A single-tree denoising prediction method and system based on three-dimensional structure of vegetation

Abstract

This application discloses a method and system for single-tree noise reduction prediction based on three-dimensional vegetation structure, belonging to the field of single-tree noise reduction prediction technology. The method includes: collecting point cloud data and noise attenuation data of single trees and classifying them; extracting vegetation structure parameters of the target point cloud from the classified point cloud data; calculating the correlation between the extracted vegetation structure parameters and the collected noise attenuation data, and obtaining the most significant influencing factor in the correlation as a model variable; constructing a single-tree noise reduction prediction model based on a support vector machine regression algorithm, and training the constructed single-tree noise reduction prediction model using different kernel functions; evaluating the prediction accuracy of the single-tree noise reduction prediction model under different kernel functions, and selecting the optimal kernel function for single-tree noise reduction prediction; wherein, the constraint calculation of the prediction model is converted into dual calculation through Lagrange multiplication to obtain the optimal kernel function. Addressing the low quantitative accuracy of single-tree noise attenuation effect prediction in existing technologies, this application improves the accuracy of quantitative prediction.

Description

Technical Field

[0001] This application relates to the field of noise reduction prediction technology for single trees, and belongs to the field of environmental noise pollution prediction and prevention technology. Background Technology

[0002] Noise pollution is a global problem, causing numerous negative impacts. Effectively mitigating urban noise has become a research hotspot both domestically and internationally. Previous studies have developed many noise reduction methods and technologies, such as synthetic sound barriers and sound-absorbing materials. However, these technologies can lead to a series of negative consequences. In recent years, the acoustic effects of Urban Green Infrastructure (UGI) have attracted widespread attention for optimizing the urban acoustic environment. The acoustic effects of UGI can be categorized into ground interference, sound absorption, and sound redirection. Specifically, the porous surface of the matrix absorbs and scatters some incident sound waves, and interference effects reduce the incident sound waves; the mechanical vibrations of vegetation and air particles convert sound energy into heat energy, leading to sound energy dissipation; vegetation alters the propagation direction of sound waves through reflection, scattering, and diffraction. Furthermore, introducing natural sounds and pleasant landscapes can eliminate noise pollution and improve public health. Based on this, utilizing UGI to mitigate urban noise is a noise pollution control approach based on natural solutions.

[0003] The morphological structure of vegetation directly affects its acoustic effects; therefore, clarifying the mechanism by which vegetation structure influences noise attenuation is a prerequisite for using UGI to mitigate urban noise. The Normalized Difference Vegetation Index (NDVI) is a common indicator reflecting vegetation cover at the regional scale, and related studies have found that noise levels decrease with increasing NDVI. However, NDVI is only a two-dimensional indicator, typically used for qualitative assessment of the relationship between noise levels and green space, and cannot reflect the impact of three-dimensional vegetation structure at the microscale. In studies on UGI morphological parameters, researchers have pointed out that factors such as leaf inclination angle, total leaf area, leaf area index, and forest belt width affect the noise reduction capacity of vegetation communities. However, these studies usually employ sampling methods, using structural information from sampling points to estimate the vegetation structure of the entire area; traditional vegetation morphology sampling methods have low accuracy.

[0004] In related technologies, such as Chinese patent document CN116976216A, a method, model, prediction method, and device for constructing a three-dimensional forest belt noise reduction prediction model are provided, belonging to the field of environmental noise pollution prediction and prevention technology. The method includes the following steps: data input, including data preparation and type and spatial division; data processing, including structural parameter inversion, cone ray tracing, model sample library variable selection and calculation, and model sample library variable combination optimization; and data output, including model construction, building a three-dimensional spatial model for complex forest belt noise reduction prediction. However, in this scheme, if the structural parameters of the trees cannot be accurately inverted, the model cannot capture sufficient tree features, thus affecting the accuracy of the noise reduction prediction. Summary of the Invention

[0005] 1. Technical problems to be solved

[0006] To address the issue of low quantitative accuracy in predicting the noise attenuation effect of individual trees in urban environments in existing technologies, this application provides a method for predicting noise reduction of individual trees in urban environments based on the three-dimensional structure of vegetation. By collecting and analyzing point clouds of individual trees, key vegetation structure parameters representing spatial range, green density, and leaf tilt angle distribution are extracted, and a support vector machine regression model is constructed to achieve a more accurate quantitative prediction of the noise attenuation effect of individual trees in urban environments.

[0007] 2. Technical Solution

[0008] The purpose of this application is achieved through the following technical solution.

[0009] One aspect of this specification provides a method for single-tree noise reduction prediction based on the three-dimensional structure of vegetation, comprising: collecting point cloud data and noise attenuation data of a single tree, and preprocessing the collected point cloud data; classifying the preprocessed point cloud data into three categories: leaves, branches, and understory vegetation; extracting vegetation structure parameters of the target point cloud from the classified point cloud data; wherein the collection points need to be able to collect the three-dimensional structural information of the single tree to be predicted from all directions; and simultaneously, the attenuation data of the single tree to environmental noise needs to be obtained as the prediction target. Preprocessing the collected point cloud data includes operations such as noise reduction and hole filling to make the point cloud data more complete and continuous. The preprocessed point cloud data is segmented into three categories: leaves, branches, and understory vegetation. Based on the local features of the point cloud, different vegetation parts are identified to prepare for subsequent extraction of structural parameters. Vegetation structure parameters are extracted from the segmented point cloud. For different parts of the point cloud, key parameters representing structural information, such as length, width, height, density, volume, and tilt angle, are measured and calculated.

