Forestry investigation planning design and analysis method based on three-dimensional laser modeling

Abstract

The invention relates to the technical field of forestry management data processing, in particular to a forestry investigation planning design and analysis method based on three-dimensional laser modeling, and the method comprises the steps: collecting laser radar point cloud data, a multispectral image and a thermal infrared image of a target area through an unmanned plane, a ground mobile platform and satellite remote sensing; establishing corresponding analysis packets based on different analysis scenes, adaptively selecting a corresponding algorithm for each analysis packet to perform data compression and dimension reduction, and then fusing the data; for each analysis packet, constructing a corresponding three-dimensional model by using the fused data, and extracting feature parameters required by forestry investigation from the three-dimensional model; and the analysis packets are fused, the extracted characteristic parameters are analyzed and constructed, and a basis is provided for forestry planning design and analysis. According to the method, the limitation of poor adaptability in traditional forestry investigation is solved through a technical link of scene analysis packet construction, differentiation processing and cross-scale fusion.

Description

Technical Field

[0001] The present invention relates to the technical field related to forestry management data processing, and in particular to a forestry survey planning design and analysis method based on three-dimensional laser modeling. Background Art

[0002] The key problem facing the current forestry resource survey is the lack of flexibility and scenario adaptability of data processing methods. Traditional methods use fixed processes to handle all types of survey tasks, which can neither meet the needs of large-scale rapid surveys nor achieve fine analysis of individual trees. In the existing forestry resource survey, if the data integrity is retained, it will lead to inefficient processing of massive data in macro-monitoring scenarios. If the data is compressed forcibly, key details in the analysis of individual tree characteristics will be lost. The black-and-white processing mode of existing technologies forces forestry workers to compromise between efficiency and accuracy, which seriously restricts the practical value of survey results. The deeper problem lies in the fragmented management and analysis of multi-source data. Existing technologies often process data collected by different technical means such as lidar, satellite images, and thermal infrared independently, and lack the ability to perform cross-scale correlation analysis. For example, it is difficult to establish a quantitative relationship between terrain feature data and the growth status of individual trees, and the tree health information reflected by thermal infrared images cannot effectively assist in the overall planning and decision-making of forest areas. This data island phenomenon not only causes a waste of information resources, but also leads to the one-sidedness of comprehensive analysis conclusions.

[0003] Therefore, the existing technical system's excessive pursuit of processing speed will sacrifice data details, while blindly retaining complete data requires a large amount of computing resources. This contradiction often puts forestry survey work into a dilemma. The conclusions drawn quickly are not reliable enough, and the high-precision analysis results lose their application value due to lack of timeliness, which ultimately affects the scientificity and timeliness of planning, design and analysis. Summary of the invention

[0004] In view of the shortcomings of the existing technology, the present invention provides a forestry survey planning, design and analysis method based on three-dimensional laser modeling to solve the limitations of "single data, rigid algorithm and fragmented scene" in traditional forestry surveys.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: A forestry survey planning, design and analysis method based on three-dimensional laser modeling comprises the following steps: collecting laser radar point cloud data, multispectral images and thermal infrared images of a target area by means of unmanned aerial vehicles, ground mobile platforms and satellite remote sensing; establishing corresponding analysis packages based on different analysis scenarios, wherein the analysis packages include analysis packages for forest resource macro-monitoring scenarios and analysis packages for tree individual feature research scenarios, wherein each analysis package adaptively selects a corresponding algorithm to perform data compression and dimensionality reduction, and then fuses the laser radar point cloud data with the multispectral image and the thermal infrared image; for each analysis package, constructing a corresponding three-dimensional model using the fused data, and extracting characteristic parameters required for forestry survey from the three-dimensional model; and fusing the two analysis packages, analyzing and constructing the extracted characteristic parameters, and providing a basis for forestry planning, design and analysis.

[0006] Furthermore, for the analysis package of forest resource macro-monitoring scenarios, data compression is to jointly adjust the lidar point cloud data, multispectral images and thermal infrared images to eliminate redundant data; data dimensionality reduction is to calculate the correlation between the original multi-source data and the macro characteristics of the forest, set a threshold for the correlation coefficient, retain the relevant feature data above the threshold, and eliminate the relevant feature data below the threshold.

[0007] Furthermore, the data fusion of the analysis package for the macro-monitoring scenario of forest resources includes: The macro terrain features and vegetation features are extracted after the lidar point cloud is denoised, the pixel DN value of the multispectral image is converted into a radiation brightness value and the spectral features are calculated. The grayscale value of the thermal infrared image is converted into an actual temperature value and the energy distribution features are extracted. During fusion, the three types of data are preliminarily aligned according to the geographic coordinates, and then secondary matching is performed based on feature similarity. Weighted fusion is used to assign weights according to the importance of the data source, and the matched feature data are fused according to the assigned weights.

[0008] Furthermore, a three-dimensional model is constructed for the macro-monitoring and analysis package of forest resources, including: The macro terrain features and vegetation features extracted from the lidar point cloud serve as the spatial skeleton. The model is given vegetation details and classification information through multispectral images. The model is given energy distribution through thermal infrared images to construct a three-dimensional model. The model surface is rendered using multispectral image texture and color information.

[0009] Furthermore, the characteristic parameters of the forest resource macro-monitoring analysis package are extracted, including: The forest coverage rate is obtained by calculating the ratio of the spatial volume occupied by the tree point cloud in the three-dimensional model to the spatial volume of the entire target area. The average tree height is obtained by obtaining the height data of the tree point cloud in the three-dimensional model and then calculating it by the weighted average method. The terrain slope is obtained by performing plane fitting analysis on the terrain point cloud in the three-dimensional model.

[0010] Furthermore, in the analysis package for the study scenario of individual tree characteristics, data compression is to calculate the covariance matrix of the lidar point cloud data, analyze the variance proportion of each dimension, select the dimension with the largest variance proportion as the core segmentation dimension, divide the interval according to the coordinate range of the core segmentation dimension, compress each interval independently, introduce distance weight to divide the nodes, cluster the point cloud in each interval based on the weight, identify the core points, retain the core points and their neighborhood points, and eliminate low-density noise points; data dimensionality reduction is to sample and split various features of the lidar point cloud data of a single tree, screen out key features, set thresholds, and retain features that exceed the threshold.

