Broad-leaved forest individual tree segmentation method and system based on branch and trunk information guidance

Abstract

The present invention relates to the technical field of terrestrial laser scanning-based forestry data processing, and disclosed are a broad-leaved forest individual tree segmentation method and system based on branch and trunk information guidance. The method comprises: taking a terrestrial laser scanning-based broad-leaved forest point cloud as a processing object, and using RANSAC cylinder fitting in combination with growth characteristics of a tree trunk to perform tree trunk detection; extracting, starting from the top of the tree trunk in a low-vegetation area, tree branches and trunk by means of segmented growth of the branches and trunk in combination with changes in the thickness of branch and trunk segments; and segmenting, starting from the tips of the branches, a tree crown leaf point cloud by means of layer-by-layer growth. The present invention achieves a stronger tree trunk detection capability under the conditions of lush low vegetation and complex terrain, and moreover, the present invention can achieve accurate segmentation of tree crowns when large and small tree crowns are intertwined and multiple tree crowns are closely surrounded. The algorithm is highly efficient and easy to use, and has great practical significance for improving the semantic understanding capability of forest scenes, and assisting in forest resource survey, vegetation ecology research, and calibration and verification of satellite remote sensing products.

Description

A method and system for segmenting single trees in broad-leaved forests based on branch information guidance Technical Field

[0001] The present invention belongs to the technical field of ground laser radar forestry data processing, and in particular relates to a broadleaf forest single tree segmentation method and system based on branch and trunk information guidance. Background Art

[0002] At present, Light Detection and Ranging (LiDAR) is an active remote sensing technology developed in recent decades. It can depict the three-dimensional structural information of forests through massive spatial points (i.e. point clouds) or waveforms. It has become the mainstream means of extracting forest structural and functional parameters.

[0003] Terrestrial Laser Scanning (TLS) can acquire high-density, high-precision point cloud data. It performs multi-station panoramic scans of forests, capturing detailed details of every tree's branches and foliage. While ground-based LiDAR data acquisition typically covers no more than one hectare, the forest parameters estimated from it are often considered ground truth and further used for calibration and validation of large-scale remote sensing estimation products. Ground-based LiDAR has been used to extract a variety of forest parameters, including diameter at breast height (DBH), tree height, tree position, trunk curve; wood volume / aboveground biomass; leaf area index (LAI) and leaf inclination distribution. Tree segmentation is a crucial prerequisite for obtaining these forest parameters through TLS. Most current tree segmentation methods are based on airborne LiDAR data. These methods typically begin by detecting treetops from point clouds or canopy height models (CHMs), followed by top-down tree segmentation using region growing or watershed algorithms. While segmentation works well for coniferous forests with distinct, loosely spaced canopies, it struggles with broadleaf forests, which have irregular, dense, and layered canopies. Ground-based LiDAR can capture information about understory branches, allowing us to first detect tree trunks based on features like cylindrical shape and upright growth, then segment the canopy from the bottom up using region growing.

[0004] Current tree segmentation methods based on TLS data include region-growing, graph-cut, and layer-wise segmentation. Region-growing methods primarily segment individual trees by gradually merging branches and leaves based on distance and angle rules, based on trunk location. These methods place high demands on data quality, for example, requiring multi-station scanning to minimize point cloud loss caused by occlusion. In the case of intertwined tree crowns, these methods also rely on extensive manual post-processing. Graph-cut methods first construct a graph for the entire point cloud or overlapping crown regions, then segment the crowns by minimizing a cost function. These methods have fewer steps and require little prior knowledge or manual interaction. However, these methods often rely on complex graph-cut algorithms, making them difficult to understand and computationally inefficient. Layer-wise segmentation methods extract crown outlines layer by layer to construct the tree's outer shape, which is then used for tree point cloud segmentation. Currently, these layer-wise segmentation methods are relatively few and incomplete. Because they primarily utilize crown outlines for segmentation and do not require high branch integrity, these methods have the potential to efficiently and accurately segment heavily occluded TLS data, and are a promising area of ​​research for future research. Technical issues

[0005] Through the above analysis, the problems and defects of the existing technology are as follows:

[0006] (1) The current single tree segmentation method based on TLS data has poor trunk detection accuracy when the understory vegetation is dense and the terrain is steep and complex.

[0007] (2) The current single-tree segmentation method based on TLS data has difficulty in accurately segmenting the canopy of broad-leaved forests with high canopy density. Technical Solutions

[0008] In response to the problems existing in the prior art, the present invention provides a method and system for segmenting individual trees in broad-leaved forests based on branch and trunk information guidance.

[0009] The present invention is achieved as follows: a broad-leaved forest single tree segmentation method based on branch information guidance, the broad-leaved forest single tree segmentation method based on branch information guidance uses the broad-leaved forest point cloud of the ground-based lidar as the processing object, uses RANSAC cylindrical fitting combined with the trunk growth characteristics to detect the trunk; starts from the top of the trunk in the low vegetation area, extracts the tree branches by growing the branches in segments and combining the changes in the thickness of the branch segments; starts from the end of the branch, the crown and leaf point cloud is segmented by layer-by-layer growth.

