A tree leaf area estimation method based on ground laser point cloud

Abstract

The application discloses a tree leaf area estimation method based on ground laser point cloud, comprising the following steps: collecting laser point cloud data of a tree through a scanner; separating branches and leaves of the tree laser point cloud data; performing single leaf segmentation on leaf point cloud; obtaining three characteristic parameters of all leaves on the tree, respectively, the number of point clouds of single leaf, the Euclidean distance of single leaf center from the scanner, and the angle between the normal vector of single leaf and the incident light line of the scanner; selecting the three characteristic parameters of multiple leaves on the tree and combining L1 regularization and L2 regularization multiple regression method to obtain the leaf area fitting value of other leaves on the tree. Compared with the direct measurement relying on manual operation, the tree leaf area estimation method can quickly and accurately estimate the real leaf area of the tree crown from the perspective of machine vision based on the scanning point cloud without damaging the leaves, so as to replace the complex manual measurement operation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of forest parameter research, and particularly relates to a tree leaf area estimation method based on ground laser point cloud. BACKGROUND

[0002] Leaves are the main organs of plants, and are important organs and media for respiration, transpiration, photosynthesis and organic matter synthesis of plants. Leaves are important indicators for measuring plant growth conditions and yield, and are also important means for crop cultivation management and pest monitoring. The characteristics (morphology, texture, etc.) and health status of leaves can usually show the growth and development state of plants. In order to extract the characteristics of plants, leaf segmentation is crucial. The leaf segmentation and identification and leaf phenotype feature extraction method based thereon can timely feedback the crop growth state to ensure crop yield, and have important practical significance. The leaf area index (LAI) is an important plant structure parameter for representing the structure of a vegetation canopy, and is an important indicator for reflecting the growth status of vegetation, and directly affects the yield of crops to some extent.

[0003] Accurate measurement of leaf area is the premise of leaf area research. The measurement of leaf area is very time-consuming and involves leaf destruction. Therefore, it is of great significance to establish a convenient and accurate method for measuring leaf area. The measurement of leaf area is divided into two categories: direct measurement and indirect measurement. Direct measurement methods include destructive sampling direct measurement method, leaf collection or establishment of allometric equation related to easily measured vegetation characteristics (such as height and diameter at breast height), etc. The direct measurement method can accurately measure the leaf area, but these methods have the disadvantages of high labor cost, time-consuming, low efficiency, etc. In order to overcome these obstacles, various methods for indirectly measuring leaf area have been developed. Common indirect methods are based on measuring canopy light transmission or gap fraction, and many optical instruments such as digital hemispherical photography (DHP), LAI-2000 plant canopy analyzer, AccuPAR intercept device, and TRAC instrument have been developed. However, the observation results of these instruments are very sensitive to environmental light and manual operation procedures.

[0004] Ground laser scanning (TLS) provides a novel indirect measurement method for estimating LAI. TLS is a method for depicting the fine structure characteristics of forests by obtaining three-dimensional laser point clouds, and has high precision and efficiency. At present, TLS has been widely used to measure many structure parameters of trees, such as tree height, diameter at breast height, crown vertical profile, and forest stem density.

[0005] Therefore, it is necessary to provide a tree leaf area estimation method based on ground laser point cloud to replace the complex manual direct measurement operation. SUMMARY

[0006] The technical problem solved by the present application is to provide a tree leaf area estimation method based on ground laser point cloud to solve the above problems of the prior art.

[0007] To achieve the above technical purpose, the technical scheme adopted by the present application is:

[0008] A tree leaf area estimation method based on ground laser point cloud, comprising the following steps:

[0009] (1) Collecting laser point cloud data of the tree by a scanner;

[0010] (2) Calculating the obtained laser point cloud data to realize branch and leaf separation of the tree laser point cloud data;

[0011] (3) Focusing on the extracted leaf point cloud, the leaf point cloud is segmented by single leaf;

[0012] (4) Obtaining three characteristic parameters of all leaves on the tree, respectively, the point cloud number of single leaf, the Euclidean distance of single leaf center from the scanner, and the angle between the normal vector of single leaf and the incident light ray of the scanner;

[0013] (5) Selecting the three characteristic parameters of multiple leaves on the tree and the real measurement value of the leaf area of the corresponding leaves as the training sample data set, and combining L1 regularization and L2 regularization multiple regression method to obtain the leaf area fitting value of other leaves on the tree.

[0014] As a further improved technical solution of the present application, the angle between the normal vector of single leaf and the incident light ray of the scanner in step (4) is:

[0015]

[0016] Where incb leaf represents the incident light ray vector of single leaf, i.e. the vector from the scanner position to the center point of single leaf point cloud, norv leaf represents the normal vector of single leaf, and θ includ represents the angle between the normal vector of single leaf and the incident light ray of the scanner.

