Super voxel segmentation-based LiDAR forest unsupervised branch and leaf separation method

Abstract

The invention discloses a super voxel segmentation-based LiDAR forest unsupervised branch and leaf separation method, and belongs to the field of excellent forest breeding, forest stand operation management and forest carbon sink metrology, and the method comprises the steps: obtaining foundation LiDAR point cloud data through a foundation laser scanner, and carrying out the preprocessing of the data; performing super voxel segmentation on the point cloud based on a boundary preserving super voxel segmentation algorithm of global energy optimization; constructing a connected network topological structure based on the point cloud super voxel representative points; the access frequency of nodes is calculated through path frequency detection to detect wood nodes, and node expansion is performed on neighborhoods of the wood nodes based on curvature and perpendicularity; detecting branch nodes by judging and combining significance features of nodes of the same path during path backtracking; optimizing a branch and leaf separation result based on DBSCAN; the method provided by the invention can improve the condition of insufficient branch node detection in path backtracking so as to realize accurate detection of forest branches and accurate separation of branch and leaf nodes.

Description

Technical Field

[0001] The present invention belongs to the fields of fine tree breeding, forest stand management and forest carbon sink measurement, and specifically relates to a method for unsupervised branch and leaf separation of trees using LiDAR based on supervoxel segmentation. Background Art

[0002] The branch and leaf structure of trees directly influences their photosynthesis and evapotranspiration, which in turn affects their relative competitive advantage and growth, ultimately impacting forest carbon, water, and nutrient cycles. Accurately extracting tree branch and leaf structural parameters is crucial for selecting high-quality trees, forest stand management, and forest carbon sequestration. Understanding the three-dimensional spatial distribution of tree canopy material and further understanding the changes in branch and leaf growth within the forest and their interactions with the environment are crucial for selecting high-quality tree varieties, sustainable forest management, and improving ecosystem quality and efficiency. Conventional forest resource surveys fail to consider detailed tree branch and leaf structure. Furthermore, parameter extraction relies primarily on field surveys and statistical analysis, which are inefficient and often lack sufficient precision, making them difficult to implement across large areas. Ground-based LiDAR data can capture detailed three-dimensional information about the forest canopy's structure, documenting the spatial distribution of branches and leaves within the canopy and their complex relationships. Furthermore, its extremely high point density facilitates the precise identification and separation of tree twigs and leaves. The precise separation of tree branches and leaves helps to achieve high-precision extraction of tree branch and leaf structural parameters.

[0003] In the existing technology, the "Improved salient feature-based approach for automatically separating photosynthetic 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 separating tree branches and leaves 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 tree branch and leaf separation. The "Single Tree Branch and Leaf Separation Based on Ground Laser Intensity Correction Data" published in Volume 48 of China Lasers uses a polynomial model to correct laser intensity data and combines color information with a random forest algorithm to separate branches and leaves. "LeWoS: Auniversalleaf-wood classification method to facilitate the 3D modelling of largetropical trees using terrestrial LiDAR", published in Volume 11 of "Methods in Ecology and Evolution", recursively uses graph segmentation to separate branch points from leaf points, and also introduces class regularization to obtain spatially smooth branch and leaf separation results. However, the above methods generally have problems such as difficulty in correcting intensity data, difficulty in determining model parameters, and easy misclassification of thin branches. They cannot obtain stable and accurate branch and leaf separation accuracy, and are therefore not conducive to the accurate identification and separation of tree branches and leaves. At the same time, there is no method that comprehensively and in-depth analyzes the three-dimensional spatial distribution characteristics of branch points and leaf points, and comprehensively considers the differences in the structural characteristics of branches and leaves to separate branches and leaves.

[0004] Moreover, existing methods are all based on the statistical characteristics or simple connection relationships of leaves and branches, and do not adopt a multi-level recognition strategy that integrates node expansion and path merging, which easily leads to missed detection of branches and leaves. Summary of the Invention

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

[0006] Technical solution: In order to solve the above technical problems, the technical solution adopted by the present invention is as follows:

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

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

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

[0010] S3: Use the random discrete noise recognition method based on hierarchical adaptive voxels to remove noise from point clouds;

[0011] S4: Supervoxel segmentation of point clouds using a boundary-preserving supervoxel segmentation algorithm based on global energy optimization;

[0012] S5: Constructing a connected network topology based on point cloud supervoxel representations;

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

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

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

[0016] Preferably, in S2, the specific content of preprocessing the ground-based lidar point cloud data is:

[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 single tree three-dimensional point cloud.

