Automatic forest number and position detection method based on HDBSCAN algorithm

Abstract

The invention provides a forest tree number and position automatic detection method based on an HDBSCAN algorithm, and the method comprises the steps: carrying out the multi-level threshold separation and denoising operation of original point cloud data, and obtaining forest tree point cloud data; performing initial forest recognition on the forest point cloud data through a density clustering algorithm after dynamic parameter optimization to obtain an initial recognition result; performing three-dimensional shape correction and adjacent target fusion on the initial recognition result to obtain a prediction recognition result; and performing multi-dimensional feature verification on the prediction identification result to obtain a final identification result. According to the method, a forest tree detection task is accurately and efficiently completed through multi-level threshold separation, an anisotropic filtering technology and multi-dimensional verification, and the detection precision and the anti-interference capability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of forest tree detection and identification, and in particular to a method for automatically detecting the number and position of forest trees based on the HDBSCAN algorithm. Background Art

[0002] With the rapid development of remote sensing technology, LiDAR (LiDAR) has become a core tool for forestry resource surveys due to its ability to acquire high-precision three-dimensional information. Traditional tree detection relies primarily on manual field measurements or visual interpretation of optical images, which can be inefficient, costly, and subjective. In recent years, automated detection methods based on point cloud data have become a research hotspot. In particular, density clustering algorithms (such as DBSCAN and HDBSCAN) have shown potential for individual tree segmentation due to their adaptability to irregularly shaped data.

[0003] The traditional DBSCAN algorithm requires manual global parameter setting and is unable to adapt to changes in stand structure. While the traditional HDBSCAN algorithm addresses some of these issues through hierarchical clustering, it still suffers from the following shortcomings: 1) It fails to consider the vertical distribution of trees, making it difficult to distinguish between different tree species and determine the true number of trees; 2) it is sensitive to point cloud density and is prone to missing detections in sparsely populated areas. Therefore, it is crucial to design an automated method for detecting tree count and location based on the HDBSCAN algorithm. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for automatic detection of the number and location of trees based on the HDBSCAN algorithm, which accurately and efficiently completes the tree detection task through multi-level threshold separation, anisotropic filtering technology and multi-dimensional verification, so as to improve detection accuracy and anti-interference ability.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A method for automatically detecting the number and location of trees in a forest based on the HDBSCAN algorithm comprises the following steps:

[0007] Perform multi-level threshold separation and denoising operations on the original point cloud data to obtain forest point cloud data;

[0008] The density clustering algorithm after dynamic parameter optimization is used to perform initial tree recognition on the tree point cloud data to obtain the initial recognition results;

[0009] Perform three-dimensional morphological correction and adjacent target fusion on the initial recognition result to obtain the predicted recognition result;

[0010] The predicted recognition results are verified by multi-dimensional features to obtain the final recognition results.

[0011] Optionally, a multi-level threshold separation and denoising operation is performed on the original point cloud data to obtain forest point cloud data, including:

[0012] The original point cloud data is separated by adaptive elevation threshold to obtain point cloud dense areas and point cloud sparse areas;

[0013] Perform virtual point completion on the sparse area of the point cloud according to density guidance to obtain the completed area;

[0014] Anisotropic density filtering is used to eliminate outlier noise points in dense point cloud areas and completed areas to obtain forest point cloud data.

[0015] Optionally, the original point cloud data is separated by an adaptive elevation threshold to obtain point cloud dense areas and point cloud sparse areas, including:

[0016] The original point cloud data is divided into multiple analysis windows with a radius of 3m;

[0017] 25% of the point cloud elevation distribution within the analysis window is used as the window benchmark value;

[0018] The window reference value is dynamically adjusted according to the maximum height difference and average slope in the analysis window to obtain the window threshold;

[0019] Perform Gaussian weighting operation on the window threshold to obtain the separation threshold;

[0020] The original point cloud data is divided into point cloud dense area and point cloud sparse area according to the separation threshold.

[0021] Optionally, virtual point completion is performed on the sparse area of the point cloud according to density guidance to obtain a completed area, including:

[0022] Extend along the gradient direction of the sparse area of the point cloud to obtain the virtual direction;

[0023] The virtual height is obtained by interpolating the height trend surface of the point cloud dense area adjacent to the point cloud sparse area;

[0024] The virtual point is determined according to the virtual direction and virtual height, and the virtual point is inserted into the sparse area of the point cloud to obtain the completed area.

