Tree structure parameter extraction and three-dimensional reconstruction method of ground laser point cloud data

Abstract

A tree structure parameter extraction and three-dimensional reconstruction method of ground laser point cloud data belongs to the technical field of tree three-dimensional modeling computer program based on laser point cloud data in forestry scientific research, takes tree point cloud as input, realizes skeleton point prediction through multiple iterations of sparse convolution-based residual U-shaped network (SpConv-based ResUNet), and constructs an initial skeleton through graph analysis, uses path smoothing of automatic driving and angle constraint connection point search for skeleton smoothing, uses artery snake to construct a branch model based on the radius estimated by the constrained B-spline curve and the skeleton, and uses a simple leaf sequence algorithm to realize virtual leaf insertion to realize tree simulation, and realizes automatic parameter extraction by using the input point cloud, the extracted skeleton and the reconstructed model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of tree three-dimensional modeling based on laser point cloud data in forestry science research, and relates to a tree structure parameter extraction and three-dimensional reconstruction method of ground laser point cloud data. BACKGROUND

[0002] The tree structure is affected by congenital inheritance, afterlife biological and non-biological factors, and produces specific physiological and ecological characteristics, which is the basis for analyzing growth dynamics, development patterns, reproduction and environmental response. However, its high complexity makes the modeling and quantification task challenging. With the application of laser scanning technology which can generate three-dimensional surface information of organisms with high throughput in forestry becoming more widespread, the expression of tree structure has developed a quantitative structure modeling method using simple geometric primitives for visualization and parameter operation. This way has become the mainstream way of tree three-dimensional modeling based on laser point cloud, and has been applied to the research of forest ecology, tree breeding, etc.

[0003] Due to the inevitable noise and occlusion problems of point cloud data, various quantitative structure models currently have problems such as low reconstruction quality, low algorithm efficiency, weak parameter extraction ability, and complex hyperparameters when processing real-world laser scanning single tree point cloud. In addition, these methods lack interactivity, which limits the analysis of branch attributes to overall statistical analysis. SUMMARY

[0004] The application provides a tree structure parameter extraction and three-dimensional reconstruction method of ground laser point cloud data to solve the technical problems in the prior art.

[0005] A tree structure parameter extraction and three-dimensional reconstruction method of ground laser point cloud data, comprising the following steps: taking the tree point cloud as input, predicting the skeleton points through multiple iterations based on a sparse convolution-based residual U-shaped network (SpConv-based ResUNet), and constructing an initial skeleton through graph analysis; using automatic driving path smoothing and angle constraint connection point search for skeleton refinement; using a constrained B-spline curve estimation radius and skeleton to construct a branch model using an arterial snake; using a simple leaf sequence algorithm to realize virtual leaf insertion to realize tree simulation; and using the input point cloud, extracted skeleton and reconstructed model to realize automatic parameter extraction.

[0006] The sparse convolution-based residual U-shaped network is embedded in a loop structure to realize point cloud contraction to obtain the skeleton points of the tree.

[0007] The graph analysis is to find the shortest path tree on the connected graph obtained by Delaunay tetrahedralization of the skeleton points, and to select the path based on the tree graph.

[0008] Path smoothing is a three-dimensional implementation of path smoothing in the Baidu autonomous driving platform Apollo, with an angle constraint in a loose branch angle size range of 15°-165°.

[0009] The estimation of the radius uses various constraint curves (power, exponential, quadratic, Bezier, B-spline) containing zero-crossing points and monotonically increasing in the value interval to output the stem form analysis result, and only B-spline is used for modeling.

[0010] The model of the arterial snake establishes a mapping relationship of the cross-sectional vertexes so as to facilitate the volume and surface area calculation.

[0011] The simple phyllotaxis algorithm can realize basic simulation of acrogyne, homophyll, whorl and fascicled phyllotaxis according to the parameterized leaf size and arrangement mode

[0012] The advantages of the method for extracting tree structure parameters and three-dimensional reconstruction of ground laser point cloud data are that the method can realize more accurate, rapid and stable parameter extraction and quantitative structure modeling, can meet the forestry research of various scenes, and can efficiently provide materials for the field of digital twin and other fields requiring tree models.