[0018] Preferably, in S3, the specific process is:

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

[0020] S32: voxelizing the slice point cloud according to the average point spacing in each point cloud 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 merging and swapping operations, and the specific process is:

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

[0024] S42: Exchange operation: After the representative point is determined, the boundary points are exchanged and the boundary points are assigned to the representative point with the smallest distance from its feature.

[0025] Preferably, in S5, the specific process of constructing the point cloud network structure diagram is:

[0026] S51: Input the supervoxel center point Q and the minimum elevation point d;

[0027] S52: Create a blank graph G;

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

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

[0030] S55: Output weighted topological network graph G;

[0031] After completing the construction of the undirected graph, the Dijkstra single-source shortest path algorithm is used to calculate the shortest path information from the lowest point in the graph G to all other nodes to generate the shortest path tree.

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

[0033]

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

[0035] As a preference, the k neighboring points of the wood node are optimized and detected, and the neighboring nodes with similar verticality and curvature to the wood node are expanded into wood nodes, specifically:

[0036]

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

[0038] Preferably, in S7, the paths are merged by using significant features, and the significant features are calculated using the following formula:

[0039]

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

[0041] Preferably, in S8, the identified wood point cloud and leaf point cloud are segmented separately using the DBSCAN clustering algorithm, the ratio of the maximum eigenvalue of each cluster is calculated, the misclassified wood points and leaf points are restored, and the branch and leaf separation results are optimized.

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

[0043] (1) Conventional forest resource surveys do not take into account the fine structure of tree branches and leaves. At the same time, the extraction of survey parameters mainly relies on field surveys and statistical analysis, which are inefficient and often not very detailed, making it difficult to promote practical application in large areas. The ground-based LiDAR data of the present 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 combination relationships. At the same time, its ultra-high point density characteristics help to accurately identify and separate tree twigs, leaves, etc. By accurately separating tree branches and leaves, it helps to achieve high-precision extraction of tree branch and leaf structural parameters.

[0044] (2) Previous methods were based on the statistical characteristics or simple connection relationships between leaves and branches. However, this method comprehensively and deeply analyzes the three-dimensional spatial distribution characteristics of branches and leaves, and then conducts targeted identification and separation based on the morphological structural characteristics of branches and leaves. Because this method takes into account the morphological structural differences between branches and leaves, it enhances the ability and accuracy of branch and leaf separation.

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

[0046] (4) The method of the present invention takes into account the influence of random noise on the effect of branch and leaf separation, and proposes a hierarchical adaptive voxel denoising method, which removes noise by judging the relationship between neighboring voxels based on the spatial arrangement of noise, thereby improving the accuracy of branch and leaf separation. At the same time, the present invention not only realizes the separation of tree branches and leaves from the perspective of morphological structure, which is beneficial to the extraction of fine structural parameters of trees, but also facilitates method transplantation (that is, it can also be applied to different forest types in different regions). The verification results show that the unsupervised separation of tree branches and leaves by the present invention is more than 10% higher than the overall accuracy of branch and leaf separation using other similar remote sensing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart of unsupervised tree branch and leaf separation based on supervoxel segmentation of the present invention;

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

[0049] Figure 3 1 is a schematic diagram of the point cloud supervoxel segmentation principle of the present invention, wherein (a) is a schematic diagram of adjacent point merging, and (b) is a schematic diagram of the result based on minimum distance exchange;

[0050] Figure 4 It is the branch and leaf separation result diagram of the present invention, wherein, (a) is the overall effect and local magnification diagram after branch and leaf separation, (b) is the leaf point cloud diagram after separation, and (c) is the branch point cloud diagram after separation. DETAILED DESCRIPTION