[0025] Optionally, anisotropic density filtering is used to eliminate outlier noise points in dense point cloud areas and complement areas to obtain forest point cloud data, including:

[0026] The valid area is constructed with any point in the point cloud dense area or the completed area as the center point; the horizontal radius of the valid area is 3m and the vertical radius is 0.5m;

[0027] The spatial proximity factor of the effective area is obtained according to the Gaussian weighted distance decay function;

[0028] The difference outliers in the effective area are obtained through the height variance of the point cloud, and the exponential penalty term is applied to the difference outliers to obtain the height consistency factor;

[0029] The composite density is obtained by normalizing the spatial proximity factor and the height consistency factor;

[0030] Eliminate points with composite density lower than the primary filtering threshold to obtain primary filtering data; the primary filtering threshold is 40% of the average density of the effective area;

[0031] The primary filtering threshold is adjusted according to the density change rate of the primary filtering data to obtain the secondary filtering threshold. The primary filtering threshold is replaced by the secondary filtering threshold for data filtering to obtain the forest point cloud data.

[0032] Optionally, the step of dynamically optimizing parameters of the density clustering algorithm includes:

[0033] Calculate the spatial dispersion and vertical concentration of forest point cloud data;

[0034] The core parameters of the HDBSCAN algorithm are updated according to the spatial dispersion and vertical concentration; the core parameters include: minimum cluster size and core distance threshold.

[0035] Optionally, the calculation formulas for the minimum cluster size and core distance threshold are: M min =[15·(2-e -η )]·ζ 0.5 ; Among them, M min is the minimum cluster size, ε c is the core distance threshold, η is the spatial dispersion, and ζ is the vertical concentration.

[0036] Optionally, the initial recognition result is subjected to three-dimensional morphological correction and adjacent target fusion to obtain a predicted recognition result, including:

[0037] Construct a triangular space topological network of the initial recognition results;

[0038] The morphological compatibility is calculated based on the triangular space topological network; the morphological compatibility includes: the deviation of the main axis direction angle, the ratio of the projected overlapping area and the similarity of the height distribution;

[0039] When the morphological compatibility is higher than the fusion threshold, the adjacent initial recognition results are fused and the morphological compatibility is recalculated until the morphological compatibility is lower than the fusion threshold.

[0040] Optionally, the calculation formula for morphological compatibility is: Among them, C l is the morphological compatibility, Sb is the projection overlap area ratio, H s is the height distribution similarity, and Δθ is the angle deviation of the principal axis direction.

[0041] Optionally, the predicted recognition result is verified by multi-dimensional features to obtain the final recognition result, specifically: a decision function is constructed based on the eigenvalues of the covariance matrix, the projection and convex hull areas, and the height standard deviation, and the predicted recognition result is verified based on the decision function to obtain the final recognition result; the expression of the decision function is: Among them, V t To verify the standard, λ1 and λ2 are the two eigenvalues of the covariance matrix, A p is the projected area, A c is the convex hull area, σ z is the height standard deviation, is the average tree height, α, β and γ are weight coefficients.

[0042] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: The present invention provides a method for automatically detecting the number and location of trees based on the HDBSCAN algorithm, the method comprising: performing multi-level threshold separation and denoising operations on the original point cloud data to obtain tree point cloud data; performing initial tree recognition on the tree point cloud data using a density clustering algorithm after dynamic parameter optimization to obtain an initial recognition result; performing three-dimensional morphological correction and adjacent target fusion on the initial recognition result to obtain a predicted recognition result; and performing multi-dimensional feature verification on the predicted recognition result to obtain a final recognition result. This method accurately and efficiently completes the tree detection task through multi-level threshold separation, anisotropic filtering technology, and multi-dimensional verification, thereby improving detection accuracy and anti-interference ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0044] Figure 1 This is a flowchart of the automatic detection of the number and position of trees in the present invention;

[0045] Figure 2 This is a multi-level threshold separation and denoising flow chart of the present invention;

[0046] Figure 3 This is a data separation flow chart of the present invention;

[0047] Figure 4 This is a flowchart of the virtual point completion of the present invention;

[0048] Figure 5 This is a flow chart of anisotropic density filtering of the present invention;

[0049] Figure 6 This is a flow chart of the three-dimensional morphology correction of the present invention. DETAILED DESCRIPTION