[0017] As a further improved technical solution of the present application, the step (5) is specifically:

[0018] (5.1) Selecting the three characteristic parameters of N leaves on the tree and the real measurement value of the leaf area of the corresponding leaves as the training sample data set;

[0019] (5.2) The leaf area fitting value is obtained by performing a quadratic linear regression fitting on the leaf, and the formula is:

[0020]

[0021] wherein θ include represents the angle between the normal vector of the single leaf and the incident light of the scanner, num leaf represents the number of point clouds of the single leaf, dist leaf represents the Euclidean distance of the center of the single leaf from the scanner, W = [w0, w1, w2, w3, w4, w5, w6, w7, w8, w9] is a polynomial coefficient vector, and y represents the leaf area fitting value of the corresponding single leaf;

[0022] (5.3) A loss function is set, and the formula is:

[0023]

[0024] wherein represents the true measurement value of the leaf area of the single leaf, y represents the leaf area fitting value of the single leaf, and ξ1 and ξ2 represent regularization parameters;

[0025] minimizing that is, the partial derivative of is taken and is equal to 0, and then the coefficient W = [w0, w1, w2, w3, w4, w5, w6, w7, w8, w9] is calculated through the training sample data set. W = [w0, w1, w2, w3, w4, w5, w6, w7, w8, w9] is recorded as W' = (w'0, w'1, w'2, w'3, w'4, w'5, w'6, w'7, w'8, w'9). i

[0026] wherein:

[0027] J(W) = P(W) + Q(W) (4);

[0028]

[0029]

[0030] minimizing J(W), that is, taking the partial derivative of J(W), and there is:

[0031]

[0032]

[0033] Let:​

[0034]

[0035] Then:

[0036]

[0037]

[0038]

[0039] The three characteristic parameters of the single leaf in the training sample set and the corresponding are input into formula (12), and W is calculated;

[0040] (5.4) The calculated W is substituted into formula (2), and the three characteristic parameters of other single leaves with unknown leaf area on the tree are substituted into formula (2), so that the leaf area fitting value of the corresponding single leaf is obtained.

[0041] The beneficial effects of the present application are:

[0042] The present application combines the related theoretical methods of computer graphics and forestry, excavates the correlation between laser point cloud and leaf area, inverses the leaf area from the point cloud number of single leaf, the angle between incident light and leaf normal vector and the distance between scanner and leaf by combining L1 and L2 regularization multivariate regression method, and further calculates the real leaf area of the crown. The method of the present application can accurately obtain point cloud data by using ground-based light radar and has low requirements for light environment configuration. Compared with direct measurement depending on artificial, the real leaf area of the crown is quickly and accurately estimated from the perspective of machine vision based on scanning point cloud without damaging the leaf, so as to replace the complex artificial measurement operation. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 Fig. (a) shows the location map of the research area;

[0044] Figure 1 Fig. (b) shows the field collection diagram of Lagerstroemia indica, Prunus serrulata, Ginkgo biloba and Cinnamomum camphora;

[0045] Figure 1 Fig. (c) shows the schematic diagram of manually measuring the target leaf area by using LI-3000C;

[0046] Figure 2 Fig. (a) shows the schematic diagram of scanning point cloud data of the experimental tree (Lagerstroemia indica);

[0047] Figure 2 Fig. (b) shows the schematic diagram of scanning point cloud data of the experimental tree (Prunus serrulata);

[0048] Figure 2Fig. 3 shows a schematic diagram of point cloud data scanned from an experimental tree (Ginkgo tree);

[0049] Figure 2 Fig. 4 shows a schematic diagram of point cloud data scanned from an experimental tree (Camphor tree);

[0050] Figure 3 Fig. 5 shows a schematic diagram of a specific experimental procedure.

[0051] Figure 4 Fig. 6 shows a schematic diagram of initial point cloud of an experimental tree;

[0052] Figure 4 Fig. 7 shows a schematic diagram of separated leaf point cloud;

[0053] Figure 4 Fig. 8 shows a schematic diagram of separated branch point cloud;

[0054] Figure 4 Fig. 9 shows a schematic diagram of results of single leaf segmentation of a tree and identification of each leaf with a random different color.

[0055] Figure 4 Fig. 10 shows a schematic diagram of Figure 4 Fig. 11 shows a partial enlarged view of part (d).