[0051] The present invention will be further illustrated below with reference to specific examples. The examples are implemented based on the technical solutions of the present invention. It should be understood that these examples are only used to illustrate the present invention and are not used 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 a method of constructing supervoxels for ground-based LiDAR point clouds, and using supervoxel representative points to construct an undirected graph of a single tree to perform shortest path analysis to obtain a shortest path tree; then, the tree nodes are detected by calculating the access frequency of the nodes through path frequency detection, and the neighborhood of the tree nodes is "node expanded" based on the two constraints of curvature and verticality; finally, the branch nodes are detected by judging the significant features of the nodes merged on the same path during path backtracing, thereby improving the situation of insufficient branch node detection in path backtracing, 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: Collect high-density lidar data through a ground-based platform;

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

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

[0057] The original ground-based lidar point cloud data is preprocessed by denoising, ground point filtering, height normalization, and single tree segmentation.

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

[0059] Then, a digital elevation model (DEM) with a resolution of 0.1 m was generated by calculating the average height of the ground points in each pixel, and the denoised point cloud was height normalized using it.

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

[0061] Table 1 Summary 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: Use the random discrete noise recognition method based on hierarchical adaptive voxels to remove noise from point clouds;

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

[0065] S32: Then, voxelize the slice point cloud according to the average point spacing in each point cloud slice;

[0066] S33: Finally, all voxels in each point cloud slice are traversed 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 of the voxel (i.e., the surrounding 26 voxels) is less than or equal to 1, the points contained in the voxel are regarded as random discrete noise points.

[0068] S4: point cloud supervoxel segmentation;

[0069] The point cloud is segmented into supervoxels using the boundary-preserving supervoxel segmentation (BPSS) algorithm based on global energy optimization.

[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 subset corresponding to the representative points is the supervoxel obtained by segmentation. The algorithm consists of merging and swapping operations, does not require initialization of seed points, and operates directly on the point cloud rather than on voxels. The objective function is:

[0071]

[0072] Among them, the objective function E(Z) adopts the form of energy function, the first term is the feature distance constraint, and the second term is the quantity constraint; C(Z) is the number of representative points; K is the expected number of representative points, which can be calculated by the supervoxel resolution R; λ is the regularization parameter that weighs the first and last two terms, and the value is automatically evaluated by the local information acceleration optimization adaptive strategy. The initial value is set to the median of the minimum feature distance between each point and its neighboring points, 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 Characteristic distance of Z ij Represents a non-representative point P j Can be represented by point P i 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 stage, the adjacent representative points r i and r j The energy loss Δ of the merger is calculated according to the following formula. If Δ>0, the merger is performed and the adjacency relationship of the representative points is updated. Otherwise, the next round of iteration is entered.

[0075] Δ=λ-c j D(r i ,r j )

[0076] Among them, λ represents the weight coefficient, c j is the representative point r jThe number of points contained in the corresponding supervoxel; D(r i ,r j ) is the representative point r i With r j The characteristic distance is calculated by the following formula:

[0077]

[0078] Among them, n p and n q is the normal vector corresponding to p and q; R represents the desired resolution.

[0079] The BPSS algorithm is used to iteratively merge supervoxels, and the characteristic distance of the supervoxel representative point is used to represent the characteristic distance of the supervoxel, thereby reducing the computational complexity. The algorithm will give priority to merging points in the smooth area, that is, points at the trunk position.

[0080] S42: After the merging operation determines K representative points, the boundary points are exchanged and assigned to the representative point with the smallest distance from its feature, that is, the two adjacent representative points r i With r j , if D(P i ,r j ) <D(P i ,r i ), then P i Reassign to r j The corresponding supervoxel. The exchange operation can better obtain the boundary of the supervoxel.

[0081] S5: Construct a connected network topology based on point cloud supervoxel representative points;

[0082] Suppose the graph is G = (V, E), where V represents the node, i.e., the supervoxel represents the point, and E is the set of edges in the graph, connecting adjacent points. In order to better indicate the growth direction of the tree, the present invention uses the distance between points as the weight of the edge and constructs 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 s is the minimum threshold for building an edge.