[0050] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0051] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0052] like Figure 1 As shown, the present invention provides a method for automatically detecting the number and location of trees based on the HDBSCAN algorithm, comprising the following steps:

[0053] Step 100: Perform multi-level threshold separation and denoising operations on the original point cloud data to obtain forest point cloud data; the specific steps are as follows: Figure 2 Shown, including:

[0054] Step 101: Separate the original point cloud data by using the adaptive elevation threshold to obtain the point cloud dense area and the point cloud sparse area; the specific steps are as follows: Figure 3 Shown, including:

[0055] Step 101.1: Divide the original point cloud data into multiple analysis windows with a radius of 3m;

[0056] Specifically, the original point cloud is projected onto a horizontal plane, and a square grid with a side length of 3 meters is created to cover the entire point cloud area. The sliding window step size of the square grid is 1.5 meters.

[0057] Step 101.2: 25% of the point cloud elevation distribution within the analysis window is used as the window benchmark value;

[0058] Specifically, an octree data structure is used to perform rapid spatial retrieval of point clouds within the analysis window, obtaining the number of point clouds, elevation distribution, and coordinate range within each window. All point clouds within each window are sorted in ascending order of elevation, with the 25th percentile of the sorted points used as the window reference value. This effectively eliminates interference from low ground points while retaining point clouds at the base of vegetation.

[0059] Step 101.3: Dynamically adjust the window reference value based on the maximum height difference and average slope within the analysis window to obtain the window threshold;

[0060] Specifically, the maximum height difference is obtained by the difference between the maximum height and the minimum height. A least squares plane is fitted based on the point cloud within the analysis window, and the average slope is determined based on the angle between the plane's normal vector and the vertical direction. The window threshold is obtained by adding the window reference value, the maximum height difference, and the average slope using the height difference compensation weight and the slope correction amount. The window threshold calculation formula is:

[0061] z th =z 25 +k1·Δz max +k2·tan(δ);

[0062] Among them, z th is the window threshold, z 25 is the window reference value, Δz max is the maximum height difference, tan(δ) is the average slope, δ is the angle between the normal vector of the plane and the vertical direction, k1 and k2 are the height difference compensation weight and slope correction amount, respectively, which are 0.15 and 0.3 in this embodiment.

[0063] Step 101.4: Perform a Gaussian weighting operation on the window threshold to obtain a separation threshold;

[0064] Specifically, a 3×3 window matrix is established with the current window as the center, and Gaussian weights are assigned to each window matrix. The window thresholds are weighted and fused using Gaussian weights to achieve a smoothing effect that eliminates sudden changes in the threshold value of a single window.

[0065] Step 101.5: Divide the original point cloud data into a point cloud dense area and a point cloud sparse area according to the separation threshold.

[0066] Specifically, if the point cloud density is greater than the separation threshold + 0.5m, the point cloud is determined as a point cloud dense area, otherwise it is determined as a point cloud sparse area.

[0067] Step 102: Perform virtual point completion on the sparse area of the point cloud according to density guidance to obtain the completed area; the specific steps are as follows: Figure 4 Shown, including:

[0068] Step 102.1: Extend along the gradient direction of the sparse area of the point cloud to obtain a virtual direction;

[0069] Specifically, the sparse area is divided into 0.5m×0.5m grid cells. The number of point clouds within each cell is counted to generate an initial density distribution map. Gaussian smoothing is then used to eliminate local fluctuation noise to obtain a continuous density field. The Sobel operator is used to perform gradient detection on the continuous density field. The horizontal gradient vector of each grid cell is calculated, and the direction with the largest gradient modulus is selected as the extension direction.

[0070] Step 102.2: Obtain a virtual height by interpolating the height trend surface of the dense point cloud area adjacent to the sparse point cloud area;

[0071] Specifically, a 1-meter-wide transition zone was extracted from the boundary between a sparse point cloud region and an adjacent dense point cloud region as an interpolation sample. A quadratic polynomial surface was fitted to the elevation distribution of the transition zone point cloud to obtain the height trend surface equation, which was solved using the least squares method. The solution was considered the virtual height.

[0072] Step 102.3: Determine a virtual point based on the virtual direction and virtual height, and insert the virtual point into the sparse area of the point cloud to obtain a completed area.