[0013] The sparse convolutional neural network is introduced into point cloud contraction and integrates methods beneficial to the modeling process, can efficiently process single-tree ground laser data in batches, supports extraction and visual analysis of 29 individual-level, 20 organ-level and 9 sample-level parameters, supports stem form analysis and horizontal spatial distribution pattern analysis, and realizes high-precision, efficient and robust parameter extraction and modeling. 2 On experimental data, the R 2 of stem length, stem volume and under-bough height are 0.999, 0.99 and 0.9 respectively, the R BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor. As shown in the drawings:

[0015] Figure 1 The technical flowchart of the present application.

[0016] Figure 2 The distribution map of the marked trees in the sample plot of the present application.

[0017] Figure 3 The laser data processing flowchart of the present application.

[0018] Figure 4 The single-tree reconstruction effect of the present application.

[0019] Figure 5 The sample plot reconstruction effect of the present application. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0021] Embodiment 1: as shown in Figure 1 , Figure 2 , Figure 3 , Figure 4 and Figure 5 , a tree structure parameter extraction and three-dimensional reconstruction method of ground laser point cloud data mainly relates to single-tree parameter extraction, tree quantitative structure modeling and forest visualization simulation technology in computer science.

[0022] A tree structure parameter extraction and three-dimensional reconstruction method of ground laser point cloud data contains the following steps: taking tree point cloud as input, realizing skeleton point prediction through multiple iterations of sparse convolution-based residual U-shaped network (SpConv-based ResUNet), and constructing an initial skeleton through graph analysis, using automatic driving path smoothing and angle constraint connection point search for skeleton smoothing, using the radius estimated based on constraint B-spline curve and the skeleton to construct a branch model using an arterial snake, and using a simple leaf sequence algorithm to realize virtual leaf insertion to realize tree simulation, using the input point cloud, the extracted skeleton and the reconstructed model to realize automatic extraction of 58 parameters (29 single-tree tree measuring factors, 20 branch attributes and 9 horizontal spatial structure parameters).

[0023] A tree structure parameter extraction and three-dimensional reconstruction method of ground laser point cloud data solves the problem that the current quantitative structure modeling method cannot realize comprehensive coordination of reconstruction quality, algorithm efficiency, parameter extraction comprehensiveness and precision, interactivity and functionality, provides a robust and easy-to-use solution for tree structure analysis and rapid reconstruction of forest scenes using point cloud data, and provides technical support for realizing precision forestry and related fields that require real tree models.

[0024] It mainly includes two aspects of point cloud contraction realized by repeated iteration of the sparse convolution-based residual U-shaped network and a novel reconstruction process of efficiently generating a three-dimensional model and parameters of the above-ground part of a tree structure, and the overall technical process is as shown in Figure 1 .

[0025] Depend on Figure 1 The process includes a training process and an inference process. The training process uses a synthetic tree dataset to train a sparse convolution-based residual U-shaped network. The inference process uses the trained network to predict the centripetal displacement of the point cloud and performs multiple iterations to shrink the point cloud to obtain skeleton points. The obtained skeleton points are then processed through a standard quantitative structural modeling workflow of initial skeleton construction – skeleton refinement – ​​radius estimation – mesh reconstruction to obtain the tree skeleton and mesh model, additionally including stem shape analysis results. This workflow integrates different methods to improve quality. Different levels of parameter extraction are achieved using the input point cloud, skeleton, and mesh, and visualization analysis is realized through a secondary development plugin of CloudCompare.

[0026] like Figure 1 As shown, the input to the entire program can be a single tree point cloud file, or an additional information in the form of a .yam l dataset. This information includes leaf images, size, angle, distribution pattern, bark texture, tree species information, and trunk curve type (quadratic / exponential / Behre / B-spline). This input information will be fed into the inference process.

[0027] The inference process first divides the input point cloud into blocks (each block contains a buffer) and performs batch voxelization, then feeds them into a trained neural network model. This model uses the attribute-containing voxels obtained through voxelization as input and predicts the direction and distance of movement using a residual U-shaped network based on sparse convolution. Based on the model's results, the predicted centripetal displacement corresponding to the point cloud can be calculated. The buffer portion is removed by masking to obtain the shrunken point cloud. This process is repeated until the iteration terminates. The final shrunken point cloud is then voxel-downsampled to obtain skeleton points, and the initial radius at each skeleton point is calculated. These skeleton points are then fed into the subsequent quantitative structure modeling process.

