Forest individual tree parameter automatic extraction method based on point cloud and deep learning
By using point cloud and deep learning methods, the problems of low efficiency and insufficient accuracy in traditional forest surveys have been solved. This method enables efficient, automated, and high-precision extraction of tree diameter at breast height (DBH) and location, overcoming the effects of image boundary truncation and vegetation occlusion, and providing higher survey efficiency and accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-27
- Publication Date
- 2026-03-10
AI Technical Summary
Traditional forest survey methods are inefficient, labor-intensive, and prone to subjective errors. They are also difficult to locate in complex terrain. Existing point cloud data segmentation algorithms cannot guarantee segmentation accuracy under high forest density, and low vegetation can cause the loss of diameter at breast height (DBH) point cloud data.
By employing a point cloud and deep learning-based approach, through multi-height layer slicing, image processing, and deep learning model training, combined with a multi-starting point translational cropping method, we overcome the interference of image boundary truncation and achieve efficient, automated, and high-precision extraction of the diameter at breast height (DBH) and position of standing trees.
It achieves efficient, automated, and robust extraction of tree parameters, overcomes interference from image boundary truncation and the influence of understory vegetation occlusion, improves survey efficiency and accuracy, and can obtain the diameter at breast height, location, and other parameters of standing trees.
Smart Images
Figure CN121640153A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of forestry investigation, remote sensing surveying and target detection, and particularly relates to a forest single tree parameter automatic extraction method based on point cloud and deep learning, which relates to a method combining ground-based laser radar point cloud data processing, computer vision and deep learning model to realize high-precision and automatic extraction of standing tree position, diameter at breast height and tree height. BACKGROUND
[0002] Forest resource investigation is the core work of forestry management, and the diameter at breast height and tree height of standing trees are key investigation factors.
[0003] The traditional field measurement method mainly relies on manual use of diameter tape, compass and other tools, and has problems such as low efficiency, high labor intensity, obvious subjective error and difficult positioning in complex terrain.
[0004] In recent years, the application of ground-based laser radar technology provides a new way for forest investigation, which can quickly obtain high-precision three-dimensional forest point cloud data. There are some single tree segmentation and parameter extraction methods based on point cloud data in the prior art, but there are still the following problems and shortcomings: first, the direct single tree segmentation algorithm for the original point cloud is complex, and the segmentation accuracy is difficult to guarantee in the case of high forest density; second, the traditional method is easily affected by low-lying vegetation (shrubs, grasses) shielding, resulting in missing point cloud at the diameter at breast height. SUMMARY
[0005] The present application provides a forest single tree parameter automatic extraction method based on point cloud and deep learning, which can effectively overcome image boundary truncation interference and realize efficient, automatic and high-precision extraction of standing tree diameter at breast height and position data.
[0006] The present application adopts the following technical solutions.
[0007] A forest single tree parameter automatic extraction method based on point cloud and deep learning, which is used for efficient, automatic and high-precision extraction of standing tree diameter at breast height and position data, and comprises the following steps: Step S1: data acquisition and preprocessing; Step S2: multi-height layer point cloud slicing; Step S3: point cloud slicing imaging and anti-truncation enhancement processing; Step S4: diameter at breast height labeling and model training; Step S5: diameter at breast height preliminary detection, position calculation and truncation elimination; Step S6: multi-height layer data fusion, false detection elimination and modeling sample labeling; Step S7: diameter at breast height prediction model construction.
[0008] Step S8: Detection and completion of breast diameter data; Step S9: Single tree segmentation and height extraction.
[0009] In step S1, the original three-dimensional point cloud data of the forest stand to be measured is obtained by a ground-based laser radar; the ground-based laser radar includes a station-mounted laser radar and a mobile laser radar; When using a station-mounted laser radar, a multi-station splicing scanning mode is adopted, and the layout of the scanning station follows the principles of no omission and full coverage and high overlap, so that each tree in the forest stand can be detected from at least two different stations to obtain complete point cloud data; When using a mobile laser radar, the operator walks along a planned path carrying or holding the device for scanning; the planned path covers the forest stand to be measured, and a closed loop or an "S" shaped path is adopted, while ensuring that the key areas of the forest stand to be measured are passed through, so as to obtain complete and uniform point cloud data.