[0073] Step 103: Eliminate outlier noise points in dense point cloud areas and completed areas through anisotropic density filtering to obtain forest point cloud data. Figure 5 Shown, including:

[0074] Step 103.1: Construct a valid region with any point in the dense point cloud region or the completed region as the center point;

[0075] Specifically, the horizontal radius of the effective area is 3m to match the typical tree crown projection range, and the vertical radius is 0.5m to match the continuity characteristics of the tree trunk. The area size is also dynamically adjusted according to the local point cloud density. If the local point cloud density is lower than 50 points / m 3 , the horizontal radius is expanded to 4m to capture potential sparse tree trunk points; if the local point cloud density is higher than 200 points / m 3 If the terrain slope is greater than 25°, the vertical radius will be expanded to 0.8m to accommodate the inclined trunk points.

[0076] Step 103.2: Obtain the spatial proximity factor of the valid area according to the Gaussian weighted distance decay function;

[0077] Specifically, the horizontal distance weight of each point from the center point is calculated with a standard deviation of 1.2m, where the weight decreases by 30% for every 1m increase in horizontal distance. The vertical distance weight of each point from the center point is calculated with a standard deviation of 0.3m, where the weight decreases to 10% for every 0.1m increase in vertical distance.

[0078] Step 103.3: Obtain the difference outliers in the valid area through the point cloud height variance, and apply an exponential penalty term to the difference outliers to obtain the height consistency factor;

[0079] Specifically, if the variance is greater than 0.5m 2 It is determined that there is a height anomaly, and the point where the actual height deviates from the average height of the area by 1.2 times the variance is marked as a difference anomaly point. The expression of the exponential penalty term is: Where Δh is the deviation between the actual height and the average height.

[0080] Step 103.4: Normalize and multiply the spatial proximity factor and the height consistency factor to obtain the composite density;

[0081] Step 103.5: Eliminate points with composite density lower than the primary filtering threshold to obtain primary filtering data;

[0082] Specifically, the primary filtering threshold is 40% of the average density of the effective area. When the composite density is within the range of the average density of the effective area and 1.2 times the average density of the effective area, if at least three of the eight adjacent points have a composite density greater than 0.6 times the average density of the effective area, the point is retained; otherwise, the point is also removed.

[0083] Step 103.6: Adjust the primary filtering threshold according to the density change rate of the primary filtering data to obtain the secondary filtering threshold. Replace the primary filtering threshold with the secondary filtering threshold to perform data filtering and obtain forest point cloud data.

[0084] Specifically, the density change rate within a local window (5m×5m) is calculated with each retention point as the center point. If the density gradient within the window is less than -0.2 / m, the primary filtering threshold is increased by 30%; if the density gradient within the window is greater than 0.1 / m, the primary filtering threshold is reduced by 20%.

[0085] Step 200: performing initial tree recognition on the tree point cloud data using a density clustering algorithm after dynamic parameter optimization to obtain an initial recognition result;

[0086] Specifically, the steps for dynamic parameter optimization of the density clustering algorithm are as follows: calculating the spatial discreteness and vertical concentration of forest point cloud data, and updating the core parameters of the HDBSCAN algorithm according to the spatial discreteness and vertical concentration; the core parameters include: minimum cluster size and core distance threshold.

[0087] Furthermore, the calculation formulas for the minimum cluster size and core distance threshold are:

[0088] M min =[15·(2-e - η)]·ζ0.5 ;

[0089]

[0090] Among them, M min is the minimum cluster size, ε c is the core distance threshold, η is the spatial dispersion, and ζ is the vertical concentration.

[0091] Step 300: Perform three-dimensional morphological correction and adjacent target fusion on the initial recognition result to obtain the predicted recognition result; the specific steps are as follows: Figure 6 Shown, including:

[0092] Step 301: constructing a triangular space topological network of the initial recognition results;

[0093] Specifically, the three-dimensional centroid coordinates of each initial recognition result (cluster) are calculated and used as nodes of the topological network. Then, the adjacent nodes are connected in three-dimensional space based on the Delaunay triangulation algorithm to form a triangular space topological network.

[0094] Furthermore, the triangular spatial topology network only retains triangle edges with a side length of less than 6 meters (the upper limit of typical tree crown spacing), eliminating invalid long-distance connections. However, for objects with overlapping vertical projections, topological connections are forcibly established even if the distance between their centroids exceeds the threshold. In areas with a terrain slope greater than 15°, the horizontal projection distance is converted to the actual distance along the slope.

