i ,P j ) represents point P i and point P j Feature distance; Z ij Indicates non-representative point P j Point P can be represented i It indicates that the following conditions are met:

[0073]

[0074] S41: The algorithm first initializes each point as a representative point and obtains the adjacency relationship; in the merging phase, it merges adjacent representative points r... i and r j Calculate the energy loss Δ of merging according to the following formula. If Δ>0, merge and update the adjacency relationship of the representative point; otherwise, proceed to the next iteration.

[0075] Δ=λ-c j D(r i ,r j )

[0076] Where λ represents the weighting coefficient, c j For example, point r jThe corresponding number of points contained in the supervoxel; D(r i ,r j (r) represents the point. i With r j The feature distance is calculated using the following formula:

[0077]

[0078] Where, n p and n q Let p and q be the normal vectors corresponding to p and q; R represents the desired resolution.

[0079] The BPSS algorithm is used to iteratively merge supervoxels. The feature distance of the supervoxel is represented by the feature distance of the point representing the supervoxel, thereby reducing the computational complexity. The algorithm will prioritize merging points in smooth regions, i.e. points at the trunk position.

[0080] S42: After determining K representative points through the merging operation, the boundary points are swapped, assigning each boundary point to the representative point with the smallest feature distance, i.e., the two adjacent representative points r. i With r j If D(P) i ,r j ) <D(P i ,r i ), then P i Reassigned to r j The corresponding hypervoxel. Swap operations can better obtain the boundaries of the hypervoxel.

[0081] S5: Construct a connected network topology based on the hypervoxel representation points of the point cloud;

[0082] Let the graph be G = (V, E), where V represents a node (i.e., a supervoxel representing a point), and E is the set of edges connecting adjacent points. To better represent the growth direction of trees, this invention uses the distance between points as the weight of the edges, constructing the edge constraint Edge(P)... i ,P j )for:

[0083]

[0084] Among them, dist(P i ,P j ) is P i With P j The Euclidean distance between them, d s This is the minimum threshold for constructing edges.

[0085] The specific steps for constructing a point cloud network structure diagram are as follows:

[0086] (1) Input the center point Q of the supervoxel and the point d with the lowest elevation among them;

[0087] (2) Create a blank graph G;

[0088] (3) Use the KNN algorithm to search for n neighboring points around each point in Q, and store the index and the corresponding point spacing r. Here, the point spacing is calculated using the Euclidean distance method, and n is set to 15.

[0089] (4) Add all points in Q to graph G. Starting from d, use r as weight and selectively add edges and weights to each point in Q and the n neighboring points searched in step (3) according to the constraints.

[0090] (5) Output the weighted topology graph G.

[0091] After constructing the undirected graph, shortest path analysis is used to detect the trunk and branches. Dijkstra's single-source shortest path algorithm is used to calculate the shortest path information from the lowest point in graph G to all other nodes, generating a shortest path tree.

[0092] S6: Path frequency detection and "node expansion" detection backbone;

[0093] By calculating the frequency of each node visited in all shortest paths of the shortest path tree, nodes with higher frequencies are identified as "wooden nodes," represented as follows:

[0094]

[0095] in, For node P i The number of times a node is visited, max(log(f)) is the highest number of times a node is visited. This process identifies nodes whose number of visits is greater than half of the highest number of times a node is visited as wood seed nodes. Wood represents a wood node.

[0096] This invention proposes a "node expansion" method based on geometric features, based on the phenomenon that the verticality and curvature changes relatively little at the trunk and branches of a tree, while the verticality and curvature differ significantly between the trunk and leaves. The method optimizes the detection of k neighboring points of a wood node, and "expands" neighboring nodes with similar verticality and curvature to the wood seed node into wood nodes, represented as follows:

[0097]

[0098] Where Verticality() represents the verticality of a node, Curvature() represents the curvature of a node, and δ is a threshold. δ is typically set to 0.075. The verticality and curvature of each node are calculated by the supervoxel represented by that node. Since eigenvalues ​​can represent the variability of data on orthogonal projection axes, they are used as the basis for quantitative measurement of the spatial arrangement of the point cloud. A covariance matrix Cov(p) is constructed for the supervoxel represented by that node. i Specifically:

[0099]

[0100] Where p is the center of the segment, and n is the total number of points in the segment. i This represents the i-th node; represents the center point of the supervoxel; T represents the transpose operation.

