A single-tree point cloud segmentation method and system for gradient closed forest
By employing adaptive DBSCAN clustering and KD-Tree search, the problem of insufficient parameter fixation and robustness in single-tree point cloud segmentation in gradient-canopied forests is solved, achieving efficient and accurate single-tree segmentation, applicable to vehicle-mounted, airborne, and ground-based lidar data.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- LANZHOU JIAOTONG UNIV
- Filing Date
- 2025-11-13
- Publication Date
- 2026-07-21
AI Technical Summary
Existing single-tree point cloud segmentation methods suffer from problems such as strong parameter fixation, poor adaptability, insufficient robustness, and high computational complexity in gradient canopy closed forests, making it difficult to operate efficiently in large-scale forest point cloud processing tasks.
An adaptive DBSCAN clustering method is adopted, which adjusts the clustering radius by calculating the vertical slice and horizontal scale. Combined with core point identification and KD-Tree nearest neighbor search, it realizes the integration of individual trees across slices and outputs the individual tree segmentation results.
It improves the accuracy and robustness of individual tree segmentation, reduces oversegmentation and undersegmentation, maintains computational efficiency, and is suitable for large-scale forest stand data processing.
Smart Images
Figure CN121685563B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of ecological monitoring technology, and in particular to a method and system for segmenting single-tree point clouds in gradient-canopied forests. Background Technology
[0002] With the development of LiDAR and UAV remote sensing technologies, acquiring information on the three-dimensional structure of forests has become an important means of forest resource surveys. In fields such as forest management, carbon sequestration assessment, and ecological monitoring, information at the individual tree level (such as tree location, tree height, diameter at breast height, and crown width) is fundamental and crucial data. Therefore, how to accurately and quickly segment individual trees from stand point cloud data is a key research issue in current forestry informatization.
[0003] Currently, the commonly used single-tree point cloud segmentation methods in academia and engineering practice mainly fall into the following categories: The first category is segmentation methods based on canopy top detection. These methods typically find local extrema in the point cloud or identify canopy apexes using a canopy height model, and then use region growing or watershed algorithms for segmentation. Their advantages are intuitive thinking and high computational efficiency, but they are prone to misjudgment when canopy overlap is severe or forest stands are dense, leading to confusion between the canopies of different trees, resulting in undersegmentation or oversegmentation. Especially in gradient-closed forests, due to the complex canopy morphology, the canopies of some trees are difficult to identify directly, significantly reducing the reliability of canopy top detection methods. The second category is segmentation methods based on clustering. A typical example is the DBSCAN (density-based spatial clustering) method. This type of method calculates the neighborhood density of points and clusters areas with higher density into a single tree. Its advantage is that it does not require pre-setting the number of clusters, making it suitable for handling complex environments. However, the traditional DBSCAN method requires manually setting a fixed neighborhood radius (ε value), while the density of forest stand point clouds varies significantly at different heights and in different spatial regions: in dense canopy areas, the point cloud spacing is small, and fixed-radius clustering may lead to oversegmentation; in sparse areas, the point cloud spacing is large, and fixed-radius clustering may result in undersegmentation, where multiple trees are merged into one class. This limitation of fixed parameters restricts the applicability of DBSCAN in gradient-canopy forests. The third category is model-fitting methods, such as using geometric morphology models (e.g., cylinders, ellipsoids, etc.) to fit the trunk or canopy, and then segmenting individual trees accordingly. These methods work well in relatively regular plantations, but in natural forests or gradient-canopy forests, due to the diversity of tree shapes and significant environmental influences, the model struggles to accurately fit the real structure, resulting in large errors and insufficient robustness. The fourth category is segmentation methods based on graph theory or energy optimization, such as using mathematical models like minimum cut and random fields for point cloud segmentation. These methods can theoretically express complex relationships well, but they have high computational complexity and cumbersome parameter settings, making them difficult to apply efficiently in large-scale forest stand data.
[0004] In summary, existing single-tree point cloud segmentation methods have the following common defects and shortcomings: (1) The parameters are highly fixed and lack adaptability: For example, the fixed radius setting of DBSCAN clustering makes it unable to adapt to both dense and sparse regions at the same time; (2) High dependence on tree structure: Crown detection and model fitting methods require relatively regular crown or trunk shapes. Once overlapping, tilted or irregular trees are encountered, the accuracy drops significantly. (3) Insufficient robustness: In environments such as gradient canopy forests with significant point cloud density gradients and complex canopy layers, existing methods often result in oversegmentation or undersegmentation, affecting overall accuracy; (4) Limited engineering applications: Some algorithms are computationally complex and difficult to run efficiently in large-scale forest point cloud processing tasks.
