TreeSpecies-PC2DT: Automated Tree Species Modeling from Point

Clouds to Digital Twins

Like Gobeawan

1 a

, Xuan Liu

1 b

, Chi Wan Lim

1 c

, Venugopalan Raghavan

1 d

, Joyjit Chattoraj

1 e

Jan Schindler

and Feng Yang

Institute of High Performance Computing, Agency for Science, Technology and Research,

1 Fusionopolis Way #16-16 Connexis, Singapore, Singapore

Manaaki Whenua - Landcare Research, Wellington, New Zealand

schindlerj@landcareresearch.co.nz

Keywords:

Digital Twin, Tree Species Model, Point Cloud Data, Tree Branch Reconstruction, Procedural Modeling,

Optimisation.

Abstract:

3D digital twin trees for a city-scale have been limited to low-resolution, static shape models due to challenges

in automation/scalability, cost performance, tree growth dynamics, species complexities and compatibilities

with simulations and virtual city platforms. To address those challenges for high-resolution tree models,

we propose an automated workﬂow of generating large-scale, lightweight, dynamic digital-twin tree species

models from point cloud data. Species digital twins are modelled as detailed hierarchical branch structures

by solving for all species proﬁle parameters through stages of branch reconstruction from point cloud data,

species proﬁling by machine learning, tropism transfer, optimisation and species growth modelling based on

botany and limited ﬁeld survey. We show that the generated high-resolution tree models can be lightweight

while representing their true species characteristics and dynamic botanical architecture (branching patterns

and growth processes).

1 INTRODUCTION

1.1 Motivation

With the rise of artiﬁcial intelligence (AI) technolo-

gies for automation tasks, the digital twin (DT) for

objects, processes or systems in real life are gain-

ing traction in manufacturing, aviation and transporta-

tion, healthcare, medicine (Barricelli et al., 2019), and

more slowly, smart city domain and city-scale DT

for urban planning, sustainability and climate change

mitigation (Mylonas et al., 2021). DT cities poses

many challenges in coverage and operationalisation

(Lei et al., 2023) due to complexities in urban life

comprising physical structures and living things (hu-

man beings, greeneries and animals). In particular,

DT cities often focus on physical structures at various

https://orcid.org/0000-0001-6501-6394

https://orcid.org/0000-0002-1760-9993

https://orcid.org/0000-0002-8319-9742

https://orcid.org/0000-0003-0831-4218

https://orcid.org/0000-0003-1910-8954

levels of details and leave out the dynamic structure of

greeneries that play essential social-economic, envi-

ronmental roles in green cities such as Singapore (Na-

tional Research Foundation, 2018). More speciﬁcally,

DT trees for a city-scale have been limited to low-

resolution/simplistic, static watertight shape models

for their cost performance and compatibilities with

simulations.

On the other hand, addressing dynamic species-

level trees and their interactions with the environ-

ments in the form of high-resolution species digital

twins will enable more comprehensive simulations

and analysis, leading to better insights and improved

urban planning. However, this poses a grand chal-

lenge in automating and maintaining the modelling

of millions of dynamically growing tree species at a

large/city scale based on remote sensing data. Firstly,

remote sensing data, such as (lower-resolution) satel-

lite imagery and (higher-resolution) LiDAR (light de-

tection and ranging) scanned point cloud, generally

contain fuzzy information of tree characteristics due

to noises, obstacles and other limitations in their ac-

quisition. Thus, they require manual pre-processing

Gobeawan, L., Liu, X., Lim, C., Raghavan, V., Chattoraj, J., Schindler, J. and Yang, F.

TreeSpecies-PC2DT: Automated Tree Species Modeling from Point Clouds to Digital Twins.

DOI: 10.5220/0012389700003660

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 19th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2024) - Volume 1: GRAPP, HUCAPP

and IVAPP, pages 81-91

ISBN: 978-989-758-679-8; ISSN: 2184-4321

Figure 1: Automated digital twin tree species model from

remote sensing data.

and patching in order to produce estimated models

with limited accuracies on the actual branch structure

of the tree species. Secondly, temporal data of tree

growths are very scarce. They can only be reliably in-

ferred from high-resolution MLS (mobile laser scan-

ning) point cloud or photogrammetry, whose acquisi-

tion is, in contrast, low in frequency (typically annu-

ally or biennially) for a city-wide campaign. Thirdly,

such high-resolution data of millions of trees have an

extremely low cost performance - they take consider-

able resources to exist dynamically on DT city plat-

forms.

