ArticlePDF Available

Evaluation of the Vegetation-Index-Based Dimidiate Pixel Model for Fractional Vegetation Cover Estimation

Authors:

Abstract and Figures

Remote sensing estimation based on the dimidiate pixel model (DPM) using vegetation indices (VIs) is a common approach for mapping fractional vegetation cover (FVC). The major drawback of DPM is that it does not consider real endmember conditions and multiple scattering between soil and vegetation. An analysis of FVC uncertainties caused by these model deficiencies is still lacking. Here, we first calculated the FVC theoretical uncertainty caused by reflectance uncertainties based on the law of prapagation of uncertainty (LPU). Then, we tested the performance of DPM using six VIs over 3-D forest scenes. We simulated both Aqua-MODIS and Landsat-OLI surface reflectance (SR) at their corresponding spatial resolutions and spectral response functions (SRFs) using a well-validated 3-D radiative transfer (RT) model which helps to separate the model and input uncertainties. We found that ratio vegetation index (RVI)- and enhanced vegetation index (EVI)-based models were most affected by sensors, followed by the normalized difference vegetation index (NDVI)-, enhanced vegetation index 2 (EVI2)-, renormalized difference vegetation index (RDVI)-, and difference vegetation index (DVI)-based models. Without considering SR uncertainties, the DVI-based model performed best (FVC absolute difference < 0.1); however, the commonly used NDVI model reached a maximum difference of 0.35. At the same time, input uncertainty increased the uncertainty of FVC retrieval. We noticed that the increase of solar zenith angle (SZA) resulted in a clear increase of retrieved FVC under the uniform distribution, which can be explained by the increased shadow proportion. Besides, model accuracy was dominated by the purity of soil (vegetation) endmember in low (high) vegetation cover area. This study provides a reference for the selection of the optimal VI for FVC retrieval based on the DPM.
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 TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 1
Evaluation of the Vegetation-Index-Based Dimidiate
Pixel Model for Fractional Vegetation
Cover Estimation
Kai Yan , Si Gao, Haojing Chi, Jianbo Qi, Wanjuan Song , Yiyi Tong ,XihanMu ,
and Guangjian Yan ,Senior Member, IEEE
Abstract Remote sensing estimation based on the dimidiate
pixel model (DPM) using vegetation indices (VIs) is a common
approach for mapping fractional vegetation cover (FVC). The
major drawback of DPM is that it does not consider real
endmember conditions and multiple scattering between soil and
vegetation. An analysis of FVC uncertainties caused by these
model deficiencies is still lacking. Here, we first calculated the
FVC theoretical uncertainty caused by reflectance uncertainties
based on the law of prapagation of uncertainty (LPU). Then,
we tested the performance of DPM using six VIs over 3-D
forest scenes. We simulated both Aqua-MODIS and Landsat-OLI
surface reflectance (SR) at their corresponding spatial resolutions
and spectral response functions (SRFs) using a well-validated
3-D radiative transfer (RT) model which helps to separate the
model and input uncertainties. We found that ratio vegetation
index (RVI)- and enhanced vegetation index (EVI)-based models
were most affected by sensors, followed by the normalized
difference vegetation index (NDVI)-, enhanced vegetation index 2
(EVI2)-, renormalized difference vegetation index (RDVI)-, and
difference vegetation index (DVI)-based models. Without con-
sidering SR uncertainties, the DVI-based model performed best
(FVC absolute difference <0.1); however, the commonly used
NDVI model reached a maximum difference of 0.35. At the
same time, input uncertainty increased the uncertainty of FVC
retrieval. We noticed that the increase of solar zenith angle (SZA)
resulted in a clear increase of retrieved FVC under the uniform
distribution, which can be explained by the increased shadow
proportion. Besides, model accuracy was dominated by the purity
of soil (vegetation) endmember in low (high) vegetation cover
area. This study provides a reference for the selection of the
optimal VI for FVC retrieval based on the DPM.
Manuscript received July 28, 2020; revised December 13, 2020; accepted
December 21, 2020. This work was supported in part by the National Natural
Science Foundation of China under Grant 41901298 and Grant 41901273;
in part by the Open Fund of the State Key Laboratory of Remote Sensing
Science under Grant OFSLRSS201924; in part by the Open Research Fund
of the Key Laboratory of Digital Earth Science, Institute of Remote Sensing
and Digital Earth, Chinese Academy of Sciences, under Grant 2018LDE002;
and in part by the Fundamental Research Funds for the Central Universities
under Grant 2652018031. (Corresponding authors: Kai Yan; Haojing Chi.)
Kai Yan, Si Gao, and Haojing Chi are with the School of Land Science
and Techniques, China University of Geosciences, Beijing 100083, China
(e-mail: kaiyan@cugb.edu.cn; gaosi_gs@163.com; chihaojing99@163.com).