[0056] Figure 5 Fig. 12 shows a schematic diagram of leaf scanning;

[0057] Figure 6 Fig. 13 shows a schematic diagram of initial point cloud of an experimental tree (Lagerstroemia indica tree);

[0058] Figure 6 Fig. 14 shows a schematic diagram of separated leaf point cloud of an experimental tree (Lagerstroemia indica tree);

[0059] Figure 6 Fig. 15 shows a schematic diagram of separated branch point cloud of an experimental tree (Lagerstroemia indica tree);

[0060] Figure 6 Fig. 16 shows a schematic diagram of initial point cloud of an experimental tree (Cherry tree);

[0061] Figure 6 Fig. 17 shows a schematic diagram of separated leaf point cloud of an experimental tree (Cherry tree);

[0062] Figure 6 Fig. 18 shows a schematic diagram of separated branch point cloud of an experimental tree (Cherry tree);

[0063] Figure 6 Fig. 19 shows a schematic diagram of initial point cloud of an experimental tree (Ginkgo tree);

[0064] Figure 6(c2) represents a schematic diagram of the point cloud of leaves after the experimental tree (ginkgo tree) was separated;

[0065] Figure 6 (c3) represents a schematic diagram of the point cloud of branches after the experimental tree (ginkgo tree) was separated;

[0066] Figure 6 The middle (d1) represents the initial point cloud diagram of the experimental tree (camphor tree);

[0067] Figure 6 (d2) represents a schematic diagram of the point cloud of leaves after the experimental tree (camphor tree) was separated;

[0068] Figure 6 The diagram in 'd3' represents a point cloud representation of the branches and trunks of the experimental tree (camphor tree) after separation.

[0069] Figure 7 Image (a1) shows the result of single-leaf segmentation of the experimental tree (crape myrtle);

[0070] Figure 7 Image (b1) shows a magnified view of the results of single-leaf segmentation of the experimental tree (crape myrtle);

[0071] Figure 7 Image (a2) shows the results of single-leaf segmentation of the experimental tree (cherry tree);

[0072] Figure 7 Image (b2) shows a magnified view of the results of single-leaf segmentation of the experimental tree (cherry tree);

[0073] Figure 7 Image (a3) ​​shows the results of single-leaf segmentation of the experimental tree (ginkgo tree);

[0074] Figure 7 Image (b3) shows a magnified view of the results of single-leaf segmentation of the experimental tree (ginkgo tree);

[0075] Figure 7 Image (a4) shows the results of single-leaf segmentation of the experimental tree (camphor tree);

[0076] Figure 7 Image (b4) shows a magnified view of the results of single-leaf segmentation of the experimental tree (camphor tree);

[0077] Figure 8 (a) represents the regression analysis graph of the measured and predicted leaf area values ​​of the experimental tree (crape myrtle);

[0078] Figure 8 (b) shows the regression analysis graph of the measured and predicted leaf area values ​​of the experimental trees (cherry trees);

[0079] Figure 8 Fig. 3 is a diagram showing the regression analysis of the measured leaf area and the predicted value of the experimental tree (Ginkgo tree);

[0080] Figure 8 Fig. 4 is a diagram showing the regression analysis of the measured leaf area and the predicted value of the experimental tree (Camphor tree). DETAILED DESCRIPTION

[0081] The specific embodiments of the present application are further described below with reference to the accompanying drawings:

[0082] Terrestrial laser scanning (TLS) can obtain dense laser point clouds of plants to precisely depict the structural parameters of forests, such as tree height and leaf area. Real leaf area is an important parameter for phenotypic research in forestry and botany, but there is currently no good measurement method. Therefore, this paper proposes a method for estimating the leaf area of trees based on terrestrial laser point clouds. First, a single-leaf segmentation algorithm for plant point clouds based on planar region growing is designed to achieve accurate single-leaf point cloud segmentation and extraction. Second, the angle between the normal vector of a single leaf and the incident laser line of the scanner, the distance between the scanner and the leaf, and the number of point clouds of a single leaf are used as input features, and the L1 and L2 regularization multiple regression method is used to invert the leaf area of each leaf in the crown, and then the real leaf area of the crown is calculated. This method takes four trees, Lagerstroemia indica, Prunus serrulata, Ginkgo biloba, and Cinnamomum camphora, as research objects. Through experiments on the four trees in the campus and comparison with the measured values, the least squares regression method and the least squares combined with L1+L2 regularization regression method are used for comparative experiments. The results show that the least squares regression method and the least squares combined with L1+L2 regularization regression method are better. Among them, the determination coefficients of the experimental analysis values and the measured values of the real single leaf area of Lagerstroemia indica and Prunus serrulata are better, which are Lagerstroemia indica (R 2 = 0.95, RMSE = 0.42 (cm 2 )), and Prunus serrulata (R 2＝ 0.92, RMSE = 1.87 (cm 2 )). Due to the existence of occlusion, the coefficients obtained by Ginkgo biloba and Cinnamomum camphora are relatively low, which are Ginkgo biloba (R 2＝ 0.83, RMSE = 1.24 (cm 2 )), and Cinnamomum camphora (R 2＝ 0.86, RMSE = 1.10 (cm 2 )). The method in this paper accurately estimates the leaf area from the perspective of computer graphics and machine vision based on laser point cloud data, providing a more promising inversion method for obtaining the leaf area of trees.