1.2 Problem Statements

This work seeks to automatically generate city-scale,

lightweight and dynamic tree models to represent the

trees at a high-resolution species level, i.e. displaying

high-ﬁdelity branch patterns and growth processes of

individual trees. To remain lightweight, the species

models should not be represented by mesh surfaces

or volumes with high count of polygons or voxels.

1.3 Literature Review

Most DT city tools seek to model (in 3D) static physi-

cal structures at various level of details, yet stop short

in modelling greeneries as low-resolution, static 3D

shape tree models (Lin et al., 2018) in virtual cities

(National Research Foundation, 2018; Government,

2019). For high-resolution trees, (Gobeawan et al.,

2018; Stava et al., 2014; Makowski et al., 2019) at-

tempted to generate dynamic tree models from re-

mote sensing data to high-resolution textured polygon

models, however they are not validated for a city-scale

implementation as their high polygon counts lead to

tendency to overload the platform rendering in real

time.

In the ﬁeld of FSPM (functional-structural plant

modelling), there are many works (Godin and Sino-

quet, 2005; Vos et al., 2010; Siev

anen et al., 2014;

Yi et al., 2018; Talle and Kosinka, 2020; Niese et al.,

2020) on modelling individual tree species along with

their growth and functions at high levels of detail, yet

very speciﬁc to certain individual or species, hence

they are not scalable. At the core of FSPM, L-system

(Prusinkiewicz and Lindenmayer, 1996) is a good

candidate to produce lightweight species models, as

it adopts a procedural modelling approach of using

a small set of rules to generate a wide variety of tree

models. However, there have been very few concerted

efforts to establish a procedural-based framework for

DT species at the city-scale.

As such, we propose an automated workﬂow of

city-scale, dynamic tree species modelling from re-

mote sensing data to digital twin tree models.

2 METHODOLOGY

Our automated workﬂow, herein codenamed as

TreeSpecies-PC2DT, starts with the point cloud data

of a tree as an input and generates a ﬁnal output of a

digital twin species model (Figure 1). The workﬂow

comprises of 5 main modules:

• branch segmentation and reconstruction for tree

measurements

GRAPP 2024 - 19th International Conference on Computer Graphics Theory and Applications

• species proﬁling

• tropism transfer

• constraint optimisation

• species growth modelling

Each digital twin tree is individually modelled

to closely match the actual trees by a set of param-

eters with different values that deﬁne their unique

species proﬁles. The values of these parameters are

derived for an individual tree by reconstructing pre-

liminary branch structures from point cloud data by

branch point classiﬁcation and branch/skeleton re-

construction. Due to fuzziness in the crown where

leaves might obstruct high-order branches (twigs), the

branch reconstruction stage is only effective in mea-

suring trunk and branch components which are not

concealed by leaves and other objects within the tree

crown. Hence, the tree measurements can be obtained

for trunk parameters such as trunk height and diame-

ter, trunk pitch and roll angles, as well as ﬁrst order

branch count and ﬁrst branching pitch and roll angles.

Subsequently, the ML (machine learning)-based

species proﬁling module will use the reconstructed

branches to predict values of other unknown tree pa-

rameters with a certain level of conﬁdence. The ML

is trained to recognise or estimate tree parameter val-

ues by learning from the big data of synthetic species

models which were generated based on true botanic

growth and branching processes incorporated in the

species growth modelling module. The knowledge

learnt is then transferred to train on reconstructed

branches derived from the scarce real point cloud

data.

Along with the former two modules above, the

tropism transfer module estimates the tree tropism,

i.e. the growth response of the tree to environment

stimuli. The tropism transfer uses the reconstructed

branch structure and existing tropism models in liter-

ature to derive the change in growth direction along

the tree stem and the stem elasticity parameter of the

tree along its proximal (trunk) and distal (branches).

Finally, the remaining unknown parameter values

(or parameter values with low conﬁdence) will be

solved by the optimisation module. The completely

solved parameter conﬁguration will be fed into the

species growth modelling module to generate the dig-

ital twin species models.

The species growth modelling module contains

a set of growth rules that dictate how the tree

grows over time. The growth rules are formu-

lated from botany knowledge and limited ﬁeld sur-

vey/measurements.

As a whole workﬂow, TreeSpecies-PC2DT is

able to automatically process city-scale individual

trees from the point cloud data into the botanically-

representative, lightweight, and dynamic digital twin

species models.

