Inversion method of real leaf area index of broad-leaved tree
By preprocessing and subdividing the ground-based lidar point cloud data, the geometric characteristics of leaf dough blocks are calculated, and the problems of complexity and accuracy sensitivity in the inversion of the real leaf area index process in the existing technology are solved, achieving efficient and robust calculation of leaf area index.
Patent Information
- Application Number
- CN202411177168.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-08-26
AI Technical Summary
The process of the prior art is complicated when inverting the real leaf area index and the accuracy is sensitive to parameter selection, making it difficult to accurately obtain the real leaf area index of broadleaf trees.
By pre-processing, leaf point extraction, coarse clustering and sub-segmenting of ground-based lidar point cloud data, the geometric characteristics of leaf dough blocks are calculated to obtain the real leaf area index, and statistical filtering, singular value filtering, DBSCAN algorithm and The subdivision method based on normal vectors reduces parameter sensitivity.
The calculation process of leaf area index is simplified, the calculation robustness and accuracy are improved, and the impact of parameter selection is reduced. It is suitable for forestry resource surveys and forest change supervision in different scenarios.
Smart Images

Figure CN119066583A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of ground-based laser point cloud data processing, and specifically is an inversion method for the real leaf area index of broad-leaved trees based on ground-based laser radar point cloud data. Background Art
[0002] As an important component of the ecosystem, broad-leaved trees play an important role in carbon storage, biodiversity and climate regulation. Leaves are the main carrier of photosynthesis and other physiological functions of broad-leaved forests. They have multiple functions such as intercepting rainfall, absorbing light and gases, and their size and density are often characterized by the leaf area index. The leaf area index is defined as half of the area of all leaves per unit surface area. It is a core parameter for characterizing vegetation structure and is widely used in models such as ecosystems, climate and land patterns.
[0003] Remote sensing is the main way to obtain leaf area index over a large area. However, the leaf area index inverted by remote sensing includes two types: effective leaf area index and true leaf area index. The effective leaf area index refers to the leaf area index of a leaf community that satisfies spatial random distribution corresponding to the same gap ratio, which is generally smaller than the true leaf area index. The main difference between the true leaf area index and the effective leaf area index is that the true leaf area index takes into account the non-random distribution of leaves in the canopy and the leaf inclination angle. Compared with the effective leaf area index, the true leaf area index can more accurately quantitatively describe the leaves and reflect the photosynthesis capacity of vegetation. Therefore, obtaining the true leaf area index is more meaningful for monitoring the carbon fixation and climate regulation capabilities of broad-leaved forests.
[0004] LiDAR is an active remote sensing technology that has developed very rapidly in recent years. It can quickly obtain the three-dimensional structural information of the forest and accurately describe the internal geometric structure information of the trees, providing a way to invert the true leaf area index. Inspired by the Beer-Lambert law, Swinehart et al. proposed to simplify the canopy into a uniformly randomly distributed canopy, and convert the interception and transmission of the laser in the canopy into attenuation in the uniformly randomly distributed canopy, and established an exponential mapping model of the leaf area index and the canopy gap rate. However, this model does not take into account the aggregation of leaves, and the inverted leaf area index is the effective leaf area index. Many scholars have modified the model to address the neglected aggregation effect, such as the finite length averaging method and the gap size distribution method.
[0005] Hu et al. further proposed the need to correct the propagation path length and proposed the path length distribution method. By calculating the propagation path length of the laser inside the canopy, the influence of the aggregation effect was further corrected, thus improving the inversion accuracy. The method based on the Beer-Lamber law is simple to invert, but the process of correcting the true leaf area index is complicated, and the inversion accuracy is seriously affected by the parameters.
[0006] To simplify the inversion process, Hosoi et al. proposed a voxel-based canopy analysis method, which divides the space into a three-dimensional space composed of cubic voxels. The real leaf area index is calculated by counting the frequency of the interaction between the laser point cloud and the voxels. This method is simple and can effectively invert the real leaf area index. However, the selection of voxel size in this method is crucial. The selection of voxel size directly affects the inversion accuracy of the real leaf area index. Although many scholars have proposed different methods for the selection of voxel size, the selection standard of voxel size is not unified.