[0083] 1. Materials and methods:

[0084] 1.1, Study area and data collection:

[0085] The point cloud data selected in this study is from the campus of Nanjing Forestry University, as shown in Figure 1 . Figure 1 In (a), the study area is located in Jiangsu Province, Nanjing City; Figure 1 In (b), the field collection of Lagerstroemia indica, Prunus serrulata, Ginkgo biloba and Cinnamomum camphora trees; Figure 1 In (c), the target leaf area is measured manually using LI-3000C. The study area is located in the transition zone between subtropical and warm temperate zones, with complex terrain and diverse ecological environment, suitable for the growth and reproduction of various plants and animals. It is a transitional zone between north and south plants, with a wide variety of biological species. In order to make this study applicable to the leaf characteristics of various tree species, we collected point cloud data of Lagerstroemia indica, Prunus serrulata, Ginkgo biloba and Cinnamomum camphora trees (as shown in Figure 2 ). Figure 2 The experimental trees and scanning laser point cloud data for this paper. Among them Figure 2 (a), (b), (c) and (d) are the point cloud data scanned by Lagerstroemia indica, Prunus serrulata, Ginkgo biloba and Cinnamomum camphora trees, respectively.

[0086] The point cloud data of the four analyzed trees was collected using the ground laser scanner FARO Focus3D X330. The FARO X330 laser scanner used in this experiment is a high-speed three-dimensional scanner with an ultra-long scanning distance. This instrument has a 300° x 360° field of view, and it can scan objects up to 330 meters away in direct sunlight. Only a few scans are needed to measure large objects in complex terrain at a distance, which can greatly improve measurement speed. The scanning rate can reach 976000 points / second, and the distance between scanning points is 3.068mm / 10m. The specific experimental process is shown in Figure 3 . In order to obtain more complete point cloud data, different ranges of collection methods were adopted according to the different shapes of tree height in this experiment. All point cloud data were scanned using FARO X330 scanner, and the detailed information of the four trees is shown in Table 1. In addition, we also used LI-3000C portable leaf area meter to manually calculate the number of leaves and measure the leaf area (as shown in Figure 1 (c)), and the results can be used as a benchmark to verify the results of our method.

[0087] Table 1 shows the detailed information of the four trees:

[0088] Indicator Loropetalum chinense Prunus serrulata Ginkgo biloba Cinnamomum camphora Height of tree (m) 2.50 3.50 10.35 16.31 Crown width north / south and east / west (m) 2.04 / 2.24 2.95 / 2.54 4.96 / 4.26 7.16 / 6.06 Scanner height (m) 1.2 1.2 1.2 1.2 Height under branch (m) 0.67 1.29 1.75 3.67 Single leaf length (cm) 5.29±1.02 11.51±2.35 4.78±0.96 7.05±1.67 Single leaf width (cm) 3.46±0.52 5.91±2.02 7.33±1.47 4.40±0.72

[0089] 1.2, Branch and leaf separation:

[0090] Obtaining accurate vegetation parameters such as leaf area requires separating branches and leaves. Branch and leaf separation is a necessary prerequisite for extracting individual tree features, and its purpose is to divide the lidar point cloud into two parts: branches and leaves. During the scanning and measurement process, the influence of the plants themselves and natural environmental disturbances such as wind disturbances can affect the accuracy and computational complexity of the subsequent model reconstruction. Therefore, it is necessary to remove invalid noise points in the experimental trees simultaneously.

[0091] This paper first removes isolated point clouds (noise points) and calculates a series of features of tree branch and leaf point cloud data, such as normal vectors and structural tensors, which are then fed into a semi-supervised support vector machine to perform branch and leaf separation operations on the scanned tree. The results of the tree branch and leaf point cloud separation experiment are as follows: Figure 4 As shown in (b) and (c). Figure 4 The experimental results of the proposed method for separating tree branches and leaves and segmenting single leaves are shown. Figure 4 In the middle (a), the initial point cloud of the experimental tree is represented. Figure 4 (b) represents the separated leaf dot clouds. Figure 4 (c) represents the separated branch point cloud. Figure 4 In the middle (d), the result of segmenting a single leaf of a tree is shown, and each leaf is identified with a different random color.

