Double-source point cloud laser collaborative inversion method for white tea leaf area concentration

Abstract

The invention relates to the technical field of cell biology, in particular to a double-source point cloud laser collaborative inversion method for white tea leaf area concentration. According to the method, the ground point cloud data and the airborne point cloud data are obtained, separation and extraction are performed through the random forest algorithm, the volume element model is established, coordinate registration is performed, and then the area density of the white tea leaves is obtained through cooperative inversion of the canopy analysis method. A dead zone generated by airborne is located at the middle lower part of a tea tree canopy, a dead zone generated by a foundation is located at the middle upper part of the tea tree canopy, and high inversion precision can be obtained by reconstructing and complementing the two curves.

Description

Technical Field

[0001] This invention relates to the field of leaf area density measurement technology, and in particular to a dual-source point cloud laser collaborative inversion method for white tea leaf area concentration. Background Technology

[0002] White tea, belonging to the Theaceae family, is an evergreen shrub or small tree and an important economic crop in my country with high economic value and benefits. Leaf area density (LAD) reflects the total leaf area per unit area and height of a tea plant, describing the growth of leaves at various heights within the canopy. It is a key factor in studying the top-down interaction between tea leaves and the entire tea garden ecosystem. Traditional methods for measuring LAD primarily rely on manual stratification, which, while highly accurate, is extremely destructive to crops, time-consuming, labor-intensive, and economically inefficient. In recent years, lidar technology has rapidly developed. Ground-based and airborne lidar, as active remote sensing technologies capable of acquiring true 3D information of targets around the clock and in all weather conditions, possess strong penetration capabilities for agricultural and forestry crops, enabling rapid and accurate acquisition of 3D structural information. This technology has been quickly applied to extracting leaf area density from agricultural and forestry canopies.

[0003] Existing technologies include: using voxel models to simulate laser beam interception through the canopy of trees, proposing a canopy analysis method (VCP) based on voxel models, and plotting leaf area density curves using the gap ratio theory based on the laser interception rate at each height level; using ground-based LiDAR data to invert leaf area density models of individual tree canopies and exploring the intrinsic relationship between voxel size and leaf area density of trees; using ground-based LiDAR data to invert the leaf area density (LAD) of individual broad-leaved trees, using voxel analysis to monitor real-time biological parameters of mango, macadamia, and avocado trees, and concluding that plant LiDAR data can dynamically reflect plant growth parameters and can serve as a basis for evaluating plant health; using the Lambert azimuth area projection method to calculate porosity and obtain parameters such as leaf area index with high accuracy; using backpack LiDAR to invert the leaf area density of apple trees, providing a reference for portable LiDAR measurement; and using airborne LiDAR... Based on canopy analysis, canopy-based LiDAR and ground-based LiDAR were used to invert the leaf area density (LAD) curves of individual trees. Airborne LiDAR point cloud data was used to establish the upper and middle LAD distribution model of individual trees, while ground-based LiDAR data was used to establish the lower and middle LAD distribution model, improving the accuracy of leaf area density inversion. Hyperspectral information was used to classify tree species, and a fusion algorithm was used to obtain the LAD of individual trees from airborne point cloud data. Based on VCP, the LAD of magnolia and birch forests was inverted using both airborne and ground-based point cloud data. This demonstrates that LiDAR data measurement methods based on volumetric models have been widely used for leaf area density inversion in agricultural and forestry crops. However, due to the relatively small height and width of shrub crops compared to trees, high leaf overlap, and small canopy width, current mature LiDAR inversion methods focus on arborescent plants. Further development and improvement are needed to rapidly and accurately estimate and validate the leaf area density of large-area shrub crops.

[0004] Therefore, taking Anji white tea as the research object and ground-based LiDAR and airborne LiDAR data as the research basis, this study investigates the construction mechanism of a voxel model of white tea trees based on point cloud segmentation and canopy analysis. Using the generated voxel model, the study further investigates the inversion method of leaf area density curve of white tea trees, and delves into the intrinsic mechanism of voxel size and contact frequency. A collaborative and complementary inversion method based on ground-based and airborne dual-source data is proposed to reconstruct a leaf area density inversion model suitable for white tea trees, providing a research foundation for the accurate estimation of LAD of large-scale shrub-type agroforestry crops in the future. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a dual-source point cloud laser collaborative inversion method for white tea leaf area concentration, which addresses the shortcomings of the prior art. The method acquires ground-based point cloud data and airborne point cloud data, and after separation and extraction, establishment of a volumetric model and coordinate registration, collaborative inversion is performed to obtain the leaf area density of white tea leaves.

[0006] To address the aforementioned technical problems, this invention provides a dual-source point cloud laser-coordinated inversion method for the area concentration of white tea leaves, comprising the following steps:

[0007] S1. Obtain ground point cloud data and airborne point cloud data of shrub-type white tea trees;

[0008] S2. Separate and extract the ground-based point cloud data and the airborne point cloud data obtained in S1;

[0009] S3. Based on the separated and extracted ground point cloud data obtained in S2 and the airborne point cloud data, the point cloud of the shrub-type white tea tree is volumetricated and a volumetric model is established.

