Content uploaded by Jianbo Qi
Author content
All content in this area was uploaded by Jianbo Qi on Oct 16, 2017
Content may be subject to copyright.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING 1
A Large-Scale Emulation System for Realistic
Three-Dimensional (3-D) Forest Simulation
Jianbo Qi, Donghui Xie, Dashuai Guo, and Guangjian Yan
Abstract—The realistic reconstruction and radiometric simula-
tion of a large-scale three-dimensional (3-D) forest scene have po-
tential applications in remote sensing. Although many 3-D radia-
tive transfer models concerning forest canopy have been developed,
they mainly focused on homogeneous or relatively small heteroge-
neous scenes, which are not compatible with the coarse-resolution
remote sensing observations. Due to the huge complexity of forests
and the inefficiency of collecting precise 3-D data of large areas, re-
alistic simulation over large-scale forest area remains challenging,
especially in regions of complex terrain. In this study, a large-scale
emulation system for realistic 3-D forest Simulation is proposed.
The 3-D forest scene is constructed from a representative single tree
database (SDB) and airborne laser scanning (ALS) data. ALS data
are used to extract tree height, crown diameter and position, which
are linked to the individual trees in SDB. To simulate the radio-
metric properties of the reconstructed scene, a radiative transfer
model based on a parallelized ray-tracing code was developed. This
model has been validated with an abstract and an actual 3-D scene
from the radiation transfer model intercomparison website and it
showed comparable results with other models. Finally, a 1 km ×
1 km scene with more than 100 000 realistic individual trees was
reconstructed and a Landsat-like reflectance image was simulated,
which kept the same spatial pattern as the actual Landsat 8 image.
Index Terms—Airborne laser scanning (ALS), three-dimen-
sional (3-D) forest reconstruction, 3-D simulation, forest.
I. INTRODUCTION
FORESTS cover approximately 31%of the land surface
across the globe and play a prominent role in the global car-
bon cycle [1]. Forests are highly complex and dynamic ecosys-
tems that comprise various species and a large number of in-
dividual trees. They are important and also widely managed
for biodiversity protection, wildlife habitat conservation, forest
products, and recreation [2]. Modeling the forest [3], [4] is an es-
sential step to assist management decisions [5]. However, most
of these models, such as the whole-stand model [6], rely on ab-
stract, conceptual, and two-dimensional (2-D) representations
of the forest canopy, which are rarely presented in a coherent
Manuscript received November 20, 2016; revised March 5, 2017 and May
2, 2017; accepted June 2, 2017. This work was supported by the National
Natural Science Foundation of China programs under Grant 41331171 and
Grant 41571341 and the National Basic Research Program of China under
Grant 2013CB733402. (Corresponding author: Donghui Xie.)
The authors are with the College of Remote Science and Engineering, Faculty
of Geographical Science, Beijing Normal University, Beijing 100875, China
(e-mail: qijb@mail.bnu.edu.cn; xiedonghui@bnu.edu.cn; 1130219602@qq.
com; gjyan@bnu.edu.cn).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/JSTARS.2017.2714423
and stimulating manner, and also complicate the translation of
this information into landscape images [7]. Consequently, new
powerful tools, such as precise 3-D landscape and forest models,
are required to model precise landscapes and environments [8].
In the remote sensing community, various methods have been
developed to extract forest parameters from remotely sensed
data [9]. However, most of these methods based on direct field
measurements are not completely and comprehensively vali-
dated due to the complexity of forest canopies and inhomo-
geneity of land surfaces [10], [11]. Thus, simulated datasets
resulting from radiative transfer models are widely used to vali-
date and analyze retrieval methods, and this simulation enables
us to precisely control the composition and spatial arrange-
ment of stand structures. It can help to understand the physical
process of interactions between the incoming solar radiation
and plants within the canopy, and it is also used to study
the radiative properties (scattering, absorption, and emission,
etc.) of specific land cover types [12] or to develop parametric
models [13].
To fully describe the scattering of radiation by vegetation
canopies, a number of 3-D radiative transfer models have been
developed [14], such as DART [15], [16] (discrete-ordinate ray
tracing), FLIGHT [17] (Monte Carlo ray tracing), and RGM
[18] (radiosity). All of them have been successfully applied to
the study of canopy properties. In [11], the DART model was
parameterized using airborne laser scanning (ALS) data and in
situ measurements to simulate the at-sensor spectral radiances,
and the results were compared with measurements of Airborne
Prism Experiment (APEX) [19]. In [20], a radiosity model, Ra-
diosity Applicable to Porous IndiviDual objects (RAPID), was
introduced for 3-D radiative transfer simulation over large-scale
heterogeneous areas. This method parameterized the single tree
crown into porous individual objects to achieve a better balance
between computing efficiency and accuracy. In order to obtain
the explicitly described 3-D structures of forest canopy, which
are the input for most 3-D models, individual tree architecture
(from leaf to crown), tree position, and tree species in a forest
area need to be predetermined.
A persistent and significant challenge in 3-D models is always
about precisely reconstructing the complex forest, which usually
contains plenty of trees, and each individual tree is constructed
with millions of leaves and branches. Thus, it is impractical to
directly reconstruct the structures of all the trees and branches.
