A single-tree point cloud segmentation method based on control point guidance and feature adaptation

Abstract

The application relates to the technical field of single-tree segmentation, and discloses a single-tree point cloud segmentation method based on control point guidance and feature self-adaptation, which comprises the following steps: acquiring laser radar point cloud data of a target stand and a plurality of sparse ground control point coordinates distributed in the target stand; optimizing parameters of a single-tree segmentation method in a preset spatial neighborhood for each ground control point to obtain a local optimal parameter set corresponding to the control point; constructing a multi-dimensional feature field reflecting spatial heterogeneity of a stand structure according to the laser radar point cloud data; generating a self-adaptive parameter field covering the whole space through feature space interpolation according to each ground control point and the corresponding local optimal parameter set; driving the single-tree segmentation method to segment the laser radar point cloud data by using the self-adaptive parameter field, and exerting position constraints on the ground control points in the segmentation process to obtain a single-tree segmentation result. The application realizes the optimal balance between cost and benefit while ensuring high precision.

Description

Technical Field

[0001] This invention relates to the field of single-tree segmentation technology, specifically to a single-tree point cloud segmentation method based on control point guidance and feature adaptation. Background Technology

[0002] As the fundamental building block of forest ecosystems, the accurate identification and parameter extraction of individual trees are crucial for forest resource surveys, biomass estimation, carbon sequestration, and the implementation of precision silviculture measures such as selective felling and thinning. Traditional manual tree measurement methods are inefficient and costly, failing to meet the demands of large-scale, high-time-sensitivity forest monitoring. LiDAR technology can directly acquire three-dimensional point cloud data of forests, providing technical support for the automated extraction of individual tree information. Researchers have developed several mainstream algorithms for individual tree segmentation based on LiDAR point clouds, but their application in complex forest stands still faces significant challenges.

[0003] Methods based on canopy height models segment 3D point clouds by projecting them onto 2D raster images. While computationally efficient, these methods suffer from information loss and smoothing effects during canopy height model generation, limiting their ability to distinguish overlapping canopies in closed stands and hindering effective detection of sub-canopy trees. Distance-based clustering methods using normalized point clouds perform clustering directly in 3D space, making better use of structural information. However, their segmentation accuracy heavily relies on spatial distance thresholds, and fixed thresholds struggle to adapt to the spatial heterogeneity of density within stands, leading to undersegmentation in dense areas and oversegmentation in sparse areas. Layer stacking methods identify individual trees by analyzing the vertical layering structure of point clouds, showing some adaptability to complex canopies. However, these methods involve multiple stages such as layering, clustering, and association, resulting in complex and coupled parameter settings. They lack mechanisms for automatically adjusting strategies across different density regions, making it difficult to achieve consistently high-precision results across the entire domain.

[0004] Recent studies have attempted to directly guide segmentation by introducing high-precision measured individual tree locations as seed points, significantly improving accuracy at corresponding points. However, this method requires field measurements of the vast majority of individual trees within the sample plot, which is costly and not feasible for large-scale implementation. Furthermore, it fails to address the adaptive segmentation problem in areas without measured points.

[0005] In summary, existing technologies face unresolved contradictions regarding the accuracy of individual tree segmentation, adaptability to complex forest stands, and cost-effectiveness for large-scale application. Specifically, the core bottleneck lies in the lack of economical and reliable external guidance information, and an intelligent mapping mechanism capable of automatically inferring and optimizing the overall segmentation strategy based on limited prior knowledge. Therefore, developing a high-precision individual tree segmentation method that can utilize sparse prior control points to achieve global parameter adaptation has become a key requirement for overcoming current technological bottlenecks and promoting the development of precision forestry. Summary of the Invention