[0007] You et al. proposed a method to directly calculate the leaf area index using the Alpha-shape method. The Alpha-shape envelope principle was used to envelop the leaf point cloud, and the size of the triangle patch of the envelope surface was counted to calculate the leaf area index. However, the choice of the Alpha radius in this method also has an important influence on the inversion result of the leaf area index and is highly sensitive. Summary of the invention
[0008] In view of the above problems or shortcomings, and to solve the problems that the existing inversion process of the true leaf area index is complicated and the inversion accuracy is sensitive to parameter selection, the present invention provides an inversion method for the true leaf area index of broad-leaved trees. Based on the geometric feature parameters of the point cloud, the leaf point cloud is coarsely clustered and then finely segmented to obtain leaf blocks that can accurately represent the leaf area, and the area of the leaf block is calculated and the true leaf area index of a single tree is obtained by statistics.
[0009] A method for inverting the true leaf area index of broad-leaved trees, the specific steps are as follows:
[0010] Step 1. Data preprocessing:
[0011] The original point cloud will have outlier noise points, so statistical filtering is used to filter the original point cloud. Due to the huge number of ground-based lidar point clouds, random sampling is used to thin the point cloud to improve computational efficiency.
[0012] The statistical filtering method requires two parameters: the number of nearest neighbors and the standard deviation multiple. First, select the nearest neighbor points of each point according to the number of nearest neighbors, and then calculate the average distance and standard deviation of each point to the nearest neighbor point. Assume that the average distance satisfies the Gaussian distribution, and regard the points whose average distance to the neighbor points is not within the range of the standard deviation multiple as noise points, and retain the points that meet the standard deviation multiple.
[0013] Step 2. Leaf point extraction:
[0014] The leaf point cloud is separated from the branch point cloud by using the branch separation operation. Since the leaves in the canopy are intertwined, the aggregated leaf point cloud will affect the accuracy of the point normal vector calculation in the subsequent calculation of the normal vector, so it is necessary to use singular value filtering on the leaf point cloud; singular value filtering will remove part of the point cloud in the intertwined area of the leaves, and the filtered leaf point cloud can accurately calculate the point normal vector.
[0015] Singular value filtering: Singular value decomposition centers each point and its neighboring points and then performs plane fitting. Singular value decomposition can obtain a set of eigenvalues and their corresponding eigenvectors, where the minimum eigenvalue represents the distance of the point from the fitting plane, and the eigenvector corresponding to the minimum eigenvalue represents the direction of the fitting plane. The eigenvalues are normalized, that is, the minimum eigenvalue is divided by the sum of the three eigenvalues. Points whose normalized values are less than a certain threshold are filtered out, and the normal vector of the filtered point is recalculated as the final normal vector of the point.
[0016] Step 3. Leaf cluster clustering:
[0017] The DBSCAN algorithm is used to perform coarse clustering on the leaf point cloud obtained in step 2 to obtain a coarse clustering cluster of multiple leaf mixtures.
[0018] Then, each coarsely clustered mass is finely segmented using a fine segmentation method based on normal vectors to obtain finely clustered leaf dough masses.
[0019] Step 4. Calculate the geometric characteristics of the leaf dough block: The geometric characteristics of the leaf dough block include the horizontal projection area S of the leaf dough block projection , the dilation distance between points and the normal vector of the point.
[0020] Step 5. Calculation of true leaf area index:
[0021] Since the real leaf is not a smooth plane, the normal vector of a single point cannot represent the normal vector direction of the leaf. It is necessary to calculate the normal vector direction of the leaf surface based on the influence area of the point as a weight. The square of the expansion distance of the point represents the area occupied by the point as the weight, and the normal vector in the leaf surface block is weighted and calculated. The weighted normal vector is used as the normal vector of the leaf surface block.