[0028] The neural network model used in the inference process is obtained by training input data that has been randomly pruned and batch-voxed on a synthetic tree dataset. Specifically, the model uses the voxelized point cloud portion of the input data to predict the direction and distance of movement, while the ground truth centripetal displacement in the input data can be decomposed into the ground truth direction and distance of movement, which can then be used to calculate the loss and update the weights.

[0029] The quantitative structural modeling process includes: initial skeleton construction, skeleton refinement, radius estimation, and mesh reconstruction.

[0030] The initial skeleton construction includes: graph construction (using Deloni tetrahedronization), shortest path tree generation (using weighted graph solution with squared distance), and path selection (using spherical domain assignment).

[0031] Skeleton refinement includes skeleton smoothing (using 3D version of automatic driving trajectory smoothing) and breakpoint connection (using angle constraint connection point search). This step outputs the tree skeleton.

[0032] Radius estimation includes initial radius filtering (using filter for rTwig), modeling special B-spline, and various curve fitting for stem shape analysis.

[0033] Mesh reconstruction includes branch and stem using arterial snake representation and leaf using simple phyllotaxis algorithm. This step outputs hierarchical model and texture model.

[0034] Based on point cloud, skeleton and mesh, multi-scale parameters can be extracted. These include:

[0035] Individual-level parameters (single-tree measurement factors): 29 parameters including DBH, tree height, volume, crown width, and height under branches.

[0036] Organ-level parameters (branch attributes): 20 parameters including branch length, branch angle, branch attachment angle, bow height, and chord length.

[0037] Plot-level parameters (horizontal spatial structure parameters): 9 parameters including crowding index, mixed degree, size ratio, angular scale, and competition index.

[0038] Visualization and interaction of these parameters are achieved through the ccEntQuery plug-in of CloudCompare secondary development, which requests and responds to the saved parameter database.

[0039] The specific operation steps are as follows:

[0040] Input

[0041] Batch processing of single-tree point cloud files is supported, and different parameters can be set for each tree in the form of a.yaml file. The configuration file needs to take a key name according to each tree, and the value is a dictionary configuration, as shown in Table 1.

[0042] Table 1 Custom configuration of each tree

[0043]

[0044]

[0045] 1. Point cloud contraction based on sparse convolution

[0046] (1) Structure of sparse convolutional neural network

[0047] The neural network takes a sparse convolution tensor (SparseConvTensor) generated by batch voxelization of the point cloud as input, which can be represented as a voxel matrix V with attributes in N rows (i.e. N voxels) and 7 columns N×7= [b, X N×3 , (IIP) N×3 ] simply, where b is batch vector, X is voxel coordinate matrix, IIP is the result matrix of voxel down-sampling (equivalent to left multiplication of a matrix Π) performed on point cloud (matrix) P containing at least N points. Vector is column vector by default, and its t-th element (i.e. t-th row) is denoted by subscript t. Matrix is denoted by subscript (t) to indicate the vector composed of its t-th row.

[0048] Each row of the matrix

[0049] V (t) = [b t , X (t) , (ΠP) (t) ] = [b t , i t , j t , k t , x t , y t , z t ], t = 1, 2,..., N,

[0050] Here b t is the batch number of the t-th voxel, i t , j t , k t are voxel coordinates or voxel indices, x t , y t , z t are the coordinates of the only sampling point in the voxel.

[0051] The operation of voxel down-sampling follows the index of the point in the point cloud, and only the first point in a voxel is retained.

[0052] The ontology of the neural network is a realization of residual UNet based on sparse convolution, which maps the 3-channel (i.e. three-dimensional coordinates of sampling points) feature map into the feature map required for input to the backbone encoder-decoder architecture through a convolution (Conv) layer containing a sparse sub-manifold convolution SubMConv3d. The encoder block uses another sparse convolution SparseConv3d for down-sampling, which is different from SubMConv3d in that only the kernel covering the input point is calculated to obtain the output point. The decoder block for up-sampling uses SparselnverseConv3d to make the output result have the same index as the corresponding down-sampled input. Residual connections are added in the steps of down-sampling and up-sampling. Finally, the output channels are integrated. There is a batch normalization BatchNorm1d followed by an activation function ReLU between layers and blocks.