[0006] To address the shortcomings of existing single-tree segmentation techniques, such as poor adaptability to complex forest stands, reliance on manual parameter adjustment, and high costs associated with high-precision methods, this invention provides a single-tree point cloud segmentation method based on sparse control point guidance and feature space adaptation. This method utilizes a small number of ground control points as "anchor points" for local accuracy and "teacher signals" for parameter optimization. By constructing an intelligent mapping model between the multi-dimensional feature space of the forest stand and the segmentation parameters, the segmentation algorithm is driven to adaptively adjust across the entire domain. This significantly reduces the reliance on data density and manual experience while maintaining high accuracy, achieving an optimal balance between cost and benefit.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A single-tree point cloud segmentation method based on control point guidance and feature adaptation includes the following steps: Acquire lidar point cloud data of the target forest stand, as well as the coordinates of multiple sparse ground control points distributed within the target forest stand; For each ground control point, the parameters of the single-tree segmentation method are optimized within a preset spatial neighborhood to obtain the local optimal parameter set corresponding to the control point; A multidimensional feature field reflecting the spatial heterogeneity of forest stand structure was constructed based on lidar point cloud data; Based on each ground control point and its corresponding local optimal parameter set, an adaptive parameter field covering the entire domain is generated through feature space interpolation. An adaptive parameter field-driven single-tree segmentation method is used to segment lidar point cloud data, and position constraints are applied to ground control points during the segmentation process to obtain single-tree segmentation results.

[0008] Furthermore, the parameters of the single-tree segmentation method are optimized within a preset spatial neighborhood for each ground control point to obtain a locally optimal parameter set corresponding to the control point, including: The preset three-dimensional spatial range of each ground control point is defined as the optimized neighborhood; Construct an objective function, which includes a distance term from the control point to the nearest seed point and a penalty term for the number of seed points in the neighborhood; The search is performed within a preset parameter space, and the parameter combination that minimizes the objective function is taken as the local optimal parameter set.

[0009] Furthermore, the multidimensional feature field includes a point cloud density field, an average height field, a height variation field, and a horizontal roughness field.

[0010] Furthermore, a multidimensional feature field reflecting the spatial heterogeneity of forest stand structure is constructed based on lidar point cloud data, including: Divide the planar space of the lidar point cloud projection into a regular grid; For each grid, a four-dimensional feature vector of the point cloud falling within it is calculated; the four-dimensional feature vector includes point cloud density, average height, height variation coefficient, and horizontal roughness. Based on the four-dimensional feature vectors obtained from all grid calculations, a continuous feature raster layer covering the entire domain is generated through spatial interpolation, forming a multi-dimensional feature field.

[0011] Furthermore, based on each ground control point and its corresponding local optimal parameter set, an adaptive parameter field covering the entire domain is generated through feature space interpolation, including: Extract the multidimensional feature vectors at the locations of each ground control point, as well as the multidimensional feature vectors at the location to be determined across the entire domain; Calculate the feature distance from the feature vector of each location to be determined to the feature vectors of all ground control points in the multidimensional feature space; The weights are determined based on the feature distance, and the local optimal parameters of all ground control points are summed in a weighted manner to obtain the predicted parameters for each location to be determined.

[0012] Furthermore, the method of segmenting lidar point cloud data using an adaptive parameter field-driven single-tree segmentation method includes: Starting with the optimized seed point, the vegetation point cloud is distributed to each individual tree using the region growth method.

[0013] Furthermore, applying positional constraints to ground control points during the segmentation process includes: Calculate the spatial distance between the ground control point and the nearest seed point; If the spatial distance is less than the preset distance threshold, the coordinates of the seed point will be replaced with the three-dimensional coordinates of the ground control point. If the spatial distance is greater than or equal to the preset distance threshold, a new seed point is added at the three-dimensional coordinates of the ground control point.

[0014] Furthermore, before obtaining the results of individual tree segmentation, the process also includes: Global conflict resolution is performed on the segmented seed points. When the growth regions of multiple seed points overlap, the seed point with the most affiliated points is retained, and the rest of the seed points are removed.

[0015] Furthermore, after acquiring the lidar point cloud data of the target forest stand, the process also includes: Ground filtering is performed on the lidar point cloud data to generate a digital elevation model; and elevation normalization is performed on the lidar point cloud data based on the digital elevation model.

