CN113222917B - DBI tree vertex detection method of airborne laser radar point cloud data CHM - Google Patents

Abstract

The invention relates to the field of laser radar data processing, discloses a DBI tree vertex detection method for a canopy height model of airborne laser radar point cloud data, and invents a method for detecting tree vertices by DBI control foreground pixels based on the canopy height canopy model , including: 1. Standardize the marker pixel generation conditions of the canopy height model; 2. Use the grayscale variation characteristics of the height canopy model to filter out the false foreground pixels; 3. Introduce the similarity judgment factor DBI for DBI-K screening of tree crowns vertex. The invention provides a tree top position detection method for a canopy height model of airborne laser radar point cloud data, which can effectively solve the threshold dependence problem of the traditional window detection tree vertex method and improve the tree vertex identification accuracy.

Description

DBI tree vertex detection method for airborne lidar point cloud data CHM

technical field

The invention relates to the field of laser radar (Light Laser Detection and Ranging, LiDAR) data processing, in particular to a DBI (Davies-Bouldin Index) tree vertex detection method of airborne laser radar point cloud data (Canopy Height Model, CHM) .

technical background

Lidar has strong penetration to the forest canopy, which is very beneficial to forest census and forest stand parameter acquisition, and can accurately identify treetop positions and single tree crowns. Airborne lidar technology provides theoretical basis and technical support for the development of precision forestry, and has broad application prospects in forest management and ecosystem research.

At present, due to the complexity of the canopy height model image generated from the airborne LiDAR point cloud, the recognition of tree vertices in the field of canopy position recognition is inaccurate. In the existing traditional methods, researchers frequently use a fixed window size to explore local maxima to detect tree tops. However, common methods based on window detection of tree vertices are easily affected by the size of the crown diameter of the tree crown. If it is too small, some trees with a larger crown radius will be allocated multiple tree vertices; otherwise, if the window size is too large, some trees with a crown radius smaller than the specified window size will not be allocated to tree vertices.

In summary, in order to solve the threshold dependence and inaccurate identification of the traditional window detection tree vertex method, it is necessary to propose a new tree vertex identification method to solve the above problems. The invention generates foreground marker pixels according to the gray level characteristics of the canopy height model, and introduces the similarity judgment factor-class spacing ratio DBI, which can solve the threshold dependence of the tree vertices of the traditional window detection height canopy model, and improve the The recognition accuracy of canopy vertices.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a tree top position detection method for airborne lidar point cloud data canopy height model, which can effectively solve the threshold dependence problem of the traditional window detection tree vertex method and improve the tree vertex recognition accuracy.

In order to achieve the purpose of the present invention, this paper proposes a DBI treetop position detection method of airborne lidar point cloud data CHM, which is realized by the following technical scheme: a DBI control foreground pixel generation based on canopy height canopy model The tree vertex method includes standardizing the marker pixel generation conditions of the canopy height model, filtering and eliminating false foreground pixels by using the grayscale variation characteristics of the canopy height model, and introducing the similarity judgment factor DBI for DBI-K (DBI-Kmeans) screening Tree canopy vertex; standard canopy height model marked pixel generation conditions: (1) the gray level values of the pixels outside the mark are lower than those inside the mark, (2) the pixels of the foreground image form a connected component, (3) the same as The pixels inside a marker have the same gray level value; the pseudo foreground pixels are filtered and eliminated by using the gray level change characteristics of the high canopy model image, and the calculated foreground pixels are combined with the gray level change characteristics of the high canopy model image. The marker pixels are further processed, and the background points that have appeared on the foreground markers are filtered according to the difference in the gray level of different images; DBI-K (DBI-Kmeans) filters the crown vertices, and performs DBI-K clustering on the foreground image. , when the DBI takes the minimum value, the optimal solution of the cluster center is obtained, which is regarded as the treetop of a single tree, and the algorithm terminates.

The beneficial effects of the present invention are: adopting a DBI based canopy height canopy model to control foreground pixels to generate tree vertices, including standard canopy height model marked pixel generation conditions, and utilizing the grayscale variation characteristics of the canopy height model to detect false The foreground pixels are filtered and eliminated, and the similarity judgment factor DBI is introduced to screen the crown vertices by DBI-K (DBI-Kmeans), which not only solves the threshold dependence problem of the traditional window detection tree vertex method, but also improves the tree vertex recognition accuracy; on the one hand , standardize the marker pixel generation conditions of the canopy height model, and use the local maxima of the canopy height model image to determine the region of interest as foreground pixels and markers, which can avoid the threshold dependence problem of perforation detection; on the other hand, The pseudo-foreground pixels are filtered and eliminated by using the gray-scale variation characteristics of the high canopy model, which overcomes the problem that the foreground marks are mixed with pseudo-maxima (foreground marks) due to texture and noise. Finally, the similarity judgment factor DBI is introduced. DBI-K is used to screen the canopy vertices, which solves the problem of unclear tree vertices based on the canopy height model, and realizes the effective identification of tree vertices.