[0010] S4. Perform coordinate registration between the ground-based point cloud data and the airborne point cloud data obtained in S1;

[0011] S5. The area density of white tea leaves is calculated by synergistically inverting ground-based point cloud data and airborne point cloud data using the canopy analysis method.

[0012] According to the present invention, a dual-source point cloud laser collaborative inversion method for white tea leaf area concentration is provided. The method for obtaining the ground-based point cloud data in S1 is as follows: according to the distance of the tea plantation, four 10m×10m tea tree vegetation plots are selected as monitoring plots to minimize the mutual shading of branches and leaves, so as to achieve all-round scanning of the ground-based laser radar and obtain the ground-based point cloud data. The ground-based point cloud data adopts Cartesian coordinates.

[0013] The method for acquiring airborne point cloud data is as follows: each sample plot is divided into four 5m×5m small quadrats, and five monitoring points are set in each quadrat. The five monitoring points are respectively set at the four corner points and the center point of the quadrat, and an airborne lidar is set at each monitoring point to obtain airborne point cloud data. The airborne point cloud data adopts geodetic projection coordinates.

[0014] According to the present invention, a dual-source point cloud laser collaborative inversion method for white tea leaf area concentration is provided. The separation in S2 is achieved by using a quadratic polynomial method to correct and fit the height and distance of the ground-based point cloud data and the airborne point cloud data, and then establishing a classification model based on the corrected data to separate the branches and leaves.

[0015] The calibration fitting process involves selecting 50 white tea trees of different shapes as the test subjects, importing their ground-based point cloud data and airborne point cloud data into a commonly used point cloud segmentation algorithm, and then outputting point cloud information that retains the effective leaf surface.

[0016] According to the present invention, a dual-source point cloud laser collaborative inversion method for white tea leaf area concentration is provided. The commonly used point cloud segmentation algorithms include: random forest classification method, normal difference method, maximum likelihood method and Euclidean distance segmentation method.

[0017] By comparing the effective leaf point cloud information, a better branch and leaf separation effect can be obtained through the random forest method.

[0018] According to the present invention, a dual-source point cloud laser collaborative inversion method for white tea leaf area concentration is provided. The method of volumetric shrub-type white tea tree point cloud described in S3 is as follows: the object to be measured is divided into a three-dimensional volume network according to the size of the volume element, and the point cloud data of the object to be measured is projected onto the volume network. Then, it is determined whether there is a laser spot in each volume element. If there is, it means that there are tea leaves in the volume element, otherwise there are no laser spots.

[0019] According to the present invention, a dual-source point cloud laser collaborative inversion method for white tea leaf area concentration is provided, wherein the optimal voxel value is determined by the maximum scanning distance, divergence angle, and exit diameter of the laser scanner.

[0020] According to the present invention, a dual-source point cloud laser collaborative inversion method for white tea leaf area concentration is provided. After the optimal voxel value is determined, the minimum value of the point cloud data in each direction is taken as the starting point and the maximum value as the ending point. The optimal voxel value is used as the interval step to obtain the voxelized coordinates of each point and establish a voxel model of white tea tree.

[0021] By projecting the point cloud data onto the xy plane and using the two-dimensional convex hull method, the canopy boundary of each horizontal layer can be quickly determined.

[0022] According to the present invention, a dual-source point cloud laser collaborative inversion method for white tea leaf area concentration is provided, wherein the coordinate registration in step S4 includes the following steps:

[0023] S401. Select point set m from the acquired ground point cloud data of the object under test, and select point set n from the airborne point cloud data;

[0024] S402. Calculate the rigid body transformation that minimizes the objective function for the corresponding point in S401, and then use the four-parameter method to find the rotation matrix R and translation vector T at this time.

[0025] S403. Substitute the R and T obtained in S402 into the point set n selected in S401 to obtain a new point set n′=Rn+T;

[0026] S404. Calculate whether the objective function of the point set n′ obtained in S403 and the point set m obtained in S401 meets the threshold requirement. If it does, stop the iteration. If it does not meet the objective function threshold requirement, return the point set n′ obtained in S403 to step S402 for a new round of iteration until the minimum sum of squared distances of the new point set obtained in S403 meets the threshold requirement and the convergence requirement is met.

[0027] According to the present invention, a dual-source point cloud laser collaborative inversion method for white tea leaf area concentration is provided. In step S5, the canopy analysis method is used to observe and analyze the inversion curves obtained by the two scanning methods, determine the blind areas of the two scanning methods, and finally reconstruct and numerically fuse the curves to obtain high-precision inversion results.

