CN-121978655-A - Automatic inversion method for standing tree cutting degree parameters based on foundation laser radar and deep learning

CN121978655ACN 121978655 ACN121978655 ACN 121978655ACN-121978655-A

Abstract

The invention provides an automatic inversion method for standing tree cutting degree parameters based on foundation laser radar and deep learning, and relates to the technical field of forest tree measurement. The method comprises the steps of obtaining an original three-dimensional laser point cloud based on multi-station scanning and obtaining a clean trunk point cloud. And (3) generating a high-density diameter-height observation data set by axial skeletonizing extraction and slicing treatment and combining circle fitting calculation. And constructing a variable parameter degree equation model which introduces a relative tree height ratio and an aspect ratio factor, and constructing a residual square sum objective function. And (3) iteratively solving an optimal shape parameter vector by using a nonlinear least square algorithm, and outputting model parameters of the target standing tree when the precision evaluation index is satisfied. The invention effectively solves the problem of measurement errors caused by the shielding of branches and leaves in the complex under-forest environment, replaces the destructive investigation mode of the traditional felling analytic tree, realizes the nondestructive, high-precision and automatic inversion of the cutting degree parameters of the standing tree, and improves the efficiency and the digital level of forest resource monitoring.

Inventors

MA FENGFENG
ZHOU XIANG
ZHANG XIE
LIU ZHENHUA
PI BING
XIAO YAQIN
SONG QINGAN

Assignees

湖南省林业科学院
湖南省自然资源事务中心

Dates

Publication Date: 20260505
Application Date: 20260113

Claims (10)