[0085] The specific steps to construct the point cloud network structure diagram are as follows:

[0086] (1) Input the supervoxel center point Q and the minimum elevation point d;

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

[0088] (3) Use the KNN algorithm to search for n adjacent 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 the graph G, starting from d, using r as the weight, and selectively add edges and weights to each point in Q and the n adjacent points searched in step (3) according to the constraints;

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

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

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

[0093] By calculating the frequency of each node visited in all the shortest paths in the shortest path tree, the nodes with higher frequency are identified as wood nodes, which can be expressed as:

[0094]

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

[0096] Based on the fact that the verticality and curvature of a tree trunk and branches vary little, while the verticality and curvature between the trunk and leaves vary greatly, this paper proposes a "node expansion" method based on geometric features. The k neighboring points of a tree node are optimized and detected, and neighboring nodes with similar verticality and curvature to the tree seed node are "expanded" into tree nodes, expressed as:

[0097]

[0098] Among them, Verticality() represents the verticality of the node, Curvature() represents the curvature of the node, and δ is the threshold. δ can generally be set to 0.075. The verticality and curvature of each node are calculated for the supervoxel represented by the node. Since the eigenvalue can represent the variability of the data on the orthogonal projection axis, the eigenvalue is used as the quantitative measurement basis for the spatial arrangement of the point cloud. The covariance matrix Cov(p) is constructed for the supervoxel represented by the 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 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 the eigenvalues ​​and eigenvectors for the calculation of verticality and curvature features. The curvature is calculated from the surface variation of the point set. The verticality and curvature are calculated as follows:

[0102]

[0103] Where 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. Detecting the neighborhood of wood nodes by this "node expansion" method can improve the recognition rate of wood nodes.

[0104] S7: path backtracking and “path merging” detection branch;

[0105] The present invention comprehensively considers the similarity between leaves and twigs and the complexity of path backtracking, and performs "path merging" through significant features. The significant features are calculated through principal component analysis, similar to the calculation method of verticality and curvature. For each point in the supervoxel where the node is located, the covariance matrix is ​​calculated, and the eigenvalues ​​λ1, λ2, λ3 (λ1≥λ2≥λ3) are obtained through singular value decomposition. The significant features can be calculated based on the eigenvalues, specifically:

[0106]

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

[0108] To solve the problem of evaluating relative significance levels, the present invention constructs a SoD index to enhance the significance level between features. SoD is calculated based on supervoxels, specifically:

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

[0110] Linearity, Planarity, and Sphericity represent linear, planar, and scattering features, respectively; SoD ranges from -1 to 1. When SoD < 0, one of the other two significant features is more significant; when SoD > 0, the linear feature is more significant. The larger the SoD value, the more likely the node is dominated by the linear feature. A threshold ξ is set to detect wood nodes, and nodes with SoD > ξ are identified as wood seed nodes. ξ is typically set to 0.9. For each shortest path extracted from the shortest path tree, backtracking from the endpoint to the base node, if a wood seed node exists in the path, merge the seed node with the adjacent nodes on the same path, and calculate whether the SoD index of the merged segment is greater than the SoD index before merging. This process is designed to merge nodes on the same branch of the path and detect wood nodes on the same branch.

[0111] Since the point clouds identified by path frequency detection and path tracing are both branch point clouds, merging the two identified point clouds yields the tree nodes for the entire tree point cloud. To avoid data redundancy, duplicate points must be removed after merging. The tree nodes are generated from supervoxel representative points. To obtain all branch points for the entire tree point cloud, each point in the original point cloud is mapped to a supervoxel representative point using the category labels from the supervoxel segmentation results. This allows the tree nodes obtained in voxel space to be mapped back to the original point cloud, yielding the entire tree's branch point cloud.

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

[0113] Based on the fact 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 and leaf point clouds separately. The ratio of the maximum eigenvalues ​​of each cluster is calculated using the same method used to calculate significant 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 branch and leaf 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. At the same time, the Class I and Class II errors of the branch and leaf separation results were calculated to further analyze the errors of the classification results.

[0115]

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

[0117]

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

[0119]

[0120] Among them, T w is the number of points correctly classified as branches, F w is the number of points misclassified as branches, T l is the number of points correctly classified as leaves, F l is 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] The method of the present invention comprehensively and in-depth analyzes the three-dimensional spatial distribution characteristics of branch and leaf points, and then performs targeted identification and separation based on the morphological structural characteristics of branches and leaves. Furthermore, a multi-level recognition strategy that integrates node expansion and path merging is employed to effectively reduce missed branch and leaf points. Case study results show that the overall accuracy of the method for branch and leaf separation is better than 94%, an improvement of approximately 10% compared to conventional methods.