[0005] Therefore, there is an urgent need for a new single-tree point cloud segmentation method that can dynamically adjust parameters according to the spatial distribution characteristics of point clouds while ensuring computational efficiency, thereby achieving robust segmentation of gradient canopy forests and providing reliable data support for the refined management of forest resources. Summary of the Invention
[0006] The purpose of this invention is to provide a single-tree point cloud segmentation method and system for gradient-canopy forests, aiming to solve the problems of over-segmentation in dense areas and under-segmentation in sparse areas in the existing fixed-radius method. It has strong adaptability, robustness and versatility, and is suitable for processing point cloud data from vehicle-mounted LiDAR, airborne LiDAR and ground-based laser systems.
[0007] The technical solution provided by this invention is as follows: A method for single-tree point cloud segmentation in gradient-canopy forests, the method comprising the following steps: S1. Noise reduction and ground point separation are performed on the original forest stand point cloud to obtain the forest stand point cloud; S2. Slice the forest stand point cloud along the Z-axis, perform preliminary DBSCAN clustering within each slice, calculate the horizontal scale of each sub-cluster on the XY plane, remove the maximum and minimum values, and use the average of the remaining scale values as the representative spatial scale of the current slice. S3. Adaptively set the clustering radius according to the representative spatial scale, complete adaptive DBSCAN clustering according to the clustering radius, and identify the core points of each generated sub-cluster. S4. Use the core point as the anchor point for subsequent cross-slice matching, integrate the cross-slice sub-clusters through KD-Tree nearest neighbor search, and output the single-tree segmentation result.
[0008] Preferably, in step S3, the clustering radius is adjusted according to the actual distribution density of the point cloud within the slice; wherein, in dense point cloud regions, the representative spatial scale is smaller, and the clustering radius is reduced accordingly; in sparse point cloud regions, the representative spatial scale is larger, and the clustering radius is automatically increased.
[0009] Preferably, in step S3, the core point is a point that contains at least a minimum number of points within its neighborhood radius, typically located on the main trunk, main branches, or in areas where the density is significantly higher than the surrounding area.
[0010] Preferably, in step S4, the single-tree segmentation result includes tree height, crown width, and diameter at breast height (DBH) parameters.
[0011] This invention further discloses a single-tree point cloud segmentation system for gradient-canopy forests, the system comprising: The point cloud data preprocessing module is used to denoise and separate the original forest stand point cloud from the ground points to obtain the forest stand point cloud. The vertical slicing and initial clustering module is used to slice the forest stand point cloud along the Z-axis, perform initial DBSCAN clustering within each slice, calculate the horizontal scale of each sub-cluster on the XY plane, remove the maximum and minimum values, and use the average of the remaining scale values as the representative spatial scale of the current slice. The adaptive radius clustering module is used to adaptively set the clustering radius according to the representative spatial scale, complete adaptive DBSCAN clustering based on the clustering radius, and identify the core points of each generated sub-cluster. The forest point cloud single tree segmentation module is used to use the core points as anchor points for subsequent cross-slice matching, integrate cross-slice sub-clusters through KD-Tree nearest neighbor search, and output single tree segmentation results.
[0012] Preferably, in the adaptive radius clustering module, the clustering radius is adjusted according to the actual distribution density of the point cloud within the slice; wherein, in dense point cloud regions, the representative spatial scale is small, and the clustering radius is reduced accordingly; in sparse point cloud regions, the representative spatial scale is large, and the clustering radius is automatically increased.
[0013] Preferably, in the adaptive radius clustering module, the core point is a point that contains at least a minimum number of points within its neighborhood radius, typically located on the main trunk, main branches, or in areas where the density is significantly higher than the surrounding area.
[0014] Preferably, in the forest point cloud single tree segmentation module, the single tree segmentation result includes tree height, crown width, and diameter at breast height (DBH) parameters.
[0015] The beneficial effects of this invention after adopting the above technical solution are as follows: First, this invention effectively captures the density differences of point clouds at different altitudes by slicing the forest stand point cloud vertically and introducing initial clustering and horizontal scale calculations within each slice. Compared with the traditional fixed-radius DBSCAN method, this invention can adaptively adjust the clustering radius according to the actual distribution of the point cloud in each slice, thereby avoiding over-segmentation in dense areas and under-segmentation in sparse areas, significantly improving the accuracy and robustness of segmentation.