[0101] Principal component analysis (PCA) is used to calculate eigenvalues ​​and eigenvectors for the calculation of perpendicularity and curvature features. Curvature is calculated from the surface variation of the point set. The methods for calculating perpendicularity and curvature are as follows:

[0102]

[0103] Wherein, Verticality represents the verticality of the node, Curvature represents the curvature of the node, λ1, λ2, and λ3 are the eigenvalues ​​calculated by principal component analysis (λ1≥λ2≥λ3), e3 is the eigenvector corresponding to λ3, and e3[2] is the third component of the eigenvector e3. By using this "node expansion" method to detect the neighborhood of the wood node, the recognition rate of the wood node can be improved.

[0104] S7: Path backtracking and "path merging" detection branch;

[0105] This invention comprehensively considers the similarity between leaves and twigs, as well as the complexity of path backtracking, and performs "path merging" through salient features. The salient features are calculated using principal component analysis, similar to the methods used to calculate verticality and curvature. For each point in the supervoxel containing the node, the covariance matrix is ​​calculated, and eigenvalues ​​λ1, λ2, and λ3 (λ1 ≥ λ2 ≥ λ3) are obtained through singular value decomposition. The salient features can then be calculated based on these eigenvalues, specifically:

[0106]

[0107] Among them, Linearity, Planarity, and Sphericity represent linear characteristics, planar characteristics, and scattering characteristics, respectively.

[0108] To address the problem of assessing relative significance levels, this invention constructs the SoD index to enhance the significance levels between features. SoD is calculated based on hypervoxels, specifically as follows:

[0109] SoD=Linearity+(1-Linearity)×[Linearity-max(Planarity,Sphericity)]

[0110] In this context, Linearity, Planarity, and Sphericity represent linear, planar, and scattering features, respectively; the SoD value ranges from -1 to 1. When SoD < 0, it indicates that one of the other two significant features is more significant; when SoD > 0, it indicates that the linear feature is more significant. The larger the SoD value, the greater the likelihood that the node is dominated by the linear feature. Wood nodes are detected by setting an empirical threshold ξ, and nodes with SoD > ξ are detected as wood seed nodes. ξ is typically set to 0.9. For each shortest path extracted from the shortest path tree, the process traces back from the endpoint to the base point. If a wood seed node exists in the path, it is merged with its adjacent nodes on the same path. The SoD index of the merged segment is then calculated to determine if it is larger than the SoD index before merging. This process merges nodes on paths within the same branch, thus detecting wood nodes on the same branch.

[0111] Since the point clouds identified by path frequency detection and path backtracking are both branch point clouds, merging the two identified point clouds yields the tree nodes of the entire tree point cloud. To avoid data redundancy, duplicate points need to be removed after merging. These tree nodes are generated from supervoxel representative points. To obtain all branch points of the entire tree point cloud, each point in the original point cloud needs to be mapped to a supervoxel representative point using the category labels of the supervoxel segmentation results. This maps the tree nodes obtained in voxel space back into the original point cloud, resulting in the branch point cloud of the entire tree.

[0112] S8: Optimize the branch and leaf separation results based on the DBSCAN clustering algorithm, and evaluate the accuracy of the separation results;

[0113] Based on the characteristic that the density of tree branch points is generally greater than that of leaf points, the DBSCAN clustering algorithm is used to segment the identified tree point clouds and leaf point clouds separately. The ratio of the maximum eigenvalue of each cluster is calculated using the same method as for calculating salient features. Clusters with a maximum eigenvalue ratio greater than 0.75 are retained to recover misclassified tree and leaf points, thereby optimizing the branch and leaf separation results.

[0114] After branch and leaf separation, the accuracy of the separation results was evaluated using four indicators: overall accuracy (OA), F1-score of leaf points, and Kappa coefficient. The accuracy evaluation of the branch and leaf separation results is shown in Table 2 below. Simultaneously, Type I and Type II errors of the branch and leaf separation results were calculated to further analyze the errors in the classification results.

[0115]

[0116] Where p represents precision, r represents recall, and p o Indicates proportional consistency, p e p represents the probability of random consistency. o p e The calculation formula is:

[0117]

[0118] The formulas for calculating Type I error T1 and Type II error T2 are as follows:

[0119]

[0120] Among them, T w To correctly classify the number of points as branches, F w T represents the number of points misclassified as branches. l To correctly classify the points as leaves, F l The number of points misclassified as leaves.

[0121] Table 2. Accuracy Evaluation of Branch and Leaf Separation Results

[0122] Branch type OA F1-wood F1-leaf kappa Simple 0.9414 0.9215 0.9470 0.8684 complex 0.9415 0.9107 0.9540 0.8647

[0123] This invention's method comprehensively and thoroughly analyzes the three-dimensional spatial distribution characteristics of branch and leaf points, and performs targeted identification and separation based on the morphological structural features of branches and leaves. Simultaneously, it employs a multi-level identification strategy that integrates node expansion and path merging, effectively reducing missed detections of branch and leaf points. Implementation case results show that the overall accuracy of branch and leaf separation using this invention's method is better than 94%, representing an improvement of approximately 10% compared to conventional methods.