[0092] 1.3. Single leaf separation:

[0093] After separating the branches and leaves, and considering the extracted leaf point cloud, since the leaves approximately exhibit planar features, this experiment designed a single-leaf segmentation algorithm based on the growth of small planar regions in plant point clouds, as follows:

[0094] Each point p in the leaf dot cloud i Spatial features include p i The spatial smoothness s i The plane normal vector n fitted to the neighborhood i Where p i The fitting plane f i =(X K×3 ,p i ,n i ) is a triplet structure, p i The K-nearest neighbor point set matrix X K×3 Calculated using iterative principal component analysis (PCA), each iteration of PCA calculates a subset belonging to p. i p in the neighborhood j Distance d of the point i,j . d i,j Defined as X K×3 point p in j The Euclidean distance to the global fitting plane, where l1 is the threshold, when d i,j >l1, move point pj from X K×3 i,j The formula is as follows:

[0095]

[0096] When some points are removed, the K-neighbor point set becomes K'-neighbor point set. When the size of X K×3 remains unchanged, the PCA iteration process will stop. The calculated fitting plane f i = (X K×3 , p i , n i ) is conducive to representing the spatial structure of each point p i . The covariance matrix C K×3 of X i also needs to be calculated to update the fitting plane, and the calculation formula is as follows:

[0097]

[0098] where X K×3 has been centered, that is, each neighborhood point is subtracted from the center point p i , λ1, λ2, λ3 (λ1> λ2> λ3) are the eigenvalues of C i . Therefore, the unit eigenvector corresponding to λ3 is the normal vector n i of the plane where p i is located, and the smoothness s i is the ratio of λ2 and λ3. The larger the ratio is, the smoother p i is around, and the more it tends to be a plane feature. The smoothness s i and the normal vector n i will be updated in each iteration.

[0099] Next, a small facet merging clustering algorithm is implemented to identify a single leaf, and the point p K×3 with the highest smoothness in the K-neighbor set X k corresponding to the center point of the leaf point cloud is defined as a seed point, and a region with the same characteristics is established by region growing with p k as the seed point. The characteristics of this region growing are as follows: 1) the distance from the neighborhood point p j to the fitting plane f j where the seed point p k is located is less than l1; 2) the angle between the neighborhood point p j and the normal vector of the fitting plane where the seed point p k is located is less than θ; 3) the Euclidean distance between the neighborhood point p j and the seed point p k is less than half of the average leaf length l2 / 2. The purpose of the above three conditions is to judge p​j With p k whether coplanar and adjacent, but a new point p j is classified into a facet, it is removed from the point cloud, neither as a "candidate" seed point, nor involved in the region growing with other seed points. By the above method, a preliminary segmentation of the point cloud of the whole plant can be obtained Facet cluster F.

[0100] Each facet F i in the facet cluster F Figure 4 is taken as a unit, region growing is performed according to the facet adjacency relationship and the coplanar characteristics between facets, and the grown multiple facets will be spliced into a large spatial structure. If the number of points covered exceeds a certain number, it is considered to be a single leaf blade segmented, thereby segmenting a single leaf blade. The single leaf segmentation effect of the algorithm in this paper is shown in Fig.

[0101] 1.4, Leaf area calculation:

[0102] In laser scanning, according to common sense, the leaf area of each leaf is related to the point spacing and the number of points in the point cloud on the leaf. The point spacing is related to the distance from the laser scanner to the leaf and the angle between the incident light beam and the normal vector of the leaf. Therefore, the number of points in a single leaf, the distance from the scanner to the leaf, and the angle between the incident light and the normal vector of the leaf are selected as the three variables of the leaf area inversion.

[0103] Since the segmentation of the single leaf point cloud has been realized in the last section, we can easily obtain the number of points in a single leaf num leaf and the Euclidean distance dist leaf of the center of the single leaf from the scanner. The angle θ include between the incident light beam and the normal vector of the leaf needs to be determined by the normal vector of the leaf and the incident angle of the laser. The normal vector of the leaf is obtained by finding the normal vector corresponding to the fitting plane of each point on the leaf, and then the normal vector of the corresponding leaf norv leaf is obtained by weighted averaging the normal vectors in different directions. The incident light vector of the leaf is the vector from the scanner position to the center point of the leaf point cloud, denoted as incb leaf . The angle θ includ between the normal vector of a leaf and the incident light vector of the scanner is calculated as:

[0104]

[0105] where incb leaf represents the incident light vector of a single leaf, i.e. the vector from the scanner position to the center point of the single leaf point cloud, norv leaf represents the normal vector of the corresponding single leaf, and θ includThis represents the angle between the normal vector of the corresponding single leaf and the incident light ray from the scanner.