1. An automatic inversion method for standing tree cutting degree parameters based on foundation laser radar and deep learning is characterized by comprising the following steps: Acquiring an original three-dimensional laser point cloud of a target standing tree based on a preset multi-station scanning scheme, and performing semantic segmentation on the original three-dimensional laser point cloud to obtain a clean trunk point cloud only comprising trunk surfaces; Calculating a tree height value of the clean trunk point cloud; Performing axial skeletonized extraction on the clean trunk point cloud to obtain a trunk central axis, and performing slicing treatment on the clean trunk point cloud along the trunk central axis at preset slicing intervals to obtain a plurality of trunk point cloud slices at different heights; Performing geometric center positioning and circle fitting calculation on each trunk point cloud slice to obtain corresponding section diameters and slice heights, and combining all the section diameters, the slice heights and the tree height values to obtain a high-density diameter-height observation data set; constructing a variable parameter degree equation model introducing a relative tree height ratio; Substituting the high-density diameter-height observation data set into the variable parameter degree equation model, and constructing a residual square sum objective function representing the deviation between an observation value and a predicted value by taking a parameter vector of a to-be-determined shape as an independent variable; And carrying out iterative solution on the residual square sum objective function by using a nonlinear least square algorithm to obtain an optimal shape parameter vector, generating a predictive cut curve based on the optimal shape parameter vector, and outputting parameters of a variable parameter cut equation model of the target standing tree when an accuracy evaluation index of the predictive cut curve meets a preset threshold.
2. The method for automatically inverting the skiving degree parameter of the standing tree based on the ground-based laser radar and the deep learning according to claim 1, wherein the method for acquiring the original three-dimensional laser point cloud of the target standing tree based on the preset multi-station scanning scheme and performing semantic segmentation on the original three-dimensional laser point cloud to obtain the clean trunk point cloud only comprising the trunk surface comprises the following steps: A plurality of laser radar scanning sites with overlapping view fields are distributed in the circumferential direction of the target standing tree so as to obtain single-station laser point clouds with multiple view angles, and all the single-station laser point clouds are unified to the same coordinate system by utilizing a point cloud registration algorithm to obtain a registration point cloud set; Denoising preprocessing and data fusion operation are carried out on the registration point cloud set, so that the original three-dimensional laser point cloud is obtained; Establishing a semantic segmentation network based on a point cloud deep learning architecture, and performing supervised training on the semantic segmentation network by utilizing a point cloud training set containing trunk labels and non-trunk labels to obtain a trained trunk semantic recognition model; Inputting the original three-dimensional laser point cloud into the trained trunk semantic recognition model, extracting high-dimensional geometric features of the original three-dimensional laser point cloud, and calculating class probability of each point cloud data point based on the high-dimensional geometric features; assigning semantic tags to each point cloud data point according to the category probabilities; extracting a point set marked as a trunk label from the original three-dimensional laser point cloud according to the semantic label to obtain a preliminary trunk point cloud; performing statistical outlier removal processing on the preliminary trunk point cloud to obtain the clean trunk point cloud only comprising trunk surfaces; The expression of the trunk semantic recognition model is as follows: ; Wherein, the For a given input point cloud Outputting probability distribution of each point belonging to trunk class; a high-dimensional geometric feature matrix for the input original three-dimensional laser point cloud data; Is as the parameter of Is a deep neural network nonlinear mapping function; The method comprises the steps of learning a weight and bias parameter set through training in a network, wherein Softmax is a normalized exponential function; predictive semantic tags for point cloud data points.
3. The automatic inversion method of standing tree skiving parameters based on foundation laser radar and deep learning according to claim 1, wherein the calculation expression of the tree height value of the clean trunk point cloud is: ; Wherein, the The tree height value of the clean trunk point cloud is calculated; Is the top end position of the tree tip; is the root neck starting position.
4. The method for automatically inverting the vertical wood cutting degree parameter based on the ground-based laser radar and the deep learning according to claim 1, wherein the axially skeletonizing extraction is performed on the clean trunk point cloud to obtain a trunk central axis, and the clean trunk point cloud is sliced along the trunk central axis at a preset slice interval to obtain a plurality of trunk point cloud slices at different heights, comprising: Performing equidistant rough layering on the clean trunk point cloud in the Z-axis direction of a global coordinate system to obtain a plurality of initial layering subsets; Calculating the geometric centroids of each initial hierarchical subset to obtain a hierarchical centroid point set, performing space curve fitting on the hierarchical centroid point set by utilizing a cubic B-spline interpolation algorithm to obtain a three-dimensional curve penetrating through the interior of the clean trunk point cloud, and marking the three-dimensional curve as the trunk central axis; equidistant discretization sampling is carried out on the central axis of the trunk according to the preset slice spacing, so as to obtain a series of axis sampling nodes which are orderly arranged; calculating a tangential vector of each axis sampling node on the central axis of the trunk, and constructing a local normal plane which passes through the axis sampling nodes according to the tangential vector and is perpendicular to the tangential vector; and setting a preset slice thickness threshold value by taking the local normal plane as a reference, retrieving and extracting all point cloud data points which are positioned at two sides of the local normal plane and have a distance smaller than the slice thickness threshold value from the clean trunk point cloud, and defining the extracted point cloud data point set as the trunk point cloud slice at the current height.
5. The method for automatic inversion of standing tree cutting parameters based on ground-based lidar and deep learning according to claim 1, wherein the performing geometric center positioning and circle fitting calculation on each trunk point cloud slice to obtain a corresponding section diameter and slice height, and combining all the section diameters, the slice heights and the tree height values to obtain a high-density diameter-height observation data set comprises: Constructing a local projection plane of each trunk point cloud slice; projecting all three-dimensional point cloud data points in the trunk point cloud slice onto the local projection plane to obtain a group of two-dimensional projection point sets; carrying out iterative regression calculation on the two-dimensional projection point set by using a least square circle fitting algorithm to obtain a fitting circle parameter with the minimum sum of distance errors; extracting a diameter value in the fitting circle parameter as the section diameter, and marking a circle center coordinate in the fitting circle parameter as the geometric center of the slice; acquiring a bottom end starting point coordinate of the trunk central axis, calculating a curve path length from the bottom end starting point to the geometric center along the trunk central axis, and defining the curve path length as the slice height; And traversing all the trunk point cloud slices, extracting the corresponding section diameters and the slice heights to form a plurality of groups of diameter-height data pairs, and packaging all the diameter-height data pairs and the tree height values according to a preset data structure to obtain the high-density diameter-height observation data set.