Fortunately, a number of site- and species-specific allometric
equations which connect biomass and leaf area (LA) to diame-
ter at breast height (DBH) or tree height have been established
1939-1404 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
2 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
[21], [22]. These equations enable us to populate the leaf area
index (LAI) measured at plot scale to each individual tree with
the aid of field-inventory data (DBH, crown diameter and tree
height). With these parameters available, we can generate a se-
ries of individual tree models that have different DBHs by some
well-developed tree modeling packages [23]–[25]. For the tree
positions, they can be manually measured using a global po-
sition system (GPS). However, it is labor-intensive and time
consuming, notably impossible in large and dangerous areas.
Recently, airborne laser scanning (ALS) has been widely used
to study the structure of forests [26], [27]. It provides a promis-
ing alternative to acquire large-scale 3-D structure and spatial
distribution data of individual trees in a relatively short time.
In this paper, we presented a so-called large-scale emulation
system (LESS) for realistic 3-D forest simulation which can
rapidly generate virtual forest with explicitly described individ-
ual tree models. In addition, a radiative transfer model based on a
parallelized ray-tracing code was developed and validated. This
system can simulate visible images, multispectral images, and
bidirectional reflectance factor (BRF) for different research pur-
poses. The entire simulation process is mostly automatic with
only a few parameters to be set by the user, and this significantly
speeds up the emulation process.
II. MATERIAL
A. Study Area
The study area is located in the Genhe Forestry Reserve
(Genhe) (120◦12to 122◦55E, 50◦20to 52◦30N), Greater
Khingan of Inner Mongolia, Northeastern China. It has a hilly
terrain with 75%forest cover, which is mainly composed of
Dahurian Larch (Larix gmelinii) and White Birch (Betula platy-
phylla Suk.). Nine forest plots (labeled as L1 ∼L9) with the
area of 45 m ×45 m were established with their locations shown
in Fig. 1.
The leaf area index (LAI) of each plot was measured with a
TRAC (Tracing Radiation and Architecture of Canopies) instru-
ment (Third Wave Engineering, ON, Canada) in August 2013.
For each plot, the positions of four corners were measured by
a differential global positioning system (DGPS) device. Each
plot was divided into 3 ×3 subplots. In these subplots, the tree
height, tree position, crown diameter, and DBH were manually
measured. The position of each individual tree was obtained by
measuring the relative distances to the plot corners. Field data
are shown in Table I.
B. Spectral Data
The spectral reflectance and transmittance of white birch
leaves were measured by a field spectroradiometer (ASD Field-
Spec 3, Analytical Spectral Devices, USA) equipped with an in-
tegrating sphere in August 2013. Some of the white birch leaves
became yellow, so both green and yellow leaves were selected
for our measurements. The barks of white birch and larch were
also collected and their reflectance spectra were obtained. The
reflectance and transmittance spectra of needles were simulated
by LIBERTY [28] and the parameters used in LIBERTY were
Fig. 1. Locations of forest plots and the DEM of the Genhe Forestry Reserve.
TAB L E I
PLOTS IN THE STUDY AREA
Plot Birch
(number)
Larch
(number)
Mean height
(m)
Mean DBH
(cm)
LAI
L1 131 220 8.4 9.9 2.52
L2 111 467 9.1 8.8 3.06
L3 11 71 7.9 7.7 2.57
L4 294 79 11.7 11.5 4.51
L5 1 327 12.7 12.9 2.96
L6 0 173 17.3 22.3 2.3
L7 18 112 10.2 13.1 1.49
L8 63 118 13.3 15.6 2.44
L9 1 585 8.4 8.7 3.16
Fig. 2. Optical properties of the larch, soil background (left) and white birch
(right).
adopted from [29], which were retrieved based on the measured
reflectance of larch needles in the Great Xingan Mountain,
Inner Mongolia, very near to our study area. All the spectral
data were resampled from 400 to 1000 nm with a step of 1 nm
(see Fig. 2).
The canopy background (the underlying soil) reflectance was
obtained by comparing the measurements of the ground with
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
QI et al.: LARGE-SCALE EMULATION SYSTEM FOR REALISTIC THREE-DIMENSIONAL (3-D) FOREST SIMULATION 3
Fig. 3. Overview of the simulation system.
measurements of an ideal and diffuse standard surface in the
same observation geometry with an ASD FieldSpec instrument.
To obtain representative spectra of the ground, we averaged
50 ASD measurements that were uniformly distributed in the
plot. The type of ground was bare earth.
C. ALS and CCD Data
Data acquisition flights with an onboard LIDAR (Leica
ALS60, Leica Geosystem AG, Heerbrugg, Switzerland) and
a charge-coupled device (CCD-Leica RCD105, Leica Geosys-
tem AG, Heerbrugg, Switzerland) were performed from Au-
gust to September 2012 above the Genhe study Area, covering
approximately 230 km2. The flight height was about 1.8 km
above ground level. This measurement used a rotation scanning
system with an angle range of ±30◦. The wavelength of the air-
borne laser sensor was 1064 nm, and the laser beam divergence
was 0.3 mrad. The vertical accuracy of the acquired ALS data
was within 0.15 m. This LIDAR+CCD mission acquired point
cloud data with an average density of 8.0 points/m2and CCD
image with 0.5 m spatial resolution.