[0010] In step S1, the obtained original point cloud data is processed for registration to obtain the overall point cloud of the sample plot to be measured; the overall point cloud is given actual accurate coordinates in the following ways: Method one: In the scanning area, three or more control points are laid out or known survey stations are used, and high-precision measurement techniques (including but not limited to real-time differential positioning technology RTK, total station, etc.) are used to obtain their actual three-dimensional coordinates; the relative coordinates of these points in the point cloud and their corresponding actual absolute coordinates are used to correct through a coordinate conversion model, and the point cloud is unified to the target coordinate system; Method two: Obtain the airborne point cloud of the same forest area by visible light unmanned aerial photogrammetry or laser radar unmanned aerial vehicle, and use the airborne point cloud as a reference to register and unify the coordinates with the ground point cloud; After obtaining the overall point cloud with actual accurate coordinates, denoising, ground point classification are performed on the overall point cloud, and a digital elevation model is generated based on the classified ground points, and then height normalization is performed on the overall point cloud.
[0011] In step S2, the height interval of 1.3 meters containing the standard breast diameter position and at least two auxiliary height layers are extracted from the normalized point cloud; The selection method of the auxiliary height layer follows the following principles: the height of the auxiliary height layer is higher than the influence height of the understory shrub and grass vegetation, and the auxiliary height layer is located in the interval where the trunk is straight and has not started branching, so as to avoid the interference of the crown structure; In step S3, each point cloud slice layer is converted into a two-dimensional raster image (such as TIFF format), and the image is uniformly scaled to a preset size; a multi-start point translation cropping method is adopted to overcome the truncation effect of the image boundary on the tree trunk: Within a fixed spatial range of each slice layer, three sets of images (e.g., JPG format) are generated, each starting from the origin (0, 0) and a preset offset (e.g., (-160, -160), (-320, -320)). During the image generation process, based on the planar distribution range of the point cloud of the slice layer in the global coordinate system, an image file name containing its geographic coordinate boundary is generated. The boundary data includes minimum X, maximum X, minimum Y, and maximum Y (e.g., 635387.652500_635394.052500_3069808.780108_3069815.180108.jpg). This establishes a precise mapping relationship between image pixel coordinates and real-world coordinates.
[0012] In step S4, the cross-sections of tree trunks in the converted images are labeled using an image annotation tool to form a training dataset; the target detection neural network model (such as the YOLO series) is trained using this training dataset to obtain the trunk diameter detection model; In step S5, the three sets of images obtained in step S3 are input into the trained detection model to obtain the bounding box and confidence score of each detected tree cross-section. Based on the center pixel coordinates of the bounding box in the image and the coordinate range information embedded in the corresponding image file name, the two-dimensional plane position of the tree cross-section in the real coordinate system is calculated. At the same time, the diameter of the tree is calculated based on the size ratio of the bounding box. The shortest distance between each detection box and the image boundary is calculated. If the distance is less than a preset threshold, the target is determined to be truncated by the image boundary, and the detection result is directly discarded. For multiple valid detection results of the same tree generated in the translation image that are not discarded, the average value of their calculated indicators is taken as the final result of the tree in this slice layer.
[0013] In step S6, the trees detected at different height levels are associated and matched; a planar distance threshold is set, and if the planar distance between the center points of the detection boxes at different height levels is less than the threshold, they are determined to belong to the same tree. For multiple detection results that are identified as the same tree, a tree is only confirmed as a valid standing tree if it is successfully detected at at least two height levels. During this process, standing trees that are successfully detected at the standard diameter at breast height (DBH) level (1.3 meters) and all preset auxiliary height levels are identified and marked as modeling samples for subsequent DBH prediction model training.