Jianbo Qi is with the College of Forestry, Beijing Forestry University,
Beijing 100083, China (e-mail: jianboqi@126.com).
Wanjuan Song is with the Aerospace Information Research Institute, China
Academy of Sciences, Beijing 100083, China (e-mail: songwj@aircas.ac.cn).
Yiyi Tong, Xihan Mu, and Guangjian Yan are with the State Key Labo-
ratory of Remote Sensing Science, Faculty of Geographical Science, Beijing
Normal University, Beijing 100101, China (e-mail: tongyiyi0311@163.com;
muxihan@bnu.edu.cn; gjyan@bnu.edu.cn).
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TGRS.2020.3048493.
Digital Object Identifier 10.1109/TGRS.2020.3048493
Index Terms—3-D radiative transfer (RT), dimidiate pixel
model (DPM), fractional vegetation cover (FVC), model eval-
uation, vegetation indices (VIs).
I. INTRODUCTION
FRACTIONAL vegetation cover (FVC), an important vari-
able for describing the quality and changes of vege-
tation in terrestrial ecosystems, refers to the proportion of
the vertically projected area of vegetation (including leaves,
stems, and branches) within the statistical area [1]–[3]. It is
commonly used in vegetation change analysis [4], ecological
environment research [5], [6], climate prediction [7], soil
erosion assessment [8], [9], among others. Therefore, accurate
estimates of FVC are of great importance.
There are two methods for measuring FVC: ground mea-
surement and remote sensing estimation. Ground measurement
is a flexible mean of extracting FVC at the point scale, inde-
pendent of atmosphere condition and model assumptions, and
with high accuracy and spatial resolution [10]. However, this
method is time-consuming, costly, and may not allow extrapo-
lating FVC to a large area quickly [11]. Moreover, vegetation
cover has obvious spatiotemporal variability, making it difficult
to monitor FVC dynamically at large spatial scales using
ground measurements. Remote sensing estimation of FVC
based on the correlation between FVC and canopy reflectance
can provide long time series and large ground coverage for
global and regional ecological monitoring [12], [13].
Remote sensing-based FVC estimation methods mainly
include: 1) empirical models; 2) physical-based models; and
3) pixel unmixing models [14]–[16].
1) Empirical models (e.g., regression models [17]) are
generally descriptive with a relatively simple form to char-
acterize the statistical relationships between a certain band,
band combination, or vegetation indices (VIs) and FVC [15].
The empirical method has great application potential due to
its high computational efficiency for large remote sensing
data sets [18]. However, this approach lacks sufficient under-
standing of physical mechanisms and highly relies on ground
sampling [19].
2) Physical-based models (e.g., radiative transfer (RT)
model [20] and geometric optical (GO) model [21]) offer
an explicit connection between canopy reflectance and
vegetation biophysical and biochemical parameters [22].
Thus, this method has a solid theoretical basis and is accurately
0196-2892 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Geosciences. Downloaded on February 08,2021 at 15:17:42 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
2IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
applicable to FVC estimation at large scales [18]. Nonetheless,
owing to the extreme complexity and “ill-posed problem”
of physical-based models, it is difficult to directly retrieve
FVC [23], [24].
3) Pixel unmixing models with a certain degree of physical
basis are simple and convenient, which can be regarded as
an improvement on the basis of physical models [25], [26].
Based on the assumption that each pixel is composed of several
components, pixel unmixing models estimate FVC at the
subpixel level as the proportion of vegetation component [27].
Among the existing pixel unmixing models, the dimidiate
pixel model (DPM) is the simplest one which assumes that
the pixel has only two components (i.e., pure vegetation and
nonvegetation) [28]. This method minimizes the influence
of the atmosphere, soil background, and other factors to
a large extent, and has been widely and successfully used
in vegetation cover extraction at both regional and global
scales [25], [27]–[29].
In addition, with the constant technical developments
in computer science, spatial data mining and knowledge
discovery (SDMKD) and machine learning have shown
their advantages in recognitions of remote sensing signals
(e.g., decision trees [30] and artificial neural networks [31]).
These approaches can accurately extract the FVC value by
training with precomputed remote sensing data sets and have
become popular for studies of FVC estimation [32], while its
practical applications are still insufficient [33].
VIs, as indirect indicators of vegetation growth characteris-
tics, make use of multiband information and are sensitive to
soil types, humidity, and temporal and spatial variations of the
atmosphere [34]. The VI-based DPM is widely used for FVC
estimation given its good accuracy and simplicity [3]. Previous
studies have shown that different VIs perform differently on
FVC calculations. The normalized difference vegetation index
(NDVI), the most widely used VI, has been used in DPM for
a long time [28], [34]. However, some studies have shown
that NDVI-based DPM may overestimate FVC, especially for
very dense canopies or open canopies with bright or dark
bare ground [18], [22]. NDVI-based models appear also to
be sensitive to soil background and scale effects [35], [36].