[0053] The output of the neural network is the predicted unit movement direction matrix and the logarithmic movement distance vector, respectively denoted by symbols

[0054] where,

[0055] With each row of them, the centripetal displacement matrix, denoted as V, can be calculated. Each row of V is the predicted centripetal displacement of the sampling point in the corresponding row of v, defined as

[0056] For row t = 1, 2,..., N

[0057] where is the t-th row of matrix V corresponding to the moving direction and the logarithmic moving distance. It is used to describe the movement of the point to the ideal central axis. Before obtaining the unit moving direction matrix, matrix row normalization is needed. The use of the logarithmic moving distance vector is not directly using the moving distance vector because the output result has a sign and spans multiple orders of magnitude from the true value, so the meaning of the logarithmic transformation here is to eliminate these anomalies.

[0058] (2) Loss function

[0059] In the training process, the total loss function used is composed of the cosine similarity loss of the unit direction and the L1 loss of the logarithmic distance, defined as:

[0060]

[0061] where, the predicted moving direction and the logarithmic moving distance The corresponding true value is denoted as D (t) , (ln d) t , which comes from the synthetic tree dataset. This dataset contains virtually generated single tree point clouds for training and the true centripetal displacement corresponding to each point. The required true logarithmic moving distance can be obtained by back calculation from the previous formula

[0062] lnd = ln(||L (1) ||2, ||L (2) ||2,..., ||L (N) ||2) T ,

[0063] and the true moving direction

[0064]

[0065] where L is the true centripetal displacement of each batch of sampling points ΠP in each round of training or evaluation; ⊙ represents the Hadamard product or element-wise product of matrices and matrices; denotes the Kronecker product or tensor product of vectors and vectors. ​​

[0066] (3) Network-embedded Iterative Point Cloud Shrinkage

[0067] The point cloud shrinkage based on sparse convolution can be represented as:

[0068]

[0069] where P (n-1) , P (n) denote the point cloud formed by the previous (n-1) and current (n) round of chunked batch points, respectively; M (n) is the mask operation in the current round, which is used to eliminate all the results in the previous round within the supplementary buffer; Π (n) denotes the sampling operation in the current round, including re-segmentation of the point cloud, chunked buffer increase, and voxel down-sampling; and are the predicted moving direction and logarithmic moving distance of the current round, respectively.

[0070] The voxel size and the cube size used for testing are twice the size used for training, which are 0.02 and 8m, respectively. This is because the trees in the synthetic tree dataset are relatively short and simple in structure. If the original size is used, the shrinkage effect will be extremely insignificant. Larger size means loss of details, which is not conducive to the representation of the skeleton of fine tree branches. For very small trees, the current size does not meet the executable condition, so the program will try to change these sizes by dividing by 2 to ensure executability.

[0071] The maximum number of iterations is set to 10, and the program will terminate early when the number of points in the point cloud is less than 1000 after a certain round of shrinkage to prevent loss of branch structure details. After the shrinkage is completed, a voxel down-sampling is performed again, and the points at this time are identified as skeleton points s. Each skeleton point is obtained by moving a surface point of a single tree, and the distance between them is regarded as the initial radius of the circular cross-section at that skeleton point.

[0072] 2. Quantitative structure modeling

[0073] (1) Initial skeleton construction

[0074] Delaunay tetrahedralization can make the graph constructed by skeleton points contain only one connected component. On this graph, the shortest path from the lowest skeleton point to other points is found using Dijkstra algorithm. To increase the preference for more proximal reachable skeleton points, the edge weight uses the square distance. At this time, the shortest path tree can be regarded as the skeleton, but there are a large number of short branches inside the branches, which should be eliminated, so path selection is performed. Before path selection, the skeleton points are orderly divided into different branches according to the length of the branches supported by each skeleton point. With the help of the initial radius, a series of spherical domains with the skeleton points on the branches as the centers are constructed in the order of priority of the branch levels, and the skeleton points that do not belong to the current branch and have not been discarded previously should be discarded in these spherical domains, and the topology of the skeleton points of each branch is reconstructed.

