AU2020104291A4

AU2020104291A4 - Singletree segmentation method based on chord angle discriminant clustering for layered LIDAR point cloud

Info

Publication number: AU2020104291A4
Application number: AU2020104291A
Authority: AU
Inventors: Yexuan Mu; Guoqing Zhou
Original assignee: Guilin University of Technology
Current assignee: Guilin University of Technology
Priority date: 2020-12-23
Filing date: 2020-12-23
Publication date: 2021-03-11
Anticipated expiration: 2028-12-23

Abstract

The invention relates to the field of single tree segmentation of LIDAR point cloud of forest remote sensing, and discloses the chord angle discriminative clustering method of layer by layer LIDAR point cloud, and invents the geometric relationship discriminant clustering method, including 1. Echo information processing of airborne lidar forest point cloud data, extracting vegetation point cloud data; 2. Layer by layer semi chord value discrimination clustering segmentation, which is generally applicable to single tree segmentation among multiple trees; 3. Step by step Layer half chord angle assisted clustering segmentation helps to segment the two partially overlapped trees in 2; 4. Set threshold for the occlusion phenomenon of two trees with "big cover and small" in 3, and cluster according to the height relationship between the big tree and the small tree; 5. Jointly judge the separated individual trees and correct or delete the misjudged ID single trees. The invention provides a segmentation method of geometric discriminant clustering based on circular chord value and angle for forest point cloud data. Compared with the CHM method, the invention can obtain single tree segmentation shape more quickly and accurately, reduce error segmentation of single tree, and contribute to forest vegetation parameter inversion and tree 3D model establishment in forestry research. 3/4 p. B Figure 4 Figure 5

Description

3/4

B

p. Figure 4

Figure 5

Singletree segmentation method based on chord angle discriminant clustering for layered

LIDAR point cloud

TECHNICAL FIELD

The invention relates to the LIDAR point cloud single tree segmentation field of forest remote

sensing, specifically uses the segmentation method of multi-echo extraction plant point cloud data

and layer by layer chord angle discrimination clustering, analyzes and solves the problem of

multi-tree clustering relationship from three aspects.

BACKGROUND

In recent years, with the implementation of a series of key forestry projects in China, the

plantation area has gradually approached 30% of the global plantation area. The development of

artificial forests can effectively solve the problems of forest consumption, wood supply shortage,

and forest ecological environment damage, and also plays an important role in improving forest

carbon sink, maintaining regional ecological balance, and accelerating forest restoration. At present,

China's plantation management has changed from simply pursuing economic benefits to focusing

on the combination of multiple benefits and sustainable development, and its monitoring purpose

has also shifted from traditional forest resource utilization to protecting the ecological environment,

maintaining forest health, and balancing forest carbon storage. Efficient and high-precision

monitoring of plantation forest resources is becoming more and more important. A single tree is the

basic unit of a forest. Its spatial structure, biophysical and chemical components are the key factors

for forest resource investigation and ecological environment modeling. Lidar can quickly and

accurately obtain single tree information, provide an effective guarantee for forest resource

monitoring and management, reasonably use various single tree segmentation algorithms to

improve single tree segmentation accuracy, and provide important algorithm and technical support

for accurately obtaining single tree spatial position, canopy structure, and mastering tree

competition and health status.

At present, the traditional single tree segmentation method usually adopts the manual

measurement method, and can only obtain the point data, so it is difficult to obtain the regional or

larger scale data; however, the most commonly used single tree segmentation method in the world is

based on the canopy height model (CHM) of the vegetation canopy. The general process of

obtaining the CHM model is to make difference between the digital surface model (DSM) and

digital elevation model (DEM) corresponding to LIDAR point cloud data. Then, the maximum

value of local CHM is located, and the crown width of a single tree can be identified in a certain

range, to extract the tree. When CHM is used for segmentation, it often results in wrong

segmentation and improper shape segmentation.

To ensure the smooth processing of the point cloud and the accuracy of single tree

segmentation, it is necessary to improve and perfect the above problems.