[0106] The smaller the angle between the blade's normal vector and the scanner's incident laser line (or the closer it is to 180°), the larger the area of ​​the blade facing the scanner, and the higher the resulting dot density. Conversely, the closer the angle between the blade's normal vector and the scanner's incident laser line is to 90°, the more the side of the blade faces the scanner, resulting in a lower dot density. Furthermore, the dot density varies with the distance between the scanner and the blade. leaf An increase in leaf area leads to a decrease in leaf point cloud size, and vice versa. Similarly, a larger leaf area results in a greater number of leaf point clouds, and vice versa. See the specific illustrations below. Figure 5 As shown. Figure 5 This is a schematic diagram of a leaf scan. The properties of the point cloud of a single leaf are related to (1) the angle θ between the scanning ray and the leaf normal vector. includ e, (2) Distance from the scanner to the center of each blade leaf and (3) the number of point clouds of a single leaf num leaf Related.

[0107] 1.5 Leaf area calculation:

[0108] For each tree, N leaves are taken as training samples. Each leaf is represented by the three feature parameters θ obtained earlier. includ ,num leaf dist leaf As input, the actual measured leaf area of ​​the corresponding leaf is... The leaf area was obtained by performing a quadratic linear regression fit on these leaves, and the polynomial was defined as follows:

[0109]

[0110] in θ include num represents the angle between the normal vector of a single leaf and the incident ray from the scanner. leaf Dist represents the number of point clouds for a single leaf. leaf W represents the Euclidean distance from the center of a single leaf to the scanner, where W = [w0, w1, w2, w3, w4, w5, w6, w7, w8, w9] is the polynomial coefficient vector, and y represents the fitted value of the leaf area of ​​the corresponding single leaf.

[0111] 1.6 L1 Regularization and L2 Regularization Fitting:

[0112] Regularization is a method of adding extra information to the original loss function in machine learning, so as to avoid overfitting and improve the generalization performance of the model. The method is to artificially add a term to the loss function, so that the parameter estimation process of the model is more conducive to convergence to the local minimum. The most commonly used regularization terms are generally two kinds, L1 norm (or L1 regularization, l1-norm) and L2 norm (or L2 regularization, l2-norm). When used, we usually add a regularization parameter ξ1, ξ2 in front of the regularization term to control the degree of regularization we need.

[0113] In this paper, a loss function containing L1+L2 regularization term is designed, which not only considers the sparsity of L1 regularization method, but also adds the operation stability and anti-overfitting performance of L2 regularization method, and can adjust the regularization parameter according to different model requirements. The Loss function using L1+L2 regularization has the expression form as shown in formula (5):

[0114]

[0115] Wherein, and y are the true measurement value and the fitting value of the algorithm in this paper respectively, and ξ1, ξ2 represent the regularization parameters;

[0116] Minimize That is, the partial derivative of is equal to 0, and then the coefficient W = [w0, w1, w2, w3, w4, w5, w6, w7, w8, w9] is calculated through the training sample data set. W = [w0, w1, w2, w3, w4, w5, w6, w7, w8, w9] is recorded as W' = (w'0, w'1, w'2, w'3, w'4, w'5, w'6, w'7, w'8, w'9);

[0117] The partial derivative of J(W) is taken to obtain W when J(W) is minimum. For convenience of calculation, J(W) is divided into two parts, that is:

[0118] J(W) = P(W) + Q(W) (6);

[0119]

[0120]

[0121] The partial derivative expressions of P(W) and Q(W) are as follows:

[0122]

[0123]

[0124] For the sake of neatness in the subsequent formulas, let:

[0125]

[0126] Minimizing J(W) yields the coefficient W that minimizes the error between the predicted and actual values ​​after regularization, i.e.:

[0127]

[0128] The three feature parameters θ in the training sample set includ ,num leaf dist leaf and the actual measured value of leaf area Input the formula (12) to calculate W;

[0129] Substitute the calculated W into formula (4), and then calculate the {θ} of the single leaf whose leaf area is to be determined on a tree of the same species as the trees in the training sample set. include ,num leaf dist leaf Substituting into formula (4), the leaf area fitting value of the corresponding single leaf can be obtained.

[0130] 2. Results and Discussion:

[0131] 2.1 Results of forest point cloud data processing:

[0132] By combining the features of leaf point cloud data, such as normal vectors and structural tensors, with a semi-supervised support vector machine using the method described in section 1.2, the trunks and leaves of four experimental trees—crape myrtle, cherry blossom, ginkgo, and camphor—were effectively and quickly separated. The separated crown and trunk point clouds showed high consistency with the real plants. Figure 6 The results of branch and leaf separation are shown, with green representing leaves and brown representing branches, indicating that the method is feasible for four experimental trees. Figure 6 This shows the effect of separating the branches and leaves of four experimental trees. The first row shows the crape myrtle and cherry blossom trees, and the second row shows the ginkgo and camphor trees. Figure 6 In the diagram, (a1), (b1), (c1), and (d1) represent the initial point clouds of the four trees. Figure 6 In the middle, (a2), (b2), (c2), and (d2) are the separated leaf point clouds. Figure 6 (a3), (b3), (c3), and (d3) are the branch point clouds after separation.

