Artificial forest single wood volume estimation method based on unmanned aerial vehicle laser radar

Abstract

The invention discloses a man-made forest single wood volume estimation method based on an unmanned aerial vehicle laser radar, and belongs to the technical field of forestry remote sensing and forest resource investigation. The method comprises the following steps: improving data quality through normalization, denoising and thinning processing; fitting a cylinder by using a least square method through the height distribution characteristics of the point cloud to extract a trunk point cloud; an improved mean shift algorithm is combined with adaptive bandwidth adjustment and an abnormal point detection mechanism to perform hierarchical clustering on tree trunk point clouds, a perfect circle is fitted by using a convex envelope algorithm to calculate a sectional area, and a single wood volume is accumulated and estimated by combining a cylinder and cone volume formula. The method solves the problem that a traditional algorithm is difficult in trunk section feature recognition, improves trunk parameter extraction precision and timber volume estimation reliability in a complex environment, supports trunk diameter extraction at any height, is suitable for subtropical man-made forests with different diameters at breast height and canopy density, and has a wide application prospect. And an efficient and accurate technical scheme is provided for forest resource investigation and carbon reserve evaluation.

Description

Technical Field

[0001] This invention belongs to the field of forestry remote sensing and forest resource survey technology, specifically involving a method for estimating the volume of individual timber in plantations based on UAV lidar, which is applicable to high-precision volume estimation of plantations such as eucalyptus and fir, as well as forest carbon storage assessment. Background Technology

[0002] Traditional forest resource surveys rely on manual measurements, which are inefficient and their accuracy is limited by the size of the sample plots. While Unmanned Aerial Vehicle Laser Scanning (UAV-LS) technology can efficiently acquire detailed information on the three-dimensional structure of forests, it has the following problems: 1. Limitations of existing algorithms: The point cloud data collected by UAV lidar has a denser canopy point cloud and a sparser trunk point cloud that is subject to noise interference. Traditional clustering algorithms (such as Kmeans and DBSCAN) are unable to accurately identify the trunk cross-sectional features, resulting in large errors in the estimation of diameter at breast height and volume. 2. Impact of data quality: Point cloud density and flight count directly affect the accuracy of trunk parameter extraction. Low-density point cloud or single-flight data are prone to estimation errors due to missing information. In addition, high-density point cloud data has a large data volume, which increases the difficulty of processing. 3. Complex environmental limitations: Tropical rainforests or complex multi-layered forests have high canopy closure, limiting laser penetration. Tree trunk point clouds are easily obscured by branches and leaves, and the canopy structure of different tree species varies significantly. Existing data processing methods are unable to effectively identify and segment individual trees, resulting in insufficient accuracy in extracting forest parameters such as tree height and diameter at breast height. Summary of the Invention

[0003] 1. Purpose of the invention This paper presents a method for estimating the volume of individual trees in plantations based on UAV lidar. By improving the mean drift algorithm and optimizing the point cloud acquisition density, the method solves the problem of chaotic trunk cross-section points, achieves the best balance between cost and accuracy in forest resource surveys, and improves the accuracy of trunk parameter extraction and the reliability of volume estimation in complex environments.

[0004] 2. Technical Solution A method for estimating the volume of individual timber in plantations based on UAV lidar includes the following steps: Step 1: Point cloud data preprocessing (1) Normalization: Ground points are classified using a progressive triangulation network (TIN), and a digital elevation model (DEM) is generated using inverse distance weighted (IDW) interpolation. Finally, the point cloud data is normalized using the generated DEM, with the following formula: ,in For normalized height, ... It is a point The absolute height. It is a point on the DEM The ground elevation of the location; (2) Noise reduction: Based on statistical methods, the points are calculated To neighboring areas Average distance of all points The calculation formula is: ,in, Point and its first Neighboring points The Euclidean distance between them; the noise threshold is determined by the median (meanD) and standard deviation (S). , among which, if Average distance of points Greater than the threshold Then it is believed If it is noise, then remove the noise point; (3) Thinning process: The point cloud is thinned from 100% to 10% using a resampling method with a gradient of 10%, as shown in the formula: ,in, This indicates the percentage of point cloud data that needs to be removed. This represents the target point cloud density, while It is the average density of the original point cloud; generate point cloud datasets with different densities, and analyze the impact of point cloud density on the accuracy of trunk parameter extraction and volume estimation.