[0022]
[0023] in is the internal normal vector of the leaf dough block, is the horizontal angle of the normal vector, θ is the zenith angle of the normal vector, r i is the influence radius of the points inside the blade dough block, is the leaf surface block normal vector modified by the influence radius of each point, is the horizontal angle of the leaf dough block normal vector, is the zenith angle of the leaf dough block normal vector.
[0024] The horizontal projection area is corrected using the zenith angle of the weighted normal vector to obtain the final leaf dough block area S leaf .
[0025] Sum the area of all leaf patches and divide by the tree's footprint S Ground , calculate the true leaf area index LAI of the tree t .
[0026]
[0027] Furthermore, a fine segmentation method based on normal vector is used to finely segment each coarse cluster to obtain finely clustered leaf dough blocks. The specific process is as follows:
[0028] (1) Starting point selection: Randomly select a point in the coarse cluster that has not been judged as the starting point.
[0029] (2) Neighboring point search: To improve the efficiency of searching for neighboring points, a K-Dtree is constructed in the leaf point cloud obtained in step 2, and a fixed search radius is used to search all points in the neighborhood of the starting point as the neighboring points of the starting point. The selection criterion for the fixed search radius is 0.1-1 times the average width of the leaf.
[0030] (3) Similar point judgment: judge whether the normal vector difference between the adjacent point and the starting point is less than a given threshold. The threshold is selected in the range of 0.1-1. If the normal vector difference between the point in the neighborhood and the starting point is less than the threshold, the adjacent point is defined as a similar point, and the similar point and the starting point are classified into one category as a fine cluster. The similar point is used as the new starting point, and the neighborhood normal vector similarity judgment is performed again. Repeat this process until there are no points with similar normal vectors in the neighborhood.
[0031] Points whose normal vectors do not differ by less than the threshold value (i.e., the normal vector differences of all points in the neighborhood are greater than or equal to the threshold value) within the neighborhood are considered invalid points. At the same time, all points whose number of points in the fine clusters is less than a fixed number are also considered invalid points. Invalid points will not participate in subsequent calculations and will be deleted. The fixed number is selected as 0.1%-5% of the average number of points on a single leaf.
[0032] (4) Extraction of fine clusters: Repeat the operations (1) to (3) until all points in the coarse clusters have been judged; at this point, leaf dough blocks that meet the calculation requirements of step 4 are extracted.
[0033] Furthermore, the method for obtaining the average number of points of a single leaf is: in the leaf point cloud processed by step 2, a plurality of representative single leaves are randomly selected, the number of points of the single leaves is counted and calculated, and the average number of points of the single leaf is obtained.
[0034] Furthermore, the horizontal projection area of the leaf dough block is calculated by projecting the three-dimensional points of the leaf dough block onto the xoy two-dimensional plane, and using the Graham convex hull algorithm to extract the two-dimensional boundary of the projected two-dimensional plane points. The area enclosed by the two-dimensional boundary is calculated as the horizontal projection area S of the leaf dough block. projection .
[0035] Furthermore, the expansion distance between points is calculated as follows: To calculate the influence range of each point, a gradually expanding virtual sphere is established with each point as the center. The initial radius r of all virtual spheres is i When the radius of each virtual sphere is 0, it starts to expand and grow at the same rate. When a virtual sphere collides with another virtual sphere (tangent), it stops growing until all virtual spheres stop growing. The final radius of each sphere is used as the expansion distance of the corresponding point (where the sphere is located).
[0036] Furthermore, the normal vector of the point is calculated: Since the point cloud obtained in step 2 is filtered by SVD to remove the interlaced areas, the normal vector of the point within the cluster can well represent the direction of the point. A fixed radius λ is used as the undetermined parameter, and all points within the fixed radius λ of the point to be estimated are used as fitting plane points with the point to be estimated as the center. The direction of the estimated fitting plane is used as the normal vector direction of the point to be estimated. The selection criterion for the fixed search radius is 0.1-1 times the average width of the leaf. The calculated normal vector may be pointing downward, while the direction of the leaf surface must be upward, so we need to flip the downward normal vector so that it points to the top of the leaf.
