Tree crown extraction apparatus and tree crown extraction method

Abstract

To provide a tree crown extraction device and a tree crown extraction method which can highly accurately and automatically extract a range of a tree crown even for a broad-leaved tree.SOLUTION: A tree crown extraction device comprises: a DEM point cloud data creation unit 31 which generates digital elevation model point cloud data from aviation laser point cloud data; a DSM point cloud data creation unit 32 which generates digital surface layer model point cloud data; a vegetation classification unit 33 which classifies vegetation point cloud data from the digital surface layer model point cloud data; a ground surface height data conversion unit 34 which converts the vegetation point cloud data into digital height model point cloud data on the basis of the digital elevation model point cloud data; a surface point cloud data creation unit 35 which creates vegetation surface point cloud data from the digital height model point cloud data; a tree crown edge / center vector prediction unit 36 which predicts a tree crown edge and a tree crown center vector on the basis of the vegetation surface point cloud data; and a tree crown polygon formation unit 37 which forms a tree crown polygon from the vegetation surface point cloud data on the basis of the tree crown edge and tree crown center vector.SELECTED DRAWING: Figure 5

Description

Technical Field

[0001] The present invention relates to a crown extraction device and a crown extraction method that automatically extract the range of a tree crown, particularly the tree crown of broad-leaved trees in a growth area, based on aerial laser point cloud data (three-dimensional point cloud data obtained by aerial laser measurement).

Background Art

[0002] Conventionally, various devices and methods have been proposed as techniques for investigating the form of trees in a forest (e.g., the position and number of trees, tree height, etc.) or detecting tree positions.

[0003] In addition, techniques for evaluating or calculating the state of a forest area are also known in order to manage forest resource information and tree felling status related to the forest area to be investigated (growth area).

[0004] Particularly in recent years, these techniques have become very important for managing trees along roads and railway lines in order to prevent fallen trees, etc.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Patent Document 2

Patent Document 3

Patent Document 4

Summary of the Invention

Problems to be Solved by the Invention

[0006] However, with previous technologies, while it was possible to automatically extract the canopy range to some extent for coniferous trees such as cedar and cypress because their canopies do not overlap, it was difficult to accurately extract the canopy range for each tree of broad-leaved trees, whose canopies overlap in a complex manner.

[0007] In other words, with broad-leaved trees, there was a challenge in detecting broad-leaved trees as individual trees within a given area, as multiple trees of the same species were often grouped together and detected as a single tree.

[0008] In particular, because broad-leaved trees have more complex crown shapes than coniferous trees, being able to accurately extract the extent of the crown for each tree and detect them as individual trees is extremely useful, for example, when managing trees along roadsides or railway tracks, in order to identify trees that pose a risk of falling or other problems.

[0009] The present invention has been made in view of the above problems, and aims to provide a tree canopy extraction device and a tree canopy extraction method that can automatically extract the extent of the tree canopy with high accuracy, even for broad-leaved trees, and are suitable for use in tree management and the like. [Means for solving the problem]

[0010] Embodiments of the present invention are tree crown extraction devices for extracting the extent of broadleaf tree crowns from within a tree growing area based on three-dimensional point cloud data obtained by aerial laser measurement, comprising: a generation unit that generates digital elevation model point cloud data in which the height of the ground is converted into three-dimensional coordinates and digital surface model point cloud data in which the height of the ground surface corresponding to the terrain is converted into three-dimensional coordinates from the three-dimensional point cloud data; a classification unit that classifies vegetation point cloud data according to the vegetation of the trees from the digital surface model point cloud data; and a generation unit that generates digital elevation model point cloud data according to the vegetation point cloud data The gist of the system is that it comprises: a conversion unit that converts data into numerical height model point cloud data corresponding to the height from the ground; a creation unit that creates vegetation surface point cloud data from the numerical height model point cloud data; a prediction unit that predicts classification point cloud data of the canopy edge and canopy center vectors based on the vegetation surface point cloud data; and a formation unit that clusters the classification point cloud data using the classification point cloud data of the canopy edge and the canopy center vectors to obtain canopy points and forms canopy polygons that include the canopy points.