[0133] Figure 7This paper presents the single-leaf segmentation results of plant point clouds using a small-plane region growing-based algorithm. Experimental results show that the canopies of the four trees exhibit good segmentation results, with each identified leaf represented by a different color. Among the four experimental trees, the algorithm achieved the best segmentation results for the crape myrtle and cherry blossom trees. This is mainly because the crape myrtle and cherry blossom trees are located near the garden and old library at the campus entrance, surrounded by buildings and less susceptible to wind disturbance, and their canopies have high porosity, resulting in less leaf shading. In contrast, the ginkgo and camphor trees have large canopies and relatively high leaf area density, and their geographical location near the main campus road results in significant leaf shading and wind interference, thus affecting the segmentation of some leaves. The specific single-leaf segmentation results of this experiment are shown in Table 2.

[0134] 2.2 Leaf area inversion results:

[0135] For each experimental tree, leaf data was randomly selected to construct a predicted leaf area, with 30% selected as training samples and 70% as test samples.

[0136] Using three parameters as input features—the angle between the normal vector of a single leaf and the incident laser line of the scanner, the distance between the scanner and the leaf, and the number of point clouds in a single leaf—this experiment compares least squares regression with the least squares fitting plus L1+L2 regularized multivariate regression method proposed in this paper to evaluate the performance of leaf area inversion. The effectiveness of the proposed method is analyzed by comparing the results indicators (coefficient of determination and root mean square error). Figure 7 The results of single-leaf segmentation for four experimental trees. Figure 7 In the image, single leaves of crape myrtle (a), cherry blossom (b), ginkgo (c), and camphor tree (d) are segmented, and the segmented leaves are assigned random colors. A magnified view of a portion of the image is shown below. Figure 7 (b1), (b2), (b3), (b4).

[0137] By comparing the coefficient of determination R of the indicators 2 The validity of this study was analyzed by quantifying the root mean square error (RMSE). Figure 8 Regression analysis of predicted and measured leaf area values ​​for some leaves of four experimental trees is presented. Figure 8 This is a regression analysis graph showing the measured and predicted leaf area values ​​of four experimental trees. The blue dots and solid blue line represent the least squares regression points and regression line, while the orange dots and solid orange line represent the least squares combined L1+L2 regression points and regression line. The blue and orange intervals represent the 90% confidence intervals. R 2 The coefficients of determination are denoted as R0 and RMSE as the root mean square error. Using least squares regression, the fitting of the four trees showed both underfitting and overfitting, with the obtained coefficients of determination for the crape myrtle tree (R0) and RMSE respectively. 2=0.91, RMSE=0.83cm 2 Cherry blossom tree (R) 2 =0.88, RMSE = 2.35cm 2 Ginkgo tree (R) 2 =0.75, RMSE = 1.57cm 2 ), camphor tree (R 2 =0.80, RMSE=1.39cm 2 In the least squares plus L1+L2 regularized multiple regression method, the estimated leaf area of ​​all four trees was higher than that of the least squares regression method. Specifically, the results were as follows: For the crape myrtle tree (R... 2 =0.95, RMSE=0.42cm 2 Cherry blossom tree (R) 2 =0.92, RMSE=1.87cm 2 Ginkgo tree (R) 2 =0.83, RMSE=1.24cm 2 ), camphor tree (R 2 =0.86, RMSE=1.10cm 2 Among them, the determination coefficient of the two large trees (ginkgo and camphor trees) was slightly lower than that of the two small trees (crape myrtle and cherry blossom trees).

[0138] This is because during laser scanning, camphor and ginkgo trees are affected by external forces such as wind, and the obstruction from above the canopy causes deviations in the segmented leaves, affecting the calculation of leaf area. It can be seen that the L1+L2 regularization method proposed in this paper has better computational stability and robustness.

[0139] Table 2 shows the analysis results of the algorithm in this paper on four experimental trees:

[0140]

[0141] Figure 8 This is a regression analysis graph showing the measured and predicted leaf area values ​​of four experimental trees. The blue dots and solid blue line represent the least squares regression points and regression line, while the orange dots and solid orange line represent the least squares combined L1+L2 regression points and regression line. The blue and orange intervals represent the 90% confidence intervals. R 2 The coefficient of determination is RMSE, which is the root mean square error.

[0142] 3. Conclusion:

[0143] In this paper, the point cloud data of Lagerstroemia indica, Prunus serrulata, Ginkgo biloba and Cinnamomum camphora are obtained by using ground LiDAR equipment, and the branch and leaf separation, leaf separation and leaf area calculation of the point cloud data of the four trees are carried out. The main innovation points of the method are as follows: first, the leaf surface point cloud extraction and plant canopy inner point cloud single leaf segmentation algorithm based on plane region growing are designed to realize the extraction of each leaf point cloud; second, the principle of tree scanning is analyzed, and the angle between the normal vector of a single leaf and the incident laser line of the scanner, the distance between the scanner and the leaf, and the number of point clouds of a single leaf are designed as three features, and the L1 and L2 regularization multiple regression method is used to obtain the true leaf area value in the tree crown.