III. METHOD
The proposed system (LESS) mainly consists of four main
modules (see Fig. 3):
1) individual tree generation;
2) individual tree detection;
3) virtual forest scene generation;
4) and radiative transfer simulation.
A. 3-D Scene Reconstruction
1) Individual Tree Generation: It is usually difficult to pre-
cisely reconstruct each individual tree in a forest, especially for
larch, which has very small needle-like leaves. In this study, 3-D
structures of individual tree were generated from the OnyxTREE
software (www.onyxtree.com). This software can generate re-
alistic 3-D trees from a variety of parameters which define the
Fig. 4. Generated trees: (a) white birch tree; (b) larch tree.
architectures of the tree. Since DBH is an important parame-
ter which is usually related to other parameters, 20 birch and
20 larch trees with different DBHs were generated. To match
them to the real trees in this study area, the tree height, tree
crown diameter, and under branch height were obtained from
field measurements. Among all the parameters, the LA of an
individual tree was the most difficult parameter to be retrieved.
To get the LA, we used allometric-based method to populate the
plot LAI to individual tree LA as presented in [30]. For com-
pleteness, we summarized the main idea of this method here.
The relationship between DBH and LA can be expressed as:
LA =aDBHb(1)
where aand bare parameters which need to be determined for
different species. For a forest plot, the LA of the whole plot is
the sum of all the individual trees. Thus,
Splot ·LAI =
N
i
LAi=a
N
i
DBHib(2)
where Splot is the area of the plot. LAiis the LA of individual
tree. Nis total number of trees in the plot. Using plot L5 and
L9, aand bof white birch were estimated as 0.746 and 1.235,
respectively. The parameters of larch were a=0.06 and b=2.36
(using plot L1 and L4). It should be noted that L1 and L4
were mixed plots with white birches and larches, the LA of
white birches in these plots were calculated by the relationship
established by L5 and L9. Some representatives of the generated
trees for the SDB are shown in Fig. 4.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
4 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
Fig. 5. Generated 3-D forest, which is a part of a 1 km ×1 km forest scene
with more than 100 000 instanced trees from SDB.
2) Individual Tree Detection: Individual trees were ex-
tracted from the canopy height model (CHM) by a watershed
algorithm, which was also presented in [30]. The CHM used in
this study was obtained by subtracting a digital elevation model
(DEM) from a digital surface model (DSM). The DSM was
directly derived from ALS first returns, while the DEM was ex-
tracted by using an easy-to-use airborne lidar filtering algorithm
[31], which was based on cloth simulation. The final resolution
of the CHM map was 0.5 m. As no multispectral data were
collected in this area, the CCD image was used to distinguish
tree species. This image was obtained in September 2012 when
the leaves of white birch became slightly yellow while larch re-
mained green. We used a statistical histogram of CIELAB color
to separate birch from larch [30].
3) 3-D Forest Generation: 3-D forest scene of this study
area was generated by choosing a proper individual tree model
from SDB, and placing it on the DEM map according to the
detected parameters of individual trees from the CHM. Thus,
a similarity factor (SF) was defined to describe the selection
criteria:
SF =w|Hd−Hm|
Hm
+(1−w)|Cd−Cm|
Cm
(3)
where Hd(Cd) is the tree height (crown diameter) of a de-
tected tree from CHM map. Hm(Cm) is the tree height (crown
diameter) of a model from SDB, wis a weight parameter that
lies between [0,1]. By default, it is set to 0.5. To generate forest,
the model with the smallest SF is selected. As the number of
individual trees in SDB is limited, a tree model may be used
multiple times. Thus, an ”instance” mechanism is adopted. If
a tree model has already been selected before the current se-
lection, an instance or reference of that model is created. This
mechanism can reduce the memory usage, especially when there
are millions of individual trees. Fig. 5 shows a generated 3-D
virtual forest, which contains more than 100 000 trees.
Fig. 6. Path tracing algorithm.
B. Reflectance Image Simulation
1) Simulation Fundamentals: The simulation system in this
study uses physically based ray-tracing as the basis to calculate
the radiances at the top of canopy (TOC). Rays are traced from
the virtual sensor into the virtual scene, and the intersections
between rays and objects in the scene are tested. The rays are
defined to be launched in parallel within a field of view (ortho-
graphic projection) that covers the whole scene. At each inter-
section point, the bidirectional reflectance distribution function
(BRDF) is used to determine the incoming direction of the ray
that is reflected toward the sensor. By this way, the contributions
and attenuations along the tracing path are summed to calculate
the radiance that light sources scatter into the sensor. For an
arbitrary point pin the scene, the reflected outgoing radiance
along direction ωocan be expressed as [32],
Lo(p, ωo)=2π
fr(p, ωi,ω
o)Li(p, ωi)|cos θi|dωi(4)
where Lo(p, ωo)is the outgoing radiance along direction ωoat
point p,fr(p, ωi,ω
o)is BRDF, Li(p, ωi)is the incident radiance
from direction ωi, it includes both direct illumination as well
as indirect illumination (i.e., multiple scattering). Usually, ωiis
considered as a differential solid angle. θiis the angle between
the incident radiance direction ωiand surface normal at point
p. The integration is applied to the upper hemisphere space at
point p. If transmittance exists, fr(p, ωi,ω
o)is simply replaced
by a bidirectional transmittance distribution function (BTDF)
ft(p, ωi,ω
o), then the incident direction ωiand ωowill be in the
different hemispheres. For convenience, BRDF and BTDF can
be combined together as f(p, ωi,ω
o), which is called bidirec-
tional scattering distribution function (BSDF). Then, the equa-
tion which considers both reflectance and transmittance will be
(5), the integration is now applied to the 4πspace at point p.