To solve this problem, Jiang et al. [29] proposed to use
the difference vegetation index (DVI) instead of NDVI in
DPM. Johnson et al. [37] found that modified soil adjusted
vegetation index (MSAVI) is less sensitive to changes in soil
brightness than NDVI. Li et al. [38] combined NDVI and ratio
vegetation index (RVI) to minimize saturation biases in an
NDVI-based model [38]. The main limitation of DPM is the
influence of background spectra variability derived from the
model assumption [22], [33]. Thus, the selection of the optimal
VI for DPM (e.g., most sensitive to vegetation coverage and
least sensitive to background factors) is the key issue in the
application of this method.
Ground-based validation is essential as the benchmark to
evaluate DPM accuracy [9], [18], [33], [38]. However, the
accuracy of this validation approach includes the uncertainty
of the ground measurement, the spatial heterogeneity-caused
uncertainty in the upscaling process from the point to the
pixel scale, and the model uncertainty caused by the DPM
with theoretical deficiency [39]. Therefore, this obviously
hampered the evaluation of the DPM uncertainty and progress
in the understanding of its deficiency. In view of the draw-
backs of ground-based validation, recent developments in
computer simulations have provided an alternative approach
and have been used for validating the retrieval of physical
and structural characteristics of vegetation, such as leaf area
index (LAI) [40], [41], fraction of photosynthetically active
radiation (fPAR) [42], surface albedo [43], solar downward
radiation [44], and so on. RT-based scene simulation can
accurately simulate multispectral and multiangle images and
radiation properties of realistic landscapes with spatial het-
erogeneity and can reduce the uncertainties of scale trans-
formations. To distinguish the input uncertainty and model
uncertainty, we employed the well-validated 3-D RT model
LargE-Scale remote sensing data and image Simulation frame-
work (LESS) [45] to simulate real forest scenes. Further-
more, this article investigated the effects of different factors
[e.g., solar zenith angle (SZA), solar azimuth angle (SAA)] on
the DPM accuracy by controlling the parameters of the simula-
tion experiment and quantitatively evaluated the performance
of different VI-based DPMs. Specifically, we: 1) calculated
the theoretical uncertainty of FVC retrieval caused by input
uncertainties based on the law of propagation of uncertainty
(LPU); 2) evaluated six VI-based DPMs (i.e., NDVI-, RVI-,
DVI-, EVI-, EVI2-, and RDVI-based model) over simulated
3-D forest scenes influenced by vegetation coverage, solar
position, vegetation clumping type, and endmember purity;
and 3) discussed the comparison of FVC uncertainty between
LPU calculation and LESS simulation experiment.
II. DATA AND METHODS
A. DPM and VI
The DPM assumes that the surface of a pixel is either vege-
tated or nonvegetated, and the spectral response of the pixel is
also linearly synthesized by these two components [28]. FVC,
as the weight, exhibits a linear relationship with the spectral
response of a mixed pixel [25]. Thus, VI as the most practical
spectral response can be represented by
VI =VIsoil ×(1FVC)+VIveg ×FVC (1)
where VI is the vegetation index of a specific mixed pixel, and
VIveg and VIsoil are the VI values of pure vegetation and bare
soil pixels, respectively. There are many ways to derive VIveg
and VIsoil from remote sensing images [46], and they can be
considered as known parameters.
Based on (1) and considering the physical constraint, we can
easily express FVC as
FVC =VI VIsoil
VIveg VIsoil
0FVC 1.
(2)
During the last decades, several VIs have been proposed.
In this article, we selected the six commonly used VIs [NDVI,
RVI, DVI, EVI, EVI2, and RDVI (see Table I)] to test the
performance of the DPM.
Authorized licensed use limited to: China University of Geosciences. Downloaded on February 08,2021 at 15:17:42 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
YA N et al.: EVALUATION OF THE VI-BASED DPM FOR FVC ESTIMATION 3
TAB LE I
SIX VISUSED IN THIS STUDY
B. Propagation of Input Uncertainty
The uncertainty of the DPM is not only related to the
model assumption but also related to the input uncertainty.
We utilized the LPU to study the propagation process of the
reflectance uncertainties into the FVC retrieval.
In the LPU, the measurement uncertainty is obtained by
taking into consideration the relative uncertainty component
of each input quantity in the model equation relating the
input quantities to the output. If x1,x2,x3,···xnare the input
quantities and the output Zis a function of the input quantities:
Z=f(x1,x2,x3,···
,xn).
The uncertainty of Z(designated by uz)represents the
estimated standard deviation of the dependent variable Z.
The general expression for uncertainty propagation is as fol-
lows [47], [48]:
u2
z=
N
i=1f
xi2
+2
N1
i=1
N
j=i+1
f
xi
f
xj
uxi,xj(3)
where uzis the uncertainty (standard deviation) of the depen-
dent variable Z;uxiis the uncertainty associated with the input
quantity xi;uxi,xj is the estimated covariance associated with
xiand xj; the partial derivatives f
xiare equal to that evaluated
at X=xi.