[0011] Furthermore, the analysis package data fusion for the study scenario of individual tree characteristics includes: The lidar point cloud data, multispectral images, and thermal infrared images of individual trees are denoised separately, and then the trunk centerline and position are determined by feature extraction of the lidar point cloud data of individual trees. The trunk diameter is then calculated, the crown layers are distinguished, and the branch features are obtained. Texture features are extracted from the multispectral images of individual trees. When matching, the tree outline is identified through the multispectral image, and then matched with the corresponding lidar point cloud through geometric position consistency. The thermal infrared and lidar data are then associated, and the geometric features of the lidar data, the multispectral texture and health index, and the temperature features of the thermal infrared are weightedly fused.

[0012] Furthermore, a three-dimensional model is constructed for the research and analysis package of individual tree characteristics, including: point cloud-based surface reconstruction based on the fused and denoised lidar point cloud data, multispectral images and thermal infrared image data, dynamic adjustment of the sphere radius parameter according to the local curvature of the tree point cloud, and texture mapping of the model surface in combination with the texture information of the multispectral image.

[0013] Furthermore, the characteristic parameters of the analysis package for individual tree characteristics were extracted, including: directly obtaining the trunk height by measuring the length of the trunk centerline in the three-dimensional model; calculating the crown shape index based on the three-dimensional shape of the crown; and obtaining the number of branch forks through statistical analysis of the topological analysis of the branch skeleton structure in the three-dimensional model.

[0014] Furthermore, the two analysis packages are integrated to analyze and construct the extracted feature parameters. Specifically, the parameters of the two analysis packages are associated through a unified geographic coordinate system, and the spatial database is used to establish an index. The feature weights are configured according to the scenario requirements, and a macro-micro parameter model is constructed to generate the corresponding scenario map, which is visualized through a three-dimensional model.

[0015] In another aspect, the present invention discloses an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned forestry survey planning, design and analysis method based on three-dimensional laser modeling.

[0016] The present invention provides a forestry survey planning, design and analysis method based on three-dimensional laser modeling, which has the following beneficial effects: Based on the technical route of "scenario-based analysis package construction-differentiation processing-cross-scale fusion", the present invention designs dynamically adaptive data processing strategies for two typical scenarios: macro-monitoring of forest resources and refined analysis of individual trees. In the macro-monitoring scenario, multi-source data joint adjustment and intelligent dimensionality reduction technology are used to significantly improve the efficiency of massive data processing while ensuring the accuracy of key statistical features. In the individual feature analysis scenario, vertical dimension segmentation and feature screening mechanisms are adopted to completely retain detailed structures such as crown hierarchy and branch topology, providing a high-fidelity data foundation for individual tree feature analysis.

[0017] Through the cross-scale data fusion mechanism, the present invention realizes the deep collaborative processing of LiDAR point cloud, multispectral image and thermal infrared data. Based on the spatial association and dynamic weight configuration of the unified geographic coordinate system, a quantitative mapping relationship between macro terrain parameters and individual tree characteristics is established. This global and local linkage analysis mode not only supports scientific decision-making in the overall planning of forest areas, but also realizes the precise positioning and health diagnosis of abnormal individual trees, significantly improving the reliability of resource assessment and disaster warning.

[0018] In addition, the present invention balances data accuracy and computing load through elastic architecture design, breaking through the passive choice of "efficiency for accuracy" or "accuracy for efficiency" in traditional methods. 3D laser modeling technology combined with multi-source data fusion achieves full process optimization from forest terrain reconstruction to single tree morphology analysis, provides adaptive decision support for dynamic monitoring of forest resources, and promotes the transformation of forestry management from experience-driven to data-intelligent driven. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of the forestry survey planning design and analysis method based on three-dimensional laser modeling of the present invention; Figure 2 This is a visualization diagram of a three-dimensional model in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] A forestry survey planning, design and analysis method based on 3D laser modeling, such as Figure 1 As shown, it includes the following steps: collecting laser radar point cloud data, multispectral images and thermal infrared images of the target area through unmanned aerial vehicles, ground mobile platforms and satellite remote sensing; in implementation, the unmanned aerial vehicle is equipped with GPS equipment, laser radar and optical camera. During the flight, the GPS equipment continuously collects the coordinates of the ground control points, the laser radar point cloud data and multispectral images of the target area, and transmits the laser radar point cloud data and multispectral images to the airborne data processing system in real time. The airborne data processing system monitors the collected data in real time; the ground mobile platform collects thermal infrared images and optical images of the target area according to the driving route; for complex terrain and drone no-fly zone data collection, the latest satellite remote sensing data is obtained to supplement. Specifically, the drone data collection is based on the preset forest area boundary (GIS vector data), generates a terrain-like flight route, dynamically adjusts the flight altitude according to the terrain undulation (5-10m from the canopy), and ensures that the laser radar point cloud density is ≥20 points / m². The laser radar and multispectral camera are synchronously collected through hardware triggering, and the GPS records POS data every 0.1 second, aligns it with the point cloud data timestamp, and transmits the laser radar point cloud and multispectral image back to the airborne processing system in real time. The ground platform is equipped with a FLIRT865 thermal imager and drives along the forest road, collecting thermal infrared images and optical images simultaneously. The satellite remote sensing data is dynamically supplemented by accessing the Sentinel-2 and WorldView-3 satellite data, automatically downloading the latest images according to the coordinates of the drone no-fly zone, and in steep slope areas (slope > 45°), integrating the DSM data generated by the satellite stereo image pair with the drone point cloud, and filling the drone blind area terrain through Poisson surface reconstruction.

[0022] Based on different analysis scenarios, corresponding analysis packages are established, including analysis packages for forest resource macro-monitoring scenarios and analysis packages for tree individual characteristics research scenarios. Each analysis package adaptively selects the corresponding algorithm to perform data compression and dimension reduction. In the implementation, for the analysis package for forest resource macro-monitoring scenarios, data compression is to perform joint adjustment processing on the lidar point cloud data, multispectral images and thermal infrared images. Specifically, for the analysis package for forest resource macro-monitoring scenarios, the lidar point cloud (including GPS / IMU positioning data), multispectral images (including POS data), and thermal infrared images (including timestamps) are first aligned through the UTC time axis; the GPS ground control point coordinates of the drone airborne system are used as a unified spatial reference to map the three types of data to the WGS84 coordinate system. The lidar point cloud and the multispectral image are matched through SIFT feature points (at least 500 pairs of points with the same name are extracted), and the affine transformation matrix is ​​calculated (error <1 pixel); the thermal infrared image is upgraded to the multispectral image resolution through bilinear interpolation, and then preliminarily aligned based on the geographic coordinate grid (10m×10m).