[0010] Specifically, point cloud data is a collection of points representing the three-dimensional shape of a target object, with each point containing XYZ coordinate information. For vegetation point clouds, it also includes attribute data recording the spectral information of the vegetation. In this application, point cloud data of a single tree can be collected by: using a handheld lidar instrument to collect point cloud data around the target tree along a closed route to obtain point cloud data of the target area. During the collection process, noise data is collected simultaneously by setting up a noise source, placing sound level meters in front of and behind the tree, playing stable white noise at different sound pressure levels, and collecting sound pressure levels before and after attenuation at three different locations on the tree. Point cloud data of the target area is obtained through point cloud calculation. The point cloud data of the area is manually cropped to obtain the point cloud data of the target tree. Noise is weakened and filtered to enhance the usable three-dimensional vegetation information for subsequent extraction of vegetation structure parameters. The ordered eigenvalues ​​of the point cloud covariance matrix are calculated, and the point cloud is divided into areal, scattered, and linear distributions according to the distribution pattern of the eigenvalues, and further divided into three types: tree trunk, canopy, and understory vegetation. This study refines and quantifies the vegetation structure information of individual trees, including crown height, crown width, canopy leaf area index, average canopy leaf inclination angle, canopy volume, diameter at breast height (DBH), branch height, understory leaf area index, and vegetation volume. The calculation of these vegetation structure parameters is implemented using Python.

[0011] The correlation between extracted vegetation structure parameters and collected noise attenuation data is calculated, and the most significant influencing factor in the correlation is selected as the model variable. Specifically, data on various extracted vegetation structure parameters, such as canopy volume, branch density, and leaf area index, are collected. Simultaneously, noise attenuation test data for corresponding individual trees are collected. Correlation analysis, such as Pearson correlation coefficient, is used to analyze the degree of correlation between each parameter and noise attenuation. The top N vegetation structure parameters with the strongest correlation to noise attenuation are selected as model variables. For example, if the leaf area index has the highest correlation coefficient with noise attenuation, it is selected as one of the model variables. The selected model variables are used as inputs, and variables that are irrelevant or weakly correlated are removed to improve the model's predictive performance. The analysis is repeated until the combination of vegetation structure parameter variables most closely related to noise attenuation is found. In this application, length information, leaf area index, and leaf tilt angle are preferred as the most significant influencing factors.

[0012] A single-tree denoising prediction model is constructed based on the support vector machine regression algorithm. The obtained model variables are used as input, and the constructed single-tree denoising prediction model is trained with different kernel functions. The prediction accuracy of the single-tree denoising prediction model under different kernel functions is evaluated, and the optimal kernel function is selected for single-tree denoising prediction. Among them, the constraint calculation of the prediction model is transformed into dual calculation through Lagrange multiplication to obtain the optimal kernel function.

[0013] The training set is constructed by collecting and extracting vegetation structure parameters as sample features and corresponding noise attenuation data as sample labels. Penalty coefficients and error parameters are set to establish loss and prediction functions. The constrained optimization problem is transformed into a dual problem using the Lagrange multiplier method. The optimal kernel function is obtained by solving the dual problem and used in the prediction function. Iterative optimization yields the support vector machine (SVM) regression prediction model. Typical kernel functions, such as linear, radial basis function, and polynomial kernels, are selected. Models are trained using different kernel functions while maintaining consistency in other parameters. The prediction errors of different models are evaluated on the test set. The prediction performance under different kernel functions is compared, and the model with the smallest error is selected. The optimal kernel function and corresponding SVM regression model are obtained. The optimized SVM model is then used to predict noise attenuation for new samples.

[0014] Specifically, by using Lagrange multiplication, the constraint calculation of the prediction model is transformed into a dual calculation to obtain the optimal kernel function. This transforms the complex constraint optimization problem into a more easily solvable dual problem, reducing computational difficulty. The dual problem has a more generalized optimal solution, which is beneficial for obtaining the global optimum rather than falling into local optima. Solving the dual problem allows for the simultaneous determination of the support vector machine's weights and biases, simplifying the computation process. The dual problem is only related to the inner product of sample points, and a kernel function can be introduced to map the samples to a higher-dimensional space, improving the approximation ability of nonlinear problems. Solving the dual problem yields the optimal kernel function, improving the model's prediction accuracy. Utilizing kernel function techniques, simple inner products can be calculated in low-dimensional space, enabling complex calculations in high-dimensional space while reducing computational load.