[0011] Another embodiment of the present invention is a tree crown extraction method for extracting the extent of a broadleaf tree crown from within a tree growing area based on three-dimensional point cloud data obtained by aerial laser measurement using a tree crown extraction device, comprising the steps of generating from the three-dimensional point cloud data a numerical elevation model point cloud data in which the height of the ground surface corresponding to a feature is converted into three-dimensional coordinates and numerical surface model point cloud data in which the height of the ground surface corresponding to a feature is converted into three-dimensional coordinates; classifying vegetation point cloud data according to the vegetation of the tree from the numerical surface model point cloud data; and based on the numerical elevation model point cloud data, The gist of this method is that it includes the steps of: converting the vegetation point cloud data into numerical height model point cloud data corresponding to the height from the ground; creating vegetation surface point cloud data from the numerical height model point cloud data; predicting classification point cloud data of the canopy edge and canopy center vectors based on the vegetation surface point cloud data; and clustering the classification point cloud data using the classification point cloud data of the canopy edge and the canopy center vectors to obtain canopy points and forming canopy polygons that include the canopy points. [Effects of the Invention]

[0012] As described above, the present invention provides a tree canopy extraction device and tree canopy extraction method that can automatically extract the canopy range with high accuracy, even for broad-leaved trees, and are suitable for use in tree management and the like. [Brief explanation of the drawing]

[0013] [Figure 1] This is a schematic diagram showing an example configuration of a tree canopy extraction system to which a tree canopy extraction device according to an embodiment of the present invention is applied. [Figure 2] This illustrates a tree crown extraction method, where (a) is a bird's-eye view showing the color-coded display of the reflectance intensity of aerial laser point cloud data for automatic tree crown extraction processing, (b) is a bird's-eye view showing the color-coded display of each tree crown in (a), and (c) is a top view showing the color-coded display of each tree crown in (a). [Figure 3] (a) is a cross-sectional view illustrating aerial laser measurement data (reflection pulses), (b) is a bird's-eye view illustrating the canopy surface data (digital surface model point cloud data / DSM) from (a), and (c) is a bird's-eye view illustrating the canopy elevation data (digital elevation model point cloud data / DEM) from (a). [Figure 4] This is a schematic diagram illustrating the extent of a tree's canopy, using broad-leaved trees as an example. [Figure 5] This is a block diagram showing an example configuration of a tree canopy extraction device. [Figure 6] This flowchart illustrates the tree canopy extraction method. [Figure 7] This is shown to provide a concrete explanation of DSM and DEM generation, where (a) is an orthomosaic image, (b) is unfiltered aerial laser measurement data when (a) is the measurement target, and (c) is filtered aerial laser point cloud data when (a) is the measurement target. [Figure 8] This is shown to provide a concrete explanation of vegetation classification, where (a) is DSM point cloud data, (b) is vegetation and other point cloud data from the DSM point cloud data in (a), and (c) is vegetation point cloud data from the DSM point cloud data in (a). [Figure 9] It is a flowchart shown to explain the flow of a process related to the conversion of vegetation point cloud data into ground height data (digital elevation model point cloud data / DCM). [Figure 10] It is for specifically explaining the flow of a process related to the conversion of vegetation point cloud data into ground height data. (a) is a diagram showing an example of vegetation point cloud data and canopy elevation data, (b) is a diagram exemplifying a method for calculating the height of a vegetation point from the ground, and (c) is a diagram showing an example of the converted ground height data. [Figure 11] It is a diagram shown to explain the extraction of feature quantities from vegetation surface point cloud data with attributes and the prediction of the classified point cloud data of the canopy edge and the canopy center vector. [Figure 12] It is shown to explain the classification of the canopy edge. (a) is a diagram exemplifying the loss function of a deep learning model for the classification of the classified point cloud data of the canopy edge, (b) is a diagram shown to explain the creation of teacher data for classifying the canopy edge and the inside in one canopy, (c) is an overhead view shown for comparing local shapes, and (d) is a cross-sectional view shown for comparing local shapes. [Figure 13] It is shown to explain the extraction of the canopy center vector. (a) is a diagram exemplifying the loss function of a deep learning model for extracting the canopy center vector, and (b) is a diagram exemplifying the canopy center vector and the canopy center point in one canopy. [Figure 14] It is a diagram shown to explain the flow of a process when obtaining canopy points by clustering using the classified point cloud data of the canopy edge and the canopy center vector. [Figure 15] It is a flowchart shown to explain the flow of a process when obtaining canopy points by clustering using the classified point cloud data of the canopy edge and the canopy center vector. [Figure 16]In the flowchart of FIG. 15, it is a diagram shown to explain a method of separating canopy edge points and internal points based on the edge prediction probability. (a) is a diagram illustrating an arithmetic expression for classifying canopy edge points and internal points, and (b) is a diagram showing an example of the process. [Figure 17] In the flowchart of FIG. 15, it is a diagram shown to explain a method of gathering internal points at the canopy center point. [Figure 18] In the flowchart of FIG. 15, it is a diagram shown to explain a method of clustering internal points and obtaining the canopy center point. [Figure 19] In the flowchart of FIG. 15, it is a diagram shown to explain the creation of a canopy polygon. [Figure 20] As a specific example, it is a diagram showing a comparison between a canopy polygon obtained by on-site investigation (visual inspection) and an automatically extracted canopy polygon.