6. The automatic inversion method of standing tree skiving parameters based on foundation laser radar and depth learning according to claim 1, wherein the expression of the variable parameter skiving equation model is: ; the diameter of the section of the trunk at any slice height; is the chest diameter value of the trunk; Is a tree high value; The absolute height of the current slice from the root neck of the trunk; in order to be a relative tree height ratio, For each element in the pending shape parameter vector.
7. The method for automatic inversion of standing tree skiving parameters based on ground-based lidar and deep learning according to claim 6, wherein substituting the high-density diameter-height observation data set into the variable parameter skiving equation model and constructing a residual square sum objective function representing the deviation between an observation value and a predicted value by using a pending shape parameter vector as an independent variable comprises: Traversing the high-density diameter-height observation data set, and extracting the tree height value and the chest diameter value corresponding to the section diameter; calculating the ratio of the height of each slice to the tree height value to obtain a relative tree height ratio, and calculating the ratio of the chest diameter value to the tree height value to obtain a height diameter ratio factor; constructing a nonlinear function framework comprising a base term and an index term; Substituting the relative tree height ratio and the height-diameter ratio factor into the index term as known quantities, and linearly combining each element in the undetermined shape parameter vector as a weighting coefficient with a polynomial of the relative tree height ratio to generate a dynamic index expression containing undetermined coefficients; substituting the chest diameter value, the tree height value, the slice height and the dynamic index expression into the nonlinear function framework to generate a diameter prediction symbol expression for the undetermined shape parameter vector; Calculating the difference value between the section diameter and the diameter prediction symbol expression in the high-density diameter-height observation data set to obtain a single point residual error item, and performing square summation operation on the single point residual error items at all slice positions to generate a residual error square sum objective function taking the undetermined shape parameter vector as a unique variable; the diameter prediction symbol expression is: ; The expression of the residual square sum objective function is: ; Wherein, the The slice height of the ith slice, N is the total number of slices; In order to determine the shape parameter vector in the pending, For a theoretical predicted diameter at the ith slice position, Is the cross-sectional diameter of the trunk at the ith slice height.
8. The automatic inversion method for the standing tree cutting parameters based on the ground-based laser radar and the deep learning is characterized in that the base term represents the geometric attenuation law of the relative height change of the trunk, and the index term represents the dynamic adjustment law of the trunk shape along with the height change.
9. The method for automatic inversion of the vertical skiving parameters based on the ground-based lidar and the deep learning according to claim 1, wherein the iterative solution is performed on the residual square sum objective function by using a nonlinear least square algorithm to obtain an optimal shape parameter vector, a predictive skiving curve is generated based on the optimal shape parameter vector, and when an accuracy evaluation index of the predictive skiving curve meets a preset threshold, parameters of a variable parameter skiving equation model of the target vertical skiving are output, and the method comprises the following steps: setting an initial iteration initial value of the parameter vector of the shape to be determined; Constructing an initial parameter vector according to the initial iteration initial value, and inputting the initial parameter vector and the residual square sum objective function into a nonlinear least square solver; calculating a jacobian matrix of the residual square sum objective function with respect to the pending shape parameter vector using the nonlinear least squares solver; Constructing an incremental normal equation based on the jacobian matrix, and locking a finally converged parameter vector into the optimal shape parameter vector by iteratively updating the initial parameter vector to minimize the function value of the residual square sum objective function until the decreasing amplitude of the function value is smaller than a preset convergence tolerance; substituting the optimal shape parameter vector back into the variable parameter cutting degree equation model, and calculating a theoretical diameter value corresponding to each slice position by taking a relative tree height ratio in the high-density diameter-height observation data set as an input variable; constructing the predictive cutting degree curve according to the theoretical diameter values of all the slice positions; extracting the true section diameter in the high-density diameter-height observation data set; Calculating the statistical difference between the true section diameter and the theoretical diameter value to obtain the precision evaluation index comprising a determination coefficient and an estimated value standard deviation; Comparing the precision evaluation index with a preset precision acceptance threshold, if the determination coefficient is larger than a first threshold and the standard deviation of the estimated value is smaller than a second threshold, judging that the model inversion is effective, extracting each element value in the optimal shape parameter vector, and outputting the element value as a parameter of the variable parameter degree equation model of the target standing tree; the expression of the delta normal equation is: ; wherein J is a jacobian matrix of a residual square sum objective function with respect to a pending shape parameter vector; a transpose of jacobian; Is a damping coefficient; Is a unit matrix; And updating the increment for the parameter vector of the current iteration step, wherein r is a residual vector consisting of single-point residual items of all slice positions.
10. The method for automatic inversion of standing tree skiving parameters based on ground-based lidar and depth learning according to claim 9, wherein constructing the predictive skiving curve from the theoretical diameter values of all slice positions comprises: establishing a one-to-one correspondence mapping relation between the theoretical diameter value and the slice height; based on the one-to-one mapping relation, packaging each group of corresponding theoretical diameter values and the slice height into a two-dimensional coordinate point to obtain a discrete theoretical coordinate point set; according to the numerical value of the slice height, performing ascending arrangement operation on the discrete theoretical coordinate point set to obtain an ordered point sequence which monotonically increases in the height direction; carrying out continuous processing on the ordered point sequence by utilizing a cubic spline interpolation algorithm, and constructing a smooth geometric locus connecting all the theoretical coordinate points; Defining the smooth geometric locus as the predicted skiving curve.