[0015] Furthermore, the preprocessed point cloud data is divided into three categories: leaves, branches, and understory vegetation. The process includes the following steps: extracting scatter features, linear features, and areal features from the point cloud data; where scatter features represent points where the distance between points is less than a threshold D; linear features represent point cloud formation lengths greater than a threshold L; and areal features represent point cloud formation areas greater than a threshold S. Based on the scatter features, a set of points where the distance between points is less than the threshold D is extracted as the leaf point cloud; based on the linear features, a set of line segments with a length greater than the threshold L and a point cloud density greater than the threshold M is extracted as the branch point cloud; and based on the areal features, a set of areal points with an area greater than the threshold S and a radius less than the threshold R is extracted as the understory vegetation point cloud.

[0016] The threshold parameters should be determined based on the distribution of the point cloud data, and statistical analysis can be used to help determine them. For the distance threshold D in the scatter feature, a value at a location of abrupt density change can be selected according to the point cloud density distribution to distinguish clustered point sets. For the length threshold L in the linear feature, the lengths of all line segments can be counted, and a value at the boundary between long and short line segments can be selected. For the area threshold S in the areal feature, the area of ​​all extracted patches can be counted, and a value at the boundary between large and small areas can be selected. For the radius threshold R when extracting understory vegetation, for example, an empirical value of 50% of the canopy width can be used to exclude non-understory areas. Multiple threshold combinations can be set for experimentation, and the combination that can better distinguish point clouds from different locations can be selected as the final threshold. Alternatively, an adaptive threshold setting method can be used to dynamically determine the threshold based on the point cloud distribution.

[0017] Furthermore, vegetation structure parameters include: length information, representing the spatial extent of vegetation in the vertical and horizontal directions, including crown height Hc, crown width Wc, diameter at breast height (DBH), and height below branches Ht; and leaf area index, representing the density of vegetation leaf cover, including the canopy leaf area index LAIc and the understory leaf area index LAI. L Leaf inclination angle, represented by the canopy average leaf inclination angle ALI; vegetation volume, representing the spatial extent of vegetation, including canopy volume Vc and understory three-dimensional green volume LVV. L .

[0018] Furthermore, the extraction of vegetation structure parameters also includes the following steps:

[0019] The target point cloud is divided into i×j×k individual elements;

[0020] i, j, and k satisfy the following formulas respectively:

[0021]

[0022]

[0023]

[0024] Among them, X min Represents the minimum coordinate value in the X direction; Y represents the minimum coordinate value in the X direction. min Represents the minimum coordinate value in the Y direction; Z represents the minimum coordinate value in the Z direction. min Δi represents the minimum coordinate value in the Z direction; Δi, Δj, and Δk represent the voxel sizes in the X, Y, and Z directions, respectively; (X, Y, Z) represents the target point cloud coordinates.

[0025] Further, calculating the length information includes the following steps: obtaining the maximum coordinate value in the Z direction of the target point cloud, as the highest point Zmax; obtaining the minimum coordinate value in the Z direction of the target point cloud, as the lowest point Zmin; calculating the coordinate difference between the highest point Zmax and the lowest point Zmin in the Z direction, as the crown height Hc; calculating the coordinate difference between the lowest point Zmin and the ground point in the Z direction, as the branch height H. t .

[0026] Further, calculating the leaf area index includes the following steps: Determine whether each voxel contains a point cloud; if so, label the corresponding voxel as M1; otherwise, label it as M2. Specifically, determine whether each voxel contains a point cloud; if it does, label it as 1; otherwise, label it as 0. Alternatively, consecutive label values ​​can be set, for example, labeling voxels containing point clouds as integers between 1 and 10, and uniformly labeling voxels not containing point clouds as 0. For each XOY plane, count the number of voxels labeled M1 (nM1(k)) and the number of voxels labeled M2 (nM0(k)); calculate the leaf area index within the XOY plane using the following formula:

[0027] Where k represents the index number of the XOY plane in the Z direction, nM1(z) represents the number of voxels marked as 1 in the k-th layer, nM0(z) represents the number of voxels marked as 0 in the k-th layer, z represents the height of the XOY plane; β represents a coefficient, where the value of β ranges from 0.9 to 1.3.

[0028] Furthermore, calculating the leaf tilt angle includes the following steps: setting the radius parameter ρ; selecting two points P. i and P j Given two points on the circle, draw a circle with radius ρ. The intersection of this circle and the point in the point cloud is determined as the third point P. k According to point P i P j and P k Define a triangle σ; calculate the normal vector n(x, y, z) of triangle σ; calculate the angle θ between the normal vector n(x, y, z) and the vertical upward unit vector (0, 0, 1), where the angle θ satisfies the following formula:

[0029]

[0030] in, This represents the normal vector of triangle σ. Let θ represent the unit vector in the Z direction; repeat the above process to calculate the included angle θ on multiple triangular faces; take the average of all included angles θ, which is the average tilt angle ALI of the blade.

[0031] Further, calculating the vegetation volume includes the following steps: uniformly dividing the point cloud data into n layers with an inter-layer distance of h; for each layer i, projecting the point cloud data of the i-th layer onto the XOY plane to generate a binary image; if the binary image contains point clouds, it is labeled as F1, otherwise it is labeled as F2; ​​counting the number Si labeled as F1 in the binary image of the i-th layer; the volume of the i-th layer is calculated using the following formula:

[0032] Vi = S i ×h

[0033] Among them, V i Let V represent the volume of the i-th layer; repeat the above steps to calculate the volume V of each layer. i By summing the volumes of all layers, we can obtain the vegetation volume of the entire point cloud.