[0010] Furthermore, the broad-leaved forest single tree segmentation method based on branch information guidance includes the following steps:

[0011] Step 1: Trunk detection: Identify and extract tree trunks surrounded by low vegetation under the forest, laying the foundation for subsequent branch extraction and crown segmentation. Trunk detection includes: relative height stratification, layer-by-layer clustering, cylinder fitting, and trunk segmentation.

[0012] Step 2: Branch Extraction: Based on the trunk extraction in the low vegetation area, we take advantage of the fact that leaves in broad-leaved forests are distributed in the outer layer of the canopy, and branches grow from thick to thin, with a sudden increase in density when the twigs transition to leaves. We extract the remaining trunks and branches by growing layer by layer from bottom to top, stopping when they encounter leaves.

[0013] Step 3: Crown segmentation: Based on branch extraction, all branch point clouds are used as seed points, and the interwoven crowns are accurately segmented through breadth-first layer-by-layer growth. The growth rate of breadth-first growth in different directions is the same. When crowns of different sizes or shapes are interwoven, directly using this growth method cannot accurately segment the crowns.

[0014] Furthermore, the step 1 specifically includes:

[0015] 1-1. First, use the cloth filter algorithm CSF in the open source software CloudCompare to separate ground points. Then, for non-ground points with a relative height less than H, layer them according to their relative ground height, with a layer spacing of 0.2m. The relative ground height is the distance from the point to the nearest ground point. Ground points are downsampled at a 0.2m point spacing.

[0016] 1-2. Perform Euclidean clustering on each layer of point cloud. The Euclidean clustering algorithm divides the point cloud into independent clusters according to the Euclidean distance, so that the closest distance between any two clusters in the cluster is greater than the set value;

[0017] 1-3. Use the random sampling consensus algorithm RANSAC to perform cylindrical fitting on each cluster. The cylindrical model is defined as , L is the cylinder axis, and the vector and a little constitute; is the radius of the cylinder, and the distance d from any point p to the surface of the cylinder model can be calculated by the following formula;

[0018] ;

[0019] The RANSAC algorithm estimates the cylindrical model by randomly selecting two points with normal lines. Points with a distance to the model surface less than a certain value are called inliers. The RANSAC algorithm continuously performs model estimation and inlier extraction within a specified maximum number of cycles. After the cycle ends, the cylindrical model with the largest number of inliers is the fitting result.

[0020] After fitting, the cylindrical models were preliminarily filtered to remove obvious non-trunk cylindrical models;

[0021] 1-4. Merge all layered cylindrical models belonging to the same tree trunk and use the merged cylindrical model to extract the corresponding tree trunk point cloud.

[0022] Furthermore, extracting the corresponding tree trunk point cloud specifically includes:

[0023] 1-4-1. Identify cylindrical models belonging to the same tree trunk: Project the centroid of each cylindrical model onto the horizontal plane, then perform Euclidean clustering on the projected centroids. The cylindrical models corresponding to the centroids belonging to the same cluster belong to the same tree trunk.

[0024] 1-4-2. Cylindrical model merging and trunk extraction: First, the cylindrical models belonging to the same trunk and with a number of not less than two are merged into an overall cylindrical model. The axis, axis point, and radius of the overall cylindrical model are respectively the vector sum of the axis of the cylindrical model to be merged, the center of gravity of the axis point of the cylindrical model to be merged, and the average value of the radius of the cylindrical model to be merged; the obtained overall cylindrical model is used to extract the trunk point cloud in the low vegetation area, that is, the formula in steps 1-3 is used to extract points whose distance from the surface of the overall cylindrical model is less than k_d*the radius of the overall cylindrical model. k_d is taken as 0.5. If k_d is too small, the trunk extraction is incomplete. If k_d is too large, the adjacent ground and low vegetation will be absorbed. The trunk point cloud with a height less than 0.8*H is removed.

[0025] Furthermore, the step 2 specifically includes:

[0026] 2-1. Initial seed branch segment generation: The point at the top of the tree trunk in the low vegetation area with a height of length is the seed branch segment, and it is divided into max_layers = ceil(length / pace) layers, with the top layer being the last layer; ceil means rounding up;

[0027] 2-2. Branch growth: Use the last layer of the seed branch segment as the seed point to perform a radius neighborhood search with a search radius of pace. Use the searched point as a new seed point to continue searching, increasing one layer at a time, until no new seed points are generated or the number of layers reaches max_layers. When the number of layers in the seed branch segment is less than max_layers, increase the radius of the first layer search to length to bridge the gap.