[0016] Furthermore, after obtaining the results of individual tree segmentation, the process also includes: Based on the segmentation results of individual trees, the tree height, crown width, and two-dimensional location information of individual trees are extracted.

[0017] The present invention has the following beneficial effects: (1) Achieving a balance between high precision and low cost: Global intelligent optimization is driven by sparse control points with approximately 22% of the sample size, significantly reducing fieldwork costs while ensuring high precision. Experiments show that, compared with traditional fixed parameter methods, the F1 score increased from 0.5944 to 0.9671, the number of correct detections increased from 499 to 661, and the number of false detections decreased by 97.5%.

[0018] (2) It has fully automatic intelligent adaptive capability: The constructed “feature-parameter” mapping model enables the algorithm to automatically adjust internal parameters according to the continuous changes in forest stand density, height and other features, which fundamentally solves the problem of poor adaptability of fixed parameters in complex forest stands and improves the robustness and universality of the method.

[0019] (3) A scalable general technical framework has been formed: the core framework of “control point guidance-feature mapping” does not depend on a specific segmentation algorithm, and can flexibly adapt to and improve the performance of various parameter-sensitive algorithms, providing a new approach to solving similar spatial heterogeneity analysis problems.

[0020] (4) The output of the operation results can be directly applied: the final output of the high-precision single tree location and attribute list can be directly used to generate precise forest management operation maps such as thinning and tending, opening up the key technology link from remote sensing analysis to on-site construction, and is highly practical. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of a single-tree point cloud segmentation method based on control point guidance and feature adaptation in this invention. Figure 2 This is a schematic diagram of the original point cloud data; Figure 3 This is a schematic diagram of the locations of the validation tree and calibration tree based on GPS measurements; Figure 4 A schematic diagram illustrating the optimization of local parameters for GPS control points; Figure 5 A schematic diagram illustrating the parameter optimization search and convergence process; Figure 6 This is a schematic diagram of the horizontal projection segmentation effect; Figure 7 This is a schematic diagram of the results of single-tree segmentation. Detailed Implementation

[0022] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0023] like Figure 1 As shown, an embodiment of the present invention provides a single-tree point cloud segmentation method based on control point guidance and feature adaptation, comprising the following steps S1 to S5: S1. Obtain lidar point cloud data of the target forest stand, as well as the coordinates of multiple sparse ground control points distributed within the target forest stand; In an optional embodiment of the present invention, step S1 performs point cloud data acquisition and preprocessing, acquires lidar point cloud data of the target forest stand, and obtains normalized vegetation point cloud through ground filtering and elevation normalization processing; at the same time, acquires high-precision three-dimensional coordinates of multiple sparsely distributed ground control points deployed in the forest stand.

[0024] In this embodiment, point cloud data acquisition and preprocessing are first performed. A drone equipped with a lidar sensor is used to acquire point clouds of the study area at a flight altitude of 80m and 40% lateral overlap, with an average density of approximately 195 points / m². 2 Ground filtering employs a cloth-based analog filtering algorithm, with cloth_resolution set to 0.8m and rigidity set to 2 to accommodate moderate slope conditions, effectively separating ground and non-ground points. Subsequently, a 0.25m resolution digital elevation model is generated using ground points through inverse distance weighted interpolation. Elevation normalization is then performed on all non-ground points, ensuring that the Z-value of each point directly reflects its ground elevation, resulting in a normalized vegetation point cloud. The point cloud data is shown below. Figure 2 As shown.

[0025] Then, ground control points were acquired. A total of approximately 694 trees were surveyed in the study area. Within nine typical 15m × 15m plots, the three-dimensional coordinates (X, Y, Z) of the trunk base center of 153 individual trees were determined in the field using a UniStrong G960 geodesy GNSS receiver in fixed solution mode, with a planar positioning accuracy of 8mm. Control points were selected considering different habitats, including high-density, low-density, and sloping terrain, representing 22% of the total number of trees in the forest area. The remaining trees served as independent validation samples to evaluate the reliability of the method. The control point distribution results are shown below. Figure 3 As shown.