[0075] (2) Skeleton refinement

[0076] Skeleton smoothing and trajectory smoothing are essentially similar, both of which reduce or eliminate collisions in paths composed of discrete sampling points. At present, the fem position deviation smoother used in the Apollo of Baidu automatic driving platform simultaneously considers smoothness, length, and offset from the original point, has fast solving speed and good effect, and is therefore modified into a three-dimensional implementation to smooth the topology of each branch respectively. At this time, the disconnection of the skeleton is caused by the fact that each branch is not connected to the connecting point of the parent branch, that is, the starting point of each branch needs to be found in the parent branch. Compared with the calculation complexity and topology modification brought by solving the best connecting point, it is more cost-effective to directly find a point on the parent branch. The nearest neighbor point to the starting point on the parent branch can be used to construct a right triangle to find the desired connecting point, and the parent branch point closest to the connecting point is used as the connecting point. However, sometimes the connecting point found is too far away, so the branch angle needs to be constrained (must be within 15°-165°), and for those that do not meet the requirements, the parent branch point that changes the growth direction of the branch the least within this angle range is set as the connecting point.

[0077] (3) Radius estimation

[0078] Noise data points are removed using a radius range filter, and then the function relationship between the trunk length from each node of the trunk to the tip and the initial radius is fitted using several commonly used dry shape curves. The filter eliminates noise with large oscillation amplitude by constraining the various proportional relationships between the trunk length from each node of the trunk to the tip and the initial radius. Power function, exponential function, quadratic function, Bezier curve, and B-spline curve are tried to fit the trunk profile, and these curves all increase the constraint of monotonically increasing within the value range and passing through the origin to prevent overfitting. The fitting effect of each curve is recorded to the file, and among them, the B-spline curve is uniquely used for subsequent model establishment because it does not overfit.

[0079] (4) Three-dimensional modeling of trees

[0080] A 3D model of the tree is built with the refined skeleton and the estimated radius. The branch structure uses an arterial snake model, which enforces the mapping of points between cross sections, facilitating efficient volume and area computation and texture UV mapping. The leaves generated with a simple parameterized phyllotaxis rule (see Table 2) are combined with bark texture to improve the visual quality of the 3D model.

[0081] Table 2 Simple phyllotaxis

[0082]

[0083] 3. Multi-level parameter extraction

[0084] The multi-level parameters supported for extraction are shown in Table 3.

[0085] Table 3 Extractable parameters

[0086]

[0087]

[0088]

[0089]

[0090] Example 2: As shown in Figure 1 , Figure 2 , Figure 3 , Figure 4 and Figure 5 , a method for tree structure parameter extraction and 3D reconstruction from terrestrial laser point cloud data, with specific implementation cases:

[0091] 3D reconstruction and parameter extraction of a typical temperate natural secondary forest sample in Qingyuan Forest Ecosystem Observation and Research Station

[0092] 1. Data collection and preprocessing

[0093] A typical temperate natural secondary forest sample including 22 tree species of coniferous, broad-leaved, liana, shrub and small tree was selected in Qingyuan Forest Ecosystem Observation and Research Station, which has been observed for a long time. The sample size is 30m x 30m, and ground laser scanning was performed in late autumn (late October) in 2020 and 2024 after leaf fall, using a multi-station ground scanner and a backpack mobile laser scanner, respectively, as shown in Figure 2 .

[0094] The obtained two-phase ground laser point cloud data is processed to obtain single tree point cloud data through the processing flow as shown in Figure 3 .

[0095] AsFigure 3 The TLS data of the sample plot in 2020 was scanned using RIEGL VZ400i, and the engineering results were spliced using the RiSCANPro (version 2.10) software to obtain the entire sample plot cloud. The scanning device used for the MLS data of the sample plot in 2024 was LiGrip H120, and the engineering was solved by SLAM using the LiDAR360MLS (version 8.0) software. The solving results were divided into multiple files, and the files were merged in LiDAR360 (version 8.0) to form a complete sample plot cloud. After obtaining the sample plot cloud, the standardization processing procedure was performed in LiDAR360:

[0096] (1) Preprocessing: including cropping the required area, coordinate system matching, downsampling, denoising, ground point classification, normalization, CSF ground point filtering, and ground point / non-ground point segmentation. Among them, cropping the required area means manually deleting points outside the sample plot, and expanding a certain range outside the actual sample plot range to prevent incomplete trees at the edge of the sample plot; coordinate system matching means manual registration of point clouds from different sources through data registration; ground point / non-ground point segmentation is realized by category extraction. The output is normalized ground point cloud and normalized non-ground point cloud.