[0034]

[0035] Where n represents the number of segmentation layers, V i Let F1 represent the volume of the i-th layer, and V represent the total vegetation volume. Specifically, the settings for F1 and F2 are the same as those for M1 and M2, and will not be repeated here.

[0036] Furthermore, a single-tree denoising prediction model is constructed based on the support vector machine regression algorithm, including the following steps: constructing a training set (x1, y1), (x2, y2), ..., (x n ,y n ), where x i Let y represent the vegetation structure parameters of the i-th training sample. i This represents the corresponding noise attenuation, where n is the number of training samples; specifically, it involves collecting point cloud data of individual trees and extracting vegetation structure parameters x. i This includes canopy volume, branch density, and leaf area index. For each point cloud sample, noise attenuation is measured to obtain the corresponding attenuation amount y. i The extracted parameter x i and attenuation y i One-to-one correspondence, forming training samples (x) i ,y i Repeat the above process to collect n point cloud samples and their attenuation values. If the number of vegetation structure parameters is consistent across the n samples, an n×m dimensional matrix X representing the parameter data and an n×1 dimensional vector Y representing the attenuation value are obtained. (X, Y) is used as the training set, where each row represents a sample and m columns represent parameters x. i and 1 column attenuation y i The size of the training set, n, is determined by the number of samples collected, typically ranging from several hundred to several thousand, to ensure effective learning of the relationships between parameters. The training set is used to train the machine learning model, i.e., to learn x. i and y iThe mapping relationship between them is used to construct an attenuation prediction model.

[0037] Construct the prediction function f(x):

[0038]

[0039] Where w represents the weight vector with a length of m, the same as the number of sample features m. b is the bias, which is a constant. Let f(x) represent the kernel function that maps the input sample x to a high-dimensional feature space. Let f(x) represent the function that predicts the input sample x, and its form is: in Represents the relationship between the weight vector w and the mapped sample The inner product. The weights w and bias b are obtained by solving the dual problem. Let f(x) be the optimal kernel function obtained during the solution process. f(x) represents the prediction model obtained through training. For a new input sample x, it is first mapped to a high-dimensional space, then multiplied by the weights and a bias is added to obtain the prediction output. The prediction output represents the noise attenuation of sample x.

[0040] Set the error variable ξ i and The error ε of the prediction function f(x) on the training set should satisfy the following formula:

[0041]

[0042] The optimal kernel function is calculated using the Lagrange multiplication method, and the optimal kernel function is calculated using the following formula:

[0043]

[0044] α i and Lagrange multipliers, K(x) i ,x j ) is the kernel function:

[0045]

[0046] Where, x i x represents the vegetation structure parameters of the i-th training sample; j Let α represent the vegetation structure parameters of the j-th training sample; the values ​​of i and j both range from 1 to n. Specifically, α i and α j Let x represent the Lagrange multipliers, which can be obtained by solving the dual problem. i and x j Let K(x) represent the vegetation structure parameters in the training samples, and let K(x) represent the features of the i-th and j-th samples, respectively. i x jK(x) represents the kernel function, which takes the features of two samples as input and outputs the kernel matrix. The kernel function value K(x) is calculated for each pair of samples by enumerating the training samples. i x j For each pair of kernel function values, a weighted sum is performed, with the weights being the corresponding Lagrange multipliers α. i and α j The summation result is the optimal kernel function, representing the optimal kernel mapping relationship between samples. Substituting the optimal kernel function into the prediction function can improve the prediction performance. Different kernel functions can yield different optimal kernels, thus allowing for the selection of different function space mappings. By using the Lagrange multiplier weighted kernel matrix method, the kernel function that best matches the training sample distribution was obtained, thereby improving the final prediction performance.

[0047] Kernel functions include linear kernels, polynomial kernels, sigmoid kernels, and radial basis function kernels. Specifically, the linear kernel is the simplest kernel function; it performs a linear mapping, mapping the input data to a higher-dimensional space. In higher-dimensional spaces, linear kernels are typically used to handle linearly separable data. The polynomial kernel performs a polynomial mapping, mapping the input data to an even higher-dimensional space. This can help handle nonlinear relationships. The parameters of the polynomial kernel include the degree of the polynomial and a constant term. The sigmoid kernel performs a sigmoid function mapping, mapping the data to the range [-1, 1]. This kernel function is commonly used in neural networks and logistic regression. The radial basis function kernel, also known as the Gaussian kernel, is one of the most commonly used kernel functions. It maps the input data to an infinite-dimensional space and performs well for nonlinear problems.