Mode for Carrying Out the Invention

[0014] The embodiments shown below are examples of devices and methods for embodying the technical idea (structure, arrangement, etc.) of the present invention, and the technical idea of the invention is not limited to the following. The technical idea of the present invention can be variously modified within the scope of the matters described in the claims. Also, it should be noted that the drawings are schematic, and the configurations of devices and systems are different from the actual ones.

[0015] Embodiment Hereinafter, a canopy extraction device and a canopy extraction method according to an embodiment of the present invention will be described with reference to the drawings.

[0016] Figure 1 is a schematic diagram showing an example of the configuration of a tree crown extraction system to which the tree crown extraction device 1 according to this embodiment is applied, and Figures 2 and 3 show an example of automatic tree crown extraction processing to explain the tree crown extraction method by the tree crown extraction device 1. Specifically, in this tree crown extraction system, the tree crown extraction device 1 automatically extracts the tree crown ranges of trees A, B, and C based on three-dimensional point cloud data (airborne laser point cloud data) from within the tree growing area obtained by airborne laser measurement.

[0017] Here, the canopy area refers to the upper part of the trunk of a broad-leaved tree, where branches and leaves grow densely, as shown in Figure 4, for example.

[0018] First, before explaining the tree canopy extraction device 1, let me briefly explain aerial laser measurement.

[0019] Aerial laser measurement, as shown in Figure 1 for example, involves using a laser measurement device (not shown) mounted on an aircraft, helicopter, or drone to irradiate laser pulses from above. The distance to the reflection point is determined by the time elapsed since the pulse irradiation, thereby acquiring height information within the pulse irradiation area, which is the target of measurement, as aerial laser measurement data.

[0020] For example, the laser pulse is adjusted so that at a distance of 1000m, the pulsed area is approximately 20cm in diameter.

[0021] In Figure 1, the laser pulses being emitted are defined as follows: the pulse reflected by tree A is the first pulse, the pulses reflected by trees B and C are the intermediate pulses, and the pulse reflected by the ground is the last pulse.

[0022] Furthermore, tree A is, for example, a broad-leaved tree in the canopy layer of a multi-layered forest where canopies and shrubs are mixed, while trees B and C are, for example, trees (broad-leaved trees, etc.) in the sub-canopy layer or shrub layer that grow under the branches of tree A in a multi-layered forest.

[0023] As shown in Figure 1, the tree crown extraction device 1 according to this embodiment consists of an input device 2, a processing unit 3, a storage device 4, and an output device 5, among others.