Lo(p, ωo)=4π
f(p, ωi,ω
o)Li(p, ωi)|cosθi|dωi.(5)
This equation is solved by a path tracing algorithm [32],
which is illustrated in Fig. 6. Assuming a ray is traced from the
sensor and the first intersection point is p. At point p, the direct
irradiance from the sun is evaluated, then an outgoing radiance
can be calculated with the BSDF of point p, this is called first
scattering. To calculate outgoing radiance induced by multiple
scattering, a new ray originated at pis traced, the direction is
determined by the BSDF of point p. If this new ray intersects
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
QI et al.: LARGE-SCALE EMULATION SYSTEM FOR REALISTIC THREE-DIMENSIONAL (3-D) FOREST SIMULATION 5
with the scene (e.g., leaves and branches) at point p1, we execute
the same operation as at point p(evaluate sun radiation and
multiple scattering), the outgoing radiation of p1is the incident
radiation of p. This process is recursively performed until it
reaches the maximum scattering order specified by the user or
when the contribution of multiple scattering is below a small
threshold. Through this way, a path between sun and sensor can
be found, which determines the amount of radiation scattered
from sun to sensor. For each pixel, many sensor rays are traced
to get an average value. It should be noted that although the
image is simulated pixel by pixel, the neighbor objects are still
considered, because the rays are allowed to traverse through the
whole scene. To implement this simulation system, a ray tracer,
named Mitsuba [33], was modified to calculate the TOC spectral
radiances.
Compared to forward tracking methods (tracing rays from sun
to sensor), this backward tracking method has a few advantages
when simulating at-sensor images: 1) it only needs to calculate
the radiation which exactly goes to the sensor, avoiding a lot of
useless computations for radiation which goes out of the scene;
2) The number of rays of each pixel is configurable, which
means that a low value (e.g., rays =2) can make the simulation
very fast with a relatively low quality. However, this mechanism
makes it possible to have a quick preview of the result to check
the correctness of parameter configurations, especially when
simulating large areas. In addition, the proposed system can
run on high-performance computing clusters with hundreds of
processing cores. This scalability also enables LESS to simulate
much larger scenes (e.g., >5km).
2) Model Intercomparison: Before applying this system to
realistic forest simulation, we first tested it with some exist-
ing models from RAMI [34], which proposes a mechanism
to benchmark various radiative transfer models by using some
well-controlled experimental conditions. A homogeneous ab-
stract canopy with two layers (HOM28 in RAMI-IV) and a
heterogeneous actual canopy (HET07_JPS_SUM in RAMI-IV)
were chosen as the testing scenes. The detailed optical and struc-
tural properties can be found at [35].
The leaf shape is defined as a disk in HOM28, we converted
it into a rectangle with the same LA, and each leaf was then
represented by two triangles only. For HET07_JPS_SUM, the
needle shape of J¨
arvselja Scots Pine is described by a scaled
sphere. In our simulation, this scaled sphere is represented as
a hexagonal prism, which keeps the same total area and needle
length with RAMI definitions. LESS treats leaves and breaches
as facet, which do not have thickness, so it is impossible to
input transmittance of needle directly, because needles are usu-
ally like cylinders. To maximumly imitate the behavior of needle
transmittance, we set outside reflectance of facets (ρout) equal to
ρneedle, which is given by RAMI (see Fig. 7). For transmittance,
both sides of the facet are set to √τneedle, and the inside re-
flectance (ρin) is 0, so there will be no multiple scattering within
the needle, then the total transmittance (τtotal) is exactly τneedle ,
which is the transmittance of a needle.
To compare with RAMI, BRF in principal plane from the
scene of HOM28 was simulated in RED and near infrared (NIR)
bands, and the scene of HET07_JPS_SUM was simulated with
Fig. 7. Structural and optical properties of needle and shoot.
Fig. 8. Comparison between measured LAI and predicted LAI.
19 bands. The observation zenith angles range from 1◦to 75◦
withastepof2
◦. The BRF was finally compared with results
from other radiative transfer models.
3) Realistic Forest Simulation: A large plot (LPLOT) with
1km×1 km (red rectangle in Fig. 1) in our study area was
reconstructed. It contains 152 848 larch trees and 9433 birch
trees. The terrain of LPLOT has a minimum elevation of 857
m, and a maximum elevation of 958 m. Each component (i.e.,
leaf, truck, and soil background) was assigned with different
optical property that was obtained in the field measurements,
and all facets in this scene were assumed to be Lambertian.