[0023] In data compression, a joint adjustment optimization model is first constructed, including: The joint adjustment objective function is defined as: Among them, L i ,S i ,T i are the observation values ​​of the i-th feature point of the lidar point cloud, multispectral image, and thermal infrared image respectively; , , is the optimized theoretical value; α, β, γ are weight distribution.

[0024] Then, the Levenberg-Marquardt algorithm is used to iteratively solve and optimize the spatial coordinates of the lidar point cloud, the radiance value of the multispectral image, and the temperature value of the thermal infrared image until the objective function converges (the residual decrease rate is <1e-5).

[0025] Then, the registered lidar point cloud is gridded, the median elevation point is retained in each grid, and redundant points are removed; the multispectral image is subjected to band correlation analysis to remove redundant bands, such as only retaining green light when the correlation between blue and green light bands is >0.9; the thermal infrared temperature data and the multispectral NDVI values ​​are regressed to remove the data dimensions of areas with insignificant temperature changes (slope <0.1); the lidar point cloud intensity values ​​are jointly clustered with the multispectral near-infrared bands to merge point cloud clusters with similar radiation characteristics.

[0026] In the embodiment, a large forest area is divided into 1km×1km sub-blocks, each block is independently subjected to joint adjustment processing, a 10m overlapping zone is set at the edge of the block, and the splicing gaps are eliminated by Kriging interpolation.

[0027] For the analysis package of forest resource macro-monitoring scenario, data dimension reduction is to calculate the correlation between the original multi-source data and the macro characteristics of the forest, set the threshold of the correlation coefficient, retain the relevant feature data above the threshold, and eliminate the relevant feature data below the threshold. Specifically, data dimension reduction is to first normalize the laser radar point cloud intensity value to [0,1], convert the multispectral image radiation brightness value into reflectance based on the radiation calibration parameters, standardize the thermal infrared temperature value into Z-score, use the processed data as input data, assign weights to the input data, and spatially grid the weighted data. The grid can be set to 10m×10m. The statistical features in each grid include the median elevation, point cloud density, and intensity variance features of the laser radar, the NDVI mean value and red edge band reflectivity features of the multispectral, and the temperature mean and day-night temperature difference features of the thermal infrared. Then, macro-feature label matching is performed to associate each grid with a preset macro-feature label (such as forest coverage and terrain slope). Specifically, the correlation coefficient between each input feature (such as NDVI mean and point cloud density) and the target macro-feature (such as coverage and slope) is calculated. The calculation formula is: r represents the correlation coefficient, which indicates the degree of linear correlation between two variables x and y. The value range of r is −1≤r≤1, where r=1 indicates a completely positive correlation, r=−1 indicates a completely negative correlation, and r=0 indicates no linear correlation. i ,y i represents the observation value of the i-th sample in the dataset. x is the input feature, including the median elevation, point cloud density, and intensity variance features of the lidar, the mean NDVI value of the multispectral spectrum, the red edge band reflectivity features, the mean temperature of the thermal infrared spectrum, and the day-night temperature difference features. i represents the i-th sample observation value of variable X (input feature), for example, in the embodiment, it may correspond to the i-th data of specific features such as the median elevation of the laser radar, the point cloud density, or the mean NDVI of the multi-spectral. y is the macroscopic feature of the target, including coverage and slope, then y i Represents the i-th sample observation value of variable Y (target macro-feature). For example, it corresponds to the i-th data of macro-features such as forest coverage and terrain slope. represents the sample mean of variable X, that is, all x i The average value of represents the sample mean of variable y, that is, all y i The average value of .

[0028] A threshold (|r|≥0.3) was set to retain strongly correlated features (such as NDVI and coverage r=0.65, point cloud density and slope r=0.41), and to remove weakly correlated features (such as thermal infrared day and night temperature difference and coverage r=0.12).

[0029] The analysis package for macro-monitoring scenarios of forest resources combines lidar point cloud data with multispectral images and thermal infrared images after data compression and dimensionality reduction. The lidar point cloud uses an outlier removal algorithm based on statistical analysis to remove noise points caused by measurement errors by calculating the average distance from the point to the K nearest neighbors and taking the mean plus 3 times the standard deviation as the threshold to extract macro-terrain features and vegetation features. Multispectral images convert pixel DN values ​​into radiation brightness values ​​based on camera radiation calibration parameters to calculate spectral features. Thermal infrared images convert grayscale values ​​into actual temperature values ​​to extract energy distribution features. When fusion is performed, the three types of data are first preliminarily aligned according to geographic coordinates, and then secondary matching is performed based on feature similarity. Weighted fusion is used to assign weights according to the importance of data sources.

[0030] In the implementation, for the noise points in the lidar point cloud, the average distance from the point to the K nearest neighbor points is calculated, and the K nearest neighbor outlier detection is used. K=50 is set to calculate the average distance d from each point to the nearest 50 neighbors. avg ; Statistical global average distance μ d With standard deviation σ d , set the threshold T = μ d +3σ d ; Remove all d avg >T noise points (such as flying birds, dust reflection points).

[0031] For multispectral image radiation correction, the pixel DN value is converted into radiation brightness L according to the camera radiation calibration parameters (gain g, offset b). The calculation formula is: L=g⋅DN+b; then the spectral features are extracted, including NDVI, SAVI, and REIP.

[0032] For thermal infrared temperature conversion, the grayscale value Gray is converted into the actual temperature T based on the sensor calibration parameters. The calculation formula is: T=k⋅Gray+c (k, c are calibration coefficients); then the energy distribution characteristics are extracted, including the temperature mean, variance, and day and night temperature difference (requires multiple data collection).

[0033] When performing data fusion, the lidar point cloud, multispectral image, and thermal infrared data are unified into the WGS84 coordinate system with a grid resolution of 10m×10m. The coordinates of the UAV GPS ground control points (±2cm accuracy) are used as a reference to correct the position deviation of the satellite data.

[0034] Secondary feature matching includes lidar and multispectral alignment, and thermal infrared and lidar alignment. The steps of lidar and multispectral alignment are: extract SIFT feature points within the grid (≥20 pairs per grid), calculate the affine transformation matrix, and the alignment error is <1 pixel; perform bilinear interpolation on the multispectral image to match the lidar point cloud resolution.

[0035] The steps for aligning thermal infrared and lidar are: based on the temperature-elevation correlation, non-rigid registration is performed by maximizing mutual information, and in steep slope areas (slope > 25°), the ICP algorithm (iterative closest point) is used to optimize the matching accuracy.