[0095] Step 302: Calculate morphological compatibility based on the triangular space topological network; morphological compatibility includes: principal axis direction angle deviation, projection overlap area ratio, and height distribution similarity;

[0096] Specifically, principal component analysis is performed on each target point cloud, and the direction of the first principal component is extracted as the principal axis. The angle θ between adjacent target principal axes is obtained, and the principal axis direction angle deviation is calculated. The calculation formula is: The target point cloud is projected onto a horizontal plane to generate a two-dimensional convex hull polygon. The ratio of the polygon overlap area to the union area is used as the projection overlap area ratio. The target point cloud is binned at 1m vertical intervals to generate a height distribution histogram. The cosine similarity between the two target histograms is used as the height distribution similarity.

[0097] More specifically, the morphological compatibility is calculated as:

[0098]

[0099] Among them, C l is the morphological compatibility, S b is the projection overlap area ratio, H sis the height distribution similarity, and Δθ is the angle deviation of the principal axis direction.

[0100] Step 303: When the morphological compatibility is higher than the fusion threshold, adjacent initial recognition results are fused and the morphological compatibility is recalculated until the morphological compatibility is lower than the fusion threshold.

[0101] It should be noted that by integrating the main axis direction and the overlapping projection features, over-segmentation caused by the interlacing of tree branches and leaves is avoided, thereby improving recognition accuracy. The influence of ground slope is taken into account in the process of constructing the triangular spatial topology network, reducing overall recognition errors.

[0102] Step 400: Perform multi-dimensional feature verification on the predicted recognition result to obtain the final recognition result.

[0103] Specifically, a decision function was constructed based on the eigenvalues of the covariance matrix, the projection and convex hull areas, and the height standard deviation. The verification criterion was obtained through the decision function. When the value of the verification criterion was greater than 0.82, the predicted recognition result was judged as a valid single tree.

[0104] More specifically, the decision function is expressed as:

[0105]

[0106] Among them, V t To verify the standard, λ1 and λ2 are the two eigenvalues of the covariance matrix, A p is the projected area, A c is the convex hull area, σ z is the height standard deviation, is the average tree height, α, β, and γ are weight coefficients, which are 0.6, 0.3, and 0.1 respectively in this embodiment.

[0107] The beneficial effects of the present invention are as follows:

[0108] 1) Accurate point cloud segmentation is achieved through multi-level threshold separation and dynamic elevation adjustment;

[0109] 2) Virtual point completion and anisotropic filtering techniques are used to improve the integrity of sparse regions and suppress noise interference;

[0110] 3) Through the parameter optimization process driven by spatial dispersion and vertical concentration, the clustering process can adapt to different tree species and improve the detection accuracy;

[0111] 4) Verifying the detection results based on multi-dimensional features greatly improves the accuracy and reliability of the final results.

[0112] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0113] The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A method for automatically detecting the number and location of trees based on the HDBSCAN algorithm, characterized in that: The steps include: Perform multi-level threshold separation and denoising operations on the original point cloud data to obtain forest point cloud data; Performing initial tree recognition on the tree point cloud data using a density clustering algorithm after dynamic parameter optimization to obtain an initial recognition result; Performing three-dimensional morphological correction and adjacent target fusion on the initial recognition result to obtain a predicted recognition result; The predicted recognition result is verified by multi-dimensional features to obtain the final recognition result.

2. The method for automatically detecting the number and location of trees based on the HDBSCAN algorithm according to claim 1, characterized in that: Perform multi-level threshold separation and denoising operations on the original point cloud data to obtain forest point cloud data, including: Separating the original point cloud data by using an adaptive elevation threshold to obtain a point cloud dense area and a point cloud sparse area; Performing virtual point completion on the sparse area of the point cloud according to density guidance to obtain a completed area; The outlier noise points in the point cloud dense area and the supplemented area are eliminated by anisotropic density filtering to obtain the forest point cloud data.

3. The method for automatically detecting the number and location of trees based on the HDBSCAN algorithm according to claim 2, characterized in that: The original point cloud data is separated by using an adaptive elevation threshold to obtain a point cloud dense area and a point cloud sparse area, including: Dividing the original point cloud data into multiple analysis windows with a radius of 3m; 25% of the point cloud elevation distribution within the analysis window is used as the window reference value; Dynamically adjusting the window reference value according to the maximum height difference and the average slope in the analysis window to obtain a window threshold; Performing a Gaussian weighted operation on the window threshold to obtain a separation threshold; The original point cloud data is divided into the point cloud dense area and the point cloud sparse area according to the separation threshold.