2.1 Input Data

The input to TreeSpecies-PC2DT workﬂow is LiDAR

(light detection and ranging) point clouds with res-

olutions ranging from high to low. The point cloud

resolution depends on the means of scanning, such as

MLS (mobile laser scanning), TLS (terrestrial laser

scanning), and ULS (UAV (unmanned aerial vehicle)

laser scanning) or ALS (aerial laser scanning).

The main input data used for the workﬂow demon-

stration in this paper is the high-resolution MLS point

cloud, which were acquired by Singapore Land Au-

thority (SLA) (Soon and Khoo, 2017) using Riegl

VMX-450 with a density of 40 points/m

. Individual

tree point clouds were extracted using tree segmenta-

tion algorithms (Gobeawan et al., 2018). A group of

MLS datasets are also placed on a plot (Figure 2) for

testing the tree modelling workﬂow.

Figure 2: Individual MLS trees on a plot.

Figure 3: Segmented individual trees from subcanopy TLS

data.

We also experimented with the subcanopy TLS

data of a complex natural forest plot in New Zealand

(Figure 3). The point cloud data was captured with

a FARO Focus3D X 330 laser scanner in the Oron-

gorongo Valley in the Greater Wellington Region,

New Zealand. The study site is particularly challeng-

TreeSpecies-PC2DT: Automated Tree Species Modeling from Point Clouds to Digital Twins

ing because of dense undergrowth including vines, a

mixture of native species and rough terrain making

accurate high-resolution measurements difﬁcult. The

point cloud was stitched together from several mea-

surements with the help of survey reﬂectors placed

in the ﬁeld, and further post-processed for noise ﬁl-

tering, ground-vegetation classiﬁcation, as well as

woody material and stem segmentation and individual

stems using point clustering. Starting at seeded stem

locations, the entire individual trees were segmented

using a region-growing algorithm.

2.2 Branch Segmentation and

Reconstruction for Tree

Measurements

This section entails branch segmentation (i.e. branch-

leaf classiﬁcation) and branch reconstruction for de-

riving tree measurements. While there have been

works in tree measurements from point cloud data

such as (Hartley et al., 2022), in general they still

require labourious onsite survey to obtain measure-

ments for higher order branches of the trees. Hence,

we focus on automated tree measurements for lower

order trunk and branches.

Given the input point cloud data of an individ-

ual tree, we perform branch-leaf classiﬁcation to ex-

tract the tree’s woody points and then branch recon-

struction to connect those woody points into a skele-

ton structure of branch centre points with their corre-

sponding radii (Figure 4).

The branch segmentation and reconstruction algo-

rithms are based on (Lim et al., 2020), producing an

MTG (multiscale tree graph) detailing the hierarchi-

cal branch points with their radius and connectivity.

Each node in the MTG ﬁle consists of a radius param-

eter to indicate the thickness of the branch. The ﬁrst

node of the MTG node is the tree base on the ground.

Using the branch connectivity and radius values, we

can derive values of an initial set of parameters for

tree measurements, such as trunk height, trunk pitch

and roll angles, trunk diameter and girth, branch pitch

and roll angles, and number of ﬁrst order branches

(Figure 5). To get appropriate measurements, it is

important to modify the MTG in (Lim et al., 2020)

to consider the actual sizes of trunk and branches

in determining the hierarchy of all nodes (stems and

twigs), i.e. parent-child relationships at all branching

points.

To do this, radii of all nodes corresponding to

each branching point are compared, starting from the

ground level of the trunk, i.e. the ﬁrst node of the

MTG. If any trunk node has a single child node, that

child node is also assigned as the trunk node. In the

Figure 4: (a) Original point cloud, (b) woody point cloud,

growth space (green), (d) hierarchy of branch centre points

and their radii (MTG format).

Figure 5: Tree measurement diagram.

Figure 6: Woody segmentation and branch reconstruction

for tree measurement of Khaya senegalensis.

GRAPP 2024 - 19th International Conference on Computer Graphics Theory and Applications

Figure 7: Species proﬁling ﬂow.

case of a branching point with multiple child nodes, a

simple test is applied to determine if any of the child

node is a trunk node too. If the combined cross-

sectional area of the child nodes is larger than the

cross sectional area of the parent node, then the child

with the largest radius assigned as the continuation of

the trunk node.

Once all the trunk nodes are assigned, tree mea-

surements can be derived (Figure 6). The trunk height

can be computed by adding up the distances between

adjacent trunk nodes. The trunk ground girth is taken

as the circumference of the trunk node closest to

the ground level. The trunk pitch angle is taken as

the smallest angle from the up vector to the trunk

axis, while the trunk roll angle is measured counter-

clockwise from the North vector to the projected trunk

axis on the horizontal plane. The number of ﬁrst or-

der branching is taken as the ﬁrst instance of multiple

child node encountered from ground up. The branch

pitch and roll angles are derived from the ﬁrst order

branches with respect to the trunk.