[0028] 2-3. Branch segment clustering and identification:

[0029] First, perform Euclidean clustering on the branch segment point set obtained by growth, and then identify and process each cluster through the following judgments:

[0030] 1) If the seed branch segment contains max_layers layers and the cluster meets one of the following conditions, then the cluster is a branch segment. The cluster is extracted and used as a seed branch segment to continue growing;

[0031] A. The cluster contains max_layers growth layers and the cluster points are less than pnk*seed branch points, where pnk is the point change coefficient and pnk is 3;

[0032] B. The number of clusters is less than max_layers and the search radius is increased less than max_long_search;

[0033] 2) If the number of seed branches is less than max_layers and the cluster contains max_layers growth layers, the cluster is a branch. The cluster is extracted and used as a seed branch to continue growing.

[0034] Furthermore, the step three specifically includes:

[0035] 3-1. Layer-by-layer growth: Starting from the end of the branch, grow one layer at a time, with a thickness of 2*point spacing, which is 20cm. Use the new growth point as the seed point and continue growing until no new growth point is generated; set the maximum number of growth layers to 10, that is, the total growth thickness is 2m;

[0036] 3-2. Post-processing: All divided and undivided points are gridded in a xoy plane with a grid side length of 10 cm. Undivided points are assigned to the tree with the closest divided point in the grid; undivided points are assigned to the tree with the closest divided point and a linear distance less than 10 cm. The remaining points that have not been divided after processing are considered noise points or non-tree points.

[0037] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the broad-leaved forest single tree segmentation method guided by branch information.

[0038] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to execute the broad-leaved forest single tree segmentation method guided by branch information.

[0039] Another object of the present invention is to provide an information data processing terminal, which is used to implement the broad-leaved forest single tree segmentation method based on branch and trunk information guidance.

[0040] Another object of the present invention is to provide a broad-leaved forest single tree segmentation system based on branch information guidance based on the broad-leaved forest single tree segmentation method based on branch information guidance, the broad-leaved forest single tree segmentation system based on branch information guidance comprising:

[0041] The tree trunk prediction module is used to process the broad-leaved forest point cloud of ground-based lidar and detect tree trunks using RANSAC cylindrical fitting combined with tree trunk growth characteristics;

[0042] The branch extraction module is used to extract tree branches starting from the top of the tree trunk in the low vegetation area by growing the branches in segments and combining the changes in the thickness of the branch segments;

[0043] The point cloud segmentation module is used to segment the crown and leaf point cloud by layer-by-layer growth starting from the branch end. Beneficial effects

[0044] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0045] First, in view of the technical problems existing in the above-mentioned prior art and the difficulty of solving the problems, closely combining the technical solutions to be protected by the present invention and the results and data during the research and development process, we will analyze in detail and deeply how the technical solutions of the present invention solve the technical problems, and some creative technical effects brought about after solving the problems. The specific description is as follows: In order to solve the problem of low accuracy of trunk detection and crown segmentation in complex broad-leaved forests (mainly referring to forests with dense low vegetation and serious interweaving of tree crowns), the present invention provides a new method for segmenting single trees in broad-leaved forests based on branch information guidance. The method uses the broad-leaved forest point cloud of a ground-based lidar as the processing object, first accurately detects the trunk through cylindrical fitting combined with the trunk growth characteristics, and then accurately segments the crown through differential regional growth guided by branch information.

[0046] Second, considering the technical solution as a whole or from a product perspective, the technical effects and advantages of the technical solution to be protected by this invention are described in detail as follows: Compared with existing single-tree segmentation methods, this invention has stronger trunk detection capabilities in conditions of dense, low vegetation and complex terrain. Furthermore, this invention can accurately segment tree crowns even when large and small crowns intersect and multiple crowns are closely enclosed. The algorithm is highly efficient and easy to use, and has significant practical significance for improving semantic understanding of forest scenes, assisting in forest resource surveys, vegetation ecology research, and the calibration and verification of satellite remote sensing products.

[0047] Third, as auxiliary evidence for the inventiveness of the claims of the present invention, it is also reflected in the following important aspects:

[0048] The technical solution of the present invention solves the technical problems that people have been eager to solve but have never been able to solve successfully:

[0049] The accuracy of individual tree segmentation in ground-based LiDAR broadleaf forest point clouds directly impacts the accuracy of subsequent estimation of tree-related parameters. Several current tree segmentation methods face technical challenges, such as difficulty accurately segmenting complex intertwined tree heights and reliance on manual correction. This paper, based on the growth structure of broadleaf forests and the characteristics of ground-based LiDAR data, proposes using branch and trunk information to guide tree height segmentation. This approach effectively addresses these issues through software system development. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] 1 is a flow chart of a method for segmenting a single tree in a broad-leaved forest based on branch information guidance according to an embodiment of the present invention;

[0051] FIG2 is a schematic diagram of a broadleaved forest tree segmentation method based on branch information guidance according to an embodiment of the present invention;

[0052] FIG3 is a diagram showing the results of single tree segmentation provided by an embodiment of the present invention: a. Magnolia forest; b. Phoebe forest;