[0124] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for unsupervised branch and leaf separation of LiDAR trees based on supervoxel segmentation, characterized by: The following steps are involved: S1: Use a ground-based laser scanner to obtain ground-based LiDAR point cloud data; S2: Preprocessing of ground-based lidar point cloud data; S3: Use the random discrete noise recognition method based on hierarchical adaptive voxels to remove noise from point clouds; S4: Supervoxel segmentation of point clouds using a boundary-preserving supervoxel segmentation algorithm based on global energy optimization; S5: Constructing a connected network topology based on point cloud supervoxel representations; S6: Detect wooden nodes by calculating the access frequency of nodes through path frequency detection, and expand the neighborhood of wooden nodes based on curvature and verticality; S7: Detecting branch nodes by judging the significant features of nodes that merge the same path during path backtracking; S8: Optimize the branch-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 using LiDAR based on supervoxel segmentation according to claim 1, characterized in that: In S2, the specific content of ground-based lidar point cloud data preprocessing 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 single tree three-dimensional point cloud.

3. The method for unsupervised branch and leaf separation of trees using LiDAR based on supervoxel segmentation according to claim 1, characterized in that: In S3, the specific process is: S31: Slice the 3D point cloud of a single tree and calculate the average point spacing in each point cloud slice; S32: voxelizing the slice point cloud according to the average point spacing in each point cloud 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 using LiDAR based on supervoxel segmentation according to claim 1, characterized in that: In S4, the algorithm process includes merging and swapping operations. The specific process is: S41: Merging operation: The boundary-preserving supervoxel segmentation algorithm iteratively merges supervoxels, and uses the characteristic distance of the supervoxel representative point to represent the characteristic distance of the supervoxel, and preferentially merges the points at the trunk position; S42: Exchange operation: After the representative point is determined, the boundary points are exchanged and the boundary points are assigned to the representative point with the smallest distance from its feature.

5. The method for unsupervised branch and leaf separation of trees using LiDAR based on supervoxel segmentation 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 supervoxel center point Q and the minimum elevation point d; S52: Create a blank graph G; S53: Use the KNN algorithm to search for n adjacent points around each point in Q, and store the index and the corresponding point distance r; S54: Add all points in Q to the graph G, starting from d, using r as the weight, and selectively add edges and weights to each point in Q and the n adjacent points searched in S53 according to the constraints; S55: Output weighted topological network graph G; After completing the construction of the undirected graph, the Dijkstra single-source shortest path algorithm is used to calculate the shortest path information from the lowest point in the graph G to all other nodes to generate the shortest path tree.

6. The method for unsupervised branch and leaf separation of trees using LiDAR based on supervoxel segmentation according to claim 1, characterized in that: In S6, by calculating the frequency of each node visit in all shortest paths in the shortest path tree, the node whose node visit frequency is greater than half of the logarithm of the highest node visit frequency is identified as a wood node, specifically: in, Represents node P i The number of times a node is visited, max(log(f)) represents the logarithm of the highest number of node visits; Wood represents a wooden node; and n represents the number of nodes.

7. The method for unsupervised branch and leaf separation of trees using LiDAR based on supervoxel segmentation according to claim 6, characterized in that: The k neighboring points of the wood node are optimized and detected, and the neighboring nodes with similar verticality and curvature to the wood node are expanded into wood nodes. Specifically: Among them, Verticality() represents the verticality of the node, Curvature() represents the curvature of the node, δ represents the threshold; P i and P j Both represent nodes.

8. The method for unsupervised branch and leaf separation of trees using LiDAR based on supervoxel segmentation according to claim 1, characterized in that: In S7, path merging is performed based on significant features. The calculation formula for significant features is: Among them, Linearity, Planarity, and Sphericity represent linear features, planar features, and scattering features respectively; λ1, λ2, and λ3 all represent eigenvalues.

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

Images