[0036] The weight allocation strategy for weighted fusion is as follows: LiDAR has the highest spatial accuracy and dominates the terrain and vegetation structure characteristics, so the LiDAR weight is 0.5; multispectral provides spectral classification information, but the resolution is low, so the multispectral weight is 0.3; thermal infrared reflects energy distribution, but is easily affected by weather interference, so the thermal infrared weight is 0.2.

[0037] Feature-level fusion generates a fusion feature vector for each grid (10m×10m). The feature vector includes the median elevation, slope, NDVI mean, temperature mean, point cloud density, and red edge reflectivity, which are calculated by weight: Eigenvalue = 0.5⋅LiDAR feature + 0.3⋅Multispectral feature + 0.2⋅Thermal infrared feature.

[0038] During implementation, dynamic weight adjustment can be performed. For example, if the quality of thermal infrared data is poor (such as cloud coverage > 50%), its weight can be reduced to 0.1 and the lidar weight can be increased to 0.6.

[0039] Construct a three-dimensional model for the macro-monitoring and analysis package of forest resources, including: Macro terrain and vegetation features are extracted from the lidar point cloud to serve as a spatial skeleton. The model is given vegetation details and classification information through multispectral images. Energy distribution is extracted through thermal infrared images. A three-dimensional model is constructed using the Poisson surface reconstruction algorithm. The model surface is rendered using the multispectral image texture and color information. The forest coverage rate is obtained by calculating the ratio of the spatial volume occupied by the tree point cloud in the three-dimensional model to the spatial volume of the entire target area. The height data of the tree point cloud in the three-dimensional model is obtained, and then the average tree height is calculated by the weighted average method. The terrain slope is obtained by plane fitting analysis of the terrain point cloud in the three-dimensional model.

[0040] The macro terrain features are extracted from the lidar points. The elevation model, slope and slope direction are calculated by the moving surface fitting method. In the specific implementation, the cloth simulation filter (CSF) algorithm is used to separate the ground points and non-ground points (vegetation, buildings). The cloth rigidity coefficient k=0.5 and the number of iterations n=100 are set to divide the point cloud into the ground layer and the non-ground layer. For steep slope areas (slope>25°), k=0.3 is adjusted to enhance the retention of terrain details. The ground point cloud is divided into 10m×10m grids, and the 20% points with the lowest elevation in each grid are retained as the reference ground points. The fitting window size is adaptively adjusted according to the complexity of the terrain. The window size is set to 5×5 grid (50m×50m) for flat areas (elevation variance <1m); the window size is set to 3×3 grid (30m×30m) for complex terrain (elevation variance ≥1m).

[0041] For each ground point in the window, the quadratic surface model is used to fit the terrain, and the calculation model is: z=ax 2 +by 2 +cxy+dx+ey+f; In the formula, a represents the quadratic curvature in the x direction, which reflects the curvature of the terrain on the x-axis; b represents the quadratic curvature in the y direction, which reflects the curvature of the terrain on the y-axis; c represents the cross curvature in the xy direction, which describes the distortion of the terrain on the xy plane; d represents the linear slope in the x direction, that is, the inclination rate of the terrain in the x-axis direction; e represents the linear slope in the y direction, that is, the inclination rate of the terrain in the y-axis direction; and f represents the reference elevation at the origin of the coordinate system.

[0042] The coefficients a, b, c, d, e, and f are solved by the least squares method; the residual threshold is set to 0.1m, and the points exceeding the threshold are considered as noise and removed. The elevation value of the fitted surface is used as the elevation of the grid center point to generate a digital surface model (DSM) with a resolution of 0.5m; for windows where fitting fails (such as no ground points), Kriging interpolation is used to complete the data from adjacent grids, and finally a DSM with a resolution of 0.5m is generated. The 3×3 Sobel operator is applied to the DSM to calculate the east-west gradient G x and the north-south gradient G y ; Then calculate the slope s. The calculation formula for the slope s is: ; The slope aspect calculation is to perform fast Fourier transform on DSM to extract the frequency domain characteristics of terrain; the calculation formula of slope aspect A (unit: degree, 0° is due north, increasing clockwise) is: ; For high-frequency noise areas (FFT amplitude > threshold), Gaussian filtering is used for smoothing and recalculation.

[0043] For extracting vegetation features from LiDAR points, based on the LiDAR point cloud after outlier removal, the CSF algorithm (k=0.5, n=100) was used again to separate ground and vegetation points, normalize the intensity values ​​of vegetation points, and calculate the height histogram. The peak intervals of the herb layer, shrub layer, and tree layer were identified by the Gaussian mixture model (GMM). The layering parameters were calculated, and the median and standard deviation of the intensity of each layer of point cloud were calculated, and the abnormal points with intensity exceeding the median ±2 standard deviation were removed. The tree layer used DBSCAN clustering (neighborhood radius 1m, minimum number of points 10) to distinguish the crowns of individual trees and remove isolated points; the shrub layer combined with multispectral NDVI values ​​(>0.4) to verify vegetation density, and NDVI <0.3 was marked as dead shrubs; the herb layer excluded non-vegetation points by thermal infrared temperature (daytime > ambient temperature +2℃). The density of each layer of point cloud was counted by gridding (10m×10m) to generate a spatial distribution map.

[0044] The Poisson surface reconstruction algorithm was used to construct the 3D model. Specifically, the LiDAR point cloud was downsampled (voxel size 0.5m) to reduce the amount of calculation. The SIFT feature points of the multispectral image were registered with the LiDAR point cloud (error <1 pixel) to ensure that the texture and geometry were aligned. The Poisson surface reconstruction algorithm was used with the LiDAR point cloud as input, and the grid fineness was controlled by adjusting the depth parameter. For densely vegetated areas, the number of iterations was increased to optimize the surface smoothness. Texture information such as NDVI and red edge band reflectivity of the multispectral image was mapped to the 3D model surface. The DN value of the multispectral image was converted to reflectivity through the radiometric calibration parameter, and then a pseudo-color map was generated based on the mean value in the grid and superimposed on the model surface to enhance the visualization effect of vegetation classification.

[0045] The forest coverage rate is calculated by extracting the tree point cloud in the three-dimensional model through threshold segmentation (such as height > 1m) and calculating the spatial volume occupied by it. The total target area volume is estimated based on the DSM range and the average terrain height.

[0046] The steps for calculating the average tree height are as follows: extract the highest point height of each individual tree crown from the tree point cloud of the three-dimensional model, combine the GMM stratification results, and assign weights to point clouds in different height ranges (such as 0.7 for the tree layer and 0.3 for the shrub layer).