[0026] S2. Optimize the parameters of the single-tree segmentation method for each ground control point within a preset spatial neighborhood to obtain the local optimal parameter set corresponding to the control point; In an optional embodiment of the present invention, step S2 optimizes the parameters of the single-tree segmentation method for each ground control point within a preset spatial neighborhood to obtain a locally optimal parameter set corresponding to the control point, including: The preset three-dimensional spatial range of each ground control point is defined as the optimized neighborhood; Construct an objective function, which includes a distance term from the control point to the nearest seed point and a penalty term for the number of seed points in the neighborhood; The search is performed within a preset parameter space, and the parameter combination that minimizes the objective function is taken as the local optimal parameter set.

[0027] In this embodiment, step S2 performs control point neighborhood parameter optimization. For each ground control point, within its preset spatial neighborhood, with the goal of correctly and uniquely segmenting the control point, the key parameters of the selected single-tree segmentation algorithm are automatically searched and optimized to obtain the local optimal parameter set corresponding to the control point.

[0028] This embodiment first defines the optimization objective and parameter space, selects the layer stacking algorithm as the basic segmentation algorithm, and selects the three parameters that have the most significant impact on its segmentation results—minimum tree spacing, layer thickness, and Gaussian smoothing standard deviation—for optimization. Each control point... The neighborhood of is defined as a cylindrical space centered at and with a radius of 2.5 meters. The objective function is then constructed. : in, The seed point set is obtained by running a layer stacking algorithm in the neighborhood using parameter combination P, and min_distance is the control point. Distance to the nearest seed point Let be the number of seed points, and α and β be the weighting coefficients. The objective is to minimize the loss, which requires generating one and only one seed point that is closest to the control point.

[0029] This embodiment then performs an optimization search for each control point. A grid search method was used to optimize the parameters within its parameter space. The parameter search range was set as follows: minimum tree spacing ∈ [2, 6] meters, layer thickness ∈ [0.5, 2], and Gaussian smoothing standard deviation (sigma) ∈ [0.5, 1.5]. After performing optimization independently for each control point, the local optimal parameter triplet {P} was obtained for each set. i_optimal The results show that the optimal values ​​of control points in high-density areas are generally larger, verifying the necessity of parameter adaptation. The principle of control point neighborhood parameter optimization is as follows: Figure 4 and Figure 5 As shown, its core principle is to find the local optimal parameters by minimizing the loss function.

[0030] S3. Construct a multidimensional feature field reflecting the spatial heterogeneity of forest stand structure based on lidar point cloud data; In an optional embodiment of the present invention, the multidimensional feature field constructed in step S3 includes a point cloud density field, an average height field, a height variation field, and a horizontal roughness field.

[0031] Step S3 involves constructing a multidimensional feature field reflecting the spatial heterogeneity of forest stand structure based on lidar point cloud data, including: Divide the planar space of the lidar point cloud projection into a regular grid; For each grid, a four-dimensional feature vector of the point cloud falling within it is calculated; the four-dimensional feature vector includes point cloud density, average height, height variation coefficient, and horizontal roughness. Based on the four-dimensional feature vectors obtained from all grid calculations, a continuous feature raster layer covering the entire domain is generated through spatial interpolation, forming a multi-dimensional feature field.

[0032] In this embodiment, step S3 involves constructing a stand spatial feature field. Based on the normalized vegetation point cloud, a multidimensional feature field reflecting the spatial heterogeneity of the stand structure is calculated; the features include at least a point cloud density field, a mean height field, a height variation field, and a horizontal roughness field.

[0033] This embodiment first divides the space into a grid: the entire sample plot is divided into a regular grid of 2 meters × 2 meters.