If there is no correlation between the input quantities (we
discuss this hypothesis in the Discussion section), (3) can be
simplified as
u2
z=f
x1
ux12
+f
x2
ux22
+··· +f
xN
uxN2
.(4)
In our study, we took different band reflectance as the input
quantities and regarded FVC as the output quantity. VI is
a quantitative value calculated from surface reflectance (SR)
(Table I). Here, we take NDVI as an example (other VIs are
given in the Appendix) for the derivation of the FVC formula
NDVI =NIR R
NIR +R.(5)
Fig. 1. Uncertainties of Landsat-OLI SR and Aqua-MODIS SR derived from
publications.
Then, the NDVI equation is introduced directly into (2) and
we get
FVC =1
NDVIveg NDVIsoil
×NIR R
NIR +RNDVIsoil.(6)
NDVIsoil and NDVIveg are constants. So, the uncertainty of
FVC (uFVC)depends on the relative uncertainties of the
red and NIR band reflectance. To simplify the approach,
we regarded the multi-band reflectance as independent vari-
ables (detailed in the Discussion section). According to (4),
uFVC is given by
u2
FVC =FVC
NIR uNIR2
+FVC
RuR2
(7)
where
FVC
NIR =1
NDVIveg NDVIsoil
×2R
(NIR +R)2
FVC
R=1
NDVIveg NDVIsoil
×2NIR
(NIR +R)2
.(8)
This article analyzed the FVC uncertainty propagation from
Operational Linear Instrument (OLI) and Moderate Resolution
Imaging Spectroradiometer (MODIS) reflectance based on
LPU and different VI-based models. We digitized the data
from published figures and replotted them in Fig. 1. The OLI
SR uncertainty was derived from [49], and the MODIS SR
uncertainty was from the science team website (https://modis-
land.gsfc.nasa.gov/ValStatus.php?ProductID=MOD09). The
SR uncertainties were obtained by comparing the SR products
with a reference obtained by the Second Simulation of the
Satellite Signal in the Solar Spectrum (6S) model with the
observations from the Aerosol Robotic Network (AERONET)
sites [49].
C. LESS RT Model
The LESS model is a newly proposed ray-tracing-based
3-D RT Model, which can accurately and efficiently simu-
late multispectral and multiangle images and radiation prop-
erties of complex realistic landscapes [45]. The accuracy
of the LESS model is validated in comparison with other
Authorized licensed use limited to: China University of Geosciences. Downloaded on February 08,2021 at 15:17:42 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
4IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
Fig. 2. Spectra of three types of soil (dark, medium and bright) and one type of vegetation (European birch) used as inputs for the LESS simulation. The
filled and unfilled curves represent the SRFs of the MODIS and Landsat-OLI sensors, respectively; the blue, orange, and crimson colors represent the blue,
red, and NIR spectral bands, respectively.
models (e.g., FLIGHT [50], RAYTYAN [51]) over several
different homogeneous and heterogeneous canopies from the
radiation transfer model intercomparison (RAMI) experiment
(http://rami-benchmark.jrc.ec.europa.eu/HTML), and also val-
idated with a publish field measurement data sets [52], [53].
The basic inputs of the LESS model are the 3-D structure
(e.g., tree height), component spectrum (e.g., leaf reflectance),
sun-sensor geometry, and illumination parameters (e.g., sky
light proportion) of the scene. LESS can simulate the trans-
mission process (absorption, reflection, and transmission) of
incident light in a scene based on ray-tracing strategy, and
output the corresponding simulated variables (such as direc-
tional reflectance, albedo, fPAR, and so on). Due to the high
computation efficiency, solid theoretical foundation, as well
as well-assessed accuracy, the output of LESS model can be
used as benchmarks for various applications in remote sensing,
forestry, photogrammetry, and so on [42], [44], [54]–[56].
Further detailed information on the LESS model can be found
on the website (http://lessrt.org/).
D. 3-D Forest Scene Simulation
The LESS model was run to simulate actual satellite obser-
vations of canopy reflectance based on the spectral response
functions (SRFs) of the MODIS and OLI sensors, and used
to calculate multiband bidirectional reflectance factors (BRFs)
of the simulated scenes. The simulated scenes are covered
by European birch and its component (i.e., leaf) spectra were
downloaded from the LOPEX93 data sets on the OPTICLEAF
website (http://opticleaf.ipgp.fr/). The soil reflectance curves
were selected from the spectral library (soils.sli file) of the
ENVI software (https://www.ittvis.com/). We sorted 25 soil
spectral curves from low to high, taking the mean values of the
first 10%, the middle 10%, and the last 10% as dark, medium,
and bright soils (Fig. 2).