Based on this virtual scene, a Landsat-like reflectance image
was simulated in RED (from 625 to 690 nm), and NIR (from
830 to 901 nm) bands with 1 nm spectral resolution and 30
m spatial resolution. The sun direction was calculated by the
overpass time of Landsat 8 and geographical location of the
study area. The final image was simulated under nadir direction
with an orthographic camera. The spectral response function
of Landsat 8 was then applied to the multispectral images to
get broadband images in RED and NIR. Finally, the normal-
ized difference vegetation index (NDVI) was calculated. The
land surface reflectance of Landsat 8 was obtained by the Fast
Line-of-sight Atmospheric Analysis of Hypercubes (FLAASH)
atmospheric correction model in ENVI 5.1. Since our study area
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
6 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
Fig. 9. Comparison with RAMI for an abstract two layers case: (a) NIR BRF. (b) RED BRF. (The figures of other models are originally from RAMI website.)
Fig. 10. Comparison with RAMI for a heterogeneous case, two bands were chosen from the simulation results: (a) band9 with a needle reflectance around 0.056.
(b) band17 with a needle reflectance around 0.52. (The figures of other models are originally from RAMI website.)
was covered by clouds during the whole summer of 2013. The
date of Landsat 8 data used in our study is May 24, 2014.
IV. RESULTS
A. Allometric Relationship
Allometric relationships between DBH and LA for larch and
white birch were established by using field inventory data of L1,
L4, L5, and L9. We then used these relationships to predict the
LAI of other plots. We found that L3 and L8 had similar LAI,
DBH distribution, and tree height distribution, but the number
of individual trees in L3 (82) was much lower than that in
L8 (181). Therefore, we inferred that the field measurement of
L3 was incorrect, and excluded it from the following analysis.
Fig. 8 shows the comparison between predicted LAI and field-
measured LAI for L2, L6, L7, and L8. The root-mean-square
error (RSME) between the predicted LAI and measured LAI is
0.42.
B. Model Intercomparison
The ray-tracing-based radiative transfer simulation was vali-
dated with other models on the RAMI website. Fig. 9 presents
the results of HOM28. It shows a relatively accurate result for
both NIR and RED BRF, since it is inconsistent with most of
the existing models. Fig. 10 shows the principal BRF of the
heterogeneous actual canopy. It can be observed that the dif-
ferences between different models are relatively large. Among
them, our model is at the middle position. BRF near hotspot of
our model is in good agreement with several other 3-D mod-
els, such as librat. However, the inconsistence also can be seen
when the zenith angles are larger than 20◦. Our BRF has a
more obvious bowl shape than other models. Although similar
ray-tracing theory has been used in other simulation models
(i.e., pbrt), the representations of 3-D scene are different. RAMI
provides a standard description of the a 3-D scene, but some
models still have simplifications due to the complex structure of
canopy, especially for the representations of needles and shoots.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
QI et al.: LARGE-SCALE EMULATION SYSTEM FOR REALISTIC THREE-DIMENSIONAL (3-D) FOREST SIMULATION 7
Fig. 11. Comparison between CCD image and generated visible image: (a) CCD image; (b) generated image.
Fig. 12. Comparison between NDVIL8and NDVIS: (a) NDVIL8; (b) NDVIS; (c) Horizontal profile 1 (P1_L8 and P1_LESS); (d) Horizontal profile 2 (P2_L8
and P2_LESS).
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
8 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
Fig. 13. Difference between simulated NDVI and Landsat 8 NDVI: (a) NDVI
difference distribution. (b) Histogram of NDVI difference.
These simplifications can result in inconsistencies between dif-
ferent models.
C. Image Simulation
Fig. 11 shows the computer-generated image of LPLOT.
Compared with the CCD image, the main distribution of in-
dividual trees in this plot is consistent with the real world, as
seen in the CCD image, especially for the isolated trees which
are easier to be segmented from CHM map. In the subplot (red
rectangle), the shadow of trees is also comparable with the CCD
image. In this virtual scene, there is a small valley around the
green rectangle in Fig. 11. The shadow produced by this valley
is well preserved. However, the classification of tree species is
not very accurate, because no multispectral data were available,
and the quality of CCD image is relatively low.
D. Spectral Simulation
Landsat-like land surface reflectance images in NIR and
RED were simulated by LESS, and NDVI from simulation
data (NDVIS) were calculated and compared with the observed
Landsat 8 NDVI (NDVIL8) (see Fig. 12). Overall, the distribu-
tion pattern of NDVISis similar to NDVIL8, both of them show
a relatively high value in areas with dense vegetation. Two hori-
zontal profiles across different places of the image were chosen
and illustrated in Fig. 12(c) and (d). It can be seen that the trends
are quite consistent. Fig. 13 illustrates the distribution of the dif-
ference between NDVISand NDVIL8. The mean value is 0.17
with a standard deviation of 0.08. The distribution of the differ-
ence showed that NDVISwas slightly larger than NDVIL8.A
possible reason is that most of the predicted LAIs based on the
allometric relationship in this study area are also overestimated
(see Fig. 8), which would result in a larger NDVIS.Themain
differences locate at areas with relatively sparse trees coverage
(the bottom of the image) and area of the valley, while in areas
with bare earth and dense vegetation, the differences are smaller,
because these areas are more homogeneous.
V. D ISCUSSION
A. 3-D Scene Generation
In this study, we used a well-established plant software pack-
age (OnyxTREE) to generate a series of individual trees, for
which the parameters were obtained by field measurements.
Fig. 14. A 30 m ×30 m forest scene: (a) RGB image, (b) NIR image. This
scene contains 30 birch trees, which are made from three single tree objects
with random translations and rotations.