[0034] This embodiment then calculates the grid feature vector for each grid cell. j Based on the normalized vegetation point cloud falling within it, the four-dimensional feature vector F is calculated. j Wherein, point cloud density F1 is the number of points per unit volume within the grid, average height F2 is the average Z value of all points within the grid, height variation coefficient F3 is the ratio of the standard deviation to the average Z value of points within the grid, and horizontal roughness F4 is the ratio of the convex hull area of ​​the point cloud projected onto the XY plane to the grid area.

[0035] In this embodiment, a continuous feature field is finally generated: using ordinary Kriging interpolation, the discrete grid feature values ​​are interpolated into four continuous feature raster layers covering the entire sample plot, with a resolution of 0.5 meters.

[0036] S4. Based on each ground control point and its corresponding local optimal parameter set, generate an adaptive parameter field covering the entire domain through feature space interpolation; In an optional embodiment of the present invention, step S4 generates an adaptive parameter field covering the entire domain through feature space interpolation based on each ground control point and its corresponding local optimal parameter set, including: Extract the multidimensional feature vectors at the locations of each ground control point, as well as the multidimensional feature vectors at the location to be determined across the entire domain; Calculate the feature distance from the feature vector of each location to be determined to the feature vectors of all ground control points in the multidimensional feature space; The weights are determined based on the feature distance, and the local optimal parameters of all ground control points are summed in a weighted manner to obtain the predicted parameters for each location to be determined.

[0037] In this embodiment, step S4 performs adaptive parameter field generation. Based on the multidimensional feature field, a mapping relationship from the feature space to the segmentation parameter space is established; specifically, each ground control point and its corresponding local optimal parameter set are used as samples, and the applicable segmentation parameters for any location without control points within the forest stand are predicted using the Kriging space interpolation method, thereby generating an adaptive parameter field that covers the entire area and is spatially continuously changing.

[0038] This embodiment first constructs a "feature-parameter" sample library: the spatial coordinates of 153 control points and their corresponding four-dimensional feature vectors F Gi and its local optimal parameter P i_optimal Together, they constitute the training sample library.

[0039] This embodiment then performs feature space interpolation: for any location q (corresponding to a 0.5-meter pixel) within the sample plot, its four feature values ​​are extracted to form a vector F. q In the characteristic space (four-dimensional Euclidean space), calculate F. q Eigenvectors F to all control points Gi distance d i Then, the inverse distance weighted method is used to predict the optimal parameters at point q. : Among them, weight , To prevent division by zero for small constants.

[0040] This embodiment finally generates a parameter field raster, traverses all pixels of the sample plot, and calculates its corresponding... Finally, three parameter field raster maps with the same resolution as the feature field are generated, representing the minimum tree spacing field, the layer thickness field, and the Gaussian smoothing standard deviation field, respectively.

[0041] S5. The LiDAR point cloud data is segmented using an adaptive parameter field-driven single-tree segmentation method, and positional constraints are applied to the ground control points during the segmentation process to obtain the single-tree segmentation results.

[0042] In an optional embodiment of the present invention, step S5, which uses an adaptive parameter field-driven single-tree segmentation method to segment the lidar point cloud data, includes: Starting with the optimized seed point, the vegetation point cloud is distributed to each individual tree using the region growth method.

[0043] Step S5, which applies positional constraints to ground control points during the segmentation process, includes: Calculate the spatial distance between the ground control point and the nearest seed point; If the spatial distance is less than the preset distance threshold, the coordinates of the seed point will be replaced with the three-dimensional coordinates of the ground control point. If the spatial distance is greater than or equal to the preset distance threshold, a new seed point is added at the three-dimensional coordinates of the ground control point.

[0044] In this embodiment, step S5 performs guided adaptive segmentation. The adaptive parameter field drives the single-tree segmentation algorithm to segment the normalized vegetation point cloud. During the segmentation process, positional constraints are applied to each ground control point to ensure that it has one and only one corresponding single-tree object in the segmentation result, and the deviation between the object's position identifier and the ground control point coordinates is less than an allowable threshold, thereby generating an initial segmentation map.