[0028] According to the present invention, a dual-source point cloud laser collaborative inversion method for white tea leaf area concentration is provided. The ground-based point cloud data and the airborne point cloud data mentioned in S1 have different blind zone locations due to different scanning methods. The blind zone generated by the airborne method is located in the middle and lower part of the tea tree canopy, while the blind zone generated by the ground-based method is located in the middle and upper part of the tea tree canopy. By reconstructing and complementing the two curves, higher inversion accuracy can be obtained.

[0029] Compared with the prior art, the present invention has the following advantages:

[0030] This invention provides a dual-source point cloud laser collaborative inversion method for white tea leaf area concentration. After applying multinomial correction to the point cloud data of white tea trees, the random forest algorithm is used to achieve high accuracy in branch and leaf separation. The precision, recall, mean precision, and overall accuracy are 0.957, 0.964, 0.960, and 0.976, respectively, indicating good separation effect.

[0031] The laser contact frequency of shrub-type white tea trees is highly correlated with the size of the set voxel. Setting the voxel too large will result in an over-calculated contact frequency, and the estimated LAI value will be greater than the measured LAI value. Setting the voxel too small will result in an under-calculated contact frequency, and the estimated LAI value will be smaller than the measured LAI value. The optimal voxel value will vary under different environmental parameters.

[0032] Airborne and ground-based lidars produce blind zones in different locations due to their different scanning methods. The blind zone produced by airborne lidar is located in the lower middle part of the tea tree canopy, while the blind zone produced by ground-based lidar is located in the upper middle part of the tea tree canopy. Reconstructing and complementing the two curves can achieve higher inversion accuracy. Attached Figure Description

[0033] Figure 1 This is a schematic diagram showing the distribution of ground-based lidar scanning stations in each sample plot of the present invention;

[0034] Figure 2 This is a schematic diagram of the volume element model in this invention;

[0035] Figure 3 This is a schematic diagram illustrating the determination of the canopy boundary using the two-dimensional convex hull method in this invention;

[0036] Figure 4 This is a diagram showing the distribution of reflection intensity of branches and leaves before correction according to the present invention;

[0037] Figure 5 This is a diagram showing the distribution of reflection intensity of branches and leaves after correction according to the present invention;

[0038] Figure 6 This is a graph showing the relationship between volume element size and contact frequency in this invention;

[0039] Figure 7 These are the LAD inversion curves for the foundation and airborne components of sample plot 1 in this invention;

[0040] Figure 8 These are the LAD inversion curves for the foundation and airborne samples of plots 1-4 in this invention;

[0041] Figure 9 These are the LAD inversion curves after the fusion of sample plots 1 to 4 in this invention;

[0042] Figure 10 This is a schematic diagram showing the estimated and measured values ​​of LAI in plots 1 to 4 of this invention. Detailed Implementation

[0043] Example 1

[0044] The experimental area selected in this embodiment is located in a tea plantation in Kuntong Township, Anji County, Zhejiang Province. This area is situated at the eastern end of the Tianmu Mountains within Anji County, with an average temperature of 15.43℃ and an average annual rainfall of 1469 mm. It has a subtropical monsoon climate and is located at 119°47'E, 30°44'N. The total planting area for white tea is 989.73 hm². 2 It is a key industrial base for white tea in Anji County.

[0045] The experiment utilized the SZT-R250 airborne lidar 3D mobile measurement system and the UA-0500 ground-based laser scanner manufactured by South Surveying and Mapping. The ground-based laser scanner used was the UA-0500 laser scanner, which features long-range measurement, high accuracy, and two built-in panoramic digital cameras, enabling it to quickly and accurately save the 3D point cloud and color texture information of the scanned object.

[0046] The 3D laser mobile measurement system is equipped with the DJI M600PRO drone flight platform, with the entire unit weighing only 2.2Kg and a maximum takeoff payload of 13.3Kg. The platform integrates an inertial navigation system and a GNSS satellite positioning system, ensuring stable flight and precise control of the measurement system. The airborne lidar measurement system has a maximum scanning rate of 100 lines / second, a maximum dynamic detection distance of 150m, and a laser emission frequency of 100KHz, ensuring the continuity of data acquisition during flight.

[0047] This embodiment provides a method for acquiring and preprocessing dual-source point cloud data of area concentration in shrub-type white tea leaves, specifically including the following steps:

[0048] Step 1: Based on the distance between tea plantations, select four 10m×10m tea tree vegetation plots as monitoring plots, minimizing mutual shading of branches and leaves, to achieve all-round scanning of the ground-based lidar and obtain complete data on the white tea tree canopy.

[0049] Step 2: Divide each sample plot into four 5m × 5m quadrats. Set up 5 monitoring points in each quadrat, placing them at the four corners and the center of the quadrat. Install an airborne lidar sensor at each monitoring point. Figure 1 As shown, in order to unify the point cloud coordinates of the 5 monitoring points, multiple target positions are set. P1 is the internal observation station that performs 360°×300° panoramic sampling of the canopy inside the sampling target. P2, P3, P4 and P5 sample the sample plot from the four corner points respectively to ensure that the sampling angles fully cover the test sample plot. The point cloud data of the 5 stations will be calibrated and stitched together later.