[0047] The terrain slope verification is to perform local plane fitting on the terrain point cloud, select a 3×3 grid neighborhood, fit the plane equation ax+by+cz+d=0 by the least squares method, calculate the normal vector of the fitted plane, and then derive the slope. For error correction, the DSM slope is compared with the plane fitting slope. If the difference exceeds 5°, recheck the ground point cloud separation parameters, adjust the rigidity coefficient or iteration number of the CSF algorithm, and ensure the accuracy of the slope calculation.

[0048] In the analysis package for the research scenario of individual tree characteristics, data compression first analyzes the variances of each dimension of the tree point cloud data through the covariance matrix, selects the vertical dimension for segmentation, simultaneously introduces distance weights to divide nodes, retains local characteristics, and evaluates the effect using the compression ratio and feature loss rate, and dynamically adjusts the construction parameters accordingly; data dimensionality reduction first samples and splits various features of the tree point cloud through a random forest, screens out key features, weights the distances of the key feature dimensions, adjusts the perplexity, and verifies the dimensionality reduction results through visualization and quantitative analysis.

[0049] Data compression in the analysis package for the research scenario of individual tree characteristics includes the steps of: Calculate the covariance matrix for the LiDAR point cloud data of a single tree (including dimensions such as three-dimensional coordinates, intensity values, echo times, etc.), and analyze the variance contributions of each dimension. It includes: First calculate the mean of each dimension, and the formula is: ; Then calculate the elements of the covariance matrix: The element C of the covariance matrix C ij is: ; Finally, the covariance matrix is expressed as: ; In the formula: Z k represents the kth sample vector of the LiDAR point cloud data of a single tree, including dimension information such as three-dimensional coordinates, intensity values, echo times, etc.; Z ik represents the specific value of the ith dimension in the kth sample. For example, when i = 1, it can represent the coordinate value, and when i = 4, it can represent the intensity value; μ i represents the mean of the ith dimension, such as calculating the average value of the vertical coordinates in all sample point cloud data; n represents the number of samples of the LiDAR point cloud data of a single tree; m represents the total number of dimensions of the LiDAR point cloud data of a single tree (such as the total number of dimensions of three-dimensional coordinates, intensity values, echo times, etc.), and μ represents the vector composed of the means of each dimension μ i that reflects the average characteristics of each dimension of the point cloud data.

[0050] By comparing the variances of each dimension, the vertical direction is preferentially selected as the segmentation benchmark because it contributes the most to the structural characteristics such as tree height and crown stratification. The hierarchical segmentation strategy is to divide the tree point cloud into intervals such as the trunk area (e.g., z ≤ 2m), the basic crown layer (2m < z ≤ 5m), and the upper crown layer (z > 5m) according to the range of the z-axis coordinates (vertical direction) of the point cloud, and each interval is compressed independently.

[0051] The steps of dividing nodes with distance weights are to first construct a local neighborhood, that is, for each point cloud p i, taking it as the center, search for neighboring points within a radius of r = 0.5m, and calculate the average distance d between this point and its neighboring points avg , and then introduce the distance weight function ω i , , where σ is the scale parameter (e.g. σ=0.2m). The larger the weight, the higher the contribution of the point to the local structure and the higher the retention priority. Then the nodes are divided, that is, the point cloud in each layer is clustered based on the weight, and the DBSCAN algorithm (neighborhood radius 0.3m, minimum number of points 5) is used to identify the core points, retain the core points and their neighborhood points, and remove low-density noise points.

[0052] The data dimensionality reduction of the analysis package for the study scenario of individual characteristics of trees includes the following steps: The multi-dimensional features of the point cloud of a single tree are extracted, including geometric features, radiation features, and topological features. The geometric features include point cloud density, local curvature, and vertical height difference. The radiation features include the laser radar intensity value and the multi-spectral NDVI mean. The topological features include the neighborhood point distribution entropy and the number of skeleton branches. Then, the random forest model is constructed using random forest training with individual parameters such as trunk diameter and crown shape index as labels. The importance score of each feature is calculated by Gini impurity, key features are screened, thresholds are set (such as the top 30% of the importance score), and features that are strongly correlated with individual parameters are retained. For example, local curvature (correlation coefficient with trunk diameter r=0.78), vertical height difference (correlation coefficient with crown shape r=0.65), etc. are retained.

[0053] Through the above data compression and dimensionality reduction strategies, the tree individual characteristics research and analysis package can effectively reduce data redundancy while retaining the morphological and physiological details of individual trees, providing lightweight and highly differentiated feature input for subsequent three-dimensional modeling and parameter extraction.

[0054] Data fusion of analysis packages for tree individual characteristics research scenarios, including: The lidar point cloud data, multispectral images, and thermal infrared images were denoised separately, and then the lidar point cloud was used to determine the trunk axis and position through PCA, the trunk diameter was calculated using the least squares method, the crown layer was distinguished by DBSCAN, and the branch characteristics were obtained through skeleton extraction and topological analysis; the multispectral image used the grayscale co-occurrence matrix to extract texture features, and the NDVI was used to judge the health of the tree. When matching, the multispectral image identified the tree outline through edge detection and morphological processing, and then matched with the lidar point cloud according to the trunk height characteristics. By establishing a temperature-diameter relationship model, the thermal infrared and lidar data were correlated.

[0055] In the specific implementation, the statistical outlier removal algorithm is used for laser radar point cloud denoising. The average distance from each point to the K nearest neighbor points (K=30) is calculated, and the mean plus 3 times the standard deviation is used as the threshold to remove noise points (such as flying birds and sensor error points). For the tree trunk area, the threshold (mean + 2 times the standard deviation) is lowered to retain details. For multispectral image denoising, non-local mean filtering is applied to reduce noise by using image block similarity. The parameters are set as a search window of 15×15 pixels, a neighborhood window of 5×5 pixels, and a grayscale difference weight of 0.8. For thermal infrared image denoising, bilateral filtering is used, taking into account both spatial distance and grayscale differences. The parameters are set as a spatial Gaussian kernel standard deviation of 3 and a grayscale Gaussian kernel standard deviation of 0.1.

[0056] LiDAR point cloud feature extraction is to perform PCA analysis on the point cloud of a single tree, extract the principal components, project the point cloud onto the principal component plane (the first two principal components), and fit a straight line as the central axis of the trunk through the RANSAC algorithm. Specific steps: Calculate the point cloud covariance matrix to obtain the eigenvector and eigenvalue. Select the direction with the largest eigenvalue as the extension direction of the trunk (Z axis). Fit a straight line to the point cloud in the Z axis direction to obtain the central axis of the trunk.