[0048] Another aspect of the embodiments of this specification provides a single-tree noise reduction prediction system based on a three-dimensional vegetation structure, used to execute the single-tree noise reduction prediction method based on a three-dimensional vegetation structure of this application. The system adopts a B / S architecture and includes modules such as a data acquisition terminal, a server, and a client. The data acquisition terminal includes sensors such as LiDAR and multispectral cameras for acquiring single-tree point clouds, images, and other data. Multi-view data is acquired using mobile platforms such as drones. The server side has modules for point cloud processing, feature extraction, and machine learning, realizing functions such as point cloud classification, vegetation parameter calculation, and noise attenuation model training. It also includes a Web service interface. The client side develops Web and mobile apps to connect to the server and realize functions such as parameter configuration, model prediction, and result display.

[0049] 3. Beneficial effects

[0050] Compared to existing technologies, the advantages of this application are:

[0051] (1) By collecting and analyzing point cloud data of individual trees and extracting key vegetation structure parameters such as spatial extent, green density, and leaf tilt angle, this method can more comprehensively and meticulously describe the characteristics of individual trees in the city. This more accurate data representation helps to construct a support vector machine regression model, thereby achieving quantitative prediction of the noise attenuation effect of individual trees in the urban environment. Compared with traditional methods, this feature extraction and modeling method based on the three-dimensional structure of vegetation can improve the accuracy and precision of prediction;

[0052] (2) By analyzing the point cloud of individual trees, specific vegetation structure parameters such as spatial range, green density, and leaf tilt angle are extracted. These parameters can more comprehensively describe the characteristics of individual trees in the urban environment, providing a more detailed and comprehensive data foundation for predicting noise attenuation effects;

[0053] (3) By constructing a support vector machine regression model and utilizing the extracted key vegetation structure parameters, this method achieves a quantitative prediction of the noise attenuation effect of individual trees in urban environments. This prediction can help urban planners, environmental managers, or researchers better understand and evaluate the role of individual trees in noise attenuation, thereby providing a more accurate reference for urban design and environmental protection.

[0054] In summary, by comprehensively utilizing advanced point cloud processing technology, comprehensive vegetation structure parameter extraction, and optimized support vector machine regression models, the quantitative prediction accuracy of the noise reduction effect of urban trees is significantly improved. Attached Figure Description

[0055] Figure 1 This is an exemplary flowchart of a single-tree noise reduction prediction method based on a three-dimensional vegetation structure according to an embodiment of this application;

[0056] Figure 2 This is a schematic diagram of point cloud acquisition of an individual tree in one embodiment of this application;

[0057] Figure 3 This is a schematic diagram showing the positional relationship between the noise source and the sound level meter monitoring point in one embodiment of this application;

[0058] Figure 4 This is a schematic diagram of single-tree noise sampling in three directions in one embodiment of this application;

[0059] Figure 5 This is a schematic diagram of point cloud classification results in one embodiment of this application;

[0060] Figure 6 This refers to the overall noise control capability of all samples in one embodiment of this application;

[0061] Figure 7 This application describes the noise control capabilities of different tree species in one embodiment. Detailed Implementation

[0062] The present application will now be described in detail with reference to the accompanying drawings and specific implementation examples.

[0063] Figure 1 This is an exemplary flowchart of a single-tree noise reduction prediction method based on a three-dimensional vegetation structure according to an embodiment of this application, such as... Figure 1 As shown, a single-tree noise reduction prediction method based on three-dimensional vegetation structure includes the following steps: data measurement and processing, including data acquisition, point cloud data preprocessing and classification, and vegetation structure parameter inversion; evaluation and analysis, including evaluation of single-tree noise reduction potential, analysis of the influence mechanism of vegetation structure on noise attenuation, and extraction of significant vegetation structure parameter indicators; and model construction, including selection of model kernel function and construction of a single-tree noise reduction model based on three-dimensional structural information. Specifically, data acquisition includes point cloud data acquisition and noise measurement data acquisition. Point cloud data acquisition uses a handheld lidar instrument to collect data around the target single tree along a closed path to obtain point cloud data of the target area.

[0064] Figure 2 This is a schematic diagram of point cloud acquisition for an individual tree in one embodiment of this application, as shown below. Figure 2 As shown, sensing devices such as LiDAR are set up around a single target tree, and data is collected using a handheld LiDAR. Canopy point cloud data is acquired from multiple perspectives, with the acquisition range of each perspective exceeding the canopy projection range. The point clouds from multiple perspectives are merged to construct the 3D point cloud data of the target tree. The point cloud data is preprocessed, including coordinate transformation, registration, filtering, and downsampling. Finally, high-density 3D point cloud data of the entire target tree is obtained, providing a data foundation for subsequent extraction of vegetation structure parameters.

[0065] Figure 3 This is a schematic diagram showing the location relationship between the noise source and the sound level meter monitoring point in one embodiment of this application. Figure 4 This is a schematic diagram of single-tree noise sampling in three directions in one embodiment of this application; the noise measurement data acquisition process is as follows: Figure 3 and Figure 4 As shown: First, a noise source was set up 1.5 meters above the ground at a height equal to the tree's trunk height. Then, sound level meters were set up at the height of the tree's centroid, one in front of and one behind the tree. Stable white noise of 55 dB (simulating the European noise threshold), 60 dB (simulating the Chinese noise threshold), and 68 dB (simulating the average traffic noise level in China) was played from the sound source. Sound pressure levels before and after attenuation were collected at three different locations on the target tree. Based on the principle that there were no other objects within 2 meters of the trees, 26 trees were selected as the research subjects, and a total of 468 sets of data were collected.