2.3 Species Proﬁling

The species proﬁling module aims to determine the

values of all species proﬁle parameters with corre-

sponding levels of conﬁdence (Figure 7). If unknown

parameters can be solved with an acceptable level

of conﬁdence, they are considered solved and subse-

quently passed to the next stage in the TreeSpecies-

PC2DT workﬂow.

The species proﬁling work has been described in

detail in (Chattoraj et al., 2022). In this work, ML-

based method of knowledge transfer learning from

abundant synthetic species data with known param-

eter values to scarce real tree data was devised to de-

tect species proﬁle parameter values of a given real

tree input.

The species proﬁle parameters and their values

(listed in Table 2) are collated from all runs and

organised into XML (extensible markup language)-

formatted:

• individual proﬁles for individual trees with their

known individual and species information

• species proﬁles for known species with their

species-speciﬁc statistics obtained from botany

knowledge and ﬁeld surveys, then bootstrapped

by newly detected values from species proﬁling

of incoming real tree data, and

• species library as a collection of species proﬁles

and individual proﬁles.

The species library are used by the species growth

modelling module (Section 2.6) to generate synthetic

species data for training the species proﬁling module

(Gobeawan et al., 2023).

2.4 Tropism Transfer

The tropism transfer module seeks to detect tropism

characteristics (such as bending curvature and elastic-

ity of tree trunk and branches) from reconstructed tree

branches and to transfer the tropism onto DT species

models. As tropism detection requires temporal data

such as tree growth snapshots which are generally not

present in individual reconstructed trees, we rely on

growth snapshots of synthetic trees or experimental

tropism data to derive ranges of values of relevant

tropism parameters in our species growth modelling

(Section 2.6).

While our preliminary work (Raghavan et al.,

2023) performs a direct transfer of the trunk bend-

TreeSpecies-PC2DT: Automated Tree Species Modeling from Point Clouds to Digital Twins

ing appearance from a tree point cloud skeleton to an

individual tree model, the tropism transfer module in

this paper expands it as a mapping of the trunk/branch

bending appearance of a tree point cloud skeleton to a

tropism model that predicts growth changes of a tree

model in response to environment/external stimuli.

Figure 8: Diagram of phototropism and gravitropism on a

stem.

Modelling of tropism, particularly the tree growth

response to light (phototropism) and gravity (gravit-

ropism) as well as its own ability to grow straight in

the absence of environment stimuli (autotropism), has

been done by deriving a mathematical representation

of the behaviour of plants (Bastien et al., 2015; Moul-

ton et al., 2020). In (Bastien et al., 2015), the tropism

is modelled as changes in angles and curvatures of the

part of the plant that is growing (Figure 8). Depend-

ing on the considered tropisms (phototropism, gravit-

ropism and/or autotropism), angles and curvatures are

linked to one or more tropism sensitivity parameters

(photosensitivity, gravisensitivity and/or propriosen-

sitivity) by partial differential equations (PDEs).

Our current work considers all three tropisms by

adopting the A

C model in (Bastien et al., 2015) (Fig-

ure 8), which is formulated as

∂C(s, t)

∂t

= −ν(A(L, t)) − A

) − βA(s, t) − γC(s, t)

(1)

where s represents the curvilinear coordinate of a

node along a stem (s = 0 at stem end at ground and s =

L at the apical end), A(s, t) is the node’s bending angle

(with respect to up vector) at s and at time t, C(s,t)

is the local curvature at s and t, A

is the direction

angle of light from down vector, while ν, β and γ are

the photo-, gravi- and proprio- sensitivity coefﬁcients

respectively.

∂A

∂s

= C, Equation 1 is complex to solve an-

alytically, hence a numerical approach is adopted.

While numerical approaches are capable of handling

integro-differential equations, Equation 2 is assumed

to ensure a better understanding of the underlying

phenomena.

A(s,t) = A(0, t) +C(s, t)s (2)

Thus, Equation 1 becomes

k+1

−C

∆t

= −ν(A

− A

) − βA

− γC

(3)

where superscript k indicates the current time step

while the subscript i indicates the current point along

the growing member.

To solve Equation 3, we adopted an explicit time-

stepping with a restriction on the value of ∆t that

can be used. Our tests indicate that the time-step ∆t

should range from 1e

−3

to 1e

−4

time units to ensure

convergences for the angle and curvature values.