[0053] 4 is a diagram of accuracy verification results provided by an embodiment of the present invention: a. Magnolia forest tree height estimation accuracy; b. Magnolia forest crown width estimation accuracy; c. Phoebe forest tree height estimation accuracy; d. Phoebe forest crown width estimation accuracy. Modes for Carrying Out the Invention

[0054] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0055] As shown in FIG1 , the broadleaved forest single tree segmentation method based on branch information guidance provided by an embodiment of the present invention includes the following steps:

[0056] S101: Using the ground-based lidar broadleaf forest point cloud as the processing object, tree trunk detection is performed using RANSAC cylindrical fitting combined with tree trunk growth characteristics;

[0057] S102: Extract tree branches starting from the top of the tree trunk in the low vegetation area by segmenting and combining the thickness changes of the branch segments;

[0058] S103: Segment the crown and leaf point cloud by layer-by-layer growth starting from the branch end.

[0059] The present invention uses the three-dimensional laser point cloud data of broad-leaved forests acquired by ground-based lidar. Based on the idea of ​​regional growth, the method goes through trunk detection, branch extraction, and crown segmentation to establish a broad-leaved forest single tree segmentation method that uses branch information to guide crown segmentation.

[0060] Example 1:

[0061] The embodiment of the present invention provides a method for segmenting individual trees in a broad-leaved forest based on branch information guidance, comprising the following steps:

[0062] Step 1: Tree trunk detection:

[0063] This step aims to identify and extract tree trunks surrounded by low vegetation, laying the foundation for subsequent branch extraction and crown segmentation. Trunk detection consists of four steps: relative height stratification, layer-by-layer clustering, cylinder fitting, and trunk segmentation.

[0064] 1-1. First, use the Cloth Simulation Filter (CSF) algorithm in the open-source software CloudCompare to separate ground points. Then, for non-ground points with a relative ground height less than H (which depends on the overall height of low vegetation and is set to 1 meter in this paper), stratify them according to their relative ground height, with a stratification interval of 0.2 meters. The relative ground height is the distance from the point to the nearest ground point. To speed up the calculation of the relative ground height, ground points are downsampled to a 0.2-meter interval.

[0065] 1-2. Perform Euclidean clustering on each layer of point cloud. The Euclidean distance clustering algorithm divides the point cloud into independent clusters based on Euclidean distance, so that the closest distance between any two clusters in the cluster is greater than a set value (here, 2 times the point spacing, which can be obtained by downsampling the spacing or counting the closest point distance).

[0066] 1-3. Use the Random Sampling Consensus (RANSAC) algorithm to perform cylinder fitting on each cluster. The cylinder model is defined as , L is the cylinder axis, and the vector and a little constitute; is the radius of the cylinder. The distance d from any point p to the surface of the cylinder model can be calculated using the following formula.

[0067]

[0068] The RANSAC algorithm estimates the cylindrical model by randomly selecting two points with normals. Points with a distance to the model surface less than a specific value (fixed to 1 cm in this invention) are called inliers. The RANSAC algorithm continuously performs model estimation and inlier extraction within a specified maximum number of cycles (fixed to 30 times in this invention). After the cycle is completed, the cylindrical model with the largest number of inliers is the fitting result.

[0069] After fitting, the cylindrical model is initially filtered to remove obvious non-trunk cylindrical models. A. The fitting radius is too large or too small. , B. Axial z component As long as a cylindrical model satisfies one of A and B, the cylindrical model will be removed. and It needs to be roughly determined according to the thickness of the tree trunk in the sample plot. = 2.5cm, = 15cm. Depending on the degree of inclination of the tree trunks in the sample plot, the trunks of the broad-leaved forest in the present invention generally grow upright, so =0.85( The value ranges from 0 to 1, with a larger value indicating a more vertical tree trunk).

[0070] 1-4. Merge all layered cylindrical models belonging to the same tree trunk and use the merged cylindrical model to extract the corresponding tree trunk point cloud.

[0071] 1-4-1. Identify cylindrical models belonging to the same tree trunk:

[0072] A tree trunk is horizontally divided into multiple layers, and a cylindrical model can be fitted in each layer. The cylindrical models belonging to the same tree trunk have close horizontal distances. Therefore, the centroid of the points inside each cylindrical model can be projected onto the horizontal plane. Then, Euclidean clustering is performed on the projected centroid points (the distance threshold is fixed at 0.1m). The cylindrical models corresponding to the centroid points belonging to the same cluster belong to the same tree trunk.