[0066] Point cloud data preprocessing specifically includes raw data processing, point cloud cropping, noise reduction, and ground point normalization. Point cloud data for the target area is obtained through point cloud processing; the area's point cloud data is manually cropped to obtain point cloud data for the target individual tree; then noise is reduced and filtered to enhance usable 3D vegetation information for subsequent vegetation structure parameter extraction.

[0067] Figure 5 This is a schematic diagram of point cloud classification results in one embodiment of this application, as shown below. Figure 5 As shown, point cloud classification involves calculating the ordered eigenvalues ​​(λ0, λ1, λ2) of the point cloud covariance matrix, classifying the point cloud into three distribution patterns: planar distribution—eigenvalues ​​in two directions are much larger than those in the third direction (λ0 ≈ λ1 >> λ2); scattered distribution—eigenvalues ​​in all three directions are similar (λ0 ≈ λ1 ≈ λ2); and linear distribution—eigenvalues ​​in one direction are larger than those in the other two directions (λ0 ≈ λ1 << λ2). The point clouds of individual trees are then divided into three types: trunk, canopy, and understory vegetation, serving as the data basis for extracting vegetation structure parameters.

[0068] Vegetation structure parameter inversion, finely quantifying the vegetation structure information of individual trees, including: crown height (H) c Crown width (W) c ), canopy leaf area index (LAI) c ), canopy mean leaf inclination (ALI), canopy volume (V) c ), diameter at breast height (DBH), height below branch (H) t Leaf area index (LAI) of understory vegetation L ) and vegetation volume (LVV) L As shown in Table 1, the above vegetation structure parameters were calculated using Python. Specifically, the canopy height of the sampled individual trees ranged from 3.02m to 9.67m, and the canopy width ranged from 3.37m to 7.29m. The canopy leaf area index ranged from 0.47 to 4.57, and the average leaf tilt angle was approximately 90° (±10°). The maximum canopy volume was 64.67m³. 3 The minimum value is only 2.82m 3 The diameter at breast height (DBH) ranges from 0.11m to 0.39m. The understory vegetation volume ranges from 0 to 2.83m³. 3 Between these ranges, the leaf area index ranged from 0 to 2.58. The vegetation structure information of individual trees is shown in Table 1.

[0069] Table 1. Vegetation structure information of individual trees

[0070]

[0071] Figure 6 To demonstrate the overall noise control capability of all samples in one embodiment of this application, such as... Figure 6As shown, the noise reduction potential of individual trees was evaluated. According to the results of the single-tree noise monitoring experiment, the average noise reduction capabilities of the 26 trees for sound sources of 55dB, 60dB, and 68dB were 6.9dB, 7.6dB, and 7.9dB, respectively.

[0072] Figure 7 To illustrate the noise control capabilities of different tree species in one embodiment of this application, such as... Figure 7 As shown, different tree species exhibit slightly different noise reduction capabilities. The noise reduction capabilities of pine (Pinus thunbergii), hackberry (Celtis sinensis), camphor tree (Cinnamomum camphora), horse chestnut (Aesculus chinensis), cassia (Osmanthus fragrans), and other species are approximately 7.2±0.5dB, 6.1±0.6dB, 6.9±0.4dB, 7.1±1.0dB, 9.0±1.0dB, and 6.3±0.5dB, respectively. Furthermore, individual trees can effectively attenuate noise above 31.5Hz, and their attenuation capability is directly proportional to the sound wave frequency. Specifically, the noise attenuation pattern peaks between 125Hz and 250Hz, then decreases slightly at 500Hz, and subsequently gradually increases in the frequency band from 500Hz to 16000Hz.

[0073] Correlation analysis was conducted using multiple linear regression to explore the relationship between vegetation structure parameters and the noise reduction capacity of individual trees. The dependent variable was noise attenuation (ΔL). Aeq The independent variables are vegetation structure parameters, including crown height (H). c Crown width (W) c ), canopy leaf area index (LAI) c ), canopy mean leaf inclination (ALI), canopy volume (V) c ), diameter at breast height (DBH), height below branch (H) t ), three-dimensional green volume of forest understory vegetation (LVV) L ) and leaf area index of understory vegetation (LAI) L Analysis results show that crown height (H) c ), canopy leaf area index (LAI) c ), canopy mean leaf tilt angle (ALI) and understory leaf area index (LAI) L The leaf area index (LAI) was significantly correlated with noise reduction capability. c The canopy height (H) has the most significant positive impact on noise reduction ability (β = 0.332, p < 0.01). Secondly, the canopy mean leaf inclination angle (ALI) also has a positive effect on the tree's noise modulation ability (β = 0.168, p < 0.01). Conversely, canopy height (H) has a more significant positive impact. c(β=-0.295, p<0.01) and leaf area index of understory vegetation (LAI) L (β=-0.438, p<0.01) showed a negative correlation with noise reduction ability. The results indicate that vegetation structure parameters affect the noise reduction ability of trees, and highlight the importance of canopy leaf area index (LAI). c The importance of the noise reduction capability is shown in Table 2 for the multiple linear regression analysis results of the embodiments of this application.