Direct feature-based foreground marker pixel extraction will result in pseudo-maxima in foreground markers due to the existence of texture and noise.

Description of drawings

Figure 1 is a classification diagram of point cloud data filtering.

Figure 2 is the DEM, DSM, and CHM diagrams generated from point cloud data.

Figure 3 is a schematic diagram of the present invention.

Figure 4 is a foreground image and foreground markers of a canopy height model generated by the present invention under normative conditions.

FIG. 5 is a foreground marker diagram after the pseudo foreground pixels are filtered and eliminated according to the present invention.

Fig. 6 is a canopy apex diagram of the DBI-K screening of the present invention.

Detailed ways

The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

Example:

Step 1) Combined with Figure 1, the point cloud data is filtered and classified. The preservation of topographic information of forest sample plots is very important. Under this premise, in order to achieve better filtering effect, the present invention adopts a progressive triangulation filtering algorithm. The following are the steps of separating ground points:

(1) Set the maximum terrain slope: set the maximum slope of the terrain displayed in the point cloud to 80°;

(2) Setting the iteration angle: the conversion angle of all allowable values between the classification point to be surveyed and the currently known classification points on the ground is 8°;

(3) Set the iterative distance: set the distance threshold between the point to be classified and the triangulation to 1.6m;

Using the above method, point cloud filter classification is performed on the experimental plot, and the separated ground point classes are displayed in brown, and the separated vegetation point classes are displayed in green.

Step 2) Combined with Figure 2, use the filtered and classified point cloud data to generate DEM, DSM, and CHM. The DSM and DEM of the experimental forest area were generated by the comprehensive use of the TIN calculation method for the non-ground laser point cloud and the ground reflected laser point cloud. After many experiments in LiDAR 360, it is found that the triangulation generated when the critical variable length is set to 1m is smoother. When the value of the insertion buffer is set to 2m, the triangulation surface contains more details, and the weight is set to 2. The canopy height model is obtained by subtracting the DEM from the DSM as follows:

CHM=DSM-DEM (1)

Step 3) With reference to Figure 3, a DBI based canopy height canopy model controls foreground pixels to generate tree vertices, and in the canopy height model marked pixel generation under standard conditions, the canopy height model is input, and the canopy height model is generated according to three standard conditions. The foreground marker pixels of the canopy height model; the false foreground pixels are filtered and eliminated by using the gray level change characteristics of the height canopy model, and the initial foreground marker pixels of the canopy height model are input, and the false foreground pixels are filtered and eliminated according to the principle of local maximum gray level pixels. _{In front} of the foreground marker pixel C; the similarity judgment factor DBI is introduced to screen the crown vertices by DBI-K, and the foreground marker pixels from which the pseudo-markers are filtered are input and used as the initial cluster center, and a new cluster is generated according to the cluster center formula. The cluster center is compared according to the DBI value of the similarity judgment factor before and after the cluster center is generated. When the DBI is the minimum value, the optimal solution of the cluster center is obtained, and the algorithm is terminated; the cluster center when the DBI is the minimum value is output as a single tree vertices.

Step 4) Combined with Figure 4, the labeled pixels of the canopy height model are generated under normative conditions. The initial foreground markers of the canopy height model are generated according to 3 canonical conditions: (1) the gray level values of the pixels outside the marker are lower than those inside the marker, (2) the pixels of the foreground image form a connected component, (3) The pixels inside the same marker have the same gray level value; the region of interest is determined by the local maximum value of the image, which is used as the foreground pixel and marker.

Step 5) With reference to Fig. 5, the gray level variation characteristics of the high canopy model are used to filter out the false foreground pixels. The canopy area of the high-level grayscale canopy model image is brighter than other areas. The relative height information of the canopy can be known through the color depth in the grayscale image. Mark) is the local grayscale maximum value pixel; because the grayscale canopy height model is represented by the gray value of the adjacent 8 pixels of the central pixel. It is necessary to make the values of R, G, and B equal to obtain a grayscale image (where R=G=B=255 is white, R=G=B=0 is black); when the foreground pixels of the canopy height model are When the sum of the R, G, and B component values takes the maximum value and the R, G, and B component values are equal, the marked pixel value takes 0, and in other cases, the marked pixel value takes 1. Expressed as the following formula:

和

分别表示初始前景标记后的冠层高度模型图像C中像素点处的R、G、B分量值。通过以上的公式可以对图像C进行转换，得到一个新的图像C′作为最终的前景标记图像。In the formula,

and

respectively represent the R, G, and B component values at the pixel points in the canopy height model image C after the initial foreground marking. Image C can be converted through the above formula to obtain a new image C' as the final foreground marker image.

Step 6) Combined with Fig. 6, DBI-K (DBI-Kmeans) filters the canopy vertices, and on the basis of generating the foreground marked pixels of the high canopy model, the similarity judgment factor is used to further screen the marked pixels; The height model two-dimensional grayscale image is clustered, and the DBI similarity judgment factor is added to the K-means clustering method to further determine the cluster center m and the number of cluster centers K value in the foreground image.