[0073] 1-4-2. Cylindrical model merging and trunk extraction:

[0074] Theoretically, simply combining the inner points of the cylindrical models belonging to the same tree trunk will form a trunk point cloud. However, due to occlusion, low vegetation attached to the trunk, and other reasons, some layers do not produce cylindrical models. However, if the number of cylindrical models belonging to the same trunk is too small (for example, only one), the cylindrical model may be an erroneous fit result generated by a non-trunk point cloud. To improve the accuracy and completeness of trunk extraction, at least two cylindrical models belonging to the same trunk are first merged into a global cylindrical model. The axis, axis point, and radius of the global cylindrical model are the vector sum of the axis of the cylindrical models to be merged (when the axial z component is negative, the axis needs to be flipped), the center of gravity of the axis point of the cylindrical models to be merged, and the average radius of the cylindrical models to be merged. Then, use the resulting global cylindrical model to extract the tree trunk point cloud within the low vegetation area (ground points and points with a height less than H relative to the ground). Specifically, use the formula in steps 1-3 to extract points whose distance from the global cylindrical model surface is less than k_d * the radius of the global cylindrical model. (In this invention, k_d is set to 0.5. A smaller k_d will result in incomplete trunk extraction, while a larger k_d will absorb the adjacent ground and low vegetation.) Finally, remove the tree trunk point cloud with a height less than 0.8 * H.

[0075] Step 2: Branch extraction:

[0076] Based on the extraction of tree trunks in low vegetation areas, this step uses the characteristics of broad-leaved forests, such as leaves distributed in the outer layer of the canopy, branches growing from thick to thin, and a sudden increase in density when twigs transition to leaves. The remaining trunks and branches are extracted by growing layer by layer from bottom to top, and growth stops when encountering leaves.

[0077] 2-1. Initial seed branch segment generation: The seed branch segment is defined as the point at the top of the tree trunk in the low vegetation area at a height of length (valued based on tree size; larger trees should have a higher value; in this case, 10 cm is used). The segment is divided into layers with max_layers = ceil(length / pace), with the top layer being the last layer. Ceil indicates rounding up.

[0078] 2-2. Branch Growth: Starting with the last layer of the seed branch segment as the seed point, a radius neighborhood search is performed with a search radius of pace (twice the point spacing). The searched points are used as new seed points and the search continues, increasing by one layer at a time, until no new seed points are generated or the number of layers reaches max_layers. (This ensures the correctness of the growth area and significantly reduces processing time by performing multiple searches to reach length.) In particular, when the number of layers in the seed branch segment is less than max_layers, the radius of the first layer search is increased to length to bridge the gap.

[0079] 2-3. Branch segment clustering and identification:

[0080] First, perform Euclidean clustering on the branch segment point set obtained by growth (the distance threshold is pace), and then identify and process each cluster through the following judgment:

[0081] 1) If the seed branch segment contains max_layers layers and the cluster meets one of the following conditions, then the cluster is a branch segment. The cluster is extracted and used as a seed branch segment to continue growing.

[0082] A. The cluster contains max_layers growth layers and the cluster point count is less than pnk * the number of seed branch points. pnk is the point count variation coefficient. Since branches grow from thicker to thinner, pnk can theoretically be fixed at 1. However, at branch bifurcations, the number of points in newly grown branches will be greater than the number of points in the seed branch, but this increase is not significant compared to the increase from thin branches to leaves. Therefore, this invention uses pnk as 3.

[0083] B. The cluster has less than max_layers growth layers and the search radius is increased less than max_long_search times (set to 50 times in the present invention to ensure that branches are fully extracted in the absence of occlusions and to reduce the degree of misextraction of leaves due to inherent gaps).

[0084] 2) If the number of seed branches is less than max_layers and the cluster contains max_layers growth layers, the cluster is a branch. The cluster is extracted and used as a seed branch to continue growing.

[0085] Step 3: Crown segmentation:

[0086] Based on branch extraction, this step uses all branch point clouds as seed points and achieves accurate segmentation of interwoven tree crowns through a breadth-first layer-by-layer growth method. Breadth-first growth has the same growth rate in different directions. When tree crowns of different sizes or shapes are intertwined, directly using this growth method cannot accurately segment the crowns. However, the present invention uses this growth method based on branch extraction. Due to the different growth starting points (branch ends), it can achieve the effect of differential growth, thereby achieving accurate segmentation of interwoven tree crowns.

[0087] 3-1. Growth layer by layer:

[0088] Since the growth process only considers spatial distance and does not pay attention to the detailed information of the point cloud, and there are many canopy leaf points and serious occlusion and missing, in order to improve the processing speed and overcome the occlusion effect, the canopy point cloud is downsampled (sampling interval 10cm) before growth, and the data is restored after the growth is completed. The growth process is: starting from the end of the branch, one layer is grown each time (the thickness is 2*point spacing, which is 20cm here), and the new growth point is used as the seed point to continue growing until no new growth point is generated. In order to avoid the crown of a single tree with insufficient branch extraction being excessively swallowed by adjacent trees, the maximum number of growth layers is set to 10, that is, the total growth thickness is 2m (20cm*10), because the thickness of the canopy leaf layer of the broad-leaved forest of the present invention does not exceed 2m overall.

[0089] 3-2, post-processing:

[0090] After the above growth, there are still a small number of scattered crown point clouds that have not been divided into corresponding individual trees. Here, these point clouds are divided in turn using two spatial proximity concepts. A. First, all divided and undivided points are gridded in the plane (xoy plane, with a grid side length of 10 cm), and then the undivided points are divided into individual trees to which the divided points in the grid with the closest vertical distance belong. B. The points that have not yet been divided are divided into individual trees to which the closest divided points belong with a straight-line distance of less than 10 cm. For the remaining points that have not been divided after processing A and B, they are regarded as noise points or non-tree points.