[0057] The trunk diameter is calculated by extracting the points closest to the central axis in the vertical direction on both sides of the central axis and fitting a circle using the least squares method.

[0058] The DBSCAN algorithm was applied to non-trunk point clouds (Z-axis height > 2m), with parameters set to a neighborhood radius of 1.5m and a minimum number of points of 15. Different levels of the crown (such as the main crown layer and the side branch layer) were distinguished based on the clustering results. A skeleton extraction algorithm based on point clouds (such as VoxelSlicing) was used to generate the branch skeleton structure. The number of skeleton nodes and branches was counted through topological analysis to identify the branch bifurcation points.

[0059] The feature extraction of multispectral images includes: calculating the gray level co-occurrence matrix (GLCM) of multispectral images, extracting texture features such as contrast, entropy, and correlation, and setting the window size to 11×11 pixels with a step size of 5 pixels. NDVI is calculated using the red band and the near-infrared band. Edge detection (Canny operator, threshold 100-200) is performed on the multispectral image, combined with morphological dilation (kernel size 3×3) and contour tracking to extract tree contours.

[0060] The correlation between thermal infrared and LiDAR is to establish a regression model between thermal infrared temperature and trunk diameter, and to realize the correlation between thermal infrared temperature data and LiDAR geometric features. In the specific implementation, more than 200 trees of different species and with a diameter at breast height of 5-50cm are selected as training samples in the sample plot, and the following data are collected simultaneously: LiDAR features: trunk diameter (D, in cm), trunk height (H, in m), point cloud intensity mean (I); Thermal infrared characteristics: mean trunk temperature (T mean , unit ℃), daytime peak temperature T peak , unit ℃), day and night temperature difference ΔT, unit ℃); Environmental parameters: Atmospheric temperature at the time of collection (T a , unit ℃), relative humidity (RH, unit %), solar radiation intensity (S, unit W / m²).

[0061] The regression model of thermal infrared temperature and trunk diameter is constructed as follows: ; in α 0- α 5 is the regression coefficient, ϵ After the above model was trained and optimized, the data of 50 trees that did not participate in the training were used to verify the model, and the absolute error between the measured diameter and the predicted diameter was calculated. The absolute error was controlled within a mean of <2 cm and a standard deviation of <1.5 cm, which met the accuracy requirements of forestry surveys.

[0062] In multi-source data matching, the trunk height matching is based on the trunk height extracted by the lidar, and the contour area of ​​the corresponding height is searched in the multispectral image, and the matching is performed through geometric position consistency. Feature-level fusion is a weighted fusion of the lidar geometric features (such as trunk diameter, crown volume), multispectral texture and health index, and thermal infrared temperature features. The weight is dynamically adjusted according to the importance of the feature (such as geometric feature weight 0.6, spectral feature 0.3, temperature feature 0.1).

[0063] A 3D model is constructed for the tree individual characteristics research and analysis package, including: Based on the fused and denoised LiDAR point cloud data, multispectral image and thermal infrared image data, a surface reconstruction algorithm based on point cloud is used. During the reconstruction process, the sphere radius parameter is dynamically adjusted according to the local curvature of the tree point cloud. At the same time, the texture information of the multispectral image is combined to perform texture mapping on the model surface. In the specific implementation, point cloud surface reconstruction includes: dynamic ball radius adjustment, and surface reconstruction based on local curvature moving least squares (MLS). For areas with high curvature (such as branch bifurcations), the ball radius is reduced to 0.1m to retain details; for flat areas (such as the surface of tree trunks), the ball radius is increased to 0.3m to improve reconstruction efficiency. The reconstructed mesh is simplified, and the QuadricEdgeCollapse algorithm is used to reduce the number of triangles while retaining geometric features. The simplification rate is controlled at 30%-40%. In texture mapping, multispectral texture fusion is used, that is, the reflectance data of the multispectral image (after radiation calibration) is mapped to the surface of the three-dimensional model. For each triangular facet, its average reflectance in the multispectral image is calculated to generate a pseudo-color map to highlight the health status of the vegetation (such as red for disease and green for health).

[0064] Extraction of characteristic parameters for tree individual characteristics research and analysis package, including: The trunk height is directly obtained by measuring the length of the trunk centerline in the three-dimensional model; the crown shape index is calculated based on the three-dimensional shape of the crown; and the number of branch forks is statistically analyzed through topological analysis of the branch skeleton structure in the three-dimensional model.

[0065] The trunk height measurement is to directly obtain the endpoint coordinates of the trunk centerline and calculate the Euclidean distance between the two points as the trunk height.

[0066] The crown shape index calculation is to calculate the three-dimensional volume and vertical projection area of ​​the crown.

[0067] The branch bifurcation count is a topological analysis of the skeleton structure, identifying nodes with a degree ≥ 3 as bifurcation points, and the number of bifurcations is the number of such nodes. Combined with the DBSCAN clustering results, different levels of bifurcations (such as main branch bifurcations and side branch bifurcations) are distinguished.

[0068] Through the above data fusion, three-dimensional modeling and feature extraction processes, the tree individual characteristics research and analysis package can analyze the morphological and physiological details of individual trees with high precision, providing key technical support for the refined management of forestry resources.

[0069] The two analysis packages are integrated to analyze and construct the extracted feature parameters. Specifically, the parameters of the two analysis packages are associated through a unified geographic coordinate system, and the spatial database is used to establish an index. The feature weights are configured according to the scene requirements, and a model of macro-micro parameters is constructed to generate the corresponding scene map, which is visualized through a three-dimensional model, such as Figure 2 shown.

[0070] The specific implementation of the fusion of the two analysis packages includes: unifying the geographic coordinate system, that is, mapping the three-dimensional model of the macro monitoring analysis package (resolution 10m×10m grid) and the individual tree data of the individual research analysis package (accuracy 0.1m level) to the WGS84 coordinate system. For the individual tree point cloud, the GPS / IMU positioning data of the lidar point cloud is rigidly aligned with the ground control points (GCPs) of the macro DSM, and the iterative closest point (ICP) algorithm is used for optimization to ensure that the position error of the individual tree is <5cm. Then, the spatial database index is established, that is, the R-tree spatial index structure is used to partition the macro grid and individual trees according to geographic coordinates. For example, 1km×1km is an index unit, and each unit stores the macro characteristics (such as coverage, slope) and the ID and location of the individual tree in the area. Establish an association table to record the macro grid ID to which each individual tree belongs, as well as the macro parameters of the grid (such as NDVI mean, terrain slope).