[0074] Table 2 Results of Multiple Linear Regression Analysis

[0075]

[0076] Kernel function selection to extract significant vegetation structure parameters (crown height (H)). c ), canopy leaf area index (LAI) c ), canopy mean leaf tilt angle (ALI) and understory leaf area index (LAI) L A Support Vector Regression (SVR) model was constructed, and the confidence scores of the prediction results from linear (LN), sigmoid (SIG), polynomial (PL), and radial basis functions (RBF) were compared to determine the most suitable kernel function for this model. As shown in Table 3, the confidence scores of SVR-LN, SVR-SIG, SVR-PL, and SVR-RBF were 77.2%, 44.7%, 15.4%, and 58.3%, respectively. Among them, SVR-LN had the highest confidence score, corresponding to the highest prediction accuracy. Therefore, in this model, the linear kernel function was determined to be the most suitable kernel function. The confidence score calculation results of the SVR kernel function in the specific embodiments of this application are detailed in Table 3.

[0077] Table 3 Confidence of SVR kernel function

[0078]

[0079]

[0080] The prediction model was constructed by extracting significant vegetation structure parameters and building a single-tree noise reduction model (SVR-LN) coupled with a linear kernel function, as shown in the following equation. To verify the universality of the SVR-LN model constructed in this application, five common greenbelt trees were re-collected as the test set, specifically including Magnolia grandflora Linn., Acer palmatum Thunb., Osmanthus fragrans., Prunus subg Cerasus sp., and Cinnamomum camphora. As shown in Table 4, the model has good predictive ability (R²). 2 =0.74, RMSE=0.38, MSE=0.15), proving the applicability and accuracy of this model.

[0081] Table 4. Measured and predicted values ​​of vegetation structure information and noise reduction capability of 5 individual trees (test set).

[0082]

[0083] This application proposes a methodology for evaluating and predicting the noise control capacity of independent trees: fine-scale inversion of individual tree structural parameters is conducted using LiDAR technology, and adaptive noise monitoring, evaluation, and prediction are performed based on tree specifications. Multiple linear regression is used to clarify the correlation between vegetation structure information and noise attenuation, and significant vegetation structure parameter indicators are extracted. The kernel function with the highest prediction accuracy is selected, and a single-tree noise reduction prediction model coupled with a linear kernel function (SVR-LN) is constructed. Starting from the smallest scale of urban green infrastructure, this application precisely quantifies the impact of tree structural characteristics on their noise blocking capacity, which is of great significance for urban planning departments in formulating and implementing strategies for optimizing the urban sound landscape through green infrastructure. Simultaneously, the constructed single-tree noise reduction prediction model compensates for the neglect of the influence of vegetation communities in current noise prediction models, thereby improving the accuracy of urban noise mapping. Based on this, this application has practical application value for urban planning and noise management, and provides guidance for the design and management of green facilities.

Claims

1. A method for single-tree noise reduction prediction based on three-dimensional vegetation structure, comprising: Collect point cloud data and noise attenuation data of individual trees, and preprocess the collected point cloud data; The preprocessed point cloud data is divided into three categories: leaves, branches, and understory vegetation. Extract vegetation structure parameters from the classified point cloud data; The correlation between the extracted vegetation structure parameters and the collected noise attenuation data was calculated, and the most significant influencing factor in the correlation was obtained as the model variable. A single-tree noise reduction prediction model is constructed based on the support vector machine regression algorithm. The obtained model variables are used as input, and the single-tree noise reduction prediction model is trained using different kernel functions. The prediction accuracy of single-tree denoising prediction models under different kernel functions is evaluated, and the optimal kernel function is selected for single-tree denoising prediction. The constraint calculation of the prediction model is transformed into a dual calculation by using Lagrange multiplication to obtain the optimal kernel function.

2. The single-tree noise reduction prediction method based on three-dimensional vegetation structure according to claim 1, characterized in that: The preprocessed point cloud data is divided into three categories: leaves, branches, and understory vegetation. The following steps are also included: Scattered features, linear features, and area features are extracted from the point cloud data respectively. Scattered features represent the feature where the distance between points is less than a threshold D; linear features represent the feature where the length of the point cloud formation is greater than a threshold L; and area features represent the feature where the area of ​​the point cloud formation is greater than a threshold S. Based on the scattered point characteristics, the set of points whose distance between points is less than the threshold D is extracted as the leaf point cloud; Based on linear features, extract line segment point sets with lengths greater than threshold L and point cloud densities greater than threshold M as branch point clouds. Based on the areal features, areal point sets with an area greater than the threshold S and a radius less than the threshold R are extracted as forest understory point clouds.

3. The single-tree noise reduction prediction method based on three-dimensional vegetation structure according to claim 1, characterized in that: Vegetation structure parameters include: Length information, representing the spatial extent of vegetation in the vertical and horizontal directions, including crown height H. c Crown width W c , diameter at breast height DBH and branch height H t ; Leaf area index, representing the density of vegetation leaf cover, includes the canopy leaf area index (LAI). c Leaf Area Index (LAI) of Understory Vegetation L ; Leaf tilt angle, expressed as the canopy average leaf tilt angle ALI; Vegetation volume represents the spatial extent occupied by vegetation, including canopy volume V. c Three-dimensional green volume (LVV) of understory vegetation L .