[0050] Step 3: The point cloud simplification method based on Voxel Grid filter can effectively reduce redundant point clouds and improve the accuracy of feature information extraction; execute the memory masking command Fence of Cyclone editing software to remove isolated abnormal point cloud data that are visible to the naked eye, and at the same time, combine manual identification to remove noise points that are not easy to identify.

[0051] Step 4: For test sites with gentle terrain and small slopes, the CSF (Cloth Simulation) algorithm is suitable for filtering the ground point cloud: The collected point cloud data is flipped to simulate cloth covering the target to be tested. The relationship between the cloth's falling elevation coordinates and the ground coordinates is observed to determine the immovable coordinates. When the system tends to stabilize, the height difference between the cloth simulated point cloud and the actual point cloud is compared to filter the ground point cloud information.

[0052] Example 2

[0053] This embodiment provides a dual-source point cloud laser collaborative inversion method for the area concentration of white tea leaves.

[0054] S1. Obtain ground point cloud data and airborne point cloud data of shrub-type white tea trees;

[0055] Point cloud data acquisition is affected by objective environment, instruments and other factors. The original point cloud data contains a lot of point cloud data that is not related to the scanning target, such as isolated point cloud data, point cloud data outside the observation area, noise data, etc. Therefore, the point cloud data should be preprocessed according to the method provided in Example 1 before it is used for calculation.

[0056] S2. Separate and extract the ground-based point cloud data and the airborne point cloud data obtained in S1, as follows:

[0057] S201. Since shrub-type crops are relatively short in height and width compared to trees, have high leaf overlap, and small crown width, traditional manual supervision methods and in-domain normal vector difference methods are not suitable for white tea. Polynomials are used to correct and fit the height and distance of point cloud data, and the corrected data is used to establish a classification model to separate branches and leaves of white tea trees, resulting in higher accuracy.

[0058] S202. Select 50 white tea trees of different forms as the test objects, obtain the point cloud data of the test objects, and then use the RGB information of the color system and the laser reflection intensity of the test objects as the classification basis. Import the sampling airborne point cloud information and the ground-based point cloud information into the commonly used point cloud segmentation algorithm, and output the effective leaf surface point cloud information. Calculate the output results to obtain the precision, recall, overall accuracy, and average precision (AP) of the test objects, which are used to evaluate the performance of the four classifiers.

[0059] accuracy In the formula, T P F represents the number of point clouds correctly classified by the system. P The number of point clouds that were incorrectly classified as part of this organization by the system;

[0060] Recall rate In the formula, F N This represents the number of point clouds from this organization that were misclassified as point clouds from other organizations.

[0061] Overall accuracy In the formula, n is the total number of point clouds in the sample;

[0062] Average accuracy

[0063] Commonly used point cloud segmentation algorithms include random forest classification, normal difference method, maximum likelihood method and Euclidean distance method;

[0064] S3. Based on the separated and extracted ground point cloud data obtained in S2 and the airborne point cloud data, the point cloud of the shrub-type white tea tree is volumetricated and a volumetric model is established.

[0065] The point cloud data of the measured object obtained in S202 is discrete and irregular. Transforming it into a regular spatial structure with voxels as the smallest unit can simplify calculations, reduce data volume, and improve inversion accuracy. The steps are as follows:

[0066] S301. Divide the object to be measured into a three-dimensional voxel network according to the size of the voxel. At the same time, project the point cloud coordinate information of the canopy of the object to be measured onto the voxel network. Determine whether there is a laser spot in each voxel. If there is a laser spot, mark the voxel as 1, indicating that there are tea leaves in the area. If there is no laser spot, mark the voxel as 0, indicating that there are no tea leaves in the area.

[0067] Among them, the selection of voxel size is a key parameter for constructing voxel model. If the voxel size is too large, there will be multiple laser spots in one voxel, which cannot accurately express the internal structure of the canopy of the object being measured, resulting in insufficient data. If the voxel size is too small, it will lead to over-segmentation of voxels, affecting the calculation of contact frequency.

[0068] The optimal voxel division is determined by factors such as the maximum scanning distance of the laser scanner, the laser divergence angle, and the diameter of the laser exit. When an object is detected in the path of the emitted laser, the distance between the two laser beams is calculated using the following two formulas:

[0069] In the formula, R is the radius, and D is the radius. i B represents the maximum scanning distance of the laser scanner. DV B is the laser divergence angle of the scanner. DM The diameter of the laser exit point;

[0070] d CD =(tanα×D) i )-2R, where d CD The distance between two adjacent laser beams is α, and the angular resolution is α.

[0071] S302. Further determine the point cloud boundary of the tea tree canopy: Starting from the minimum value of the point cloud data of the measured object in each direction and ending at the maximum value, using the size of the voxel as the interval step, obtain the voxelized coordinates of each point as (i, j, k), and establish a voxel model from the white tea point cloud information.