[0037]
[0038] Where N is the point normal vector, N Z is the z-component of the point normal vector, is the point normal vector after updating the direction.
[0039] The present invention first performs denoising and thinning processing on the point cloud to reduce the point cloud density of the ground-based lidar point cloud and shorten the operation time; then performs a branch-leaf separation operation and performs singular value filtering on the acquired leaf point cloud to obtain a leaf point cloud with accurate point normal vector calculation; then adopts the DBSCAN coarse clustering method to obtain the coarse clustering clusters of the leaves, and uses the fine segmentation method based on the normal vector to perform fine segmentation on the coarse clustering clusters to obtain leaf point cloud clusters of appropriate size; then calculates the geometric features of the leaf point cloud clusters, and calculates the leaf area of the leaf point cloud through the geometric features; finally, the leaf areas of all clusters are counted to obtain the true leaf area index.
[0040] In summary, compared with the existing true leaf area index calculation method, the leaf area index calculation process of the present invention is simple, and is insensitive to the selection of parameters. The algorithm has high robustness and is simple to apply in different scenarios. It can effectively help forestry resource surveys, forest change supervision and many other fields. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 It is a schematic diagram of the process of the present invention;
[0042] Figure 2 It is a blade point cloud clustering diagram of the embodiment;
[0043] Figure 3 Figure 1 is a graph of cluster leaf area calculation for an example. DETAILED DESCRIPTION
[0044] The following example further illustrates the present invention in detail by simulating data of 10 single trees and combining with the accompanying drawings: the development environment is PyCharm, the programming language is Python 3.7, and the software packages used mainly include open3d, numpy, and sklearn.
[0045] Original point cloud data acquisition: This embodiment acquires simulated broadleaf tree point cloud data based on the LESS platform. LESS is a three-dimensional radiation transmission model based on ray tracing that can simulate remote sensing data and images of large-scale and real 3D scenes. You only need to input a predetermined tree model, plan the distribution of the tree model, select different lidar platforms and corresponding parameters in LESS, and you can acquire high-precision three-dimensional point cloud data.
[0046] Ten citrus tree models (CISI1-CISI10) in the HET14_WCO_UND scene of RAMI VI were selected as pre-selected tree models. The tree heights of the models ranged from 2.83m to 4.11m, and the single-sided leaf area ranged from 11.4913m 2 ~42.8818m 2 There are differences in leaf characteristics (mean leaf length 8.8 cm ± 1.9 cm, leaf curl 20%). These models have some variability and can represent broadleaf trees with different growth conditions.
[0047] The tree model was placed in the center of the scene, and a ground-based lidar was used to perform multi-station scanning of the tree model. The angular spacing of the multi-station scan was 90°, and each scan was 10m away from the center of the tree model. The zenith angle resolution and horizontal angle resolution of the scan were 0.01°. The acquired multi-station scanning data were spliced to obtain the initial complete broadleaf tree point cloud data.
[0048] Table 1 Broadleaf tree model parameter details:
[0049]
[0050] Step 1: Data preprocessing: Statistical filtering is used to filter and denoise the original point cloud. In the statistical filtering, twice the standard deviation is selected as the filtering standard, and 40% random sampling is selected to thin the point cloud.
[0051] Step 2: Extracting leaf points: This embodiment uses a branch-leaf separation operation based on graph theory to separate the leaf point cloud from the branch point cloud; and uses singular value filtering on the separated leaf point cloud. The resulting point cloud is as follows: Figure 2 (a) shown.