[0024] Input device 2 is a keyboard or mouse, and is used by the operator when operating the tree canopy extraction device 1. Although not shown in the figures, input device 2 may also include an acquisition device that takes in aerial laser measurement data (raw data) acquired by the aircraft 6, for example via a hard disk, as well as a scanner or a communication device connected to the internet.

[0025] The processing unit 3 is constructed using a computer (PC) and programs executed on the PC, and is used, for example, to automatically extract the canopy range of trees A, B, and C based on aerial laser point cloud data corresponding to the growing areas of trees A, B, and C, which are the targets of measurement, by executing various programs.

[0026] The aerial laser point cloud data is three-dimensional point cloud data (original point cloud data) corresponding to the aerial laser measurement data, which has been converted into three-dimensional coordinates using a Cartesian coordinate system by the processing unit 3.

[0027] The storage device 4 consists of a hard disk drive, either built into or attached to the PC, and stores, for example, various programs and deep learning models.

[0028] Here, a deep learning model is an AI (Artificial Intelligence) technology that eliminates the need for "feature design," which is essential in conventional machine learning, by automatically extracting features from large amounts of data and learning from them.

[0029] Output device 5 is a display or printer and is used for screen display and printout of individual tree crown polygons based on the crown ranges of trees A, B, and C, which are automatically extracted by the processing unit 3. Output device 5 may also include a communication device that communicates with a cloud or mobile terminal connected to the internet.

[0030] Aerial laser point cloud data, which is obtained by converting aerial laser measurement data into 3D coordinates, includes attribute information such as height information and reflectivity corresponding to the canopy of broadleaf trees with complex shapes, as shown in Figure 2, for example. By enabling the extraction of the canopy range individually and with high precision, it becomes easy to display each canopy in 2D (top view) or 3D (bird's-eye view) color coding.

[0031] In Figure 2, (a) is a bird's-eye view illustrating the color-coded display of reflectance intensity in aerial laser point cloud data, (b) is a bird's-eye view showing an example of color-coded display for each tree canopy in (a), and (c) is a top view showing an example of color-coded display for each tree canopy in (a).

[0032] Furthermore, the height information that can be obtained includes, for example, as shown in Figure 3, DSM (Digital Surface Model) point cloud data that converts the height of geographical features such as trees and structures into three-dimensional coordinates as canopy surface data (digital surface model point cloud data), as well as DEM (Digital Elevation Model) point cloud data that converts the height of the ground into three-dimensional coordinates as canopy elevation data (digital elevation model point cloud data).

[0033] In Figure 3, (a) is a cross-sectional view showing an example of aerial laser measurement data (reflection pulse), (b) is a bird's-eye view illustrating the DSM point cloud data of (a), and (c) is a bird's-eye view illustrating the DEM point cloud data of (a).

[0034] The tree canopy extraction device 1 according to this embodiment will be described in more detail below.

[0035] Figure 5 shows the configuration of the tree canopy extraction device 1 in detail. The processing unit 3, for example, is the CPU (Central Processing Unit) of a PC and functions as a DEM point cloud data creation unit (generation unit) 31, a DSM point cloud data creation unit (generation unit) 32, a vegetation classification unit (classification unit) 33, a ground height data conversion unit (conversion unit) 34, a surface point cloud data creation unit (creation unit) 35, a tree canopy edge / center vector prediction unit (prediction unit) 36, and a tree canopy polygon formation unit (formation unit) 37.

[0036] The DEM point cloud data creation unit 31 and the DSM point cloud data creation unit 32 perform noise reduction and filtering processing on the aerial laser measurement data provided by the aircraft 6 using specific software to generate DEM point cloud data and DSM point cloud data.

[0037] The vegetation classification unit 33 classifies the DSM point cloud data into vegetation point cloud data and other point cloud data using a deep learning model, such as the feature extraction and data classification method (e.g., KPConv), which will be described later.

[0038] The ground height data conversion unit 34 generates DCM (Digital Canopy Model) point cloud data as ground height data for vegetation (numerical height model point cloud data) by converting the vegetation point cloud data by ground height (h) based on the vegetation point cloud data and the DEM point cloud data.