[0072] In coordinates

[0073] In coordinates

[0074] In coordinates In the formula, (i, j, k) are the volume element coordinates of the corresponding (x, y, z) respectively, and (x min y min , z min ( ) represents the starting coordinates of the voxel, and (△x, △y, △z) represents the determined size of the voxel.

[0075] In actual measurements, due to the irregularity of the white tea tree canopy, the obtained point cloud data contains a lot of invalid voxels. In order to accurately calculate the contact frequency and eliminate the errors and interference of invalid point clouds, this embodiment projects the white tea point cloud data onto the xy plane and uses the two-dimensional convex hull method Graham Scan to determine the canopy boundary of each horizontal layer.

[0076] S4. Perform coordinate registration between the ground-based point cloud data and the airborne point cloud data obtained in S1;

[0077] Ground-based LiDAR acquires point cloud data of the measured object using Cartesian coordinates, while airborne LiDAR acquires point cloud data of the measured object using geodetic projection coordinates.

[0078] To achieve the collaborative inversion of ground-based and airborne point cloud data, the point cloud data of the two coordinate systems need to be accurately registered. The four-parameter method is used to set several reference points for coarse registration of the coordinates. On this basis, the improved iterative nearest point ICP algorithm is used for further fine registration.

[0079] Let the Cartesian coordinates of the reference point obtained from the foundation be (x, y), and the geodetic projection coordinates of the reference point obtained from the airborne system be (x', y'). According to the plane coordinate transformation formula, we can obtain formula (1):

[0080]

[0081] Let a = (1+m)cosα and b = (1+m)sinα. Substituting them into formula (1) and simplifying the calculation, we get... By transforming the matrix, the parameters to be solved are transformed into the same matrix, thus obtaining...

[0082]

[0083] Using the indirect adjustment method v=BX-L, Δx, Δy and rotation angle α can be obtained, and the coordinates on the z-axis can be calculated based on the average height difference between the two coordinate systems;

[0084] Meanwhile, based on the ICP algorithm, initial corresponding points are selected for registration, and then the neighboring regions of the corresponding points are selected for iterative calculation according to the objective function. When the value of the objective function is minimized, the current values ​​of the rotation matrix and translation matrix are saved as the optimal solution.

[0085] The above objective function is In the formula, R is the rotation matrix and T is the translation vector;

[0086] The process for further fine-tuning the ICP algorithm described above is as follows:

[0087] S401. Select point set m from the point cloud data of the object under test acquired by ground-based LiDAR, and select point set n from the point cloud data of the object under test acquired by airborne LiDAR.

[0088] S402. Calculate the rigid body transformation that minimizes the objective function for the corresponding point in S401, and then use the four-parameter method to find the rotation matrix R and translation vector T at this time.

[0089] S403. Substitute the R and T obtained in S402 into the point set n selected in S401 to obtain a new point set n′=Rn+T;

[0090] S404. Calculate whether the objective function of the point set n′ obtained in S403 and the point set m obtained in S401 meets the threshold requirement. If it does, stop the iteration. If it does not meet the objective function threshold requirement, return the point set n′ obtained in S403 to step S402 for a new round of iteration until the minimum sum of squared distances of the new point set obtained in S403 meets the threshold requirement and the convergence requirement is met.

[0091] S5. The area density of white tea leaves is calculated by synergistically inverting ground point cloud data and airborne point cloud data using the canopy analysis method.

[0092] First, based on the voxel model canopy analysis method, the leaf area density of the tested tea trees was calculated using both ground-based and airborne point cloud data. The calculation formula is as follows:

[0093]

[0094] Where α(θ) is a correction factor for the blade tilt angle and the laser incident tilt angle, usually taken as 1.1;

[0095] ΔH represents the horizontal layer thickness, which is the size of a single voxel in the voxel model.

[0096] It is the laser contact frequency, which represents the ratio of the number of voxels penetrated in layer k to the total number of voxels.

[0097] The total number of volume elements in layer k is n1(k) + n p (k);

[0098] The number of voxels intercepted by the laser in the tested object is n1(k);

[0099] The number of volume elements in the tested object that are penetrated by the laser is n p (k);

[0100] Ground-based LiDAR scans white tea trees from the ground up. However, because the scanning position of ground-based LiDAR is fixed and the canopy of shrub-type white tea trees is complex, it is difficult to obtain large-area complete canopy point cloud data of tea trees.

[0101] Airborne LiDAR scanning, unlike ground-based LiDAR scanning, scans white tea trees from top to bottom. The scanning characteristics of the two methods complement each other well. However, simply covering and superimposing the inversion values ​​of the two methods has a limited effect on improving inversion accuracy. Therefore, volumetric models are established for ground-based point cloud data and airborne point cloud data respectively. Based on the canopy analysis method, the inversion curves obtained by the two scanning methods are observed and analyzed to determine the blind spots of the two scanning methods. Finally, the curves are reconstructed and the values ​​are fused to obtain high-precision inversion results.