[0052] The branch-leaf separation method based on graph theory is to construct a graph on a single tree point cloud, query the shortest path from all points to the root node on the graph; segment the single tree point cloud into homogeneous and feature clusters at multiple scales according to the path length; identify regular tree clusters with cylindrical or linear features in the cluster clusters, and regard these tree clusters as initial tree points; use the initial tree points as seed points, and extract tree points in irregular clusters through the region growing algorithm; the initial tree points and the tree points extracted by region growing constitute the final tree points, and the remaining point clouds are regarded as leaf points.
[0053] Step 3, leaf cluster clustering: In this embodiment, 1 cm is selected as the fixed search radius, 0.5 is selected as the difference threshold of the normal vector, and the number of points for cluster determination is 10.
[0054] First, implement rough clustering of leaf point cloud, such as Figure 2 (b) shows the fine segmentation of the leaf point cloud. Figure 2 (c) as shown.
[0055] Step 4: Calculation of geometric features of clusters: Figure 3 As shown, in this embodiment, 1 cm is selected as the fixed search radius.
[0056] Step 5: Calculate the real leaf area index.
[0057] Accuracy evaluation: Calculate the RMSE and accuracy of the inverted LAIt and the true LAIt as the evaluation criteria. The calculation process is as follows:
[0058]
[0059] Among them LAI i LAIt stands for inversion, L i represents the real LAIt, and n is the number of samples. The broadleaf tree LAIt inversion of the ground-based laser radar data implemented in this embodiment has an average accuracy test result of 72.17%.
[0060] Table 2 Inversion accuracy table:
[0061]
[0062] It can be seen from the above embodiments that the method provided by the present invention effectively segments the leaf point cloud, calculates the geometric characteristic parameters of the segmented leaf clusters, and realizes the calculation of the leaf area index by geometric parameters that are insensitive to parameters. The present invention can solve the problem that the traditional leaf area calculation method based on Beer's law is sensitive to parameters, selects the parameters of the leaf scale size and does not need to perform too many parameter tests, so as to realize the accurate calculation of the leaf area index.
Claims
1. A method for inverting the true leaf area index of broad-leaved trees, characterized in that: The specific steps are as follows: Step 1. Data preprocessing: Statistical filtering is used to filter the original point cloud to remove outlier noise points; random sampling is then used to thin the point cloud; Step 2. Leaf point extraction: The leaf point cloud is separated from the branch point cloud by using the branch-leaf separation operation; then the leaf point cloud is filtered by singular value filtering to remove part of the point cloud in the interlaced area of the leaves; Step 3. Leaf cluster clustering: The DBSCAN algorithm is used to perform coarse clustering on the leaf point cloud obtained in step 2 to obtain a coarse clustering cluster of multiple leaf mixtures; Then, each coarse cluster is finely segmented using a fine segmentation method based on normal vectors to obtain finely clustered leaf dough blocks; Step 4. Calculate the geometric characteristics of the dough block: The geometric characteristics of the leaf dough block include the horizontal projection area S of the leaf dough block projection , the expansion distance between points and the normal vector of the point; Step 5. Calculation of true leaf area index: The influence area of the point is used as the weight to calculate the direction of the leaf surface normal vector. The square of the expansion distance of the point represents the area occupied by the point as the weight, and the normal vector in the leaf surface block is weightedly calculated. The obtained weighted normal vector is used as the normal vector of the leaf surface block. in is the normal vector of the inner point of the leaf dough block, is the horizontal angle of the normal vector, θ is the zenith angle of the normal vector, r i is the influence radius of the points inside the blade dough block, is the leaf surface block normal vector modified by the influence radius of each point, is the horizontal angle of the normal vector of the leaf dough block, is the zenith angle of the blade dough block normal vector; The horizontal projection area is corrected using the zenith angle of the weighted normal vector to obtain the final leaf dough block area S leaf ; Sum the area of all leaf patches and divide by the area of the tree S Ground , calculate the true leaf area index LAI of the tree t ; 2. The inversion method of the real leaf area index of broad-leaved trees as claimed in claim 1, characterized in that: The fine segmentation method in step 3 obtains finely clustered leaf dough blocks, and the specific process is as