[0039] The surface point cloud data creation unit 35 divides the DCM point cloud data into a grid, designates the highest point with the highest ground elevation data within the grid as the representative point, and generates attributed vegetation surface point cloud data, using its reflection intensity and the altitude difference between the representative point (highest point) and the lowest point as attribute information.

[0040] The canopy edge / center vector prediction unit 36 ​​performs feature extraction from attributed vegetation surface point cloud data using a deep learning model (e.g., KPConv), and calculates classification point cloud data of the canopy edge and predicts the canopy center vector using a deep learning model (e.g., multilayer perceptron (MLP)).

[0041] The canopy polygon formation unit 37 uses the predicted classification point cloud data of the canopy edge and the canopy center vector to cluster the classification point cloud data to obtain canopy points, calculate polygons that enclose the canopy points, and finally form individual canopy polygons.

[0042] The storage device 4 includes, for example, a point cloud data storage unit 41 for storing various types of point cloud data, a program storage unit 42 for storing programs for operating the arithmetic processing unit 3 and trained deep learning models, and a parameter storage unit 43 for storing calculation parameters, thresholds, training parameters, and the like.

[0043] Next, with reference to Figure 6, the tree crown extraction method using the tree crown extraction device 1 will be described.

[0044] As shown in Figure 6, for example, when aerial laser measurement data acquired by the aircraft 6 is temporarily stored in the point cloud data storage unit 41 of the memory device 4, first, the DEM point cloud data is created by the DEM point cloud data creation unit 31 and the DSM point cloud data is created by the DSM point cloud data creation unit 32 (step S01).

[0045] In the DEM point cloud data creation unit 31 and the DSM point cloud data creation unit 32, using specific software pre-stored in the program storage unit 42 of the storage device 4, noise reduction processing and filtering processing, such as those shown in Figures 7(b) and (c), are performed on the aerial laser measurement data read from the point cloud data storage unit 41. This process creates DEM point cloud data, which represents the height of the ground excluding features such as trees and artificial structures in 3D coordinates, and DSM point cloud data, which represents the height of the ground surface including features such as trees and artificial structures in 3D coordinates. The DEM point cloud data and DSM point cloud data are then stored in a dedicated area of ​​the point cloud data storage unit 41.

[0046] In Figure 7, (a) is the orthomosaic image, (b) is the unfiltered aerial laser measurement data when (a) is the measurement target, and (c) is the filtered original point cloud data (DEM point cloud data, DSM point cloud data) when (a) is the measurement target.

[0047] Next, using the deep learning model (e.g., KPCocv) stored in the program storage unit 42 of the memory device 4, the DSM point cloud data read from the point cloud data storage unit 41 is classified into vegetation point cloud data corresponding to tree vegetation and other point cloud data (step S02).

[0048] Specifically, in the vegetation classification unit 33, as shown in Figures 8(a) to 8(c), for example, vegetation point cloud data (x,y,z)i corresponding only to tree vegetation is classified from the DSM point cloud data. This vegetation point cloud data (x,y,z)i is temporarily stored in a dedicated area of ​​the point cloud data storage unit 41.

[0049] Next, in the ground elevation data conversion unit 34, the vegetation point cloud data (x,y,z)i read from the point cloud data storage unit 41 is converted into vegetation ground elevation point cloud data (x,y,h)i based on the DEM point cloud data (step S03).

[0050] Specifically, the ground height data conversion unit 34 first searches for the nearest ground points (about 6 points) to the ground position coordinates (x,y) of a vegetation point (x,y,zt) from the vegetation point cloud data and the DEM point cloud data, as shown in Figure 10(a) (step S31 in Figure 9).

[0051] Then, as shown in Figure 10(b), for example, the plane (ground) of the nearest neighboring ground point is predicted by linear regression (step S32 in Figure 9).

[0052] Furthermore, as shown in Figure 10(b), for example, the height from the ground (x,y,zg) is calculated by dropping a vertical line onto the plane predicted from the vegetation point (x,y,zt) (step S33 in Figure 9).