SUMMARY

The purpose of the invention is to provide a method for single tree segmentation of LIDAR

point cloud data, which can effectively distinguish the subordination relationship of multiple tree

point clouds, improve the accuracy rate of single tree segmentation, avoid the situation of wrong

segmentation, and provide a strong guarantee for the correct extraction of single tree segmentation

parameters in the later stage.

To realize the purpose of the invention, the invention proposes a multi-tree clustering

segmentation method based on chord angle relationship discrimination: Taking the highest point of

the target tree and the neighbor tree as the vertex, establishing the spatial geometric relationship

between the target tree and the neighbor tree, drawing a circle with the vertex of the target tree as

the center of the circle, the horizontal distance between the sample point and the vertex of the target

tree is half chord length, Then half chord pair is defined as half chord angle according to 1 / 2 of

circle center angle. Similarly, the neighbor tree establishes a corresponding chord angle relationship.

From the highest point to the lowest point, the above method is used to discriminate and cluster the

two trees layer by layer. This method is suitable for the terrain with slope, and the point cloud data

is normalized and segmented by single tree clustering. This method can cluster the 3D point cloud

by chord angle discrimination in spatial geometry, which provides a new idea for the correct segmentation of a single tree.

The growth state of the tree changes according to the terrain and other environmental factors.

The present invention considers three situations in the clustering relationship of adjacent trees of

many trees.

Case 1: clustering segmentation of hierarchical semi chord discriminant is generally

applicable to single tree segmentation among multiple trees;

Case 2:it uses half chord angle to assist clustering segmentation. In case 1, there may be a

special neighbor tree relationship to continue the segmentation, that is, the two partially overlapped

trees are actually one tree;

Case 3: it compresses to two-dimensional clustering, ignores the index in X (or y) direction,

and performs threshold re judgment for the single tree in case 2. Since two temporary trees with

"wide coverage and small size" may appear in case 2, this method sets the maximum threshold and

minimum threshold of the number of point clouds and assists method 2 to identify the small trees

covered. In this case, according to the trunk height of the big tree, the method sets the maximum

threshold and minimum threshold of the number of point clouds. The relationship between (HA)

and small tree height (HB) was used to judge the attribute of the sample points.

The beneficial effect of the invention is: it is carried out the single tree segmentation by using

multi-tree clustering method based on chord angle relationship discrimination, which mainly

includes two parts: extracting plant point cloud and clustering discrimination. On the one hand, extracting vegetation can not only reduce the number of footpoints participating in clustering but

also enhance the filtering effect by using multiple echo information of point cloud data. On the

other hand, the geometric chord angle discrimination of three-dimensional point cloud used

clustering can effectively distinguish the subordination of multi-tree point cloud and improve the

accuracy of single tree segmentation.

BRIEF DESCRIPTION OF THE FIGURES

Figure 1 is a flow chart of the overall division of the present invention;

Figure 2 shows the result of multi-echo vegetation point cloud extraction;

Figure 3 is a schematic diagram of semi chord length discriminant clustering;

Figure 4 is a schematic diagram of discriminative clustering of half chord angle;

Figure 5 is a schematic diagram of height threshold discriminant clustering;

Figure 6 is the resulting graph of the initial semi chord discriminant clustering;

Figure 7 shows the result of correcting the error code tree.

DESCRIPTION OF THE INVENTION

Step 1. Analyze the characteristics of data echo

As shown in the flow chart of Figure 1, it firstly summarizes the echo information of the data

in the invention, and the echo of different times can reflect different ground feature information.

According to the characteristics of multiple echoes of the LIDAR point cloud in forest area, the first

echo mainly comes from the dense and tall plant canopy and leaves close to the canopy, the

intermediate echo mainly comes from the branches or low vegetation of high vegetation, and the

last echo comes from the laser footpoints reflected from the earth's surface. For urban areas, the

single echo data mainly come from the surface, the top or wall of artificial buildings (including

overpasses and bridges across rivers), and a small number of vegetation points; the first echo comes

from the canopy of vegetation and the edge of artificial buildings (including overpasses); the

intermediate echo mainly comes from the branches and leaves of vegetation and the facade of

buildings; The last echo mainly comes from the surface, and some of them are from the complex

roof of buildings and the branches and leaves of low vegetation layer.