follows: (1) Starting point selection: Randomly select a point in the coarse cluster that has not been judged as the starting point; (2) Neighboring point search: A K-Dtree is constructed in the leaf point cloud obtained in step 2, and all points in the neighborhood of the starting point are searched with a fixed search radius as the neighboring points of the starting point; the selection standard of the fixed search radius is 0.1-1 times the average width of the leaf; (3) Similar point judgment: judge whether the normal vector difference between the adjacent points and the starting point is less than a given threshold. The threshold selection range is 0.1-1. If the normal vector difference between the points in the neighborhood and the starting point is less than the threshold, the adjacent point is defined as a similar point, and the similar point and the starting point are classified into one category as a fine cluster. Take the similar point as the new starting point and perform the neighborhood normal vector similarity judgment again; repeat this process until there are no points with similar normal vectors in the neighborhood; Points whose normal vectors do not differ by less than a threshold in the neighborhood are considered invalid points. All points whose number of points in the fine clusters is less than a fixed number are also considered invalid points. Invalid points will not participate in subsequent calculations and will be deleted. The fixed number is selected as 0.1%-5% of the average number of points on a single leaf. (4) Extraction of fine clusters: Repeat steps (1) to (3) until all points in the coarse clusters have been judged; at this point, leaf dough blocks that meet the calculation requirements of step 4 are extracted.
3. The inversion method of the real leaf area index of broad-leaved trees as claimed in claim 2, characterized in that: The method for obtaining the average number of points of a single leaf in step (3) is as follows: in the leaf point cloud processed in step 2, a plurality of representative single leaves are randomly selected, the number of points of the single leaves is counted and calculated, and the average number of points of the single leaves is obtained.
4. The inversion method of the real leaf area index of broad-leaved trees as claimed in claim 1, characterized in that: The horizontal projection area calculation of the leaf dough block: project the three-dimensional points of the leaf dough block onto the xoy two-dimensional plane, extract the two-dimensional boundary of the projected two-dimensional plane points using the Graham convex hull algorithm, and calculate the area surrounded by the two-dimensional boundary as the horizontal projection area S of the leaf dough block. projection .
5. The inversion method of the real leaf area index of broad-leaved trees as claimed in claim 1, characterized in that: The expansion distance between points is calculated as follows: a gradually expanding virtual sphere is established with each point as the center, and the initial radius of all virtual spheres is r i When the value of θ is 0, the radius of each virtual sphere starts to expand at the same rate and stops growing when one virtual sphere becomes tangent to another sphere, until all virtual spheres stop growing. The final radius of each virtual sphere is used as the expansion distance of the corresponding point.
6. The inversion method of the real leaf area index of broad-leaved trees as claimed in claim 1, characterized in that: The normal vector calculation of the point: using a fixed radius λ as an undetermined parameter, taking the point to be estimated as the center, all points within the fixed radius λ of the point to be estimated are used as fitting plane points, and the direction of the estimated fitting plane is used as the direction of the normal vector of the point to be estimated; The selection criteria of the fixed search radius is 0.1-1 times of the average width of the blade; the normal vector calculated to be in the downward direction is flipped to point to the top of the blade; Where N is the point normal vector, N Z is the z-component of the point normal vector, is the point normal vector after updating the direction.
7. The inversion method of the real leaf area index of broad-leaved trees as claimed in claim 1, characterized in that: In step 2, a branch-leaf separation operation based on graph theory is used to separate the leaf point cloud from the branch point cloud.
Citation Information
Patent Citations
Tree leaf area estimation method based on ground laser point cloud
CN115994939A
Vegetation aggregation index estimation method based on Poisson distribution and point cloud clustering
CN116503727A
Cited By
Tree point cloud branch and leaf separation method and system based on multi-dimensional super voxel random forest, terminal equipment and medium
CN120524353A
A method, system, terminal equipment, and media for tree point cloud branch and leaf separation based on multidimensional hypervoxel random forest.
CN120524353B