[0102] Example 3

[0103] Evaluation of the accuracy model for branch and leaf separation

[0104] The laser reflection intensity distribution of branches and leaves was obtained by multiple samplings of branches and leaves in four quadrat plots after correcting for laser incident height and distance using a quadratic polynomial. Figure 4 and Figure 5 As shown, Figure 4 This is the reflection intensity distribution map before correction. Figure 5 This is the corrected reflection intensity distribution map;

[0105] By comparison Figure 4 and Figure 5 It can be seen that before the correction, the reflection intensities of the trunk and leaves of the four quadrats overlapped to varying degrees. After correction by the correction model, the reflection intensities of the trunk and leaves were well distinguished and stabilized in a certain fixed range, with a large difference between the reflection intensities of the trunk and leaves.

[0106] The corrected data was imported into four point cloud segmentation methods for comparison. The branch and leaf separation results of the four algorithms are shown in Table 1 below:

[0107] Table 1 Evaluation Table of Branch and Leaf Separation Accuracy

[0108] Classification Algorithm P Re AP OA Random Forest 0.957 0.964 0.960 0.976 Normal difference method 0.729 0.801 0.748 0.782 Maximum likelihood method 0.757 0.695 0.736 0.729 Euclidean distance partitioning method 0.842 0.837 0.822 0.827

[0109] As shown in the table above, the random forest algorithm achieved the best separation effect. The precision (P), recall (Re), mean precision (AP), and overall accuracy (OA) of 50 white tea plants were 0.957, 0.964, 0.960, and 0.976, respectively, which were higher than the branch and leaf separation accuracy results of the other three algorithms. This indicates that the random forest algorithm can achieve better branch and leaf separation results for low-growing shrub-type white tea trees.

[0110] Example 4

[0111] The influence of volumetric size on the estimation of LAD (Lower Dimension) of white tea trees.

[0112] To determine the optimal voxel values, this embodiment varied the voxel size of the point cloud data of white tea trees in the sample plots. Ten different voxel sizes (1mm, 2mm, 4mm, 8mm, 10mm, 15mm, 20mm, 30mm, 40mm, and 50mm) were selected between 1 and 50mm. The laser contact frequencies at six height layers were calculated for each size, and the relationship between voxel size and contact frequency was analyzed. The relationship between voxel size and contact frequency is shown in the figure below. Figure 6 As shown;

[0113] according to Figure 6 It can be seen that the laser contact frequency of white tea trees is highly related to the size of the volume element. When the volume element is large, there are multiple light spots inside the volume element. According to formula (2), the number of volume elements intercepted by the laser in n1(k) increases, while the total number of volume elements remains unchanged. Therefore, the contact frequency also increases.

[0114] When the volume element is small, n1(k) will gradually approach the number of laser points, and due to the subdivision of the volume element, the number of volume elements n that are penetrated will increase. p (k) will gradually increase, resulting in an abnormally low final contact frequency;

[0115] When the volume element increases from 1 mm to 50 mm, the contact frequencies increase from 0.078, 0.124, 0.054, 0.145, 0.046 and 0.087 to 0.647, 0.812, 0.571, 0.829, 0.481 and 0.659. The contact frequencies of the six height layers of the quadrat tend to stabilize when the volume element value is set to 10 mm, which is close to the result derived from formula (2).

[0116] Therefore, the choice of voxel size is crucial for contact frequency calculation, and selecting the optimal voxel size is of great significance to the accuracy of LAD inversion.

[0117] Example 5

[0118] LAD experimental verification and accuracy evaluation.

[0119] Based on the correlation between contact frequency and volumetric size, the contact frequency tends to stabilize when the volumetric size is 10 mm. Using canopy analysis, LAD curve inversion was performed on the white tea trees in plot 1, and the results are as follows: Figure 7 As shown;

[0120] like Figure 7As shown, the dashed curve is obtained from airborne LiDAR data, and the solid curve is obtained from ground-based LiDAR data. It can be seen that the two inversion curves are roughly similar in shape. The airborne leaf area density inversion result is greater than the ground-based inversion result in the upper part of the canopy, while the ground-based inversion result is greater than the airborne result in the lower part of the canopy.

[0121] Therefore, it can be seen that due to the differences in scanning methods between airborne and ground-based methods, their scanning blind zones are also different. By observing and determining the blind zone locations of the two scanning methods through experiments, and combining the original data of both methods to numerically fuse the curves, the scanning blind zone can be eliminated to a certain extent.

[0122] To further achieve synergistic and complementary inversion between the two methods, this embodiment employs a random cropping method, randomly cropping tea trees at different heights within four sample plots. Inversion models are then re-established for these four tea tree sample plots, and leaf area density curves for both airborne and ground-based applications are plotted. Figure 8 As shown;

[0123] observe Figure 8 It can be seen that no matter how high the canopy of the tea trees in the sample plot is cut, the intersection of the two inversion curves will not change, indicating that for the same test object, the intersection of the foundation and airborne LAD inversion curves does not change with the change of canopy structure.