Step 2: it extracted vegetation points by echo characteristics

This patent mainly aims at forest data processing. This patent uses the forest LIDAR point

cloud data contained 22019 points, including 4 echoes in total, to complete this work. Among them,

there are 15365 laser spots in the first echo and 1190 laser spots in the last echo. After that, it uses

lidar 360 software to classify LIDAR point cloud data according to the echo times of points cloud

and the echo characteristics of forest lidar data targeted by this patent. It summarized the method of extracting vegetation point cloud data and simplified a vegetation point cloud extraction formula, which is expressed as follows:

Vegetation point cloud = first echo point + intermediate echo point - last echo point

Step 3: it carries out the clustering of semi chord value discriminant among multiple trees

As shown in Figure 3, it defines the highest point of the target tree as vertex A, and the highest

point of the neighbor tree as vertex B, draws circles A and B with the connecting lines of sample

points A and B as the radius, and the half chords are 01C and0 2D. Generally, the crown diameter is

1.5m when the trunk is 50mm thick, and the crown diameter is 3m when the trunk is 125mm.

Therefore, referring to the clustering principle in Fig. 3, iterative clustering is started with the vertex

A of the target tree and vertex B of neighbor tree B. When the horizontal distance X between a point

and the top of the target tree x > 3 / 2 (i.e. 1.5m), the point is directly identified as tree point B;

when the horizontal distance x from a is x < 1.5/2 (i.e. 0.75), it is directly determined as tree point

A; when 0.75 < x < 1.5, it is necessary to judge whether the horizontal distance of the point is

closer to point A or point B. As shown in Fig. 4, the half chord angles of A and B tree respectively

are 01 and 02. If the sample point is closer to tree B, that is, 02C <0 1 C, the sample point

belongs to tree B, otherwise it belongs to tree A.

Step 4: semi chord angle assisted clustering

Based on step 3, the case of 0.75m < x < 1.5m is supplemented. As shown in Figure 4, tree A

and tree B have local overlapping points. If the half chord length relationship of two circles with a

and B as the vertices is 02C < 0 1C, there is a half chord angle 92>01. If the elevation

relationship between tree A and tree B is Ha > Hb, then tree B belongs to tree A.

Step 5: "big cover small" height threshold discriminant clustering

In step 4, the threshold value of the tree segmented from the single tree is re-judged. Since

there may be obstacles in step 4 that "big trees cover small trees", the method assists step 4 to

identify the small trees covered. Set the maximum threshold and minimum threshold of the number

of point clouds, the minimum threshold is 2137.505, and the maximum threshold is 2459.683,

When it is greater than the minimum threshold and less than the maximum threshold, it is regarded as "big cover small" obstacle. In this case, the relationship between the height of a big tree trunk (HA) and a small tree height (HB) is ha & Hb. Therefore, the threshold value is set according to the tree trunk ha. The points higher than the threshold belong to tree a, and the points lower than the threshold belongs to tree B.

Step 6. Clear and correct the wrong log ID:

According to the method of step 3, the LIDAR point cloud is segmented by single tree clustering. The result of segmentation is shown in Figure 6. There are only a few points in some ID trees in the figure. The difference between the elevation of the vertex and the surrounding ID tree is more than 20m, and the horizontal distance (x) is 0.75m < x < 1.5m. Therefore, the ID of these trees is removed, and the chord angle discriminant clustering method in step 4 is used to re-cluster the trees with ID removed Cut.

Step 7. calculate the segmentation accuracy rate:

According to the clearing and correction of the wrong ID tree in step 6, the correct single tree segmentation result diagram in Figure 7 is obtained. There are 61 trees in the forest experimental data to verify the accuracy of the invention, and 59 single tree trees are obtained by the chord angle discrimination clustering method proposed in the invention, so the segmentation accuracy of the invention is as high as 97%. Therefore, the invention can effectively distinguish the subordination

relationship of multi-tree point clouds, improve the accuracy rate of single tree segmentation, avoid the situation of wrong segmentation, and provide a strong guarantee for the correct extraction of single tree segmentation parameters in the later stage.