[0014] In step S7, a robust diameter at breast height (DBH) prediction model is constructed using the data from the modeling samples identified in step S6. The model is based on the taper law of tree trunks (i.e., the relationship between trunk diameter and height). It aims to effectively predict the DBH value missing in the point cloud due to understory shading by utilizing the trunk diameter detected at higher, unobstructed heights. The steps include the following: Step S7.1: Data preparation and scenario adaptation; The trunk diameter DBH of the 1.3-meter standard diameter at breast height layer was extracted as the dependent variable in the modeling sample. Based on the stand structure characteristics, two auxiliary height layers (h1, h2) were preset. The height selection followed the following principles: the auxiliary height layers were higher than the influence height of the understory shrubs and grasses, and the auxiliary height layers were located in the interval where the trunks were straight and had not yet started to branch, in order to avoid interference from the canopy structure. The trunk diameters d_h1 and d_h2 detected in these two auxiliary height layers were used as independent variables to construct the training dataset. Step S7.2: Model training; Two types of regression models were trained and compared to adapt to different stand structures and accuracy requirements. Both types of regression models were evaluated using K-fold cross-validation. The coefficient of determination R² and root mean square error RMSE were used as core indicators to quantify the goodness of fit and prediction accuracy of the models. Regression Model Option 1: Linear Model, specifically: using multiple linear regression as the basic model, expressed by the formula: Where DBH represents the diameter at breast height (DBH), d_h1 and d_h2 represent the trunk diameters detected at heights h1 and h2, respectively; β0 is the intercept of the regression model, and β1 and β2 are the partial regression coefficients of the independent variables d_h1 and d_h2, respectively; the regression coefficients β0, β1, and β2 are estimated using the least squares method, and the optimization objective is to minimize the sum of squared residuals between the predicted and actual detected values. This model assumes a stable linear relationship between diameter at breast height (DBH) and auxiliary height diameter, and is suitable for plantations or monoculture communities with straight trunks and uniform taper. Regression Model Option 2: Tree Model; Specifically, for complex forest stand scenarios, machine learning models such as random forest regression and gradient boosting decision tree (GBDT) are introduced. By constructing multiple decision trees and integrating their prediction results, the complex nonlinear relationships and interaction effects between features d_h1, d_h2 and the target DBH are automatically captured. This model is more suitable for natural forests or mixed forests with large variations in trunk morphology and complex taper patterns. When the linear assumption does not hold, it can provide more accurate predictions.
[0015] In this step, by comparing and analyzing the performance of two different regression models, and taking into account R², RMSE and model complexity, the optimal circumference prediction model is determined based on the highest prediction accuracy and the best stability.
[0016] In step S8, for a standing tree that is confirmed to be valid, the diameter calculated in step S6 at the standard diameter at breast height (1.3 meters) is used directly as the diameter at breast height value. If the diameter at breast height value is missing, data completion is performed: the diameters d_h1 and d_h2 detected at a higher height are substituted into the diameter at breast height prediction model trained in step S7 to calculate the diameter at breast height value.
[0017] In step S9, the actual plane position of each valid standing tree is used as a seed point, and the complete point cloud of a single tree is separated by a single tree segmentation method based on the seed point. Based on the segmented single tree point cloud, the difference between the Z coordinates of its highest and lowest points is taken as the tree height.
[0018] After separating the complete 3D point cloud of a single tree, the tree height and crown width parameters are obtained from the point cloud data.
[0019] This invention employs the aforementioned structured process based on multi-height layer slicing, image-based deep learning detection, and model completion, and has the following advantages: 1. High efficiency and automation: The complex 3D point cloud segmentation problem is transformed into a 2D image target detection problem. By utilizing a mature deep learning framework, the computational complexity is greatly reduced, and the batch and automated extraction of forest parameters is realized, which significantly improves the efficiency of the survey.
[0020] 2. High precision and robustness: By segmenting from multiple starting points, interference from image boundary truncation is overcome; by fusing and cross-validating detection results from multiple height layers, false detections and missed detections that may occur in single-layer detection are effectively eliminated, thus improving the reliability and robustness of the results.
[0021] 3. Strong resistance to shading: The innovative data completion mechanism (taper model estimation) can effectively address the problem of missing diameter-at-breast height (DBH) point cloud due to shading by understory vegetation, ensuring data integrity in complex forest environments.
[0022] 4. Precise positioning: By embedding global coordinate information into the image file name, a mapping from the image pixel space to the actual measurement space is established, enabling precise positioning of the tree location.