Description

Automatic inversion method for standing tree cutting degree parameters based on foundation laser radar and deep learning Technical Field The invention relates to the technical field of forest measurement, in particular to an automatic inversion method for a standing tree cutting degree parameter based on foundation laser radar and deep learning. Background The standing tree cutting degree equation is a mathematical model describing the change rule of trunk diameter along with height, and is the basis for accurately calculating the standing tree volume, biomass and yield. The current construction or calibration of the degree of sharpness equation is generally based on related forestry standards, and mainly adopts a traditional analytical wood investigation method. The method requires felling down the sample wood, manually measuring the diameters of the strip skin and the peel according to the sections, and then fitting parameters by using statistical software. While existing binary volume models or simple degree of cut equations can satisfy the basic estimation, it is often necessary to employ variable parametric degree of cut equations that contain relative tree height ratios when dealing with trunk morphology with significant nonlinear features. With the development of precision forestry, the requirements on timeliness and nondestructivity of single wood parameter measurement are higher and higher. The foundation laser radar technology is an important means for forestry investigation because of the capability of rapidly acquiring the high-density three-dimensional point cloud data of the forest. Meanwhile, a deep learning algorithm in the field of computer vision breaks through in the aspect of three-dimensional point cloud semantic segmentation, so that automatic separation of trunks and branches and leaves from complex forest scenes is possible. The tree trunk diameter sequence is directly extracted by utilizing the foundation laser radar data so as to invert the parameters of the cutting equation, thereby being a necessary trend for replacing the traditional destructive sampling. However, the existing forest tree measurement technology based on the foundation laser radar still has obvious defects. Firstly, the degree of automation of data processing is low, the precision is limited, in a complex under-forest environment, the crown shielding and the branch adhesion cause large point cloud noise, the traditional geometric fitting algorithm is difficult to accurately separate trunks and branches and leaves, the upper diameter extraction error is large, and the fitting precision of higher-order parameters in a cutting degree equation is directly influenced. Secondly, model inversion lacks pertinence, the existing method focuses on direct calculation of volume, a mechanism for combining high-density point cloud data with a specific mathematical model conforming to a regional standard is lacking, standardized model parameters conforming to forestry operation management requirements cannot be output, and therefore measurement results are difficult to directly apply to an existing forest resource management system. Thirdly, the sampling cost is high and the environment is damaged, the traditional method relies on lumbering, forest resources are damaged, the sample collection period is long, the cost is high, and the cutting degree parameters of specific tree species are difficult to update rapidly in a large range. Disclosure of Invention In order to overcome the defects of the prior art, the invention aims to provide an automatic inversion method for standing tree cutting degree parameters based on foundation laser radar and deep learning, which solves the problems that the traditional analytic tree method damages resources and has low efficiency, and the traditional laser radar technology has poor segmentation precision in a complex environment and is difficult to invert standardized model parameters in the prior art. In order to achieve the above object, the present invention provides the following solutions: an automatic inversion method for standing tree cutting degree parameters based on foundation laser radar and deep learning comprises the following steps: Acquiring an original three-dimensional laser point cloud of a target standing tree based on a preset multi-station scanning scheme, and performing semantic segmentation on the original three-dimensional laser point cloud to obtain a clean trunk point cloud only comprising trunk surfaces; Calculating a tree height value of the clean trunk point cloud; Performing axial skeletonized extraction on the clean trunk point cloud to obtain a trunk central axis, and performing slicing treatment on the clean trunk point cloud along the trunk central axis at preset slicing intervals to obtain a plurality of trunk point cloud slices at different heights; Performing geometric center positioning and circle fitting calculation on each trunk point cloud slice to obta