[0016] Secondly, this invention calculates the representative scale of the slice by removing extrema and then averaging, effectively reducing the interference of outliers or extreme subclusters on the radius setting. This robust scale estimation method enables the invention to maintain high stability in complex forest stand environments and reduces segmentation errors caused by uneven point cloud distribution.
[0017] Furthermore, this invention combines core point identification with KD-Tree cross-slice search to achieve single-tree integration across slices. This method not only ensures the vertical integrity of the tree crown but also effectively avoids erroneous connections between different trees in upper and lower slices, further improving the continuity and reliability of single-tree segmentation results.
[0018] Furthermore, the method of this invention has clear steps and low computational complexity. The algorithms used (DBSCAN clustering and KD-Tree search) are mature point cloud processing tools, which are easy to implement and apply in engineering practice. Compared with complex energy optimization or graph theory methods, this invention can maintain high computational efficiency in large-scale point cloud data, meeting the needs of rapid processing of large-area forest stands in forest resource surveys.
[0019] In summary, this invention not only solves the problem of insufficient robustness of existing single-tree segmentation methods in gradient-canopy forests, but also improves segmentation accuracy and stability through adaptive parameter adjustment and cross-segment integration strategies. The technical solution of this invention is highly versatile and can be widely applied to single-tree segmentation of point cloud data from UAV LiDAR, airborne LiDAR, and ground-based LiDAR, demonstrating significant potential for widespread promotion and application. Attached Figure Description
[0020] Figure 1 This is a schematic diagram of the steps of the method of the present invention; Figure 2 This is an experimental sample map of the method of the present invention; Figure 3 This is a diagram showing the experimental results of the method of the present invention. Detailed Implementation
[0021] The technical solution of the present invention will be further described in detail below with reference to specific embodiments, but this does not constitute any limitation on the present invention.
[0022] This invention discloses a method for segmenting single-tree point clouds in gradient-canopy forests, such as... Figure 1 As shown, the method includes the following steps: S1. Noise reduction and ground point separation are performed on the original forest stand point cloud to obtain the forest stand point cloud.
[0023] In step S1, the original forest stand point cloud is preprocessed to obtain the forest stand point cloud, ensuring data quality and reducing complexity.
[0024] S2. Slice the forest stand point cloud along the Z-axis, perform preliminary DBSCAN clustering within each slice, calculate the horizontal scale of each sub-cluster on the XY plane, remove the maximum and minimum values, and use the average of the remaining scale values as the representative spatial scale of the current slice. In step S2, based on slicing the forest stand point cloud in the vertical direction (Z-axis), the density differences of point clouds at different height levels are effectively captured by introducing initial clustering and horizontal scale calculation within the slices.
[0025] After obtaining the horizontal scale of each sub-cluster, the maximum and minimum values are first removed to eliminate the interference of extreme outliers. Then, the average of the remaining scale values is calculated as the representative spatial scale of the current slice.
[0026] Compared with the traditional fixed-radius DBSCAN method, this invention can adaptively adjust the clustering radius according to the actual distribution of point clouds in each slice, thereby avoiding oversegmentation in dense areas and undersegmentation in sparse areas, significantly improving the accuracy and robustness of segmentation, and showing stronger adaptability, especially in complex forest stand environments.
[0027] Furthermore, this invention employs a mean-based approach after removing extrema, effectively reducing the interference of outliers or extreme subclusters on the radius setting. This robust scale estimation method enables the invention to maintain high stability even in complex forest stand environments, reducing segmentation errors caused by uneven point cloud distribution.
[0028] S3. Adaptively set the clustering radius according to the representative spatial scale, complete adaptive DBSCAN clustering according to the clustering radius, and identify the core points of each generated sub-cluster.
[0029] In step S3, the representative spatial scale is multiplied by a preset scaling factor to obtain the clustering radius adaptively set for the slice. During this process, the clustering radius can be flexibly adjusted according to the actual distribution density of the point cloud within the slice: in dense point cloud regions, the representative spatial scale is smaller, and the clustering radius decreases accordingly, effectively avoiding over-segmentation; in sparse point cloud regions, the representative spatial scale is larger, and the clustering radius automatically increases, thereby preventing under-segmentation and achieving an adaptive response of the clustering parameters to the local density features of the point cloud.