[0023] 5. Comprehensive Functionality: This method can not only accurately obtain diameter at breast height (DBH) and location, but also further obtain the complete three-dimensional point cloud of a single tree through seed point segmentation, thereby obtaining more parameters such as tree height and crown width, forming a complete single tree parameter extraction scheme. Attached Figure Description
[0024] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Appendix Figure 1 This is a flowchart illustrating the extraction method in an embodiment of the present invention; Appendix Figure 2 This is a schematic diagram of the anti-truncation enhancement processing at multiple starting positions in the extraction method of this invention embodiment; Appendix Figure 3 This is a schematic diagram of multi-height layer data fusion and false detection removal in the extraction method of this invention (Figure a is a point cloud preview and scanning route; Figure b is a schematic diagram of target detection and extraction of tree trunks; Figure c is a schematic diagram of the results after filtering three height layers). Detailed Implementation
[0025] As shown in the figure, an automatic extraction method for forest tree parameters based on point cloud and deep learning is proposed. The method is used for efficient, automated and high-precision extraction of diameter at breast height (DBH) data and location data of standing trees in forests, and includes the following steps; Step S1: Data acquisition and preprocessing; Step S2: Multi-height layer point cloud slicing; Step S3: Point cloud slice imaging and anti-truncation enhancement processing; Step S4: Diameter at breast height (DBH) annotation and model training; Step S5: Preliminary measurement of chest diameter, location calculation, and truncation / removal; Step S6: Multi-height layer data fusion, false detection removal, and modeling sample identification; Step S7: Construction of the diameter at breast height (DBH) prediction model.
[0026] Step S8: Detection and completion of chest diameter data; Step S9: Individual tree segmentation and tree height extraction.
[0027] In step S1, the original three-dimensional point cloud data of the forest stand to be measured is acquired by ground-based lidar; the ground-based lidar includes stationary lidar and mobile lidar. When using a station-mounted lidar, a multi-station stitching scanning method is adopted. The layout of the scanning stations follows the principles of full coverage without omission and high overlap, ensuring that every standing tree in the forest can be detected from at least two different stations to obtain complete point cloud data. When using a mobile lidar, the operator carries or holds the device along a planned path to scan the forest stand to be measured. The planned path covers the forest stand to be measured and adopts a closed loop or "S" shaped path, while ensuring that the key areas of the forest stand to be measured are traversed by the path to obtain complete and uniform point cloud data.
[0028] In step S1, the acquired raw point cloud data is registered to obtain the overall point cloud of the sample land to be tested; the overall point cloud is assigned actual precise coordinates in the following way; Method 1: Within the scanning area, set up three or more control points or utilize known survey points, and use high-precision measurement technology (including but not limited to real-time differential positioning technology RTK, total station, etc.) to obtain their actual three-dimensional coordinates; use the relative coordinates of these points in the point cloud and their corresponding measured absolute coordinates to perform correction through a coordinate transformation model, and unify the point cloud to the target coordinate system; Method 2: Obtain airborne point clouds of the same forest area through visible light UAV photogrammetry or lidar UAV, and use the airborne point cloud as a reference to register and unify the coordinates with the ground-based point cloud; After obtaining the overall point cloud with accurate actual coordinates, it is denoised, ground points are classified, and a digital elevation model is generated based on the classified ground points. Then, the overall point cloud is height normalized.
[0029] In step S2, from the normalized point cloud, the height range containing the standard diameter at breast height (DBH) of 1.3 meters and at least two auxiliary height layers are extracted according to height. The selection of the auxiliary height layer follows these principles: the height of the auxiliary height layer is higher than the influence height of the understory shrubs and grasses, and the auxiliary height layer should be located in the section where the trunk is straight and has not yet begun to branch, so as to avoid interference from the canopy structure; In step S3, each point cloud slice layer is converted into a two-dimensional raster image (such as TIFF format), and the image is uniformly scaled to a preset size; a multi-starting-point translation cropping method is used to overcome the truncation effect of image boundaries on the tree trunk: Within a fixed spatial range of each slice layer, three sets of images (e.g., JPG format) are generated, each starting from the origin (0, 0) and a preset offset (e.g., (-160, -160), (-320, -320)). During the image generation process, based on the planar distribution range of the point cloud of the slice layer in the global coordinate system, an image file name containing its geographic coordinate boundary is generated. The boundary data includes minimum X, maximum X, minimum Y, and maximum Y (e.g., 635387.652500_635394.052500_3069808.780108_3069815.180108.jpg). This establishes a precise mapping relationship between image pixel coordinates and real-world coordinates.
[0030] In step S4, the cross-sections of tree trunks in the converted images are labeled using an image annotation tool to form a training dataset; the target detection neural network model (such as the YOLO series) is trained using this training dataset to obtain the trunk diameter detection model; In step S5, the three sets of images obtained in step S3 are input into the trained detection model to obtain the bounding box and confidence score of each detected tree cross-section. Based on the center pixel coordinates of the bounding box in the image and the coordinate range information embedded in the corresponding image file name, the two-dimensional plane position of the tree cross-section in the real coordinate system is calculated. At the same time, the diameter of the tree is calculated based on the size ratio of the bounding box. The shortest distance between each detection box and the image boundary is calculated. If the distance is less than a preset threshold, the target is determined to be truncated by the image boundary, and the detection result is directly discarded. For multiple valid detection results of the same tree generated in the translation image that are not discarded, the average value of their calculated indicators is taken as the final result of the tree in this slice layer.