4. The method for automatically detecting the number and location of trees based on the HDBSCAN algorithm according to claim 2, characterized in that: Performing virtual point completion on the sparse area of the point cloud according to density guidance to obtain a completed area, including: Extending along the gradient direction of the sparse area of the point cloud to obtain a virtual direction; Obtaining a virtual height by interpolating a height trend surface of the point cloud dense area adjacent to the point cloud sparse area; A virtual point is determined according to the virtual direction and the virtual height, and the virtual point is inserted into the sparse area of the point cloud to obtain the completed area.

5. The method for automatically detecting the number and location of trees based on the HDBSCAN algorithm according to claim 2, characterized in that: Eliminating outlier noise points in the point cloud dense area and the supplemented area through anisotropic density filtering to obtain the forest point cloud data, including: A valid area is constructed with any point in the point cloud dense area or the supplemented area as the center point; the horizontal radius of the valid area is 3m and the vertical radius is 0.5m; Obtaining a spatial proximity factor of the valid area according to a Gaussian weighted distance decay function; Obtaining the difference outliers in the valid area through the point cloud height variance, and applying an exponential penalty term to the difference outliers to obtain a height consistency factor; Normalizing the spatial proximity factor and the height consistency factor to obtain a composite density; Eliminate points whose composite density is lower than the primary filtering threshold to obtain primary filtering data; the primary filtering threshold is 40% of the average density of the effective area; The primary filtering threshold is adjusted according to the density change rate of the primary filtering data to obtain a secondary filtering threshold, and the primary filtering threshold is replaced by the secondary filtering threshold to perform data filtering to obtain the forest point cloud data.

6. The method for automatically detecting the number and location of trees based on the HDBSCAN algorithm according to claim 1, characterized in that: The steps of dynamically optimizing the parameters of the density clustering algorithm include: Calculating the spatial dispersion and vertical concentration of the forest point cloud data; The core parameters of the HDBSCAN algorithm are updated according to the spatial dispersion and the vertical concentration; the core parameters include: a minimum cluster size and a core distance threshold.

7. The method for automatically detecting the number and location of trees based on the HDBSCAN algorithm according to claim 6, characterized in that: The calculation formulas for the minimum cluster size and the core distance threshold are: min =[15·(2-e -η )]·ζ 0.5 ; Among them, M min is the minimum cluster size, ε c is the core distance threshold, η is the spatial dispersion, and ζ is the vertical concentration.

8. The method for automatically detecting the number and location of trees based on the HDBSCAN algorithm according to claim 1, characterized in that: Performing three-dimensional morphological correction and adjacent target fusion on the initial recognition result to obtain a predicted recognition result, including: Constructing a triangular space topological network of the initial recognition results; The morphological compatibility is calculated based on the triangular space topological network; the morphological compatibility includes: the deviation of the main axis direction angle, the ratio of the projected overlapping area and the similarity of the height distribution; When the morphological compatibility is higher than a fusion threshold, the adjacent initial recognition results are fused and the morphological compatibility is recalculated until the morphological compatibility is lower than the fusion threshold.

9. The method for automatically detecting the number and location of trees based on the HDBSCAN algorithm according to claim 8, characterized in that: The calculation formula of the morphological compatibility is: Among them, C l is the morphological compatibility, S b is the projection overlap area ratio, H s is the height distribution similarity, and Δθ is the angle deviation of the principal axis direction.

10. The method for automatically detecting the number and location of trees based on the HDBSCAN algorithm according to claim 1, characterized in that: The predicted recognition result is subjected to multi-dimensional feature verification to obtain the final recognition result, specifically: a decision function is constructed based on the covariance matrix eigenvalues, projection and convex hull areas, and height standard deviations, and the predicted recognition result is verified based on the decision function to obtain the final recognition result; the expression of the decision function is: Among them, V t To verify the standard, λ1 and λ2 are the two eigenvalues of the covariance matrix, A p is the projected area, A c is the convex hull area, σ z is the height standard deviation, is the average tree height, α, β and γ are weight coefficients.

Images