Assuming that the values of the tropism sensitiv-

ity parameters, ν, β and γ are obtained from available

growth data (either from ﬁeld experiments or analyt-

ics from synthetic tree data), the angles and curvatures

of all points along the stem for each time step can be

computed.

By solving all the tropism parameters above, we

can generate species models with the same tropism

behaviour of a tree input (Figure 9). Based on the in-

herent characteristics of each species, the exhibited

tropism effect will be different even with the same

speciﬁed values for the tropism parameters (refer to

species proﬁle parameters for tropism in Table 2).

(a) (b)

Figure 9: A tropism is transferred to four tree species: (a)

Archontophoenix alexandrae, (b) Khaya senegalensis, (c)

Samanea saman, (d) Hopea odorata.

GRAPP 2024 - 19th International Conference on Computer Graphics Theory and Applications

Table 1: Optimised parameters for various tree samples.

2.5 Optimisation

The ﬁnal round of solving for remaining parameters

with unknown values or low conﬁdences from branch

reconstruction and species proﬁling modules is done

by optimisation. The optimisation module utilises a

genetic algorithm (GA) method to ﬁnd the best con-

ﬁguration of species proﬁle parameters within speci-

ﬁed physical constraints (Gobeawan et al., 2021; Lim

et al., 2023). An initial database is generated from

a population of possible parameter conﬁguration so-

lutions over ranges of possible values appended with

cost function evaluated for each solution. The algo-

rithm then pairs a selection of the best solutions with

a number of random solutions from the database, cre-

ating parent solutions. A successive generation of 4

child solutions are generated from the parent solution

by making slight adjustments to their values. These

successive generations along with their correspond-

ing cost function elevations, are appended to update

the database. This process continues iteratively until

a predeﬁned number of generations is reached or the

cost meets the speciﬁed threshold.

The cost function of the optimisation measures

a linear weighted distances between the solution

and the physical constraints. Theses physical con-

straints include both macroscopic shape (i.e.crown

shape and dimension) and microscopic structure of

the target tree (i.e. trunk bending, measurements,

growth space). Speciﬁcally, the growth space refers

to a set of voxels within a uniform grid in the 3D

space derived from input point cloud data. Details

of the cost function are described in (Gobeawan et al.,

2021).

Table 1 shows the optimised parameters for indi-

vidual trees of various species. It has 3 categories

of parameters: (1) solved by branch reconstruction:

trunk pitch angle, trunk roll angle, trunk height, num-

ber of ﬁrst order branches and branch pitch angle, (2)

solved by species proﬁling: phyllotaxis and monopo-

dial/sympodial branching and (3) solved by optimisa-

tion: the rest of all parameters. Any solved parame-

ters with low conﬁdence levels in the previous mod-

ules (branch reconstruction and species proﬁling) will

be re-solved by optimisation.

Figure 10: Optimised species models (black) within growth

spaces (green).

TreeSpecies-PC2DT: Automated Tree Species Modeling from Point Clouds to Digital Twins

Figure 11: Comparison of running times for each iteration

with and without parallelisation.

In addition, if the tree comes with identiﬁed

species, the optimisation search space can be further

reduced to within smaller ranges of parameter values

speciﬁc to that species.

The optimisation process is set to terminate by a

limited period of time or when the error E falls be-

low 10%, whichever is earlier. Their corresponding

tree models with given growth space (voxel spacing

of 50cm, minimum of 5 points per voxel) are shown

in Figure 10.

The optimisation process is also highly paral-

lelised for improving system efﬁciency and running

time. Figure 11 shows that the running time decreases

linearly with the increase in the number of computing

processors used.

2.6 Species Growth Modelling

The species growth modelling module aims to gen-

erate dynamically growing DT species models (Fig-

ure 12) based on the input values of species proﬁle pa-

rameters. The mechanism and formulation of growth

rules in this module is described in details in (Gob-

eawan et al., 2021). Further updates to the list of

species proﬁle parameters are shown in Table 2 to in-

clude tropism parameters. Additional growth param-

eters can be added to the list in order to simulate more

phenomena and interactions between trees and envi-

ronments.

In addition, synthetic models can be generated by

varying the values of species proﬁle parameters.

3 RESULTS AND DISCUSSION

We tested high-resolution MLS data of Singapore

trees and low-resolution TLS data of New Zealand

trees, both individually and as a scene of trees, in the

TreeSpecies-PC2DT workﬂow.