[0091] The broadleaved forest single tree segmentation system based on branch information guidance provided by an embodiment of the present invention includes:

[0092] The tree trunk prediction module is used to process the broad-leaved forest point cloud of ground-based lidar and detect tree trunks using RANSAC cylindrical fitting combined with tree trunk growth characteristics;

[0093] The branch extraction module is used to extract tree branches starting from the top of the tree trunk in the low vegetation area by growing the branches in segments and combining the changes in the thickness of the branch segments;

[0094] The point cloud segmentation module is used to segment the crown and leaf point cloud by layer-by-layer growth starting from the branch end.

[0095] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art will appreciate that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0096] The present invention will be further described in detail below using two broadleaf forest plots as examples with reference to the accompanying drawings. The development environment for this embodiment is Microsoft Visual Studio 2013, the external library PCL1.8.0, and the programming language C++.

[0097] The embodiment of the present invention provides a method for segmenting individual trees in a broad-leaved forest based on branch information guidance, comprising the following steps:

[0098] Step 1: Data collection and preprocessing:

[0099] The broad-leaved forest data collected in this paper involve two forest types, one of which is a magnolia forest located on the Qingshui River campus of the University of Electronic Science and Technology of Chengdu. The trees are of different sizes and irregular shapes, and the trees are arranged relatively randomly. There are both densely connected areas and clearly spaced areas between the crowns. There are many small trees close to the bottom of the large trees, which increases the difficulty of crown segmentation. The ground is flat and covered with grass about 20 cm thick. The TLS data was collected in May 2020, and the collection equipment was Leica C10. A total of 4 measuring stations were arranged, with one main station located in the center of the sample plot and three sub-stations located on the periphery of the sample plot to reduce the shielding effect.

[0100] The other is a nanmu forest located next to the Baoshikou Reservoir in Longquan Mountain, Chengdu. This forest is artificially planted but has long been unattended. The trees grow evenly and neatly, with heavily interwoven crowns, dense understory vegetation, and highly undulating terrain. These characteristics present significant challenges for trunk detection and crown segmentation. TLS data were collected in May 2022 using a Leica C10. Due to the strong understory obstruction, five observation stations were deployed: a main station in the center of the plot and four substations on the periphery.

[0101] Leica Cyclone software was used to splice, denoise, and downsample the collected data. Details of the preprocessed point cloud data are shown in Table 1. Next, the method presented in this paper was used to perform individual tree segmentation on the preprocessed data from the two broadleaf forest plots.

[0102] Table 1 Data details of broad-leaved forest plots

[0103] Plot area (m2) Number of trees Tree height (m) DBH (cm) Crown width (m) Number of points Spacing between points (m) Magnolia forest 30*30743.84-9.976.4-30.11.75-5.095,118,0300.02 Phoebe forest 20*20238.53-12.525.2-20.93.05-6.8310,890,4070.01

[0104] Step 2: perform tree trunk detection as described in step 1 of the present invention.

[0105] Step 3: Extract branches and trunks according to step 2 of the invention.

[0106] Step 4: perform tree crown segmentation as described in step 3 of the invention.

[0107] Step 5: Accuracy evaluation:

[0108] After the above steps one to four, the final result of the trunk point cloud extraction is shown in Figure 3. Accuracy evaluation was performed from two aspects: single tree detection and crown segmentation. In terms of single tree detection, there are actually 74 complete trees with a breast diameter greater than 5 cm in the magnolia forest, and 23 trees in the Longquanshan nanmu forest. The method of the present invention successfully detected all the single trees in the two sample plots. In the magnolia forest, 3 more trunks were detected, which were actually pedestrians; in the nanmu forest, 1 more incomplete trunk was detected at the boundary of the study area. The results show that the method of the present invention has excellent single tree detection capabilities for magnolia sample plots with sparse understory vegetation and flat terrain, and for nanmu sample plots with lush understory vegetation and large terrain undulations.

[0109] For crown segmentation, trees with intertwined crowns were selected from within the plot and carefully manually segmented as a reference baseline (30 trees were selected in the magnolia forest and 15 trees were selected in the nanmu forest). Reference trees were matched with algorithm-derived trees by trunk position, and the crown segmentation performance was comprehensively evaluated based on the degree of point cloud matching and the results of individual tree parameter extraction. In the magnolia forest, the proposed method achieved an IoU of 0.53 to 1.00 and a mIoU of 0.82. The RMSEs of tree height and crown width estimated from the individual tree segmentation results were 0.42 m (5%) and 0.41 m (12%), respectively (see Figures 4a and 4b). In the nanmu forest, the proposed method achieved an IoU of 0.81 to 0.96 and a mIoU of 0.88. The RMSEs of tree height and crown width estimated from the individual tree segmentation results were 0.05 m (0.5%) and 0.34 m (7%), respectively (see Figures 4c and 4d). The results show that the method of the present invention can accurately segment the intertwined crowns, whether it is a magnolia forest with large differences in tree size or a nanmu forest with uniform tree growth but closely surrounded crowns.