[0005] Step 2: Trunk Extraction (1) Point cloud density distribution: Analysis of point cloud height distribution function and its normalization expression ,pass ,in, The height of the significant peak point is represented by the point cloud data below the crown-trunk transition zone, excluding ground points. The expression for the normalized point cloud height distribution function is: ,in, Indicates the height after normalization The percentage of point cloud density, and These represent the minimum and maximum heights of the point cloud in the sample plot, respectively; the point cloud of the tree trunk is extracted using the following formula: ,in, This represents a point in a point cloud. It is a point The height value; (2) Trunk morphology fitting: The point cloud data was processed using the least squares method, and a cylindrical fitting algorithm was applied to simulate the cylindrical model of the tree trunk. The objective function was: ,in, Point The distance to the surface of the cylinder. It is the centerline of the cylinder. It is the radius of the cylinder. It is the height of the cylinder.

[0006] Step 3: Trunk parameter extraction and volume estimation (1) Layering and partitioning: The tree trunk point cloud is layered at 2-meter height intervals, and the geometric center of the XY plane of each layer is calculated. The tree trunk tilt error is corrected by the normal vector. Then, the point cloud is assigned to different sectors according to the angle of each point relative to the center point. The formula for calculating the geometric center of each layer is as follows: ,in,( , () represents the XY coordinates of points in a point cloud grouped at the same height. It is the total number of points in this height group; angle The formulas for the sectors assigned to points are as follows: , ,in, This refers to the number of partitions selected; (2) Tree trunk edge point determination: The mean shift algorithm is used to effectively cluster the point cloud and identify the density centers representing the tree trunk edge by setting an appropriate Gaussian kernel function K and bandwidth h; the expression of the kernel function K is: ,in, It is a vector radiating outwards from the core point. It is the spatial dimension. It refers to bandwidth, used to control the width of the kernel function; the formula for calculating bandwidth h is: ,in, This represents the set of input data points. It is a quantile between 0 and 1, which determines the estimated bandwidth relative to the point cloud distribution; the highest density cluster centers typically represent the actual location of the tree trunk, and their expression is: ,in, Indicates the cluster center The density estimate at that location. It represents the total number of points in the point cloud. It is the first in point cloud One point, It's a kernel function. This refers to bandwidth. To improve the accuracy of clustering results, an outlier handling mechanism is introduced. When the Euclidean distance between a mean-drift cluster center and its nearest neighbor exceeds 1.5 times the median deviation, the outlier cluster center is removed. The expression for this mechanism is: ,in, It is the median density of all cluster centers. These are the average absolute deviations of these densities relative to the median; (3) Cross-section fitting: The Convex Hull algorithm is used to find all vertices that form the convex hull from the extracted edge points. They are sorted counterclockwise to form a convex polygon to fit the shape of the tree trunk cross-section. The least squares method is used to optimize the fitting of the perfect circle and calculate the cross-sectional area. The center of the circle is r is the radius of the fitted circle; the objective function is: ; (4) Trunk volume calculation: The trunk volume is obtained by summing the volume of the top and non-top portions; the main body of the trunk is calculated using the cylinder volume formula. Where h is the layer height, the treetop portion (<2m) is optimized using the cone volume formula, which is: ,in This represents the actual height of the treetop.

[0007] 3. Advantages of the present invention (1) Algorithm improvement: The improved mean shift algorithm improves the clustering accuracy of tree trunk point clouds in complex forest stands through adaptive bandwidth and outlier detection mechanism. Compared with traditional algorithms (Kmeans, DBSCAN), the determination coefficients R² of eucalyptus and Chinese fir volume estimation are increased to 0.929 and 0.927, respectively, and the absolute errors are reduced to 14.59% and 17.96%, respectively. (2) Multi-factor optimization: The influence of point cloud density and flight frequency on the accuracy of trunk parameter extraction was systematically analyzed. The optimal data acquisition strategy with point cloud density of not less than 50% and multiple flights (≥2 times) was proposed, which significantly improved data processing efficiency and estimation accuracy. (3) Strong adaptability: The algorithm supports the extraction of trunk diameter at any height. Especially at the standard breast height of 1.3 meters, the relative errors of eucalyptus and fir are controlled within 2.41% and -4.05% respectively, realizing high-precision estimation of trunk diameter at different heights. Attached Figure Description

[0008] Figure 1 Flowchart of a method for estimating the volume of a single log based on UAV lidar; Figure 2 Schematic diagram of point cloud height distribution and trunk extraction of eucalyptus and cedar trees; Figure 3Comparison of volume estimation accuracy between the improved mean drift algorithm and the traditional algorithm; Figure 4 : The effect of different point cloud densities on the number of sample wood tests and volume accuracy; Figure 5 : The effect of different flight sorties on the number of sample wood tests and volume accuracy. Detailed Implementation

[0009] The features of the present invention will be further illustrated below through examples, but the claims of the present invention are not limited in any way.