However, the highest uncertainty is the LA of individual tree,
this uncertainty is produced either by the effective LAI mea-
surements or allometric relationships used in this study. In the
future, terrestrial laser scanner (TLS) may be a good alternative
to estimate LA of both individual tree and plot [36]. More im-
portantly, TLS can obtain very accurate 3-D branch structures
of trees. In this way, we can generate a virtual tree which is
almost identical to real trees.
B. Spectral Simulation
The absolute values of NDVISand NDVIL8show an apparent
offset. This may be caused by a number of reasons. Since we
only used a limited number of individual tree models and the
regenerated 3-D forest was not really identical to the real one,
and this structural difference may produce the gap between
NDVISand NDVIL8. Another reason was that the selection of
individual trees from SDB was only based on the tree height
and crown diameter, but LA, which is a very important factor
that influences NDVI, was not considered currently. In fact, the
predicted LAIs in most plots, excluding L8, were overestimated,
which means that the produced individual tree model may have
higher LA than a real tree and in turn produced a higher NDVI.
Additionally, the Landsat data were obtained in May, but the
scene in August was reconstructed, the LAI in May was lower
than the LAI in August, which caused that NDVISwas larger
than NDVIL8. However, the simulation is still useful because
we have predetermined all the “ground truth” (such as LAI)
of the regenerated scene, which is important to support some
quantitative remote sensing studies, such as the validation of
land surface parameter retrieval algorithms.
C. Ray-Tracing Sampling
Since ray-tracing mainly relies on random sampling of image
plane; thus, the sampling count (SC) per square meter may have
a large influence on the simulation accuracy, especially in highly
heterogeneous areas. To study the influences of SC, we simu-
lated the nadir BRF in NIR band of a 30 m ×30 m scene (see
Fig. 14) under different SC values, which range from 20m−2to
2Nm−2(N=0, 1,..., 10). For each SC, the simulations were
carried out 50 times. The mean, minimum, maximum value,
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
QI et al.: LARGE-SCALE EMULATION SYSTEM FOR REALISTIC THREE-DIMENSIONAL (3-D) FOREST SIMULATION 9
Fig. 15. Influence of sampling count (SC): (a) Simulation accuracy under different SC, BFRmean, BRFmin, BRFmax represents the mean, minimum, and maximum
value of the 50 simulations for each SC, 95%CI is the confidence interval that indicates each simulated BRF has a possibility of 0.95 to lie in this interval.
(b) Simulation time of first-scattering and multi-scattering under different SC (excluding the time for reading and writing files).
and 95%confidence interval (CI) are shown in Fig. 15, which
also shows the corresponding simulation time (under a laptop
with 8 cores and 16 GB memory). It can be observed that BRF
tends to be stable while SC increases, however, the difference
is only 0.1% between SC =64 and SC =1024. This means
that the accuracy of SC =64 is comparable with higher SC, but
the simulation time is significantly reduced (from 23.1 to 1.5 s).
Although BRFmin and BRFmax have a significant difference un-
der very low SC, the 95%CI is relatively small, i.e., there is
a possibility of 95%that the simulated BRF value will lie in a
small interval, which becomes smaller when SC increases.
This figure also indicates that the simulation of multiscatter-
ing takes up most of the computing resources, because vegeta-
tion usually has very complex foliage structures and relatively
high reflectance in NIR band. If the reflectance is low (e.g.,
RED band), we can set the tracing paths shorter to reduce the
scattering times, which will in turn reduce multiscattering and
increase computing efficiency.
VI. CONCLUSION
In this study, we proposed a large-scale emulation system
to generate 3-D forest scene with explicitly described individ-
ual trees and simulate reflectance images. Allometric equations
were used to populate plot LAI to individual tree LA. With these
parameters, we constructed a SDB which contained a series of
individual trees. This SDB can be reused to build a larger forest
scene and can be used in the remote sensing and forest manage-
ment research. We also developed a parallelized radiative trans-
fer model based on ray-tracing techniques, and it provided a
relatively efficient way to simulate images of large and complex
scene. This model has been validated with other 3-D radiative
transfer models from RAMI, and a high consistency with most
of them proved the model’s feasibility. Moreover, we have re-
leased LESS to the public and it has an easy-to-use graphic user
interface (GUI) [37]. This software can simulate multispectral,
multiangle, and multiresolution images. We hope this software
can help remote sensing researchers to better understand the
complex interactions between vegetation and radiation, and also
possibly ease their modeling and validation work.
ACKNOWLEDGEMENT
The authors would like to thank K. Yan and Y. Chen for
providing us with field measurement data. Special thanks also to
Dr. T. Yin and Prof. J.-P. Gastellu-Etchegorry, for their valuable
advice on improving this paper.
REFERENCES
[1] J. K. Winjum, R. K. Dixon, and P. E. Schroeder, “Estimating the global
potential of forest and agroforest management practices to sequester car-
bon,” in Natural Sinks CO2, J. Wisniewski and A. E. Lugo, Eds. The
Netherlands: Springer, 1992, pp. 213–227.
[2] M. J. Twery and A. R. Weiskittel, “Forest-m modelling,” in Environmental
Modelling: Finding Simplicity Complexity, 2nd ed. Hoboken, NJ, USA:
Wiley, 2013, pp. 379–398.