[0124] Because the scanning blind zone is often determined by the scanning characteristics of the instrument, the incident angle of the airborne scan is from top to bottom. However, the dense leaves below the canopy of shrub crops intertwine and block each other, causing interference to the airborne scan and creating more blind zones below the intersection point.

[0125] When scanning the ground, the scanning angle is fixed, which cannot penetrate the leaf surface of the top layer of the canopy, resulting in missed detection of the leaf surface above the canopy. Therefore, the leaf area obtained above the intersection point is smaller than the measured value.

[0126] The inversion curve is divided into upper and lower parts based on the altitude of the intersection point. The upper part of the altitude layer uses airborne inversion results, and the lower part uses ground-based inversion results. The two curves are then fitted and reconstructed. The final inversion result is as follows: Figure 9 As shown;

[0127] To further verify the accuracy of LAD co-inversion, the measured leaf area index of four quadrats was obtained using an LAI-2200 canopy analyzer. A 90° cover cap was used during the measurement, and the measurement was performed once every 90° for a total of 4 measurements. The average value was taken as the measured value of LAI.

[0128] Different areas were selected in four quadrats, and the leaf area density values ​​obtained from airborne, base, and fusion curves were integrated to calculate the corresponding discrete LAI estimates.

[0129] The estimated LAI values ​​were compared with the measured LAI values, and curve fitting was performed. The fitting results are as follows: Figure 10 As shown;

[0130] From an individual perspective, the estimation accuracy R of the leaf area index obtained from a single foundation is... 2 The value is 0.824–0.933, which is higher than the airborne estimation accuracy R. 2 The range is 0.693 to 0.838. For low-growing shrub-type crops, due to their short height and high overlap of leaves at the bottom of the canopy, the advantage of airborne scanning of the height layer is not so obvious. However, by setting up multiple ground stations at appropriate observation angles to perform multi-angle scanning of shrub-type white tea trees, higher estimation accuracy can be obtained.

[0131] Overall, the accuracy R of the LAI estimation obtained through collaborative inversion is... 2 The accuracy rates were 0.930–0.963 and RMSE were 0.101–0.118, respectively, which are higher than the estimation accuracy using only ground-based and airborne methods. The fusion curve LAI estimation value is highly correlated with the measured LAI value, indicating that the fusion curve improves the accuracy of estimating the leaf area density of white tea leaves.

[0132] This invention uses the SZT-R250 three-dimensional laser mobile measurement system and the UA-0500 ground-based laser scanner to sample point cloud data of shrub-type white tea trees, and reconstructs the LAD inversion curve of white tea trees based on the complementary characteristics of the two scans, thereby achieving the synergistic inversion of leaf area density of shrub-type crops.

[0133] The experimental group compared common point cloud segmentation algorithms and selected the random forest algorithm to perform preprocessing such as branch and leaf separation on the canopy point cloud data. The obtained information was then used for voxel modeling to explore the intrinsic relationship between the voxel size of white tea trees and the laser contact frequency. A collaborative inversion model was established based on Beer-Lambert's law.

[0134] After applying multinomial correction to the point cloud data of white tea trees, the random forest algorithm was used to separate branches and leaves with high accuracy. The precision, recall, mean precision and overall accuracy were 0.957, 0.964, 0.960 and 0.976, respectively, showing good separation effect. However, whether this method is applicable to other shrub-type crops of the same type needs further research.

[0135] The laser contact frequency of shrub-type white tea trees is highly correlated with the size of the set voxel. Setting the voxel too large will result in an over-calculated contact frequency, and the estimated LAI value will be greater than the measured LAI value. Setting the voxel too small will result in an under-calculated contact frequency, and the estimated LAI value will be smaller than the measured LAI value. The optimal voxel value will vary under different environmental parameters.

[0136] Ground-based and airborne lidars produce blind zones in different locations due to their different scanning methods. The blind zone produced by airborne lidar is located in the lower middle part of the tea tree canopy, while the blind zone produced by ground-based lidar is located in the upper middle part of the tea tree canopy. Reconstructing and complementing the two curves can achieve higher inversion accuracy.

[0137] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Any simple modifications, alterations, and equivalent changes made to the above embodiments based on the inventive essence shall still fall within the protection scope of the present invention.

Claims

1. A dual-source point cloud laser-coordinated inversion method for the area concentration of white tea leaves, characterized in that, Includes the following steps: S1. Obtain ground point cloud data and airborne point cloud data of shrub-type white tea trees; S2. Separate and extract the ground-based point cloud data and the airborne point cloud data obtained in S1; S3. Based on the separated and extracted ground point cloud data obtained in S2 and the airborne point cloud data, the point cloud of the shrub-type white tea tree is volumetricated and a volumetric model is established. S4. Perform coordinate registration between the ground-based point cloud data and the airborne point cloud data obtained in S1; S5. The area density of white tea leaves is calculated by synergistically inverting ground-based point cloud data and airborne point cloud data using the canopy analysis method.