[0031] In step S6, the trees detected at different height levels are associated and matched; a planar distance threshold is set, and if the planar distance between the center points of the detection boxes at different height levels is less than the threshold, they are determined to belong to the same tree. For multiple detection results that are identified as the same tree, a tree is only confirmed as a valid standing tree if it is successfully detected at at least two height levels. During this process, standing trees that are successfully detected at the standard diameter at breast height (DBH) level (1.3 meters) and all preset auxiliary height levels are identified and marked as modeling samples for subsequent DBH prediction model training.
[0032] In step S7, a robust diameter at breast height (DBH) prediction model is constructed using the data from the modeling samples identified in step S6. The model is based on the taper law of tree trunks (i.e., the relationship between trunk diameter and height). It aims to effectively predict the DBH value missing in the point cloud due to understory shading by utilizing the trunk diameter detected at higher, unobstructed heights. The steps include the following: Step S7.1: Data preparation and scenario adaptation; The trunk diameter DBH of the 1.3-meter standard diameter at breast height layer was extracted as the dependent variable in the modeling sample. Based on the stand structure characteristics, two auxiliary height layers (h1, h2) were preset. The height selection followed the following principles: the auxiliary height layers were higher than the influence height of the understory shrubs and grasses, and the auxiliary height layers were located in the interval where the trunks were straight and had not yet started to branch, in order to avoid interference from the canopy structure. The trunk diameters d_h1 and d_h2 detected in these two auxiliary height layers were used as independent variables to construct the training dataset. Step S7.2: Model training; Two types of regression models were trained and compared to adapt to different stand structures and accuracy requirements. Both types of regression models were evaluated using K-fold cross-validation. The coefficient of determination R² and root mean square error RMSE were used as core indicators to quantify the goodness of fit and prediction accuracy of the models. Regression Model Option 1: Linear Model, specifically: using multiple linear regression as the basic model, expressed by the formula: Where DBH represents the diameter at breast height (DBH), d_h1 and d_h2 represent the trunk diameters detected at heights h1 and h2, respectively; β0 is the intercept of the regression model, and β1 and β2 are the partial regression coefficients of the independent variables d_h1 and d_h2, respectively; the regression coefficients β0, β1, and β2 are estimated using the least squares method, and the optimization objective is to minimize the sum of squared residuals between the predicted and actual detected values. This model assumes a stable linear relationship between diameter at breast height (DBH) and auxiliary height diameter, and is suitable for plantations or monoculture communities with straight trunks and uniform taper. Regression Model Option 2: Tree Model; Specifically, for complex forest stand scenarios, a machine learning model combining random forest regression and gradient boosting decision tree (GBDT) is introduced. By constructing multiple decision trees and integrating their prediction results, the complex nonlinear relationships and interaction effects between features d_h1, d_h2 and the target DBH are automatically captured. This model is more suitable for natural forests or mixed forests with large variations in trunk morphology and complex taper patterns. When the linear assumption does not hold, it can provide more accurate predictions.
[0033] In this step, by comparing and analyzing the performance of two different regression models, and taking into account R², RMSE and model complexity, the optimal circumference prediction model is determined based on the highest prediction accuracy and the best stability.
[0034] In step S8, for a standing tree that is confirmed to be valid, the diameter calculated in step S6 at the standard diameter at breast height (1.3 meters) is used directly as the diameter at breast height value. If the diameter at breast height value is missing, data completion is performed: the diameters d_h1 and d_h2 detected at a higher height are substituted into the diameter at breast height prediction model trained in step S7 to calculate the diameter at breast height value.
[0035] In step S9, the actual plane position of each valid standing tree is used as a seed point, and the complete point cloud of a single tree is separated by a single tree segmentation method based on the seed point. Based on the segmented single tree point cloud, the difference between the Z coordinates of its highest and lowest points is taken as the tree height.
[0036] After separating the complete 3D point cloud of a single tree, the tree height and crown width parameters are obtained from the point cloud data.