[0144] Through the analysis of the four trees in the campus with different sizes, the experimental results show that the determination coefficients R 2 of the manual leaf area measurement value and the predicted value of the smaller Lagerstroemia indica and Prunus serrulata are higher than 0.92, and the determination coefficients R 2 of the area measurement value and the predicted value obtained by the single leaf segmentation algorithm are still higher than 0.80 due to the existence of leaf shielding and leaf cluster aggregation in the high Ginkgo biloba and Cinnamomum camphora. Compared with the direct measurement depending on artificial, the method proposed in this paper quickly and accurately estimates the true leaf area of the tree crown from the perspective of machine vision to replace the complex manual measurement operation.

[0145] In the future, the following two points will be improved and optimized:

[0146] (1) When obtaining the point cloud of the experimental tree, although the crown of the experimental tree is not seriously shielded, there is still a shielding problem in the place where the crown is relatively dense, therefore, how to design a compensation mechanism to reasonably evaluate the crown shielding effect is a problem to be solved. (2) For different tree crowns, there are problems such as leaf cluster overlap and branch and leaf mutual shielding, how to improve the existing single leaf segmentation accuracy for different tree crown laser point cloud data is a problem worthy of further study.

[0147] The protection scope of the present application includes but is not limited to the above embodiments, and the protection scope of the present application is subject to the claims, any substitution, deformation, improvement of the present technology easily thought by the person skilled in the art falls within the protection scope of the present application.

Claims

1. A method for estimating tree leaf area based on ground laser point cloud, characterized in that: The method comprises the following steps: (1) collecting laser point cloud data of the tree by a scanner; (2) calculating the obtained laser point cloud data to realize branch and leaf separation of the tree laser point cloud data; (3) performing single leaf segmentation on the leaf point cloud for the extracted leaf point cloud; (4) obtaining three characteristic parameters of all leaves on the tree, respectively, the point cloud quantity of single leaf, the Euclidean distance of single leaf center from the scanner, and the angle between the normal vector of single leaf and the incident light of the scanner; (5) selecting the three characteristic parameters of multiple leaves on the tree and the real measured value of the leaf area of the corresponding leaves as training sample data set, and combining L1 regularization and L2 regularization multiple regression method to obtain the leaf area fitting value of other leaves on the tree; The step (5) is specifically: (5.1) Selecting three characteristic parameters of N leaves on a tree and the true measurement value of the leaf area of the corresponding leaves as a training sample data set; (5.2) performing secondary linear regression fitting on the leaves to obtain the leaf area fitting value, and the formula is: wherein θ include represents the angle between the normal vector of a single leaf and the incident light of the scanner, num leaf represents the number of point clouds of a single leaf, dist leaf represents the Euclidean distance from the center of a single leaf to the scanner, W = [w0, w1, w2, w3, w4, w5, w6, w7, w8, w9] is a polynomial coefficient vector, and y represents the fitted value of the leaf area of the corresponding single leaf; (5.3) setting a loss function, and the formula is: wherein represents the true measurement value of leaf area of a single leaf, y represents the fitted value of leaf area of a single leaf; and ξ1, ξ2 both represent regularization parameters. minimizing i.e. partial derivative, and make its partial derivative equal to 0, and then calculate the coefficient W = [w0, w1, w2, w3, w4, w5, w6, w7, w8, w9] through the training sample data set, and W = [w0, w1, w2, w3, w4, w5, w6, w7, w8, w9] is recorded as W' = (w'0, w'1, w'2, w'3, w'4, w'5, w'6, w'7, w'8, w'9); Wherein: J(W)=P(W)+Q(W) (4); Minimizing J(W), that is, taking the partial derivative of J(W), we have: Let: Then: The three feature parameters in the training sample set and the corresponding are input into formula (12) to calculate W. (5.4) substituting the calculated W into formula (2), and substituting the three characteristic parameters of the single leaf of the other leaf area to be solved on the tree into formula (2), so as to obtain the leaf area fitting value of the corresponding single leaf.

2. The method of claim 1, wherein: The angle between the normal vector of the single leaf and the incident light of the scanner in step (4) is: where incb leaf represents the incident light vector of a single leaf, i.e. the vector from the scanner position to the center point of the single leaf point cloud, norv leaf represents the normal vector of a single leaf, θ includ represents the angle between the corresponding single leaf normal vector and the scanner incident light.

Images