We calculated the broadband reflectance and transmittance
using SRFs of MODIS and OLI sensors with the following
equation:
R=
λmax
λ=λmin
SλRλλmax
λ=λmin
Sλ
T=
λmax
λ=λmin
SλTλλmax
λ=λmin
Tλ
(9)
where Rand Tare the broadband reflectance and transmit-
tance, respectively, λmin and λmax are the wavelength lower
boundary and upper boundary detected by the sensor, and Sλ
is the SRF value of the sensor at the wavelength λ.Rλand
Tλare the narrowband reflectance and transmittance derived
from the spectral curves.
E. Experimental Design
This study focused on the influence of five factors on
FVC retrieval: FVC magnitude (Exp. I), solar position
(Exp. II), vegetation clumping type (Exp. III), endmember
purity (Exp. IV), and input uncertainty (Exp. V). To illustrate
the roles of these factors, we designed five experiments
following the principle of a single variable.
The purpose of Exp. I was to analyze the influence of
vegetation cover on FVC retrieval. The range of FVC truth
designed in LESS was 0.1–0.9 with an increment of 0.2
(Fig. 3). Exp. II was used to analyze the influence of the solar
position, including two key variables, SZA and SAA. Exp. III
served to analyze the influence of vegetation distribution.
We simulated three scenarios (Fig. 4). Exp. IV studied the
influence of endmember purity on DPM accuracy. We assumed
that the scene with FVC truth of 0, 0.1, 0.2, and 0.3 was
soil endmembers with the purity of 100%, 90%, 80%, and
Authorized licensed use limited to: China University of Geosciences. Downloaded on February 08,2021 at 15:17:42 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
YA N et al.: EVALUATION OF THE VI-BASED DPM FOR FVC ESTIMATION 5
TAB LE I I
EXPERIMENTAL PARAMETER DESIGN
Fig. 3. 3-D forest scenes with five different FVC magnitudes.
(a) FVC =0.1. (b) FVC =0.3. (c) FVC =0.5. (d) FVC =0.7.
(e) FVC =0.9. They are all in uniform distribution.
70%, respectively, and the scene with FVC truth of 1, 0.9,
0.8, and 0.7 was vegetation endmembers with the purity of
100%, 90%, 80%, and 70%, respectively. Exp. V was based
on Exp. I, adding OLI and MODIS reflectance uncertainties,
and studying the effect of reflectance uncertainty on the DPM
accuracy. All experiments consider multiple scattering with
simulated reflectance of 400–900 nm, and model of individual
trees was from the 3-D Object of LESS. The scene sizes of
OLI and MODIS were 30 30 m and 500 500 m, respec-
tively. The experimental parameters are shown in Table II, and
the research process is shown in Fig. 5.
III. RESULTS
A. Uncertainty of FVC Based on LPU
To analyze the FVC uncertainty caused by the input
reflectance, we calculated the FVC theoretical uncertainty of
Fig. 4. 3-D forest scenes with three different vegetation distribution
types. (a) Uniform distribution. (b) “Half-half” distribution. (c) Row structure
distribution.
six VI-based models by propagating the SR uncertainty. The
plots of the OLI and MODIS FVC theoretical uncertainty
versus band reflectance were shown in Figs. 6 and 7, respec-
tively. Overall, we found that the FVC theoretical uncertainty
of MODIS was higher than that of OLI. For RVI- and
EVI-based models, the FVC uncertainties of OLI and MODIS
were almost consistent. Besides, we found DVI-based model
performed best with FVC uncertainty ranging from 0 to 0.2 for
OLI [Fig. 6(b)], and the FVC uncertainty was least affected by
changes in reflectance. For NDVI-based model, the FVC theo-
retical uncertainty decreased faster as the red band reflectance
increased, which indicates that the main source of uncertainty
is in the red band. RDVI performed similar to NDVI, but the
overall uncertainty was lower and less affected by changes in
reflectance. For RVI- and EVI-based models, the denominator
of the partial derivative in (4) approached zero for a given
reflectance, leading to very large FVC uncertainty; for this
reason, we put a 0–1 limit on the FVC uncertainty of the
LPU. As shown in Fig. 6, the range of FVC uncertainties
of the RVI- and EVI-based models (colorbar ranging from
0 to 1) was larger than that of NDVI- (colorbar ranging from
Authorized licensed use limited to: China University of Geosciences. Downloaded on February 08,2021 at 15:17:42 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
6IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
Fig. 5. Flowchart of the designed experiments. SR, VI, LPU, and DPM represent SR VI, LPU, and DPM, respectively. LESS is a 3-D RT model for the
realistic scene simulation. Green indicates the study of SR uncertainty based on LPU, red indicates the study of LESS simulation experiments, and blue
indicates the study of LESS simulation experiments adding input uncertainty.