[0010] The study area was selected as the eucalyptus and Chinese fir plantations in Gaofeng Forest Farm, Nanning City, Guangxi Zhuang Autonomous Region. A scheme combining Ovi Map technology and Real-time Dynamic Positioning (RTK) system was adopted to conduct detailed spatial positioning and tree measurement of 126 trees in the eucalyptus plot and 106 trees in the Chinese fir plot. Tree height, diameter at breast height, slope and canopy closure of the plots were measured and recorded in the field. Point cloud data (average point cloud density of 448 points / m²) was obtained by DJI L2 UAV LiDAR. The flight altitude was 80 meters, and three flights were carried out. The flight path was a star pattern, the flight speed was 5 m / s, and the repeated scanning mode was adopted. The FOV was 70° horizontally and 15° vertically.

[0011] 1. Perform data preprocessing on the raw point cloud data according to the method described in step 1, including merging and classifying point cloud data from each flight, normalization, and noise reduction. =5) and dilution (100% to 10%, 10 groups in total).

[0012] 2. Determine the peak point of the crown-trunk transition zone according to the method described in step 2. View the point cloud density of eucalyptus and cedar in different height ranges using LiDAR360 software, then calculate the percentage of point cloud coverage for eucalyptus and cedar in different height ranges in Excel, and draw a schematic diagram of the point cloud height distribution and trunk extraction for eucalyptus and cedar samples. Figure 2 ).

[0013] 3. Following the methods described in steps 2 and 3, write code for an improved mean-shift method using Python. Export the point cloud coordinates (including X, Y, and Z) of the 126 eucalyptus trees and 106 cedar trees (after individual tree segmentation) in LiDAR360 software as .csv files. Each .csv file contains all the 3D point cloud coordinates of one tree. Then run the improved mean-shift method program, setting the input path to the .csv file containing the point cloud coordinates of the eucalyptus and cedar trees. This will yield the tree heights of the 126 eucalyptus trees and the 106 cedar trees, the tree diameters at 1.3m and 2m, and the volume of the individual tree trunks.

[0014] 4. Based on the measured data of the sample plots and the binary volume tables of eucalyptus and fir, the volume of each individual tree was obtained. The volume was then compared with the results estimated by the improved mean drift algorithm to evaluate the accuracy of the full-density, full-flight volume estimation and the influence of different point cloud densities and flight times on the accuracy of trunk parameter extraction. The results showed that: (1) Under the full density and full-flight conditions of point cloud data, the R² of the volume calculated by the improved mean drift algorithm for eucalyptus and fir was 0.929 and 0.927, respectively, with absolute errors of 14.58% and 17.96%, and relative errors of 4.95% and -1.78%, respectively. The study showed that the algorithm has good stability under different forest stand conditions and the model estimation accuracy is high. (2) The point cloud density of eucalyptus and fir had little impact on the accuracy of extraction progress before it was diluted to 50% of the original density. The increase in overall point cloud density was positively correlated with accuracy and the number of sample trees detected. However, when the point cloud density was below 50%, especially when it was only 10%, the accuracy dropped significantly, with the maximum absolute error reaching 86.44%. Studies have shown that excessive thinning can seriously affect the model's estimation ability, and the obtained parameter estimation data is no longer reliable. (3) The absolute error of the full-flight point cloud data is at least 5.43% higher than that of other different flight combinations, while the absolute error of the single-flight data is likely to be up to 34.2% higher than that of the full-flight point cloud data due to its limitations in capturing tree trunk morphology. This indicates that single-flight collection will miss some tree trunk information, leading to a decrease in the accuracy of model prediction. (4) Under the condition of full-flight point cloud data density at a height of 1.3 meters, the absolute errors of the improved mean-shift algorithm in estimating diameter at breast height are 12.14% and 14.19%, respectively, and the relative errors are 2.41% and -4.05%, respectively. This finding highlights the impact of height stratification on the accuracy of diameter at breast height estimation.

[0015] It should be noted that the embodiments described above are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art and related fields based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention. Modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and should all be covered within the scope of the claims of the present invention.