[0030] In step S3, a core point refers to a point within its neighborhood radius that contains at least a minimum point count threshold, typically located on the trunk, main branches, or areas with significantly higher density than the surrounding region. By identifying and preserving these core points, the main structure and vertical axis orientation of a single tree can be effectively characterized.
[0031] S4. Use the core point as the anchor point for subsequent cross-slice matching, integrate the cross-slice sub-clusters through KD-Tree nearest neighbor search, and output the single-tree segmentation result.
[0032] In step S4, the present invention uses the core point as the anchor point for subsequent cross-slice matching to ensure that the main structure of the canopy and the main spatial axis are preserved and continued during the integration process, thereby avoiding the break of trunk information caused by sparse branches and leaves or missing point clouds, and ensuring the integrity and continuity of the three-dimensional structure of a single tree.
[0033] Building upon this foundation, this invention combines core point identification with KD-Tree cross-slice search to achieve single-tree integration across slices. This method not only ensures the vertical integrity of the tree crown but also effectively avoids erroneous connections between different trees in upper and lower slices, thereby improving the continuity and reliability of single-tree segmentation results.
[0034] In addition, cross-slice sub-cluster integration can achieve the vertical continuity of trees, and finally output the single tree segmentation results, which can be used to extract parameters such as tree height, crown width, and diameter at breast height.
[0035] Accordingly, the present invention further discloses a single-tree point cloud segmentation system for gradient-canopy forests, the system comprising: The point cloud data preprocessing module is used to denoise and separate the original forest stand point cloud from the ground points to obtain the forest stand point cloud. The vertical slicing and initial clustering module is used to slice the forest stand point cloud along the Z-axis, perform initial DBSCAN clustering within each slice, calculate the horizontal scale of each sub-cluster on the XY plane, remove the maximum and minimum values, and use the average of the remaining scale values as the representative spatial scale of the current slice. The adaptive radius clustering module is used to adaptively set the clustering radius according to the representative spatial scale, complete adaptive DBSCAN clustering based on the clustering radius, and identify the core points of each generated sub-cluster. The forest point cloud single tree segmentation module is used to use the core points as anchor points for subsequent cross-slice matching, integrate cross-slice sub-clusters through KD-Tree nearest neighbor search, and output single tree segmentation results.
[0036] The technical solution of the system of the present invention is the same as that of the method described above, and will not be repeated here.
[0037] To objectively evaluate the application effect of the method of the present invention, in this embodiment, five sample plots with different canopy closure were selected. Each sample plot was 50×50 m in size, and there were a total of 477 trees with a height between 3 and 8 meters in the five sample plots, including species such as cedar, Chinese scholar tree, larch, and spruce. The cloud cover at each sample location under different canopy closure levels is shown in the figure. Figure 2 As shown in Table 1, the canopy density and the actual number of trees are as follows.
[0038] Table 1. Experimental plot data 1 0.18 78 2 0.43 51 3 0.51 80 4 0.72 109 5 0.79 159 .
[0039] To evaluate the tree segmentation effect, this invention uses five quantitative indicators: Recognition Rate (R), Recall (r), Precision (p), Harmonic Mean (F1 Score, F), and Point Cloud Completeness C. The calculation formulas are as follows: (1) (2) (3) (4) (5) Where N is the actual number of trees in the sample plot, n is the number of detected trees, TP represents the number of correctly segmented trees, FN is the number of missed detections, FP is the number of false detections, and M... 算法 M represents the number of point clouds in the segmentation result obtained by the method in this paper. 人工 This represents the number of point clouds resulting from manual segmentation.
[0040] The recognition rate reflects the proportion of detected trees out of the actual trees; a higher value indicates stronger detection capability. The recall rate measures the proportion of correctly identified trees among all actual trees; a higher value indicates fewer missed detections. The precision rate represents the proportion of correctly identified trees out of all detection results; a higher value indicates fewer false detections. The harmonic mean F comprehensively considers both recall and precision and is an important indicator for evaluating overall segmentation performance; a value closer to 1 indicates better segmentation results. Point cloud completeness represents the proportion of correctly segmented single-tree point clouds out of manually segmented single-tree point clouds; a higher value indicates more accurate single-tree point cloud segmentation.
[0041] Experimental results are as follows Figure 3 And as shown in Table 2.