0 to 0.25), DVI-, RDVI-, and EVI2-based models (colorbar
ranging from 0 to 0.05). For RVI-based model, the FVC
uncertainty was much higher in the case of extremely low
red band reflectance and high NIR band reflectance than
in other cases. The EVI-based model determined by three
bands [Fig. 6(c)] was distinct from the other indices, showing
higher FVC uncertainty in low red reflectance and high blue
reflectance.
B. Uncertainty of FVC Based on LESS simulation
In Exp. I, we simulated scenes with FVC truth of 0.1–0.9
over three soil types and calculated the corresponding FVC
retrievals. The histograms in Fig. 8 show the absolute dif-
ference (AD =FVC retrieval - FVC truth) for different VIs
over different soil types. We observed that the FVC retrieval
variations of OLI and MODIS for most cases were highly
consistent. Overall, DVI-, EVI-, and EVI2-based models per-
formed better, both in AD magnitudes and its trend changed by
different FVC truth (the absolute values of ADs in most cases
were below 0.1). Fig. 8(b), (e), and (f) shows increasing trends
in FVC retrievals with soil brightness, indicating that the DVI-,
EVI-, and EVI2-based models were sensitive to soil brightness.
DVI-, EVI-, and EVI2-based models performed best over dark
soil background (AD 0). NDVI- [Fig. 8(a)] and RDVI-based
[Fig. 8(d)] models performed similarly, both overestimating
FVC and showing the opposite trend of other indices as the
soil brightness changed. The NDVI-based model was very
sensitive to soil brightness, especially in the case of low
vegetation cover (FVC =0.1: AD of dark soil =0.13; AD of
medium soil =0.8; AD of bright soil =0.6). Compared with
the NDVI-based model, the FVC retrievals of the RVI-based
model were smaller and its sensitivity to soil brightness did
not decrease with FVC truth [Fig. 8(c)]. The RVI-based model
generally underestimates under low vegetation cover (FVC
truth <=0.5) and overestimates under high vegetation cover
(FVC truth >=0.7). Besides, RVI-based model was not sen-
sitive to soil brightness under low and high vegetation cover,
while it was significantly sensitive under medium vegetation
cover (FVC truth =0.3 0.7).
Fig. 8 also compares the FVC retrievals considering the
input uncertainties and FVC truth. The result shows that
RVI-based model was most sensitive to reflectance uncertainty;
especially when the FVC truth =0.5, the standard deviation is
up to 0.23. The DVI-based model had the lowest sensitivity to
reflectance uncertainty, and the maximum standard deviation
only can reach 0.067. Besides, we noted that the EVI2-based
model was less sensitive to reflectance uncertainty than EVI
(standard deviation range of EVI: 0.042–0.090; standard
deviation range of EVI2: 0.036–0.071). When FVC truth =
0.9, the standard deviation of the NDVI-based model is small
(about 0.038), which means that the NDVI-based model
is not sensitive to input uncertainty under high vegetation
cover.
Authorized licensed use limited to: China University of Geosciences. Downloaded on February 08,2021 at 15:17:42 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
YA N et al.: EVALUATION OF THE VI-BASED DPM FOR FVC ESTIMATION 7
Fig. 6. FVC uncertainty propagated from the SR uncertainties of OLI based on (a) NDVI-based model, (b) RVI-based model, (c) EVI-based model, (d)
DVI-based model, (e) RDVI-based model, and (f) EVI2-based models. Colors correspond to FVC uncertainty magnitudes. Black lines and dots represent soil
lines. For EVI, the reflectance increment is 0.005 in red and blue bands and 0.015 in NIR band. For other VIs, the reflectance increment is 0.02 in red and
NIR bands.
Fig. 7. FVC uncertainty propagated from the SR uncertainties of MODIS based on (a) NDVI-based model, (b) RVI-based model, (c) EVI-based model, (d)
DVI-based model, (e) RDVI-based model, and (f) EVI2-based models. Colors correspond to FVC uncertainty magnitudes. Black lines and dots represent soil
lines. For EVI, the reflectance increment is 0.005 in red and blue bands and 0.015 in NIR band. For other VIs, the reflectance increment is 0.02 in red and
NIR bands.
C. Influence of the Solar Position
We studied the influence of the solar position in Exp. II.
In the case of uniform distribution, the shadow area increased
with the increase in SZA, and the change trends of the FVC
retrievals from different models were approximately the same
as the change of the shadow area (Fig. 9). To further study
the effect of sun incident angle on the accuracy of the model
under different vegetation distributions, we designed Exp. III
Authorized licensed use limited to: China University of Geosciences. Downloaded on February 08,2021 at 15:17:42 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
8IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
Fig. 8. Absolute difference (AD =FVC retrieval FVC truth) between FVC retrieval and FVC truth. The colorbars represent AD without considering
reflectance uncertainties, and the black dots and vertical lines represent the mean values and standard deviations of the result after adding the reflectance
uncertainties. The green series is for OLI, and the blue series is for MODIS. The same color series from dark to light indicates dark soil, medium soil, and
light soil types, respectively. Panels a, b, c, d, e and f are for NDVI-, DVI-, RVI-, RDVI-, EVI-, and EVI2-based models, respectively.
based on Exp. II and the result was shown in Fig. 10. In the
scene with uniform vegetation distribution, the AD of FVC
calculated by different VI-based models increased with SZA.