Claims
1. A method for automatically extracting forest individual tree parameters based on point cloud and deep learning, characterized in that: The method comprises the following steps when extracting the diameter at breast height data and position data of single tree in forest stand: Step S1: data acquisition and preprocessing; Step S2: multi-height layer point cloud slicing; Step S3: point cloud slice imaging and anti-truncation enhancement processing; Step S4: diameter at breast height labeling and model training; Step S5: diameter at breast height preliminary detection, position calculation and truncation elimination; Step S6: multi-height layer data fusion, false detection elimination and modeling sample identification; Step S7: diameter at breast height prediction model construction; Step S8: diameter at breast height data detection and completion; Step S9: single tree segmentation and tree height extraction.
2. The method according to claim 1, wherein the method comprises the following steps: In step S1, the original three-dimensional point cloud data of the measured stand is obtained by ground-based laser radar; the ground-based laser radar includes a station-mounted laser radar and a mobile laser radar; When using the station-mounted laser radar, a multi-station splicing scanning mode is adopted, and the layout of the scanning station follows the principles of no omission and full coverage and high overlap, so that each tree in the stand can be detected from at least two different stations to obtain complete point cloud data; When using the mobile laser radar, the operator walks along the planned path while carrying or holding the device; the planned path covers the measured stand and adopts a closed loop or "S" shape path, while ensuring that the key areas of the measured stand are passed through to obtain complete and uniform point cloud data; In step S1, the obtained original point cloud data is subjected to registration processing to obtain the overall point cloud of the measured sample plot; The overall point cloud is given actual accurate coordinates in the following ways: Method one: in the scanning area, three or more control points or known survey points are laid out, and the actual three-dimensional coordinates thereof are obtained by using high-precision measurement technology; the relative coordinates of these points in the point cloud and the corresponding measured absolute coordinates are used to correct the point cloud to the target coordinate system through a coordinate conversion model; Method two: airborne point cloud of the same forest area is obtained by visible light unmanned aerial photogrammetry or laser radar unmanned aerial vehicle, and the airborne point cloud is used as a reference to register and unify the coordinates with the ground point cloud; After obtaining the overall point cloud with actual accurate coordinates, the point cloud is subjected to denoising, ground point classification, and a digital elevation model is generated based on the classified ground points, and then the overall point cloud is subjected to height normalization processing.
3. The method according to claim 2, wherein the method comprises the following steps: 1) obtaining the point cloud data of the forest; 2) preprocessing the point cloud data; 3) extracting the single tree parameters from the point cloud data based on the deep learning model. In step S2, the height interval containing the standard diameter at breast height position 1.3 meters and at least two auxiliary height layers are extracted from the normalized point cloud; The selection method of the auxiliary height layer follows the following principles: the height of the auxiliary height layer is higher than the influence height of the understory shrub and grass vegetation, and the auxiliary height layer is located in the interval where the trunk is straight and has not started branching, so as to avoid the interference of the crown structure.
4. The method according to claim 3, wherein the method comprises the following steps: In step S3, each point cloud slice layer is converted into a two-dimensional raster image, and the image is uniformly scaled to a preset size; a multi-start point translation cropping method is adopted to overcome the truncation effect of the image boundary on the tree trunk: In the fixed spatial range of each slice layer, the original point (0, 0) and a preset offset are used as the starting points to generate three sets of images; In the imaging process, according to the planar distribution range of the slice layer point cloud in the global coordinate system, an image file name containing the geographical coordinate boundary is generated, the boundary data includes minimum X, maximum X, minimum Y and maximum Y information, so as to establish the accurate mapping relationship between the image pixel coordinates and the actual coordinates.
5. The method according to claim 4, wherein: In step S4, the image labeling tool is used to label the tree trunk cross section in the converted image to form a training data set; the target detection neural network model is trained by using the training data set to obtain a trunk diameter detection model; In step S5, the three sets of images obtained in step S3 are input into the trained detection model to obtain the boundary box and the confidence of each detected tree cross section; according to the center pixel coordinates of the boundary box in the image, combined with the coordinate range information embedded in the image file name, the two-dimensional plane position of the tree cross section in the actual coordinate system is calculated; at the same time, according to the size proportion relationship of the boundary box, the diameter of the tree is calculated; the shortest distance between each detection box and the image boundary is calculated; if the distance is less than a preset threshold, it is determined that the target is truncated by the image boundary, and this detection result is directly discarded; for multiple valid detection results of the same tree in the translated image, the average value of the calculation indicators is taken as the final result of the tree in the slice layer.