[0053] Thus, as shown in Figure 10(c), for example, the DSM point cloud data is ultimately converted into DCM point cloud data (vegetation ground height data) of ground height h. This DCM point cloud data is then stored in a dedicated area of ​​the point cloud data storage unit 41.

[0054] Next, in the surface point cloud data creation unit 35, for example, attributed vegetation surface point cloud data is generated from the DCM point cloud data read from the point cloud data storage unit 41, with attributes such as the reflectance intensity of representative points (step S04).

[0055] The attributed vegetation surface point cloud data is temporarily stored in a dedicated area of ​​the point cloud data storage unit 41.

[0056] In creating attributed vegetation surface point cloud data, first, the XY plane of the DCM point cloud data of ground height h read from the point cloud data storage unit 41 is divided into, for example, a 50cm × 50cm grid. Points belonging to each of these divided grids are extracted, and the highest point of the extracted points is used as the representative point. The following grid attributes (features) are used as input (x, y, h, r, n, Δz, (Rn)j, (Ln)k)i for processing by the deep learning model in the next step.

[0057] Specifically, as attributes of the grid, for example, the coordinates (x,y,h) of input i were taken as the coordinates of the representative point, the reflection intensity (r) was taken as the reflection intensity of the representative point, the number of points belonging to the grid was n, and the altitude difference between the highest and lowest points belonging to the grid was Δz.

[0058] Furthermore, the number of points belonging to the grid for each reflection number is defined as (Rn)j (j=1,2,...,Rmax) (for example, if the reflection numbers of the points range from 1 to 5, then j will be 1,2,...,5), and the number of points belonging to the grid for each level of ground height is defined as (Ln)k (k=1,2,...,Lmax) (for example, if the levels of ground height are divided into 2m intervals: 0-2m, 2m-4m, 4m-6m,..., 18-20m, 20m and above, then k will be 1,2,...,11).

[0059] Next, in the canopy edge / center vector prediction unit 36, as shown in Figure 11, for example, feature extraction processing by KPConv and canopy edge prediction (canopy edge classification) and center vector prediction (canopy center vector regression) by MLP are performed on the attributed vegetation surface point cloud data read from the point cloud data storage unit 41 (step S05).

[0060] For canopy edge prediction, as shown in Figures 12(a) and (b), for example, it is treated as a two-class classification problem in which each point in the vegetation surface point cloud data is classified as either 0 or 1. A class is defined as the output ci (0 for edge, 1 for interior), and a human-created ground truth canopy polygon is scaled down to R%, with the area representing the interior of the canopy being used as training data, and the point cloud contained within that area being defined as interior canopy points. A Cross Entropy function with Dice Loss added is set as the loss function.

[0061] As is clear from Figures 12(c) and (d), the local shape differs between the edges and the interior of the canopy polygon.

[0062] In the canopy center vector prediction, as shown in Figures 13(a) and (b), for example, the regression problem predicts the direction of the canopy center point to which each point in the vegetation surface point cloud data belongs. A two-dimensional vector (u,v) is defined as the output (u,v)i, a unit (center) vector from the point of interest to the center point is set as the training data, and a Cosine Loss is set as the loss function.

[0063] The classification point cloud data of the canopy edge (edge ​​prediction probability) and the canopy center vector predicted by the canopy edge / center vector prediction unit 36 ​​may be stored, for example, in the memory device 4.

[0064] Next, in the canopy polygon formation section 37, clustering of canopy points is performed using the predicted classification point cloud data of the canopy edge and the canopy center vector (step S06).

[0065] The clustering process here will be described in detail later, but for example, as shown in Figure 14, Point Gathering is used to repeatedly move internal and edge points and search for neighboring points. When the direction of the canopy center vector of a neighboring point changes significantly, the movement stops, and finally, individual canopy polygons are created based on the grouped points.

[0066] Figure 15 is a flowchart illustrating the process of clustering canopy points using the canopy margin prediction probability and canopy center vector of the classification point cloud data at the canopy margin (see step S06 in Figure 6).