[0045] This embodiment first modifies the main loop of the layer stacking algorithm: the standard layer stacking algorithm is modified. When the algorithm processes a point (x, y, z) in three-dimensional space, it reads the dynamic parameter value corresponding to that position in real time from the above parameter field grid map using bilinear interpolation based on its planar coordinates (x, y), and uses it for clustering and tracking decisions of the current layer.

[0046] This embodiment then enforces control point constraints: the above adaptive layer stacking algorithm is run to obtain the initial seed point set S. initial .

[0047] Finally, for each ground control point G in this embodiment... i In S initial Find its nearest neighbor seed point S in the middle near .

[0048] If distance(G) i , S near If S < 1.0 meter, then S near Replace the coordinates with G i The coordinates.

[0049] If distance(G) i , S near If ) >= 1.0 meter, then in G i Add a new seed point directly at the coordinates.

[0050] This step ensures that each control point has a unique and precisely located corresponding point in the seed point set. The horizontal projection segmentation effect is as follows: Figure 6 As shown.

[0051] This embodiment also includes performing global optimization and outputting results. A global consistency check and optimization are performed on the initial segmentation map to handle potential boundary conflicts or minor fragments, outputting the final high-precision individual tree segmentation results along with the corresponding individual tree positions and height parameters.

[0052] This embodiment first performs conflict resolution, checking all seed point pairs. If the distance between any two points is less than 1 meter, it is considered a conflict. The number of points "belonging" to each pair of conflicting seed points in the point cloud is calculated (region growing is performed with the seed point as the center). The seed point with more belonging points is retained, and the other is merged or deleted.

[0053] This embodiment then performs final segmentation and parameter extraction: starting with the optimized seed point set, the entire vegetation point cloud is segmented into individual trees using the region growing method.

[0054] The final output of this embodiment includes: 1) a single-tree segmentation vector boundary map; 2) a single-tree attribute table, containing a unique ID for each tree, tree height based on segmentation point cloud computing, crown width, and its precise two-dimensional location (from control points or corrected seed points). The final segmentation result is as follows: Figure 7 As shown.

[0055] To verify the effectiveness of this invention, the results of this embodiment were compared with those of the traditional fixed-parameter layer stacking method. Layer stacking was performed using the global average parameters of 9 sample plots. The global average parameters of the 9 sample plots were set with a minimum tree spacing of 2.5 m, a Gaussian smoothing standard deviation of 0.7, and a layer thickness of 2 m.

[0056] The effectiveness of single-tree recognition based on point cloud data is evaluated by calculating the detection rate r, accuracy p, and metric F.

[0057] in, It refers to the number of trees present in the test area. It represents the number of trees missed by the algorithm. It is the number of trees that are not present in the sample plot but were detected.

[0058] As shown in Table 1, compared with the traditional method, the present invention, using only 22% of control points, achieves a 32.5% increase in detection rate r, a 93.9% jump in accuracy p, an overall 62.7% increase in measurement F score, and a 97.5% and 83.1% decrease in false detection and false negative detection, respectively, significantly improving the accuracy and reliability of single-tree segmentation.

[0059] Table 1: Accuracy Verification Results Chart This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0060] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0061] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0062] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0063] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A single-tree point cloud segmentation method based on control point guidance and feature adaptation, characterized in that, Includes the following steps: Acquire lidar point cloud data of the target forest stand, as well as the coordinates of multiple sparse ground control points distributed within the target forest stand; For each ground control point, the parameters of the single-tree segmentation method are optimized within a preset spatial neighborhood to obtain the local optimal parameter set corresponding to the control point; A multidimensional feature field reflecting the spatial heterogeneity of forest stand structure was constructed based on lidar point cloud data; Based on each ground control point and its corresponding local optimal parameter set, an adaptive parameter field covering the entire domain is generated through feature space interpolation. An adaptive parameter field-driven single-tree segmentation method is used to segment lidar point cloud data, and position constraints are applied to ground control points during the segmentation process to obtain single-tree segmentation results.