[0110] The accuracy of the crown segmentation of the method of the present invention is mainly affected by the accuracy of the branch extraction in the previous step. The branches and leaves of the nanmu forest are obviously layered, and the branches are fully exposed to the field of view under the forest, while the leaves of the magnolia forest have a larger span in the vertical direction, resulting in a large number of branches being blocked. In addition, compared with the magnolia forest, the nanmu forest is smaller in area and has more scanning stations. The above two factors make the branches of the nanmu forest fully portrayed in the point cloud, while the branches of the magnolia forest are seriously blocked and the twigs are not clear. Therefore, the branch extraction effect of the nanmu forest is much better than that of the magnolia forest, and naturally the crown segmentation accuracy of the nanmu forest is better than that of the magnolia forest. Although the point cloud quality of the magnolia forest is slightly poor, resulting in insufficient branch extraction, it can also robustly complete the crown segmentation and achieve good segmentation accuracy.

[0111] This example uses the proposed method to process ground-based LiDAR data from two broadleaved forest plots. Following the method steps, the same parameters were used to achieve individual tree segmentation for both plots. The segmentation results were then tested for accuracy from the perspectives of individual tree detection and crown segmentation. The results demonstrate that the proposed method achieves excellent segmentation results for both plots, which vary in tree species composition, understory density, terrain conditions, and point cloud quality. Furthermore, the computer configuration used was Windows 7, an Intel(R) Core(TM) i5-7500 (quad-core), and 8GB of RAM. The magnolia forest contained over five million points, and the nanmu forest contained over ten million points. The complete processing time for each plot was approximately 10-20 minutes, demonstrating the high efficiency of this method and its ability to meet the needs of individual tree segmentation in hectare-scale forests.

[0112] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A broad-leaved forest single tree segmentation method based on branch information guidance, characterized in that: The broad-leaved forest single tree segmentation method based on branch information guidance takes the ground-based lidar broad-leaved forest point cloud as the processing object, uses RANSAC cylindrical fitting combined with the trunk growth characteristics to detect the trunk; starts from the top of the trunk in the low vegetation area, extracts the tree branches by growing the branches in segments and combining the changes in the thickness of the branches; starts from the end of the branch, the crown and leaf point cloud is segmented by layer-by-layer growth.

2. The method for segmenting a single tree in a broad-leaved forest based on branch information guidance according to claim 1, characterized in that: The broad-leaved forest single tree segmentation method based on branch information guidance comprises the following steps: Step 1: Tree trunk detection: Identify and extract tree trunks surrounded by low vegetation under the forest, laying the foundation for subsequent branch extraction and crown segmentation; tree trunk detection includes: relative height stratification, layer-by-layer clustering, cylinder fitting, and tree trunk segmentation; Step 2: Extraction of branches and trunks: Based on the extraction of tree trunks in low vegetation areas, we use the characteristics of broad-leaved forest leaves distributed in the outer layer of the canopy, branches growing from thick to thin, and the density suddenly increasing when the twigs transition to leaves. We extract the remaining trunks and branches by growing layer by layer from bottom to top, and stop growing when encountering leaves. Step 3: Crown segmentation: Based on branch extraction, all branch point clouds are used as seed points, and accurate segmentation of intertwined crowns is achieved through breadth-first layer-by-layer growth. The growth rate of breadth-first growth in different directions is the same. When crowns of different sizes or shapes are intertwined, directly using this growth method cannot accurately segment the crowns.

3. The method for segmenting a single tree in a broad-leaved forest based on branch information guidance according to claim 2, characterized in that: The step 1 specifically includes: 1-1. First, use the cloth filter algorithm CSF in the open source software CloudCompare to separate the ground points, and then stratify the non-ground points whose relative height to the ground is less than H according to the relative ground height, with a stratification interval of 0.2m; the relative ground height is the distance from the point to the nearest ground point, and the ground points are downsampled at a point spacing of 0.2m; 1-2. Perform Euclidean clustering on each layer of point cloud. The Euclidean clustering algorithm divides the point cloud into independent clusters according to the Euclidean distance, so that the shortest distance between any two clusters in the cluster is greater than the set value; 1-3. Use the random sampling consensus algorithm RANSAC to perform cylinder fitting on each cluster. The cylinder model is defined as , L is the cylinder axis, and the vector and a little constitute; is the radius of the cylinder, and the distance d from any point p to the surface of the cylinder model can be calculated by the following formula; ； The RANSAC algorithm estimates the cylindrical model by randomly selecting two points with normal lines. Points whose distance to the model surface is less than a certain value are called inliers. The RANSAC algorithm continuously performs model estimation and inlier extraction within the specified maximum number of cycles. After the cycle ends, the cylindrical model with the largest number of inliers is the fitting result. After fitting, the cylindrical models were initially filtered to remove obvious non-trunk cylindrical models; 1-4. Merge all layered cylindrical models belonging to the same tree trunk, and use the merged cylindrical model to extract the corresponding tree trunk point cloud.