Under the same SZA, AD hardly changed with the SAA
[Fig. 10(a)]. In the case of the “half-half” clumped distribution,
the AD of FVC was symmetrical along the 90–270line, and
the FVC was the largest when the SAA was 90and 270,and
the SZA was 70[Fig. 10(b)]. In the case of the row structure
distribution, the AD was symmetrical along the 0–180line,
and the FVC retrieval was the largest when the SAA is 0and
180, and the SZA is 70[Fig. 10(c)]. The change trends of
the FVC retrievals in different vegetation distributions were
also approximately the same as the change of the shadow
area (Fig. 11).
Overall, the uncertainties of VI-based models were sorted
from low to high: DVI <EVI <EVI2 <RDVI <RVI <
NDVI. The NDVI-based model showed a large uncertainty,
but was the least affected by the solar position. Under the
“half-half distribution,” the maximum and minimum differ-
ence of absolute difference was about 0.3; the RVI-based
model was greatly affected by solar angle, and the max-
imum and minimum difference of absolute difference was
the largest (0.65) under “half-half distribution. For DVI-,
RDVI-, EVI-, and EVI2-based models, the ADs were relatively
small, but they were more affected by the solar position
than NDVI.
D. Influence of Endmember Purity
In Exp. IV, we studied the influence of endmember purity on
the model accuracy. The result shows that AD of FVC retrieval
was greatly affected by soil purity in low vegetation coverage
Fig. 9. (a) Variation of the four-component ratio as a function of SZA (FVC
is fixed at 0.5, the solar azimuth is 0). (b) Absolute difference (AD) between
FVC retrieval and FVC truth at different SZA.
areas (Fig. 12), while is greatly affected by vegetation purity
in high vegetation coverage areas, especially for the RVI-based
model. This may be explained by the proportions of vegetation
and soil in the scenes.
Authorized licensed use limited to: China University of Geosciences. Downloaded on February 08,2021 at 15:17:42 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
YA N et al.: EVALUATION OF THE VI-BASED DPM FOR FVC ESTIMATION 9
Fig. 10. Variation of absolute difference (AD =FVC retrieval - FVC truth) with the change of SZA for three vegetation distribution types. (a) Uniform
distribution. (b) “Half-half distribution. (c) Row structure distribution.
Fig. 11. Proportion of shadows in the scene changes with the sun position. The radial direction represents the SZA. The axial direction represents SAA.
(a) Uniform distribution. (b) “Half-half” distribution. (c) Row structure distribution.
We noted that for the NDVI-based model, in low and
medium vegetation cover areas, the AD was the smallest
when the soil purities were 82% and 70%, respectively.
In high vegetation cover areas, the AD was positive, and the
minimum value was 0.19, indicating that FVC retrieval was
overestimated. For RVI-based model, FVC retrieval under low
vegetation cover was underestimated. In medium vegetation
cover areas, the AD was the smallest when the vegetation
purity is 70%. Besides, we noted that the error trends of the
DVI-, EVI-, RDVI-, and EVI2-based models were similar.
When the purity of the endmembers was equal, the absolute
differences of these four VI-based models ordered from large
to small were: RDVI >EVI2 >EVI >DVI.
IV. DISCUSSION
A. Discussion About Results
In the present study, we investigated the effect of input
uncertainty on the DPM accuracy (Part A of the Results).
We found that DVI-, RDVI-, and EVI2-based models were
greatly influenced by input uncertainty when NIR SR =0.1.
This striking observation could be explained in Fig. 1: when
SR truth is around 0.1, the absolute uncertainty of the MODIS
NIR band is abruptly elevated. Beyond that, as elaborated
in Part B of Results, RVI-based model was most susceptible
to different vegetation cover among all six VI-based models.
The standard deviation of RVI-based model was significantly
higher under high and middle vegetation cover than under
low vegetation cover. Thus, this characteristic should be taken
note of when RVI-based model is applied for FVC estimation.
This finding is consistent with Guo and Zeng [57]. We also
found that EVI-based model was affected by the reflectance
uncertainty more than EVI2-based model, which is caused by
the fact that EVI2 replaces the blue band with the red band
whose atmosphere correction uncertainty is much smaller (see
Fig. 1) [49].
We also studied the effect of different soil types and
different FVC on the DPM accuracy in Part B of the Results.
The FVC retrieved by DVI-based model increased with soil
Authorized licensed use limited to: China University of Geosciences. Downloaded on February 08,2021 at 15:17:42 UTC from IEEE Xplore. Restrictions apply.