[0124] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for unsupervised branch and leaf separation of LiDAR forest based on super voxel segmentation, characterized in that: Includes the following steps: S1: Use a ground-based laser scanner to acquire ground-based LiDAR point cloud data; S2: Preprocessing of point cloud data from ground-based lidar; S3: A random discrete noise identification method based on hierarchical adaptive voxels is used to remove noise from the point cloud; S4: A boundary-preserving supervoxel segmentation algorithm based on global energy optimization is used to perform supervoxel segmentation on the point cloud; S5: Constructing a connected network topology based on point cloud hypervoxel representation points; S6: Detect wood nodes by calculating the access frequency of the nodes through path frequency detection, and expand the neighborhood of the wood node based on curvature and verticality. By calculating the frequency of each node visited in all shortest paths of the shortest path tree, nodes whose visit frequency is greater than half the logarithm of the node with the highest visit frequency are identified as "wooden nodes". Specifically: ； wherein, representing a node the number of times, ) represents the pair of nodes with the highest number of visits; representing a tree node; n represents the number of nodes; Optimize the detection of k neighboring points of the wood node, and expand neighboring nodes with similar perpendicularity and curvature to the wood node as wood nodes, specifically: ； in, () indicates the perpendicularity of a node. () represents the curvature of a node. Indicates the threshold; and Both represent nodes; S7: Detect branch nodes by judging the saliency features of nodes that merge the same path during path backtracking; S8: Optimize the branch and leaf separation results based on the DBSCAN clustering algorithm, and evaluate the accuracy of the separation results.

2. The method for unsupervised branch and leaf separation of trees based on supervoxel segmentation using LiDAR according to claim 1, characterized in that: In S2, the specific preprocessing of ground-based lidar point cloud data is as follows: The original ground-based lidar point cloud data is preprocessed by denoising, ground point filtering, height normalization, and single-tree segmentation to obtain a complete three-dimensional point cloud of a single tree.

3. The method for unsupervised branch and leaf separation of trees based on supervoxel segmentation using LiDAR according to claim 1, characterized in that: In S3, the specific process is as follows: S31: Slice the 3D point cloud of a single tree into points and calculate the average point spacing in each point cloud slice; S32: Voxelize the point cloud slice based on the average point spacing in each slice; S33: Traverse all voxels in each point cloud slice to identify random discrete noise points.

4. The method for unsupervised branch and leaf separation of trees based on supervoxel segmentation using LiDAR according to claim 1, characterized in that: In S4, the algorithm process includes merge and swap operations, specifically as follows: S41: Merging operation: The boundary-preserving supervoxel segmentation algorithm iteratively merges supervoxels, using the feature distance of the supervoxel representative point to represent the feature distance of the supervoxel, and prioritizes merging points at the trunk position; S42: Exchange operation: After determining the representative point, perform an exchange operation on the boundary points, assigning the boundary point to the representative point with the smallest feature distance.

5. The method for unsupervised branch and leaf separation of trees based on supervoxel segmentation using LiDAR according to claim 1, characterized in that: In S5, the specific process of constructing the point cloud network structure diagram is as follows: S51: Input the center point Q of the supervoxel and the point d with the lowest elevation among them; S52: Create a blank graph G; S53: Use the KNN algorithm to search for n neighboring points around each point in Q, and store the index and the corresponding point spacing r; S54: Add all points in Q to graph G. Starting from d, use r as the weight and selectively add edges and weights to each point in Q and the n neighboring points searched in S53 according to the constraints. S55: Output a weighted topology graph G; After constructing the undirected graph, Dijkstra's single-source shortest path algorithm is used to calculate the shortest path information from the lowest point in graph G to all other nodes, and the shortest path tree is generated.

6. The method for unsupervised branch and leaf separation of trees based on supervoxel segmentation using LiDAR according to claim 1, characterized in that: In S7, path merging is performed using saliency features. The formula for calculating saliency features is as follows: ； Among them, Linearity, Planarity, and Sphericity represent linear characteristics, planar characteristics, and scattering characteristics, respectively; , , All of these represent eigenvalues.

7. The method for unsupervised branch and leaf separation of trees based on supervoxel segmentation using LiDAR according to claim 1, characterized in that: In S8, the DBSCAN clustering algorithm is used to segment the identified tree point cloud and leaf point cloud respectively, calculate the ratio of the maximum feature value of each cluster, recover the misclassified tree points and leaf points, and optimize the branch and leaf separation results.

Images