Claims

1. A method for estimating the volume of individual timber in plantations based on UAV lidar, characterized in that, Includes the following steps: (1) Point cloud data preprocessing: a. The progressive triangulation (TIN) method is used to classify ground points, and the inverse distance weighted interpolation (IDW) method is combined to generate a digital elevation model (DEM) to normalize the point cloud data of the sample plots; b. Based on statistical methods, noise points are identified and removed by calculating the average Euclidean distance and standard deviation of the spatial neighborhood of the point cloud. c. Use a resampling method to thin the original point cloud from 100% to 10% at a gradient of 10% to generate point cloud datasets of different densities; (2) Trunk extraction: a. Based on the point cloud density distribution characteristics, analyze the point cloud height distribution function P(h) and its normalized expression. Identify the transition region from the tree crown to the trunk, and extract the point cloud data below this region, excluding ground points, as the trunk point cloud; b. The point cloud data is processed by fitting a cylinder using the least squares method to obtain a cylindrical model that describes the geometry of the tree trunk; (3) Trunk parameter extraction and volume estimation: a. The preprocessed tree trunk point cloud is layered according to height intervals. For the geometric center of the point cloud in the XY plane of each layer, the tree trunk tilt error is corrected by the normal vector to determine the layer center point. Then, the point cloud is assigned to different sectors in each height layer according to the angle of each point relative to the center point. b. Mean drift clustering is used to process the point cloud within each angle division. An appropriate Gaussian kernel function and bandwidth are set to effectively cluster the point cloud and identify the density centers representing the tree trunk edge positions. c. The improved mean-shift algorithm introduces an adaptive clustering bandwidth adjustment mechanism based on point cloud density, which allows the algorithm to reduce bandwidth in high-density regions. Combined with a gradient-based outlier detection mechanism, when the Euclidean distance between the mean-shift cluster center and the nearest neighbor center exceeds 1.5 times the median deviation, the outlier cluster center is removed. d. Using the Convex Hull algorithm, all vertices forming the convex hull are found from the extracted edge points and sorted counterclockwise to form a convex polygon to fit the shape of the tree trunk cross-section. Then, the least squares method is used to optimize the fitting of a perfect circle, and the center of the circle is determined by optimizing the objective function. Given the radius r of the perfect circle, calculate the cross-sectional area of ​​the tree trunk and extract the diameter parameters at heights of 1.3 meters and 2 meters. e. Based on the trunk height, the volume of a single piece of wood is estimated using the cylindrical volume formula, while the volume of the treetop is optimized using the conical volume formula based on the measured height.

2. The method according to claim 1, characterized in that, The formula for the average Euclidean distance mentioned in step (1) is: ,in, Point and its first Neighboring points The Euclidean distance between them, where k represents the number of neighboring points.

3. The method according to claim 1, characterized in that, In the thinning process described in step (1), the formula for the point cloud density thinning ratio is as follows: ,in For the target point cloud density, Average density of the original point cloud.

4. The method according to claim 1, characterized in that, The formula for the point cloud height distribution function mentioned in step (2) is: ,in, Indicates the height after normalization The percentage of point cloud density, Represents the point cloud height distribution function. and Let represent the minimum and maximum heights of the point cloud in the sample plot, respectively; the height of the significant peak point in the transition area between the canopy and trunk is: ,in, The height represents the significant peak point, at which the rate of increase in point cloud quantity reaches its maximum, indicating the transition region from the canopy to the trunk; the extracted point cloud of the trunk portion is: ,in, This represents a point in a point cloud. It is a point The height value.

5. The method according to claim 1, characterized in that, The least squares fitting cylinder algorithm described in step (2) yields cylinder parameters such as centerline, radius, and height. The main structural feature expression of the tree trunk is: ,in, Point The distance to the surface of the cylinder. It is the centerline of the cylinder. It is the radius of the cylinder. It is the height of the cylinder.

6. The method according to claim 1, characterized in that, The height interval mentioned in step (3) is 2 meters, and the geometric center formula is: ,in( , ) represents the XY coordinates of the midpoint of the point cloud within the same height layer, and N represents the total number of point clouds within the layer; angle Represented as: The allocated sectors are represented as follows: ,in, This is the number of partitions selected.

7. The method according to claim 1, characterized in that, The adaptive Gaussian kernel function bandwidth adjustment mechanism described in step (3) is formulated as follows: ,in, This represents the set of input data points. It is a quantile between 0 and 1, which determines the estimated bandwidth relative to the point cloud distribution. The formula for the Gaussian kernel function K is: ,in, It is a vector radiating outwards from the core point. It is the spatial dimension. It is bandwidth, used to control the width of the kernel function.

8. The method according to claim 1, characterized in that, The expression for the gradient-based outlier detection mechanism described in step (3) is: Remove Abnormal cluster centers >1.5, among which, It is the median density of all cluster centers. These are the average absolute deviations of these densities relative to the median. It is the cluster density center, and the formula is: ,in, It represents the total number of points in the point cloud. It is the first in point cloud One point. It's a kernel function. Bandwidth determines the scope of the kernel function.

9. The method according to claim 1, characterized in that, The formula for the objective function mentioned in step (3) is: The cross-sectional area of ​​the tree trunk is: , where r is the radius of the least squares fitted circle.

10. The method according to claim 1, characterized in that, The formula for the volume of the cylinder mentioned in step (3) is: Where S is the cross-sectional area of ​​the trunk in each layer, h is the layer height, and the formula for optimizing the treetop volume is: ,in This represents the actual height of the treetop.

Images