Given an individual tree of known species Khaya

senegalensis from MLS input data, the TreeSpecies-

PC2DT workﬂow generates a DT species model with

a good ﬁt with the actual scanned tree (Figure 13).

Figure 12: Clockwise from top left: Growth snapshots of a

typical tree of species Tabebuia rosea grows from 1 to 17

years old, with a growth space reference (green) at 11 years

old.

For an individual tree of unknown species from low-

resolution TLS input data, the workﬂow generates a

very simple DT model correspondingly, while captur-

ing the essential bending and architecture of the input

tree (Figure 14).

While time benchmark was not done in this exper-

iment, we show the planting of DT species models in

a plot for both MLS and TLS data in Figures 15 and

16 and observed stark difference in loading time for

DT models (in seconds) and original point clouds (in

minutes) on the same computer system. This shows

the potential of lightweight DT species models for

large-scale city platforms.

4 CONCLUSIONS

The proposed high-resolution, species-level represen-

tation DT models in MTG format is very lightweight

yet representing fundamental species characteristics

of branching patterns and dynamic growth, making

them suitable for large scale simulations. While their

visual appearance as not as realistic as high-polygon-

count trees with textures, the DT species models cor-

respond to the actual tree branch structure in real-

ity. Meshing such DT species models as in (Lim

et al., 2020) potentially turns the DT models into

simulation-ready models - enabling many environ-

mental simulations and analysis without using expen-

sive high-resolution tree models.

Future works for TreeSpecies-PC2DT are planned

for testing with more types of LiDAR data while eval-

uating the performance of DT species on virtual cities,

in addition to expansion to include root of the trees

in order to allow wider environment simulations us-

ing whole trees for tree health and safety management

purposes.

GRAPP 2024 - 19th International Conference on Computer Graphics Theory and Applications

Table 2: Species proﬁle parameters.

No. Parameter Name Format Unit Description

A Growth process

1 bud lifespan ≥ 0 year Lifespan of a bud, 0 if indeterminate

2 rhythmic growth period ≥ 0 year Period of rhythmic growth to produce

1 GU, zero if continuous growth

3 growth unit shape a choice - Unspeciﬁed (U), acrotonic (A),

mesotonic (M), or basitonic (B)

4 diameter growth model a choice - Relative growth rate (RGR): linear (L),

exponential (E), power law (P),

monomolecular (M), 3-param logistic (3PL),

4-param logistic (4PL), Gombertz (G)

5 initial diameter > 0 meter Non-zero minimum branch diameter

6 diameter growth ranges t:day, Variables for RGR equation with respect

variables r, β, K, L M:cm, to growth model

B Branching process

7 branching rhythm a choice - Continuous (C), rhythmic (R), or diffuse (D)

8 rhythmic branching pattern xxx - Binary rhythmic branching pattern of 0s and 1s

9 terminal/lateral a choice - New branches are formed by lateral

branching buds (L) or apical split (T)

10 branching delay ≥ 0 year New branches grow out immediately (0)

or after some delay

11 monopodial/sympodial a choice - Apical stem remains dominant with emergence of

branching lateral branches (M) or stops growing with

emergence of dominant lateral branches (S)

12 no. of dominant apices ≥ 1 - Sole or multiple dominant apices (scaffolds)

13 phyllotaxis type a choice - Alternate (1), opposite (2), or whorled (n)

14 divergence angle 0 ≤ θ ≤ 180 degree Angle between two branches from adjacent rows

C Tropism One or multiple entries

15 tropism type a choice - Gravitropism, phototropism, or autotropism

16 proximal response a choice - Trunk responds towards/against stimulus,

orthotropically/plagiotropically

17 distal response a choice - Branches respond towards/against stimulus,

orthotropically/plagiotropically

18 plagiotropic angle a range, degree For plagiotropy, at an angle from stimulus

0 ≤ θ ≤ 90

19 elasticity a range, - Ease to bend

0 < E < 1

20 source format a vector - Field, point, plane, or volume to respond to

21 pruning diameter ≥ 0 meter To prune branches smaller than pruning diameter

E Constraints

22 growth space a volume - Physical space occupied by the tree

23 age a range, ≥ 1 year Simulation age for the species

24 trunk ground girth a range, > 0 m Circumference of trunk closest to ground level