[0067] In clustering of tree canopy points, first, as shown in Figures 16(a) and (b), for example, the tree canopy points are divided into tree canopy edge points and interior points based on the edge prediction probability (step S61).

[0068] Then, the internal points are gathered by Point Gathering (step S62). Here, for example, as shown in Figure 17, first, each internal point i(x_i,y_i) is selected as a point to be moved, and after the selected point is moved by the crown center vector, the nearest neighbor internal point j(x_j,y_j) at the destination is searched for and becomes the point to be moved next. In this way, the selection and movement of the point to be moved is repeated for all internal points i(x_i,y_i). Note that around the crown center point, there is a large variation in the direction of the crown center vector for internal points i(x_i,y_i), and this is considered the condition for the iteration to end.

[0069] In this way, the moved internal points i(x_i,y_i) are clustered using an algorithm such as DBSCAN, as shown in Figure 18, and the average position of each cluster is set as the center point of the tree canopy (0,1,2,3,4) (step S63).

[0070] Once the center point of the tree canopy is determined, the tree canopy edge points are then gathered by Point Gathering, similar to the case of the internal point i(x_1,y_1) as explained earlier (step S64).

[0071] Then, the nearest internal point j(x_j,y_j) of the moved canopy edge point is searched for, and its distance is calculated.

[0072] In this way, if the calculated distance is within, for example, a pre-stored threshold, the moved canopy edge point is considered a valid canopy edge point and is assigned to the cluster to which its nearest neighbor internal point j(x_j,y_j) belongs. Invalid canopy edge points are discarded as noise.

[0073] In this way, by ultimately calculating the convex hull of the vegetation point cloud data for each cluster through grouping (step S65), it becomes possible to create individual canopy polygons, for example, as shown in Figure 19 (step S66).

[0074] The canopy polygons formed in the canopy polygon formation unit 37 may be stored in, for example, the memory device 4.

[0075] Furthermore, it is possible to omit the storage of the various point cloud data acquired in the storage device 4 into the point cloud data storage unit 41 at each stage, thereby creating a continuous series of processes.

[0076] Specific example Next, a specific example of the tree canopy extraction device 1 according to this embodiment will be described.

[0077] Figure 20 shows an example of output from the output device 5, comparing tree canopy polygons 51, which were automatically extracted using the tree canopy extraction device 1, with tree canopy polygons 52, which were created based on on-site surveys (visual inspection), using the growing area of ​​broad-leaved trees planted along roadsides such as highways as the measurement target.

[0078] As shown in Figure 20, for example, in areas where multiple trees are planted in rows, the automatically extracted canopy polygons 51 can accurately extract the canopy range of more trees as individual trees, compared to the canopy polygons 52 obtained through on-site surveys.

[0079] In other words, even if the trees being measured are broad-leaved trees, it becomes possible to accurately extract the canopy range for each tree.

[0080] In particular, in some growing areas, there are instances where the canopy range of multiple trees is mistakenly extracted as the canopy range of a single tree. However, since deep learning models are used for vegetation classification and prediction of canopy center vectors based on classification point cloud data of the canopy margin, further improvements in accuracy can be expected by increasing the quality of training data and accumulating learning.

[0081] Therefore, when managing trees along roadsides or railway tracks, this method is extremely useful for identifying trees that pose a risk, such as oversized trees or trees that might fall onto the road or railway tracks.

[0082] As described above, according to this embodiment, even for broad-leaved trees whose canopy range is more complex than that of coniferous trees, the canopy range can be automatically extracted with high accuracy, and a canopy extraction device and canopy extraction method suitable for use in tree management can be provided.

[0083] In other words, a tree crown extraction device that extracts the crown range of broad-leaved trees from within a tree's growing area based on aerial laser point cloud data obtained by converting raw data (aerial laser measurement data) from aerial laser measurement into 3D coordinates will be able to extract the crown range of broad-leaved trees more accurately. This will make it possible to efficiently and automatically extract individual trees, even if they are broad-leaved trees. Therefore, when managing trees along roadsides or railway lines, it will be extremely useful in extracting the crown range of specific broad-leaved trees from among many trees, and will facilitate rational management of risk trees and other such trees.