4. The method for segmenting a single tree in a broad-leaved forest based on branch information guidance according to claim 3, characterized in that: The extracting of the corresponding tree trunk point cloud specifically includes: 1-4-1. Identify the cylindrical models belonging to the same tree trunk: project the centroid of each point inside the cylindrical model onto the horizontal plane, and then perform Euclidean clustering on the projected centroid points. The cylindrical models corresponding to the centroid points belonging to the same cluster belong to the same tree trunk. 1-4-2. Cylindrical model merging and trunk extraction: First, the cylindrical models belonging to the same trunk and with a number of not less than two are merged into an overall cylindrical model. The axis, axis point, and radius of the overall cylindrical model are respectively the vector sum of the axis of the cylindrical model to be merged, the center of gravity of the axis point of the cylindrical model to be merged, and the average value of the radius of the cylindrical model to be merged; the obtained overall cylindrical model is used to extract the trunk point cloud in the low vegetation area, that is, the formula in step 1-3 is used to extract points whose distance from the surface of the overall cylindrical model is less than k_d*the radius of the overall cylindrical model. k_d is taken as 0.

5. If k_d is too small, the trunk extraction is incomplete. If k_d is too large, the adjacent ground and low vegetation will be absorbed. The trunk point cloud with a height less than 0.8*H is removed.

5. The method for segmenting a single tree in a broad-leaved forest based on branch information guidance according to claim 2, characterized in that: The step 2 specifically includes: 2-1. Initial seed branch segment generation: The point with the height length of the top of the tree trunk in the low vegetation area is the seed branch segment, and it is divided into max_layers = ceil (length / pace) layers, and the top layer is set as the last layer; ceil means rounding up; 2-2. Branch growth: Take the last layer of the seed branch as the seed point to search the radius neighborhood, the search radius is pace, and use the searched points as new seed points to continue searching, increasing one layer at a time, until no new seed points are generated or the number of layers reaches max_layers; when the number of layers contained in the seed branch is less than max_layers, increase the radius of the first layer search to length to cross the gap; 2-3. Branch segment clustering and identification: First, perform Euclidean clustering on the branch segment point set obtained by growth, and then identify and process each cluster through the following judgments: 1) If the seed branch segment contains max_layers layers and the cluster meets one of the following conditions, the cluster is a branch segment, and the cluster is extracted and used as a seed branch segment to continue growing; A. The cluster contains max_layers growth layers and the number of cluster points is less than pnk*the number of seed branch points, where pnk is the coefficient of point variation and pnk is 3; B. The cluster has less than max_layers growth layers and the search radius increases less than max_long_search times; 2) If the number of seed branches is less than max_layers and the cluster contains max_layers growth layers, the cluster is a branch. The cluster is extracted and used as a seed branch to continue growing.

6. The method for segmenting a single tree in a broad-leaved forest based on branch information guidance according to claim 2, characterized in that: The step three specifically includes: 3-1. Layer-by-layer growth: Starting from the end of the branch, grow one layer at a time, with a thickness of 2*point spacing, which is 20cm. Use the new growth point as the seed point to continue growing until no new growth point is generated; set the maximum number of growth layers to 10, that is, the total growth thickness is 2m; 3-2. Post-processing: All divided and undivided points are gridded in the xoy plane with a grid side length of 10 cm. Undivided points are divided into single trees with the closest vertical distance to the divided points in the grid; undivided points are divided into single trees with a linear distance less than 10 cm and the closest divided points; the remaining points that have not been divided after processing are regarded as noise points or non-tree points.

7. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the broad-leaved forest single tree segmentation method guided by branch information as described in any one of claims 1 to 6.

8. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the broad-leaved forest single tree segmentation method based on branch information guidance as claimed in any one of claims 1 to 6.

9. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the broad-leaved forest single tree segmentation method based on branch and trunk information guidance as described in any one of claims 1 to 6.

10. A broad-leaved forest single tree segmentation system based on branch information guidance based on the broad-leaved forest single tree segmentation method based on branch information guidance according to any one of claims 1 to 6, characterized in that: The broad-leaved forest single tree segmentation system based on branch information guidance includes: The tree trunk prediction module is used to detect tree trunks using the ground-based lidar broad-leaved forest point cloud as the processing object, using RANSAC cylindrical fitting combined with the tree trunk growth characteristics; The branch extraction module is used to extract tree branches from the top of the tree trunk in the low vegetation area by growing the branches in sections and combining the changes in the thickness of the branch segments; The point cloud segmentation module is used to segment the crown and leaf point cloud by layer-by-layer growth starting from the branch end.

Images