Assuming that the foreground image of the given sample canopy height model contains the dataset A={a _i }; the n sample datasets in A are initially divided into K different clusters, to ensure that the similarity within the cluster is high, and the outside of the cluster is high. Similarity is low. Therefore, the core theory of the K-means clustering method can be described as: cluster the dataset A into a set B composed of K clusters, B={b _i }; the intra-cluster centers of each cluster subset can be obtained for:

In the formula, A represents the sample data object, and _{Ni represents the number of samples in the cluster b i .}

When the value of the similarity judgment factor DBI is the smallest, the DBI similarity judgment factor is output, and the optimal solution of the number of clusters K is obtained.

In the formula, a _i represents each data object in the ith class, m _i represents the cluster center of the ith class, N _i represents the number of data objects in the ith class, and Wi represents the number of data objects in the _ith class. The degree of dispersion of each data object, W _j represents the degree of dispersion of each data object in the jth class, and D _i,j represents the Euclidean distance between the centroids of the ith class and the jth class.

In formula (1.2), D _i,j represent the distance between classes, and the expression formula is as follows:

D _i,j =||m _i -m _j || (5)

In formula (1.2), _Wi represents the intra-class distance, and the expression formula is as follows:

If DBI _new < DBI _last , a new cluster center can be formed, otherwise the algorithm terminates. At the end of the iteration, the tree vertices and the number K are determined, and the cluster center generated by the iteration termination is used as the treetop point, the clustering K value is used as the number of single trees in the target image, and the filtered cluster center pixel is used as the description of each tree. The seed point (top point) position of the tree crown.

In the formula, && is the logical relationship 'and', m' is the new center set, m _1...j ... are the cluster centers m ₁ to m _i , and m _1...i are the cluster centers m ₁ to m _j . When DBI takes the minimum value, the optimal solution is obtained, and the number and position of tree vertices are determined at this time.

The above only describes the preferred embodiments of the present invention in conjunction with the accompanying drawings. It should be pointed out that for those skilled in the art, various changes and modifications can be made within the scope of the appended claims. These improvements and Deformation should also be regarded as the protection scope of the present invention.

Claims (3)

1. The DBI tree vertex detection method of airborne laser radar point cloud data CHM comprises the following specific steps:

step 1) filtering and classifying point cloud data and generating a Digital Elevation Model (DEM), a Digital Surface Model (DSM) and a Canopy Height Model (Canopy Height Model, CHM);

step 2) standardizing the foreground mark pixel generation conditions of the canopy height model according to the pixel gray level of the height canopy model (CHM) to generate an initial foreground mark of the canopy height model, which specifically comprises the following steps:

(1) the gray level values of the pixels outside the canopy height model image markers are all lower than those inside the markers;

(2) forming a connected component by pixel points of the foreground image of the canopy height model image;

(3) the canopy height model image has the same gray level value with the pixel point inside the same mark;

generating a canopy height model image C after the initial foreground marking according to the 3 standard conditions;

step 3) filtering and removing the pseudo foreground pixels by utilizing the gray value change characteristics of the pixels in eight adjacent domains of the central pixel based on the initial foreground mark of the height canopy model;

step 4) introducing a similarity judgment factor DBI, and carrying out crown vertex screening of a DBI-K method on the canopy height model foreground mark, wherein the method specifically comprises the following steps:

clustering the input two-dimensional gray level images of the canopy height model, and generating clustering center m and clustering center number K values of the foreground images of the canopy height model by utilizing a K-means clustering method introducing a similarity judgment factor DBI (direct binary input) for multiple times of iteration; if DBI_new＜DBI_lastForming a new cluster center, otherwise, terminating the algorithm, and taking the cluster center m generated by iteration termination as a treetop point T_pointThe value K of the number of the clustering centers is used as the number of the single plants of the target image, and the screened clustering center pixels are used for describing the treetop point T of each crown_pointThen the tree vertices are labeled and displayed visually in three dimensions.

2. The method according to claim 1, characterized in that said step 1) is in particular:

in order to ensure the integrity of terrain confidence and the filtering and classifying effect of laser radar point cloud, the method adopts a progressive triangulation network filtering algorithm, filtering and classifying the experimental sample plot data by setting a maximum gradient, an iteration angle and an iteration distance, performing coloring display on the point cloud after filtering and classifying, calculating the non-ground laser point cloud and the ground reflected laser after filtering and classifying by using a TIN method to obtain DSM and DEM which respectively generate experimental forest zones, and subtracting the DEM from the DSM to obtain CHM.

3. The method according to claim 1, characterized in that said step 3), in particular:

the method utilizes R, G, B values to obtain gray level image of the canopy height model, when C of foreground pixel of the canopy height model_R、C_G、C_BThe sum of the component values is taken to be the maximum value and the requirement R, G, B that the component values are equal, the pixel value of the foreground pixel is marked to be 0, otherwise the pixel value is 1 in all cases, resulting in a new image C' as the final foreground marked image.

Images