[0042] Table 2 Experimental Results 1 77 0.99 0.99 1.00 0.99 0.96 2 48 0.95 0.93 1.00 0.96 0.93 3 75 0.94 0.94 0.98 0.96 0.95 4 101 0.93 0.93 0.99 0.96 0.91 5 146 0.92 0.92 0.98 0.95 0.91 .
[0043] Depend on Figure 3 As shown in Table 2: (1) The method of the present invention has an average recall rate of 0.94 and an average precision rate of 0.99 in 5 typical sample plots, and shows strong robustness, especially in areas with dense canopy and intertwined branches; (2) In the sample plots of the study area, the segmentation accuracy F of this method decreased with the increase of canopy closure, but the average F value was still 0.91 at high canopy closure, indicating excellent segmentation performance. The results show that this method has stronger adaptability and stability under different stand conditions, especially under high canopy closure data; (3) The adaptive clustering radius strategy significantly improves the completeness and accuracy of single tree extraction, and the average completeness of the segmentation point cloud (C=0.93) is relatively high, which is suitable for the fine structure analysis of high-density forest stands.
[0044] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.
Claims
1. A method for segmenting single-tree point clouds in gradient-canopy forests, characterized in that, The method includes the following steps: S1. Noise reduction and ground point separation are performed on the original forest stand point cloud to obtain the forest stand point cloud; S2. Slice the forest stand point cloud along the Z-axis, perform preliminary DBSCAN clustering within each slice, calculate the horizontal scale of each sub-cluster on the XY plane, remove the maximum and minimum values, and use the average of the remaining scale values as the representative spatial scale of the current slice. S3. Adaptively set the clustering radius according to the representative spatial scale, complete adaptive DBSCAN clustering according to the clustering radius, and identify the core points of each generated sub-cluster. S4. Use the core point as the anchor point for subsequent cross-slice matching, integrate the cross-slice sub-clusters through KD-Tree nearest neighbor search, and output the single-tree segmentation result.
2. The method as described in claim 1, characterized in that, In step S3, the clustering radius is adjusted according to the actual distribution density of the point cloud within the slice; wherein, in dense point cloud regions, the representative spatial scale is smaller, and the clustering radius is reduced accordingly; in sparse point cloud regions, the representative spatial scale is larger, and the clustering radius is automatically increased.
3. The method as described in claim 1, characterized in that, In step S3, the core point is a point that contains at least a minimum number of points within its neighborhood radius, typically located on the main trunk, main branches, or areas with significantly higher density than the surrounding area.
4. The method as described in claim 1, characterized in that, In step S4, the results of the single-tree segmentation include tree height, crown width, and diameter at breast height (DBH) parameters.
5. A single-tree point cloud segmentation system for gradient-canopy forests, characterized in that, The system includes: The point cloud data preprocessing module is used to denoise and separate the original forest stand point cloud from the ground points to obtain the forest stand point cloud. The vertical slicing and initial clustering module is used to slice the forest stand point cloud along the Z-axis, perform initial DBSCAN clustering within each slice, calculate the horizontal scale of each sub-cluster on the XY plane, remove the maximum and minimum values, and use the average of the remaining scale values as the representative spatial scale of the current slice. The adaptive radius clustering module is used to adaptively set the clustering radius according to the representative spatial scale, complete adaptive DBSCAN clustering based on the clustering radius, and identify the core points of each generated sub-cluster. The forest point cloud single tree segmentation module is used to use the core points as anchor points for subsequent cross-slice matching, integrate cross-slice sub-clusters through KD-Tree nearest neighbor search, and output single tree segmentation results.
6. The system as described in claim 5, characterized in that, In the adaptive radius clustering module, the clustering radius is adjusted according to the actual distribution density of the point cloud within the slice; wherein, in dense point cloud regions, the representative spatial scale is small, and the clustering radius is reduced accordingly; in sparse point cloud regions, the representative spatial scale is large, and the clustering radius is automatically increased.
7. The system as described in claim 5, characterized in that, In the adaptive radius clustering module, the core point is a point that contains at least a minimum number of points within its neighborhood radius, typically located on the main trunk, main branches, or in areas where the density is significantly higher than the surrounding area.
8. The system as described in claim 5, characterized in that, In the forest point cloud tree segmentation module, the tree segmentation results include tree height, crown width, and diameter at breast height (DBH) parameters.