25 trunk pitch angle a range, degree Angle between up vector and trunk axis

0 ≤ θ ≤ 180 closest to ground

26 trunk roll angle a range, degree Counter-clockwise (CCW) angle from

0 ≤ θ < 360 North to trunk on ground

27 trunk height a range, ≥ 0 meter Distance from ground to ﬁrst branching point

28 no. of 1

order branches a range, ≥ 0 - At ﬁrst branching point from bottom

29 branch pitch angle a range, degree From parent’s head down to start of branch

0 ≤ θ ≤ 180

30 branch roll angle a range, degree Rotate CCW around parent’s head

0 ≤ θ < 360

31 diameter growth rate a range, - Normalised growth rate; ﬁrst three decimals

0 ≤ g ≤ 1 are mapped to diameter growth variables

32 number of new nodes a range, > 0 /year Number of nodes a bud produces in a year

33 internode length a range, ≥ 0 meter Length of a segment between adjacent nodes

TreeSpecies-PC2DT: Automated Tree Species Modeling from Point Clouds to Digital Twins

Figure 13: Generating DT species of Khaya senegalensis from MLS data: (a) MLS point cloud, (b) segmented branch point

cloud, (c) reconstructed branch (red) within growth space (green), (d) DT species (yellow) within growth space (green), (e)

DT species (yellow) superimposed with reconstructed branch (red) within growth space (green).

Figure 14: Generating a DT tree of unidentiﬁed species

from TLS data: (a) TLS point cloud, (b) segmented branch

point cloud (brown), (c) reconstructed branch (red) within

growth space (green), (d) DT tree (black) within growth

space (green).

Figure 15: Comparison of a scene of lightweight DT species

(top) with a scene of much denser MLS point clouds of the

same tree species (bottom).

Figure 16: Comparison of a scene of lightweight DT palms

(left) with a scene of denser TLS point clouds of palms

(right).

ACKNOWLEDGEMENTS

This research/project is supported by the Catalyst:

Strategic Fund from Government Funding, adminis-

tered by the Ministry of Business Innovation & Em-

ployment, New Zealand under contract C09X1923, as

well as the National Research Foundation, Singapore

under its Industry Alignment Fund – Pre-positioning

(IAF-PP) Funding Initiative. Any opinions, ﬁndings

and conclusions or recommendations expressed in

this material are those of the author(s) and do not re-

ﬂect the views of National Research Foundation, Sin-

gapore.

GRAPP 2024 - 19th International Conference on Computer Graphics Theory and Applications

REFERENCES

Barricelli, B. R., Casiraghi, E., and Fogli, D. (2019). A

survey on digital twin: Deﬁnitions, characteristics,

applications, and design implications. IEEE Access,

7:167653–167671.

Bastien, R., Douady, S., and Moulia, B. (2015). A uni-

ﬁed model of shoot tropism in plants: Photo-, gravi-

and propio-ception. PLoS computational biology,

11:e1004037.

Chattoraj, J., Yang, F., Lim, C. W., Gobeawan, L., Liu,

X., and Raghavan, V. S. (2022). Knowledge-driven

transfer learning for tree species recognition. In 2022

17th International Conference on Control, Automa-

tion, Robotics and Vision (ICARCV), pages 149–154.

Gobeawan, L., Chattoraj, J., Yang, F., Lim, C. W., Liu,

X., and Raghavan, V. S. G. (2023). Knowledge-

based learning for plant phenotyping. In Chen, T.-W.,

Kahlen, A. F. K., and St

utzel, H., editors, 10th Inter-

national Conference on Functional-Structural Plant

Models (FSPM2023), pages 140–141.

Gobeawan, L., Lin, E. S., Tandon, A., Yee, A. T. K., Khoo,

V. H. S., Teo, S. N., Su, Y., Lim, C. W., Wong, S. T.,

Wise, D. J., Cheng, P., Liew, S. C., Huang, X., Li,

Q. H., Teo, L. S., Fekete, G. S., and Poto, M. T.

(2018). Modeling trees for Virtual Singapore: From

data acquisition to CityGML models. International

Archives of the Photogrammetry, Remote Sensing and

Spatial Information Sciences, XLII-4/W10:55–62.

Gobeawan, L., Wise, D. J., Wong, S. T., Yee, A. T. K.,

Lim, C. W., and Su, Y. (2021). Tree species modelling

for digital twin cities. Transactions on Computational

Science XXXVIII, pages 17–35.

Godin, C. and Sinoquet, H. (2005). Functional–structural

plant modelling. New Phytologist, 166(3):705–708.

Government, N. S. W. (2019). Nsw spatial digital twin.

Hartley, R. J. L., Jayathunga, S., Massam, P. D., De Silva,

D., Estarija, H. J., Davidson, S. J., Wuraola, A., and

Pearse, G. D. (2022). Assessing the potential of

backpack-mounted mobile laser scanning systems for

tree phenotyping. Remote Sensing, 14(14).