2. The single-tree point cloud segmentation method based on control point guidance and feature adaptation according to claim 1, characterized in that, For each ground control point, the parameters of the single-tree segmentation method are optimized within a preset spatial neighborhood to obtain a locally optimal parameter set corresponding to the control point, including: The preset three-dimensional spatial range of each ground control point is defined as the optimized neighborhood; Construct an objective function, which includes a distance term from the control point to the nearest seed point and a penalty term for the number of seed points in the neighborhood; The search is performed within a preset parameter space, and the parameter combination that minimizes the objective function is taken as the local optimal parameter set.

3. The single-tree point cloud segmentation method based on control point guidance and feature adaptation according to claim 1, characterized in that, The multidimensional feature field includes a point cloud density field, an average height field, a height variation field, and a horizontal roughness field.

4. The single-tree point cloud segmentation method based on control point guidance and feature adaptation according to claim 1, characterized in that, A multidimensional feature field reflecting the spatial heterogeneity of forest stand structure was constructed based on lidar point cloud data, including: Divide the planar space of the lidar point cloud projection into a regular grid; For each grid, a four-dimensional feature vector of the point cloud falling within it is calculated; the four-dimensional feature vector includes point cloud density, average height, height variation coefficient, and horizontal roughness. Based on the four-dimensional feature vectors obtained from all grid calculations, a continuous feature raster layer covering the entire domain is generated through spatial interpolation, forming a multi-dimensional feature field.

5. The single-tree point cloud segmentation method based on control point guidance and feature adaptation according to claim 1, characterized in that, Based on each ground control point and its corresponding local optimal parameter set, an adaptive parameter field covering the entire domain is generated through feature space interpolation, including: Extract the multidimensional feature vectors at the locations of each ground control point, as well as the multidimensional feature vectors at the location to be determined across the entire domain; Calculate the feature distance from the feature vector of each location to be determined to the feature vectors of all ground control points in the multidimensional feature space; The weights are determined based on the feature distance, and the local optimal parameters of all ground control points are summed in a weighted manner to obtain the predicted parameters for each location to be determined.

6. The single-tree point cloud segmentation method based on control point guidance and feature adaptation according to claim 1, characterized in that, The method of segmenting LiDAR point cloud data using an adaptive parameter field-driven single-tree segmentation method includes: Starting with the optimized seed point, the vegetation point cloud is distributed to each individual tree using the region growth method.

7. The single-tree point cloud segmentation method based on control point guidance and feature adaptation according to claim 1, characterized in that, Applying positional constraints to ground control points during the segmentation process includes: Calculate the spatial distance between the ground control point and the nearest seed point; If the spatial distance is less than the preset distance threshold, the coordinates of the seed point will be replaced with the three-dimensional coordinates of the ground control point. If the spatial distance is greater than or equal to the preset distance threshold, a new seed point is added at the three-dimensional coordinates of the ground control point.

8. A single-tree point cloud segmentation method based on control point guidance and feature adaptation according to claim 1, characterized in that, Before obtaining the results of individual tree segmentation, the following steps are also included: Global conflict resolution is performed on the segmented seed points. When the growth regions of multiple seed points overlap, the seed point with the most affiliated points is retained, and the rest of the seed points are removed.

9. A single-tree point cloud segmentation method based on control point guidance and feature adaptation according to claim 1, characterized in that, After acquiring the lidar point cloud data of the target forest stand, the following steps are also included: Ground filtering is performed on the lidar point cloud data to generate a digital elevation model; and elevation normalization is performed on the lidar point cloud data based on the digital elevation model.

10. A single-tree point cloud segmentation method based on control point guidance and feature adaptation according to claim 1, characterized in that, After obtaining the results of individual tree segmentation, the following is also included: Based on the segmentation results of individual trees, the tree height, crown width, and two-dimensional location information of individual trees are extracted.

Images