[0097] (2) Normalized ground point processing: including denormalization. The output is the ground point cloud.

[0098] (3) Normalized non-ground point processing: including denoising, single tree segmentation, denormalization, point cloud extraction to multiple files according to tree ID, manual correction and interpretation. Among them, single tree segmentation uses point cloud segmentation; manual correction and interpretation are used to assign or match sample tree numbers to the coarse single tree segmentation results split into individual files and correct the misclassified and missed points. The output is all single tree point clouds in the sample plot.

[0099] 2. Reconstruction quality

[0100] Figure 4 , Figure 5 Visual display is performed from the perspectives of single tree reconstruction effect and sample plot reconstruction effect.

[0101] Figure 4 The global and local hierarchical model reconstruction effects of a Japanese larch at two perspectives, the overall and local reconstruction effects of multiple tree species, and the two-period map models of one tree of each of 22 tree species are displayed.

[0102] Figure 5 The effects of simultaneously displaying hierarchical models of all trees in the sample plot and the constructed digital twin scene are displayed.

[0103] 3. Parameter extraction

[0104] Table 4 presents the overall performance of several important individual tree measurement factors on the two periods of data for the 22 trees. Figure 4

[0105] Table 4 several individual tree measurement factors

[0106]

[0107]

[0108]

[0109] Table 5 presents the results for several important branch attributes.

[0110] Table 5 several branch attributes

[0111]

[0112] The above description is only the preferred embodiment of the present application, and is not used to limit the present application, any modification, equivalent replacement and improvement within the spirit and principle of the present application should be included in the protection scope of the present application.​

Claims

1. A method for extracting tree structure parameters and three-dimensional reconstruction of ground laser point cloud data, characterized in that, The method comprises the following steps: taking a tree point cloud as input, predicting skeleton points through a sparse convolution-based residual U-shaped network (SpConv-based ResUNet) by multiple iterations, taking a voxelized attribute-containing voxel as input, predicting a moving direction and a moving distance through the sparse convolution-based residual U-shaped network, calculating a predicted centripetal displacement of the point cloud according to a result of the model, removing a buffer zone through a mask to obtain a contracted point cloud, repeating the process until a last contracted point cloud is obtained after voxel down-sampling to obtain skeleton points, and calculating an initial radius at each skeleton point, constructing an initial skeleton through graph analysis, searching for a connection point through path smoothing and angle constraint, using a fem position deviation smoother algorithm to smooth topology of each branch, constructing a right triangle with a nearest neighbor point of a parent branch to a starting point to find an expected connection point, taking a parent branch point closest to the connection point as the connection point, for a connection point that is too far away, constraining a branch angle, for a connection point that does not satisfy the branch angle constraint, taking a parent branch point with a minimum change in growth direction within an angle range as the connection point, estimating a radius based on a constrained B-spline curve, using an arterial snake to construct a branch model, and using a simple phyllotaxis algorithm to insert virtual leaves to realize simulation of a tree, and automatically extracting parameters by using the input point cloud, the extracted skeleton and the reconstructed model.

2. The method of claim 1, wherein, The graph analysis is to find a shortest path tree on a connected graph obtained through Delaunay tetrahedralization of the skeleton points, and to select a path based on the tree graph. 3.The method of claim 1, wherein, The branch angle ranges from 15° to 165°.

4. The method of claim 1, wherein, The model of the arterial snake establishes a mapping relationship of cross-section vertices for volume and surface area calculation.

5. The method of claim 1, wherein, The simple phyllotaxis algorithm realizes basic simulation of acrogeny, homophyll, cyclosis and fascicled phyllotaxis according to parameterized leaf size and arrangement.

Images