[3] D. Alder, “Growth modelling for mixed tropical forests,” Tropi-
cal Forestry Papers (United Kingdom), 1995. [Online]. Available:
http://agris.fao.org/agris-search/search.do?recordID=GB9618283
[4] M. Battaglia, P.J. Sands, and S. G. Candy, “Hybrid growth model to predict
height and volume growth in young Eucalyptus globulus plantations,”
Forest Ecol. Manag., vol. 120, no. 1, pp. 193–201, 1999.
[5] J. K. Vanclay, Modelling Forest Growth and Yield: Applications to Mixed
Tropical Forests. Wallingford, U.K.: CAB Int., 1994.
[6] R. O. Curtis, G. W. Clendenen, D. J. DeMars, and P. N. Forest, ANew
Stand Simulator for Coast Douglas-Fir: DFSIM User’s Guide. U.S. Dept.
Agriculture, Forest Service, Pacific Northwest Forest and Range Experi-
ment Station, Portland, 1981.
[7] P. Paar, “Landscape visualizations: Applications and requirements of 3D
visualization software for environmental planning,” Comput., Environ-
ment Urban Syst., vol. 30, no. 6, pp. 815–839, Nov. 2006. [Online]. Avail-
able: http://linkinghub.elsevier.com/retrieve/pii/S0198971505000566
[8] S. R. Sheppard, “Guidance for crystal ball gazers: Developing a code of
ethics for landscape visualization,” Landscape Urban Plann., vol. 54, no.
1, pp. 183–199, 2001.
[9] G. Zheng and L. M. Moskal, “Retrieving leaf area index (LAI)
using remote sensing: Theories, methods and sensors,” Sensors,
vol. 9, no. 4, pp. 2719–2745, Apr. 2009. [Online]. Available:
http://www.mdpi.com/1424-8220/9/4/2719/
[10] M. B´
eland, J.-L. Widlowski, R. A. Fournier, J.-F. Cˆ
ot´
e, and M. M.
Verstraete, “Estimating leaf area distribution in savanna trees from terres-
trial lidar measurements,” Agricultural Forest Meteorol., vol. 151, no. 9,
pp. 1252–1266, 2011.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
10 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
[11] F. D. Schneider et al., “Simulating imaging spectrometer data: 3D for-
est modeling based on LiDAR and in situ data,” Remote Sens. Environ.,
vol. 152, pp. 235–250, 2014.
[12] J.-L. Widlowski, J.-F. Cˆ
ot´
e, and M. B´
eland, “Abstract tree crowns in 3d
radiative transfer models: Impact on simulated open-canopy reflectances,”
Remote Sens. Environ., vol. 142, pp. 155–175, 2014.
[13] R. Darvishzadeh, A. Skidmore, M. Schlerf, and C. Atzberger, “Inversion of
a radiative transfer model for estimating vegetation LAI and chlorophyll
in a heterogeneous grassland,” Remote Sens. Environ., vol. 112, no. 5,
pp. 2592–2604, 2008.
[14] M. I. Disney, P. Lewis, and P. North, “Monte Carlo ray tracing in opti-
cal canopy reflectance modelling,” Remote Sens. Rev., vol. 18, no. 2–4,
pp. 163–196, 2000.
[15] J.-P. Gastellu-Etchegorry, V. Demarez, V. Pinel, and F. Zagolski, “Model-
ing radiative transfer in heterogeneous 3D vegetation canopies,” in Satel-
lite Remote Sensing (ser. Satellite Remote Sensing). International Society
for Optics and Photonics, 1995, pp. 38–49.
[16] J.-P. Gastellu-Etchegorry et al., “Discrete anisotropic radiative transfer
(dart 5) for modeling airborne and satellite spectroradiometer and lidar
acquisitions of natural and urban landscapes,” Remote Sens., vol. 7, no. 2,
pp. 1667–1701, 2015.
[17] P. R. J. North, “Three-dimensional forest light interaction model using a
Monte Carlo method,” IEEE Trans. Geosci. Remote Sens., vol. 34, no. 4,
pp. 946–956, Jul. 1996.
[18] W. Qin and S. A. Gerstl, “3-D scene modeling of semidesert vegetation
cover and its radiation regime,” Remote Sens. Environ., vol. 74, no. 1,
pp. 145–162, 2000.
[19] M. E. Schaepman et al., “Advanced radiometry measurements and earth
science applications with the airborne prism experiment (apex),” Remote
Sens. Environ., vol. 158, pp. 207–219, 2015.
[20] H. Huang, W. Qin, and Q. Liu, “RAPID: A radiosity ap-
plicable to porous indiviDual objects for directional re-
flectance over complex vegetated scenes,” Remote Sens. En-
viron., vol. 132, pp. 221–237, May 2013. [Online]. Available:
http://linkinghub.elsevier.com/retrieve/pii/S0034425713000254
[21] A. Komiyama, J. E. Ong, and S. Poungparn, “Allometry, biomass, and pro-
ductivity of mangrove forests: A review,” Aquatic Botany, vol. 89, no. 2,
pp. 128–137, Aug. 2008. [Online]. Available: http://www.sciencedirect.
com/science/article/pii/S0304377007001829
[22] D. P. Turner, S. A. Acker, J. E. Means, and S. L. Garman, “As-
sessing alternative allometric algorithms for estimating leaf area of
Douglas-fir trees and stands,” Forest Ecol. Manage., vol. 126, no. 1,
pp. 61–76, Feb. 2000. [Online]. Available: http://www.sciencedirect.
com/science/article/pii/S0378112799000833
[23] J.-F. Barczi et al., “Amapsim: A structural whole-plant simulator based
on botanical knowledge and designed to host external functional models,”
Ann. Botany, vol. 101, no. 8, pp. 1125–1138, 2008.