6. The method of claim 5, wherein the method comprises: In step S6, the trees detected in different height layers are associated and matched; a plane distance threshold is set, and if the center point plane positions of the detection boxes in different height layers are less than the threshold, they are determined to belong to the same tree; For multiple detection results determined as the same tree, only when they are successfully detected in at least two height layers, they are finally confirmed as valid standing trees; in this process, those standing trees that are successfully detected in the standard breast height layer and all preset auxiliary height layers are identified and marked as modeling samples for subsequent breast diameter prediction model training.
7. The method of claim 6, wherein the method comprises: In step S7, the data of the modeling samples identified in step S6 are used to construct a breast diameter prediction model, which is based on the tapering rule of the tree trunk and uses the trunk diameter detected at a higher and unobstructed height to predict the breast diameter value caused by the point cloud missing due to undergrowth obstruction. The method comprises the following steps: Step S7.1, data preparation and scene adaptation; In the modeling sample, the trunk diameter DBH of the standard breast height layer is extracted as the dependent variable, based on the stand structure characteristics, two auxiliary height layers (h1, h2) are preset, the height selection follows the following principles: the auxiliary height layer is higher than the influence height of the undergrowth shrub and grass vegetation, and the auxiliary height layer is located in the interval where the trunk is straight and has not started branching to avoid the interference of the crown structure, the trunk diameters d_h1 and d_h2 detected at the two auxiliary height layers are used as independent variables to construct a training data set; Step S7.2, model training; Two types of regression models are trained and compared to adapt to different stand structures and precision requirements, both types of regression models use K-fold cross-validation for performance evaluation, and the coefficient of determination R² and the root mean square error RMSE are used as core indicators to quantify the fitting degree and prediction accuracy of the model; The regression model scheme one is a linear model, specifically: a multiple linear regression is used as a basic model, and is expressed by a formula as follows: wherein, DBH represents a diameter at breast height value, d_h1 and d_h1 represent trunk diameters detected at heights h1 and h2 respectively, β0 is an intercept of the regression model, β1 and β2 are partial regression coefficients of independent variables d_h1 and d_h2 respectively, the least square method is used to estimate regression coefficients β0, β1 and β2, and minimizing the residual sum of squares between a predicted value and an actual detected value is used as an optimization target; This model scheme assumes that there is a stable linear relationship between the diameter at breast height and the auxiliary height diameter, and is suitable for artificial forests or single tree species communities with straight trunks and uniform degree of cutting The regression model scheme two is a tree model, specifically: for a complex forest stand scene, machine learning models of random forest regression and gradient boosting decision tree GBDT are introduced, a plurality of decision trees are constructed, and prediction results thereof are integrated, a complex nonlinear relationship and an interaction effect between the features d_h1 and d_h2 and the target DBH are automatically captured, This model scheme is more suitable for natural forests or mixed forest scenes with large trunk shape variation and complex cutting degree rules; In this step, by comparing and analyzing the performance of two different regression models, the R², RMSE and model complexity are comprehensively considered, the highest prediction accuracy and the best stability are used as the standard, and the best diameter at breast height prediction model is determined.
8. The method according to claim 7, wherein the method comprises the following steps: 1) obtaining the point cloud data of the forest; 2) preprocessing the point cloud data; 3) extracting the single tree parameters from the point cloud data based on the deep learning model. In step S8, for the standing tree confirmed to be valid, first, the diameter calculated at the standard diameter layer in step S6 is directly used as the diameter at breast height value; if the diameter at breast height value is missing, data completion is performed: the diameters d_h1 and d_h2 detected at higher height layers are substituted into the diameter at breast height prediction model trained in step S7 to calculate the diameter at breast height value.
9. The method according to claim 8, wherein: In step S9, the real plane position of each valid standing tree determined is used as a seed point, a single tree segmentation method based on the seed point is used to separate the complete point cloud of a single tree; based on the single tree point cloud obtained by segmentation, the difference between the Z coordinates of the highest point and the lowest point of the single tree point cloud is taken as the tree height of the single tree.
10. The method of claim 9, wherein the method comprises: After the complete three-dimensional point cloud of a single tree is separated, the tree height and crown width parameters are obtained from the point cloud data.