Claims

1. The method is characterized by the following steps:

1) According to the characteristics of multiple echoes of LIDAR point cloud data in forest area,

vegetation point cloud data are extracted;

2) Based on the analysis of the spatial geometric relationship between the target tree and the

neighbor tree, three kinds of single tree clustering segmentation models are established: semi chord

value discriminant clustering, semi chord angle assisted discriminant clustering, and "big cover

small" height threshold discriminant clustering. The three clustering segmentation methods are

applied to three kinds of multi-tree relationships: general tree relationship, local overlapping tree

relationship, and large cover small tree relationship;

3) Correct the wrong judgment and clear the wrong single tree ID;

4) Compare the number of trees between the segmentation results and the sample data, and

calculate the segmentation accuracy.

2. according to claim 1, the characters of this method in step 1) specifically comprises:

the first echo mainly comes from the dense and tall plant canopy and leaves close to the

canopy, the intermediate echo mainly comes from the branches or low vegetation of high vegetation,

and the last echo comes from the laser footpoints reflected from the earth's surface. The extracted

vegetation point cloud is summarized as the following formula:

According to the echo feature formula, vegetation point cloud data are extracted.

3. According to claim 1, the characterized of the method in step 2) specifically comprises:

(1)The spatial geometric relationship between the target tree A and the neighbor tree B is

established by taking the highest point of the target tree and the neighbor tree as the vertex. The

circle is drawn by taking the vertex A of the target tree as the center of the circle and the connecting

line between the sample point and the center of the circle as the radius. The horizontal distance 01C between the sample point and the top of the target tree is the half chord value, and the half chord angle is 01. In the same way, the neighbor tree establishes the corresponding relationship between the half chord value 0 2 D and the half chord angle 02, and classifies the canopy layer by layer from the highest to the lowest. According to the empirical value, two half chord thresholds are set. When 1C is greater than the maximum threshold, the sample points are assigned to the neighbor tree, and when 0 1C is less than the minimum threshold, the sample points are assigned to the target tree;

(2) when the minimum threshold value is less than 0 1 C and the maximum threshold value is less than the maximum threshold value, there are local overlapping points between the target tree A and the neighbor tree B, which needs to be intervened by the half chord angle and. For the unified sample points, when 0 1C > 0 2D, < 0 If the relationship between the target tree and the neighbor tree height has Ha > Hb, then the local sample points are assigned to the target tree, that is, in this case, the neighbor tree B belongs to the target tree A;

(3) Based on the above two steps, the height threshold discriminant clustering method is used to cluster and segment the "big cover small" situation. The minimum threshold is 2137.505, and the maximum threshold is 2459.683. When the minimum threshold < the number of point clouds < the maximum threshold, it is regarded as the "big cover small" obstacle. In this case, the big tree is regarded as the target tree a, the trunk height is ha, the small tree is regarded as the neighbor tree B, the tree height is Hb, and the height relationship is Ha > Hb. Therefore, the threshold value is set according to the tree trunk Ha, and the points higher than the threshold belong to tree A, and the points lower than the threshold belong to tree B.

4. According to claim 1, the characterized of the method in step 3) specifically comprises:

The method for clearing and correcting the wrong single tree ID, which is characterized in that the initial single tree clustering segmentation is carried out according to the layer by layer semi chord value discrimination clustering method in claim 3, and then the auxiliary clustering segmentation is carried out by using the layer by layer half chord angle auxiliary clustering method and height threshold discrimination clustering method in claim 3, and the segmentation of step (1) of claim 3 does not meet the threshold range The chord angle discriminant clustering method and the height threshold discriminant clustering method in steps (2) and (3) of claim 3 are used to re cluster and segment the trees with clear ID. after that, the wrong ID tree is corrected and deleted.

5. According to claim 1, the characterized of the method in step 4) specifically comprises:

According to the removal and correction of the wrong ID tree in claim 4, the correct single tree

ID result is obtained after the correct segmentation is completed. By comparing the result under the

method of the invention with the number of trees in the sample data, the accuracy rate of single tree

segmentation can be calculated as high as 97%.

FIGURES 1/4

Figure 1