4. The single-tree noise reduction prediction method based on the three-dimensional structure of vegetation according to claim 3, characterized in that: Extracting vegetation structure parameters also includes the following steps: The target point cloud is divided into i×j×k individual elements; i, j, and k satisfy the following formulas respectively: X min Represents the minimum coordinate value in the X direction; Y represents the minimum coordinate value in the X direction. min Represents the minimum coordinate value in the Y direction; Z represents the minimum coordinate value in the Z direction. min Δi represents the minimum coordinate value in the Z direction; Δi, Δj, and Δk represent the voxel sizes in the X, Y, and Z directions, respectively; (X, Y, Z) represents the target point cloud coordinates.

5. The single-tree noise reduction prediction method based on three-dimensional vegetation structure according to claim 4, characterized in that: Calculating the length information involves the following steps: Obtain the maximum coordinate value in the Z direction in the target point cloud, and use it as the highest point Z. max ; Obtain the minimum coordinate value in the Z direction in the target point cloud, and take it as the lowest point Z. min ; Calculate the highest point Z max and the lowest point Z min The coordinate difference in the Z direction is taken as the crown height H. c ; Calculate the minimum point Z min The difference between the coordinates of the point and the ground in the Z direction is used as the height H below the branch. t .

6. The single-tree noise reduction prediction method based on three-dimensional vegetation structure according to claim 4, characterized in that: Calculating the leaf area index involves the following steps: Determine whether the segmented voxels contain point clouds. If they do, mark the corresponding voxel as M1; otherwise, mark the corresponding voxel as M2. For each XOY plane, count the number of voxels labeled M1 (nM1(k)) and the number of voxels labeled M2 (nM0(k)); The leaf area index in the XOY plane is calculated using the following formula: k represents the index number of the XOY plane in the Z direction, nM1(z) represents the number of voxels labeled M1 in the k-th layer, nM0(z) represents the number of voxels labeled M2 in the k-th layer, z represents the height of the XOY plane, and β represents the coefficient.

7. The single-tree noise reduction prediction method based on three-dimensional vegetation structure according to claim 4, characterized in that: Calculating the blade tilt angle involves the following steps: Set the radius parameter ρ; Select two points P from the target point cloud data. i and P j Assuming two points on the circle, draw a circle with radius ρ. The intersection of this circle with the point cloud is determined as the third point P. k ; According to point P i P j and P k Define a triangle σ; Calculate the normal vector n(x,y,z) of triangle σ; Calculate the angle θ between the normal vector n(x,y,z) and the vertically upward unit vector (0,0,1). The angle θ satisfies the following formula: This represents the normal vector of triangle σ. Represents the unit vector in the Z direction; Repeat all the steps to calculate the included angle θ on the multiple triangular faces; The average angle θ of all the included angles is taken as the average tilt angle ALI of the blade.

8. The single-tree noise reduction prediction method based on three-dimensional vegetation structure according to claim 4, characterized in that: Calculating vegetation volume involves the following steps: The point cloud data is uniformly divided into n layers, with a distance of h between the layers; For each layer i, the point cloud data of the i-th layer is projected onto the XOY plane to generate a binary image; If the binary image contains a point cloud, it is labeled as F1; otherwise, it is labeled as F2. Count the number S labeled F1 in the i-th layer binary image. i ; The volume of the i-th layer is calculated using the following formula: You=S i ×h Among them, V i Represents the volume of the i-th layer; Repeat all steps to calculate the volume V of each layer. i ; Summing up the volumes of all layers yields the vegetation volume of the entire point cloud: n represents the number of segmentation levels, V i Let V represent the volume of the i-th layer, and let V represent the total volume of the vegetation.

9. The single-tree noise reduction prediction method based on three-dimensional vegetation structure according to claim 1, characterized in that: A single-tree noise reduction prediction model is constructed based on the support vector machine regression algorithm, including the following steps: Construct a training set (x1, y1), (x2, y2), ..., (x n ,y n ), where x i Let y represent the vegetation structure parameters of the i-th training sample. i This represents the corresponding noise attenuation amount, where n is the number of training samples; Construct the prediction function f(x): w represents the weight vector, and b represents the bias. Represents the kernel function; Set the error variable ξ i and The error ε of the prediction function f(x) on the training set should satisfy the following formula: The optimal kernel function is calculated using the Lagrange multiplication method, and the optimal kernel function is calculated using the following formula: α i and Lagrange multipliers, K(x) i ,x j ) is the kernel function: x i x represents the vegetation structure parameters of the i-th training sample; j Represents the vegetation structure parameters of the j-th training sample; the values ​​of i and j are both from 1 to n; Kernel functions include linear kernels, polynomial kernels, sigmoid kernels, and radial basis kernels.

10. A system based on the single-tree noise reduction prediction method based on the three-dimensional structure of vegetation as described in any one of claims 1 to 9.

Images