[0071] Dynamic configuration of feature weights is mainly based on weight allocation driven by scenario requirements. First, a scenario weight profile is defined to support users to dynamically adjust feature weights according to application objectives (such as forest health assessment and resource planning).

[0072] In the specific implementation, according to the core objectives of forestry survey, four typical application scenarios are defined, and weights are allocated according to the importance differences between macro characteristics (M) and individual characteristics (I). The feature classification and initial weight configuration are shown in Table 1: Application Scenario Core Goals Macroscopic characteristics (M) Individual characteristics (I) Basic weight distribution (M:I) Forest Health Assessment Identify tree diseases, pests, and growth stress Grid NDVI mean, temperature mean, day and night temperature difference Individual tree NDVI anomaly, crown shape index, thermal infrared temperature variance 0.3:0.7 Carbon stock estimation Assessment of regional vegetation biomass and carbon sink capacity Average tree height, vegetation volume, and coverage Trunk diameter, trunk height, wood density (predicted values) 0.6:0.4 Resource planning (harvesting / planting) Optimizing forest spatial distribution and economic value assessment Terrain slope, point cloud density, red edge reflectivity Diameter at breast height (1.3m diameter of trunk), tree species classification, growth potential index 0.5:0.5 Disaster risk assessment (wind / fire) Identification of high-risk areas and tree resilience Slope, aspect, vegetation density Crown wind resistance index (based on the number of skeleton branches), thermal infrared hot spot intensity 0.4:0.6 During implementation, the feature weights for different scenarios can be configured in detail as follows: In the pest and disease monitoring scenario, individual characteristics are 0.7 and macro characteristics are 0.3; Among them, the weight distribution of individual characteristics of 0.7 is as follows: single tree NDVI abnormal value (deviation from the mean of healthy trees, weight 0.3), crown shape index (main canopy volume / vertical projection area, weight 0.25), thermal infrared temperature variance (reflecting the thermal stability of the trunk cortex, weight 0.15), pest and disease markers (based on multispectral texture feature classification results, weight 0.1) The weight distribution of the macro feature 0.3 is as follows: grid NDVI mean (regional health baseline, weight 0.15), day and night temperature difference (environmental stress factor, weight 0.1), vegetation density (grid point cloud number, weight 0.05).

[0073] If the NDVI abnormal value of a single tree exceeds 3σ (standard deviation), the weight of the individual feature will be automatically increased to 0.8 and the macro feature will be reduced to 0.2, thus strengthening the identification priority of the abnormal individual.

[0074] In the carbon stock estimation scenario, the macroscopic characteristics are 0.6 and the individual characteristics are 0.4; Among them, the weight distribution of the macro feature 0.6 is as follows: average tree height (based on GMM stratified weighted average, weight 0.25), vegetation volume (tree point cloud volume in the three-dimensional model, weight 0.2), coverage (proportion of tree point cloud volume in the grid, weight 0.15), NDVI mean (indirect indicator of biomass, weight 0.1); The weights of individual features (0.4) are: trunk diameter (DBH, weight 0.2), trunk height (central axis length, weight 0.15), LiDAR intensity value (wood density prediction, weight 0.05), For dense forest areas with canopy density > 0.9, the macro-feature weights were fine-tuned to 0.65 (increasing the vegetation volume weight to 0.25) to reduce the impact of individual feature measurement errors.

[0075] In the resource planning scenario, macro-individual equilibrium configuration (0.5:0.5).

[0076] The weight distribution of macro-features is as follows: terrain slope (affecting felling cost, weight 0.2), point cloud density (vegetation richness, weight 0.15), red edge reflectivity (forest growth stage, weight 0.15); The weight distribution of individual characteristics is as follows: diameter at breast height (core indicator of economic value, weight 0.25), tree species classification (based on multispectral spectral characteristics, weight 0.15), growth potential index (combined with NDVI and annual increment of tree height, weight 0.1).

[0077] When the user sets "economic value priority", the combined weights of "diameter at breast height" and "tree species classification" in the individual characteristics are increased to 0.4 (originally 0.4), and the weight of "red edge reflectivity" in the macro characteristics is increased to 0.2 (reflecting the maturity of the tree species).

[0078] The weight optimization algorithm uses the analytic hierarchy process (AHP) combined with the entropy weight method, combining subjective and objective weighting. The subjective weight is determined by expert scoring, and the objective weight is calculated based on the contribution of feature variance. The final weight is the weighted average of the two.

[0079] The macro-micro parameter model is constructed by first extracting the fusion feature set, including macro parameters (such as coverage, slope, NDVI mean) and individual parameters (such as trunk height, crown shape index, and pest index). The gradient boosting tree (XGBoost) or deep neural network (such as ResNet) is used to build a cross-scale model, with the input as fusion features and the output as comprehensive evaluation results (such as forest health level and resource distribution prediction). During the training process, the model hyperparameters are optimized through cross-validation to ensure the balance between the model in macro trends and individual anomaly identification. The Apriori algorithm is used to mine strong association rules between macro and micro features. For example: if the slope of a grid is >30° and the average crown shape index of trees in the grid is <0.8, the wind resistance of trees in this area is low (support >0.6, confidence >0.7). If the grid NDVI mean is <0.4 and more than 50% of the trees have NDVI <0.3, there may be pests and diseases in this area (support >0.5, confidence >0.8).

[0080] In the specific implementation, the macro-micro parameter model algorithm uses XGBoost gradient boosting tree, and the parameter configuration is: learning rate 0.1, tree depth 6, subsampling 0.8, column sampling 0.7. The evaluation index is root mean square error (RMSE) ≤ 0.15. The training data is a fusion data set containing 2000 sets of macro features and 50,000 individual features.

[0081] Generate macro-micro fusion thematic maps, including: health assessment map, with macro grid as the unit, superimpose the health index of individual trees (such as NDVI value, pest and disease markers), and use heat map to show the distribution of health status. Resource planning map, combining macro coverage, average tree height and economic value of individual trees (such as tree species, diameter at breast height), marking high-value areas. Using the graduated symbol method, the key parameters of individual trees are represented by symbols of different sizes and colors on the macro map (such as the larger the trunk diameter, the larger the symbol size). Load the fused model in a 3D GIS platform (such as Cesium, ArcGIS Pro) to achieve the following functions: multi-scale switching, through zooming operations, smoothly transition from the macro forest panorama to the details of a single tree, and display the macro and individual parameters of the corresponding position in real time. Dynamic query, click any position in the 3D model, and the macro characteristics of the area (such as coverage, slope) and the individual characteristics of the nearest 5 trees (such as height, health index) will pop up. Simulation analysis, based on the fusion model, simulate the impact of different management strategies (such as selective felling, pest and disease control) on forest resources and visualize the difference in results.