[24] S. Griffon and F. de Coligny, “AMAPstudio: An editing and sim-
ulation software suite for plants architecture modelling,” Ecologi-
cal Modelling, vol. 290, pp. 3–10, Oct. 2014. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S030438001300536X
[25] J. Weber and J. Penn, “Creation and rendering of realistic trees,” in Proc.
22nd Annu. Comput. Graph. Interactive Techn., 1995, pp. 119–128. [On-
line]. Available: http://doi.acm.org/10.1145/218380.218427
[26] K. Lim, P. Treitz, M. Wulder, B. St-Onge, and M. Flood, “LiDAR remote
sensing of forest structure,” Progress Phys. Geography, vol. 27, no. 1,
pp. 88–106, 2003.
[27] M. Van Leeuwen and M. Nieuwenhuis, “Retrieval of forest structural
parameters using LiDAR remote sensing,” Eur. J. Forest Res., vol. 129,
no. 4, pp. 749–770, 2010.
[28] T. P. Dawson, P. J. Curran, and S. E. Plummer, “LIBERTY-Modeling the
effects of leaf biochemical concentration on reflectance spectra,” Remote
Sens. Environ., vol. 65, no. 1, pp. 50–60, 1998.
[29] L. Wang et al., “Reflectance features of water stressed Larix gmelinii
needles,” Forestry Studies China, vol. 11, no. 1, pp. 28–33, 2009.
[30] J. Qi, D. Xie, and G. Yan, “Realistic 3D-simulation of large-scale
forest scene based on individual tree detection,” in Proc. IEEE Int.
Geosci. Remote Sens. Symp., Jul. 2016, pp. 728–731. [Online]. Available:
http://ieeexplore.ieee.org/document/7729183/
[31] W. Zhang et al., “An easy-to-use airborne lidar data filtering method
based on cloth simulation,” Remote Sens., vol. 8, no. 6, 2016,Art. no. 501.
[Online]. Available: http://www.mdpi.com/2072-4292/8/6/501
[32] J. T. Kajiya, “The rendering equation,” in ACM SIGGRAPH Comput.
Graph., vol. 20, no. 4, pp. 143–150, 1986.
[33] W. Jakob, “Mitsuba physically based renderer. mitsuba-renderer.org,”
2010.
[34] J. L. Widlowski et al., “Third radiation transfer model intercomparison
(RAMI) exercise: Documenting progress in canopy reflectance models,”
J. Geophys. Res., Atmospheres (1984-2012), vol. 112, no. D9, 2007.
[35] “RAMI, RAMI-IV - EXPERIMENTS - European Commission.”
2017. [Online]. Available: http://rami-benchmark.jrc.ec.europa.
eu/HTML/RAMI-IV/RAMI-IV.php. Accessed on: Sep. 27, 2017.
[36] W. Yan, D. Xie, X. Wang, Q. I. Jianbo, W. Zhang, and G. Yan, “Simulation
and analysis of point clouds from a terrestrial laser scanner,” J. Remote
Sens., vol. 19, no. s1, pp. 391–399, 2015.
[37] J. Qi, “Home—LESS Documentation,” Mar. 2017. [Online]. Available:
http://ramm.bnu.edu.cn/projects/less/
Jianbo Qi is currently working toward the Ph.D. de-
gree in cartography and geographic information sys-
tem at Beijing Normal University, Beijing, China.
He is also a joint Ph.D. student at CESBIO Labora-
tory,Paul Sabatier University, France.
His research interests include 3-D radiative trans-
fer modeling, realistic forest scene simulation, and
vegetation parameter retrieval.
Donghui Xie received the Ph.D. degree in remote
sensing and geographic information systems from
Beijing Normal University, Beijing, China, in 2005.
From 2005 to 2007, she was a Postdoctoral Re-
search Associate with Beijing Normal University. She
is currently with the State Key Laboratory of Remote
Sensing Science, School of Geography, Beijing Nor-
mal University. Her research interests include canopy
radiative transfer modeling and biophysical parame-
ter retrieval of vegetation.
Dashuai Guo is currently working toward the Mas-
ter’s degree in cartography and geographic informa-
tion system at Beijing Normal University, Beijing,
China.
His research interests are mainly in individual trees
modeling and realistic forest scene simulation.
Guangjian Yan received the Ph.D. degree in carto-
graphy and geographic information system from the
Institute of Remote Sensing Applications, Chinese
Academy of Sciences, Beijing, China, in 1999.
He is currently a Professor with the State Key Lab-
oratory of Remote Sensing Science, School of Geog-
raphy, Beijing Normal University, Beijing. He has
published more than 160 papers. His main research
interests are multiangular remote sensing, vegetation
remote sensing, radiation budget, scale effect, and
scale correction of remote sensing.