[0084] Furthermore, this method is not limited to cases where multiple trees are planted in rows, but can also be applied to cases where multiple trees are planted in a planar manner or where trees grow naturally. [Explanation of symbols]

[0085] 1 Tree crown extraction device 2 Input devices 3. Arithmetic Processing Unit 4 Storage device 5. Output device 6 flying objects 31 DEM point cloud data creation unit (generation unit) 32 DSM point cloud data creation unit (generation unit) 33. Vegetation Classification Section (Classification Section) 34 Ground elevation data conversion unit (conversion unit) 35. Surface Point Cloud Data Creation Unit (Creation Unit) 36. Canopy Edge / Center Vector Prediction Unit (Prediction Unit) 37. Canopy polygon formation area (formation area) 41 Point cloud data storage unit 42 Program Storage Unit 51 Auto-extracted tree canopy polygons 52 Tree canopy polygons surveyed in the field

Claims

1. A tree crown extraction device that extracts the extent of a broadleaf tree crown from within a tree growing area based on three-dimensional point cloud data obtained by aerial laser measurement, A generation unit generates, from the aforementioned three-dimensional point cloud data, a numerical elevation model point cloud data in which the height of the ground is converted into three-dimensional coordinates, and a numerical surface model point cloud data in which the height of the ground surface corresponding to the features is converted into three-dimensional coordinates. A classification unit that classifies vegetation point cloud data corresponding to the tree vegetation from the numerical surface model point cloud data, A conversion unit that converts the vegetation point cloud data into numerical height model point cloud data corresponding to the height from the ground, based on the aforementioned numerical elevation model point cloud data, A creation unit that creates vegetation surface point cloud data from the aforementioned numerical height model point cloud data, A prediction unit predicts classification point cloud data of the tree canopy margin and tree canopy center vectors based on the aforementioned vegetation surface point cloud data, A forming unit that uses the classification point cloud data of the tree canopy edge and the tree canopy center vector to cluster the classification point cloud data to obtain tree canopy points and to form tree canopy polygons that include the tree canopy points, A tree canopy extraction device characterized by being equipped with the following features.

2. The tree crown extraction device according to claim 1, characterized in that the classification unit classifies the numerical surface model point cloud data into vegetation point cloud data and non-vegetation point cloud data using a deep learning model.

3. The tree crown extraction device according to claim 1, characterized in that the creation unit creates vegetation surface point cloud data having the highest point of the grid dividing the XY plane of the numerical height model point cloud data as a representative point, and the feature quantities of the grid as attributes.

4. The tree crown extraction device according to claim 1, characterized in that the prediction unit predicts the classification point cloud data of the tree crown edge and the tree crown center vector using a deep learning model.

5. A tree crown extraction method for extracting the extent of a broadleaf tree crown from within a tree growing area based on three-dimensional point cloud data obtained by aerial laser measurement, using a tree crown extraction device according to any one of claims 1 to 4, The process involves generating, from the aforementioned three-dimensional point cloud data, a numerical elevation model point cloud data in which the ground height is represented in three-dimensional coordinates, and a numerical surface model point cloud data in which the ground surface height corresponding to a feature is represented in three-dimensional coordinates. A step of classifying vegetation point cloud data according to the tree vegetation from the numerical surface model point cloud data, A step of converting the vegetation point cloud data into numerical height model point cloud data corresponding to the height from the ground, based on the aforementioned numerical elevation model point cloud data, The process of creating vegetation surface point cloud data from the aforementioned numerical height model point cloud data, A step of predicting classification point cloud data of the canopy margin and canopy center vector based on the aforementioned vegetation surface point cloud data, The process involves clustering the classification point cloud data of the tree canopy edge and the tree canopy center vector to obtain tree canopy points, and forming a tree canopy polygon that includes the tree canopy points. A method for extracting tree canopy, characterized by comprising the following features.

Images