Lei, B., Janssen, P., Stoter, J., and Biljecki, F. (2023). Chal-

lenges of urban digital twins: A systematic review and

a delphi expert survey. Automation in Construction,

147:104716.

Lim, C. W., Gobeawan, L., Wong, S. T., Wise, D. J., Cheng,

P., Poh, H. J., and Su, Y. (2020). Generation of tree

surface mesh models from point clouds using skin sur-

faces. In 15th International Conference on Computer

Graphics Theory and Applications, pages 83–92.

Lim, C. W., Liu, X., Gobeawan, L., Raghavan, V. S. G.,

Chattoraj, J., and Yang, F. (2023). Species model

parameterisation. In Chen, T.-W., Kahlen, A. F. K.,

and St

utzel, H., editors, 10th International Conference

on Functional-Structural Plant Models (FSPM2023),

pages 68–69.

Lin, E.-S., Teo, L.-S., Yee, A.-T.-K., and Li, Q.-H. (2018).

Populating large scale virtual city models with 3D

trees. In 55th IFLA World Congress, Singapore.

Makowski, M., H

adrich, T., Scheffczyk, J., Michels, D. L.,

Pirk, S., and Pałubicki, W. (2019). Synthetic silvicul-

ture: Multi-scale modeling of plant ecosystems. ACM

Trans. Graph., 38(4).

Moulton, D., Oliveri, H., and Goriely, A. (2020). Multiscale

integration of environmental stimuli in plant tropism

produces complex behaviors. Proceedings of the

National Academy of Sciences (PNAS), 117:32226–

32237.

Mylonas, G., Kalogeras, A., Kalogeras, G., Anagnostopou-

los, C., Alexakos, C., and Mu

noz, L. (2021). Digital

twins from smart manufacturing to smart cities: A sur-

vey. IEEE Access, 9:143222–143249.

National Research Foundation (2018). Vir-

tual Singapore. NRF website

https://www.nrf.gov.sg/programmes/virtual-

singapore.

Niese, T., Pirk, S., Albrecht, M., Benes, B., and Deussen,

O. (2020). Procedural urban forestry.

Prusinkiewicz, P. and Lindenmayer, A. (1996). The Al-

gorithmic Beauty of Plants. Springer-Verlag, Berlin,

Heidelberg.

Raghavan, V. S. G., Gobeawan, L., Lim, C. W., Liu, X.,

Chattoraj, J., and Yang, F. (2023). Detecting plant

tropism from lidar data. In Chen, T.-W., Kahlen,

A. F. K., and St

utzel, H., editors, Book of Abstracts

of the 10th International Conference on Functional-

Structural Plant Models (FSPM2023), pages 123–

124.

Siev

anen, R., Godin, C., Dejong, T., and Nikinmaa, E.

(2014). Functional-structural plant models: A grow-

ing paradigm for plant studies. Annals of botany,

114:599–603.

Soon, K. H. and Khoo, V. H. S. (2017). CITYGML

MODELLING FOR SINGAPORE 3D NATIONAL

MAPPING. In ISPRS - International Archives of the

Photogrammetry, Remote Sensing and Spatial Infor-

mation Sciences, volume XLII-4-W7, pages 37–42.

Copernicus GmbH.

Stava, O., Pirk, S., Kratt, J., Chen, B., M

zch, R., Deussen,

O., and Benes, B. (2014). Inverse procedural mod-

elling of trees. Comput. Graph. Forum, 33(6):118–

131.

Talle, J. and Kosinka, J. (2020). Evolving l-systems in

a competitive environment. In Magnenat-Thalmann,

N., Stephanidis, C., Wu, E., Thalmann, D., Sheng,

B., Kim, J., Papagiannakis, G., and Gavrilova, M.,

editors, Advances in Computer Graphics, pages 326–

350, Cham. Springer International Publishing.

Vos, J., Evers, J. B., Buck-Sorlin, G. H., Andrieu, B.,

Chelle, M., and de Visser, P. H. B. (2010). Func-

tional–structural plant modelling: a new versatile tool

in crop science. Journal of Experimental Botany,

61(8):2101–2115.

Yi, L., Li, H., Guo, J., Deussen, O., and Zhang, X. (2018).

Tree growth modelling constrained by growth equa-

tions. Computer Graphics Forum, 37(1):239–253.

TreeSpecies-PC2DT: Automated Tree Species Modeling from Point Clouds to Digital Twins