2. The dual-source point cloud laser collaborative inversion method for white tea leaf area concentration according to claim 1, characterized in that, The method for obtaining the ground point cloud data described in S1 is as follows: according to the distance of the tea plantation, four 10m×10m tea tree vegetation sample plots are selected as monitoring sample plots to minimize the mutual shading of branches and leaves, so as to achieve all-round scanning of the ground-based lidar and obtain ground point cloud data. The ground point cloud data adopts Cartesian coordinates. The method for acquiring airborne point cloud data is as follows: each sample plot is divided into four 5m×5m small sample squares, and five monitoring points are set in each sample square. The five monitoring points are respectively set at the four corner points and the center point of the sample square, and an airborne lidar is set at each monitoring point to obtain airborne point cloud data. The airborne point cloud data adopts geodetic projection coordinates.

3. The dual-source point cloud laser collaborative inversion method for white tea leaf area concentration according to claim 1, characterized in that, The separation described in S2 is achieved by using a quadratic polynomial method to correct and fit the height and distance of the ground-based point cloud data and the airborne point cloud data, and then using the corrected data to establish a classification model for branch and leaf separation. The calibration fitting process involves selecting 50 white tea trees of different shapes as the test subjects, importing their ground-based point cloud data and airborne point cloud data into a commonly used point cloud segmentation algorithm, and then outputting point cloud information that retains the effective leaf surface.

4. The dual-source point cloud laser collaborative inversion method for white tea leaf area concentration according to claim 3, characterized in that, The commonly used point cloud segmentation algorithms include: random forest classification, normal difference method, maximum likelihood method, and Euclidean distance segmentation method; By comparing the effective leaf point cloud information, a better branch and leaf separation effect can be obtained through the random forest method.

5. The dual-source point cloud laser collaborative inversion method for white tea leaf area concentration according to claim 3, characterized in that, The method for creating point clouds of shrub-type white tea trees using volumetric methods described in S3 is as follows: the object to be measured is divided into a three-dimensional volumetric network according to the size of the volumetric elements. At the same time, the point cloud data of the object to be measured is projected onto the volumetric network. Then, it is determined whether there are laser light spots in each volumetric element. If there are, it means that there are tea leaves in that volumetric element; otherwise, there are not.

6. The dual-source point cloud laser collaborative inversion method for white tea leaf area concentration according to claim 5, characterized in that, The optimal voxel value is determined by the maximum scanning distance, divergence angle, and exit diameter of the laser scanner.

7. The dual-source point cloud laser collaborative inversion method for white tea leaf area concentration according to claim 6, characterized in that, After the optimal voxel value is determined, the minimum value of the point cloud data in each direction is taken as the starting point and the maximum value as the ending point. The optimal voxel value is used as the interval step to obtain the voxelized coordinates of each point and establish the voxel model of the white tea tree. By projecting the point cloud data onto the xy plane and using the two-dimensional convex hull method, the canopy boundary of each horizontal layer can be quickly determined.

8. The dual-source point cloud laser collaborative inversion method for white tea leaf area concentration according to claim 6, characterized in that, The coordinate registration described in S4 includes the following steps: S401. Select point set m from the acquired ground point cloud data of the object under test, and select point set n from the airborne point cloud data; S402. Calculate the rigid body transformation that minimizes the objective function for the corresponding point in S401, and then use the four-parameter method to find the rotation matrix R and translation vector T at this time. S403. Substitute the R and T obtained in S402 into the point set n selected in S401 to obtain a new point set n′=Rn+T; S404. Calculate whether the objective function of the point set n′ obtained in S403 and the point set m obtained in S401 meets the threshold requirement. If it does, stop the iteration. If it does not meet the objective function threshold requirement, return the point set n′ obtained in S403 to step S402 for a new round of iteration until the minimum sum of squared distances of the new point set obtained in S403 meets the threshold requirement and the convergence requirement is met.

9. The dual-source point cloud laser collaborative inversion method for white tea leaf area concentration according to claim 1, characterized in that, The canopy analysis method described in S5 observes and analyzes the inversion curves obtained by the two scanning methods, determines the blind zone of the two scanning methods, and finally reconstructs and numerically fuses the curves to obtain high-precision inversion results.

10. The dual-source point cloud laser collaborative inversion method for white tea leaf area concentration according to claim 1, characterized in that, The ground-based point cloud data and airborne point cloud data mentioned in S1 have different blind spots due to different scanning methods. The blind spot generated by the airborne method is located in the middle and lower part of the tea tree canopy, while the blind spot generated by the ground-based method is located in the middle and upper part of the tea tree canopy. Reconstructing and complementing the two curves can achieve higher inversion accuracy.

Images