[0082] The fusion of the two analysis packages can effectively integrate macro and micro data, provide cross-scale and multi-dimensional decision support for forestry resource management, and achieve refined management goals from the global to the local.

[0083] It is obvious that the method of the present invention can be implemented by a computer program, and the computer program for implementing the method of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that when the computer program is executed by the processor, the functions / operations specified in the flow chart and / or block diagram are implemented. The computer program can be executed entirely on the machine, partially on the machine, partially on the machine as a stand-alone software package and partially on a remote machine, or entirely on a remote machine or server.

[0084] Therefore, it can be understood that the present invention discloses an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned forestry survey planning, design and analysis method based on three-dimensional laser modeling.

[0085] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A forestry survey planning, design and analysis method based on three-dimensional laser modeling, characterized in that: The method comprises the following steps: collecting laser radar point cloud data, multispectral images and thermal infrared images of a target area by means of unmanned aerial vehicles, ground mobile platforms and satellite remote sensing; establishing corresponding analysis packages based on different analysis scenarios, wherein the analysis packages include analysis packages for forest resource macro-monitoring scenarios and analysis packages for tree individual feature research scenarios, wherein each analysis package adaptively selects a corresponding algorithm to perform data compression and dimensionality reduction, and then fuses the laser radar point cloud data with the multispectral images and thermal infrared images; for each analysis package, constructing a corresponding three-dimensional model using the fused data, and extracting characteristic parameters required for forestry investigation from the three-dimensional model; and fusing the two analysis packages, analyzing and constructing the extracted characteristic parameters, and generating corresponding scene maps, thereby providing a basis for forestry planning, design and analysis.

2. The forestry survey planning, design and analysis method based on three-dimensional laser modeling according to claim 1 is characterized in that: For the analysis package of forest resource macro-monitoring scenarios, data compression is to jointly adjust the lidar point cloud data, multispectral images and thermal infrared images to eliminate redundant data; data dimension reduction is to calculate the correlation between the original multi-source data and the macro characteristics of the forest, set the threshold of the correlation coefficient, retain the relevant feature data above the threshold, and eliminate the relevant feature data below the threshold.

3. The forestry survey planning, design and analysis method based on three-dimensional laser modeling according to claim 2 is characterized in that: Data fusion of analysis packages for macro-monitoring scenarios of forest resources, including: The macro terrain features and vegetation features are extracted after the lidar point cloud is denoised, the pixel DN value of the multispectral image is converted into a radiation brightness value and the spectral features are calculated. The grayscale value of the thermal infrared image is converted into an actual temperature value and the energy distribution features are extracted. During fusion, the three types of data are preliminarily aligned according to the geographic coordinates, and then secondary matching is performed based on feature similarity. Weighted fusion is used to assign weights according to the importance of the data source, and the matched feature data are fused according to the assigned weights.

4. The forestry survey planning, design and analysis method based on three-dimensional laser modeling according to claim 3 is characterized in that: Construct a three-dimensional model for the macro-monitoring and analysis package of forest resources, including: The macro terrain features and vegetation features extracted from the lidar point cloud serve as the spatial skeleton. The model is given vegetation details and classification information through multispectral images. The model is given energy distribution through thermal infrared images to construct a three-dimensional model. The model surface is rendered using multispectral image texture and color information.

5. The forestry survey planning, design and analysis method based on three-dimensional laser modeling according to claim 4 is characterized in that: Feature parameter extraction for forest resource macro monitoring and analysis package, including: The forest coverage rate is obtained by calculating the ratio of the spatial volume occupied by the tree point cloud in the three-dimensional model to the spatial volume of the entire target area. The average tree height is obtained by obtaining the height data of the tree point cloud in the three-dimensional model and then calculating it by the weighted average method. The terrain slope is obtained by performing plane fitting analysis on the terrain point cloud in the three-dimensional model.

6. The forestry survey planning, design and analysis method based on three-dimensional laser modeling according to claim 1 is characterized in that: In the analysis package for the study scenario of individual characteristics of trees, data compression calculates the covariance matrix of the lidar point cloud data, analyzes the variance proportion of each dimension, selects the dimension with the largest variance proportion as the core segmentation dimension, divides the interval according to the coordinate range of the core segmentation dimension, compresses each interval independently, introduces distance weight to divide the nodes, clusters the point cloud in each interval based on the weight, identifies the core points, retains the core points and their neighborhood points, and eliminates low-density noise points; data dimensionality reduction samples and splits various features of the lidar point cloud data of a single tree, screens out key features, sets thresholds, and retains features that exceed the threshold.

7. The forestry survey planning, design and analysis method based on three-dimensional laser modeling according to claim 6 is characterized in that: Data fusion of analysis packages for tree individual characteristics research scenarios, including: The lidar point cloud data, multispectral images, and thermal infrared images of individual trees are denoised separately, and then the trunk centerline and position are determined by feature extraction of the lidar point cloud data of individual trees. The trunk diameter is then calculated, the crown layers are distinguished, and the branch features are obtained. Texture features are extracted from the multispectral images of individual trees. When matching, the tree outline is identified through the multispectral image, and then matched with the corresponding lidar point cloud through geometric position consistency. The thermal infrared and lidar data are then associated, and the geometric features of the lidar data, the multispectral texture and health index, and the temperature features of the thermal infrared are weightedly fused.

8. The forestry survey planning, design and analysis method based on three-dimensional laser modeling according to claim 7 is characterized in that: A three-dimensional model is constructed for the tree individual characteristics research and analysis package, including: point cloud-based surface reconstruction based on the fused and denoised lidar point cloud data, multispectral images and thermal infrared image data, dynamic adjustment of the sphere radius parameter according to the local curvature of the tree point cloud, and texture mapping of the model surface combined with the texture information of the multispectral image.

9. The forestry survey planning, design and analysis method based on three-dimensional laser modeling according to claim 1 is characterized in that: The two analysis packages are integrated and the extracted feature parameters are analyzed and constructed. Specifically, the parameters of the two analysis packages are associated through a unified geographic coordinate system, and the spatial database is used to establish an index. The feature weights are configured according to the scene requirements, and a macro-micro parameter model is constructed to generate the corresponding scene map, which is visualized through a three-dimensional model.

10. An electronic device, characterized in that: include: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, wherein the instructions are executed by the at least one processor to enable the at least one processor to perform the method of claim 9.

Images