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 TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
Fig. 12. FVC absolute difference (AD =FVC retrieval FVC truth) as a function of endmember purity. The three layers represent FVC magnitudes for
each VI-based model. Panels a, b, c, d, e and f are for NDVI-, DVI-, RVI-, EVI-, EVI2-, and RDVI-based models, respectively.
brightness, and with larger increments under middle vegetation
cover than low and high vegetation cover. These findings
corroborate the ideas of Ding et al. [3], who found that DVI
increased nearly linearly with an increase in soil reflectance
under low vegetation cover and increases exponentially with
the great levels of vegetation cover due to the nonlinear mixing
of the NIR reflectances of vegetation and soil. Besides, we also
observed that NDVI- and RVI-based models were sensitive to
changes in soil brightness, which is confirmed by Huete and
Jackson [35]. In addition, the NDVI-based model saturated
readily under high vegetation cover can be seen obviously.
Chen et al. [58] have demonstrated that this saturation is
related to the spectral ranges of red and near-infrared channels.
After comprehensive consideration of LPU calculation and
LESS simulation, DVI-based model showed excellent stability
and accuracy than the other five VIs, which is in good
agreement with Xu and Shen [59].
B. Drawbacks of the DPM
The DPM assumes that the surface of a pixel is either
vegetated or nonvegetated, and the spectral response of the
pixel is also linearly synthesized by these two components.
However, this assumption does not hold in the real scenario,
and in turn affected the accuracy of FVC calculated by DPM.
For example, when the incident light is off-nadir, the pixel
is composed of four components, which are sunlit soil,
shadow soil, sunlit vegetation, and shadow vegetation [29].
In Section III, we have evaluated absolute difference of
different VI-based models and their changes with the increase
of SZA from 0to 70, and discovered that the AD of
the NDVI-based model was minimally changed with SZA.
We can induce from this phenomenon that NDVI is least
sensitive to shadow. The ground-based measurement result
of Zhang et al. [60] also evidenced it. Multiple endmember
spectral mixture model (MESMM) has been used to address
the case where the pixel contains multiple endmembers [61].
Another drawback in DPM is the absence of consideration
of multiple scattering between vegetation and soil [62]. In real-
world scenarios, most field plots are patchy, of complex spatial
structures, and the effects of multiple scattering can hardly be
ignored [63]. Hence, the actual interactions between endmem-
ber abundances and spectra were inherently nonlinear, which
leads to unavoidable biases of FVC retrieved by DPM. In the
Exp V, we fixed vertical irradiation of sunlight to exclude the
effect of a pixel containing multiple endmembers, and then
examined the influence of endmember purity on the model
accuracy. Our results suggest that the accuracy of DPM does
not always continually increase as the increase in endmember
purity owing to the existence of theoretical deficiency. Among
Authorized licensed use limited to: China University of Geosciences. Downloaded on February 08,2021 at 15:17:42 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
YA N et al.: EVALUATION OF THE VI-BASED DPM FOR FVC ESTIMATION 11
Fig. 13. Comparison of FVC uncertainty from LPU calculation and
simulation experiment. The number at the end of the legend represents the
correlation coefficient (R2). The size of the point represents the true FVC.
all six VI-based models, the NDVI-based model was most
affected by multiple scattering. At present, several nonlinear
mixture models have been put toward reducing the impacts of
multiple scattering [64]–[66].
C. Comparison of FVC Uncertainty From LPU and
Simulation Experiment
In the Methods and Results sections, we assumed that
there is no correlation between the input quantities (spec-
tral SR) for the LPU calculation. However, confirming
the study obtained by Jiang et al. [29], the correlation
between different bands certainly exists for most ground
targets. Failure to account for this correlation may result
in a certain deviation of FVC theoretical uncertainty cal-
culated by LPU. Meanwhile, we found that the EVI-based
model was affected by the reflectance uncertainty more than
the EVI2-based model. Moreover, during data processing,
we employed physical constraints on broadband reflectance
(0 <SR <1) and FVC retrieval (0 <=FVC <=1)
for the simulation experiments, but employed a constraint on
FVC uncertainty (0 <=FVC uncertainty <=1) in the LPU
only. Furthermore, it is notable that the LPU calculation only
contains input uncertainty, whereas the simulation experiments
contain both input uncertainty and model uncertainty of LESS.
Here, we discuss the consistency between the two
approaches for FVC uncertainty analysis used in this study.
The results of the comparison are presented in Fig. 13. It can
be seen that the FVC uncertainty from LPU calculation and
that from simulation experiments were very close to each other
for DVI- and EVI2-based models. For NDVI- and RVI-based
models, the results under low vegetation cover are close
between these two approaches; however, the result of LPU
calculation was significantly greater than that of simulation
experiments under high and middle vegetation cover. Espe-
cially for RVI, this difference under high vegetation cover was
even up to about 0.2. This is caused by the uncertainty of the
LESS model itself and the correlation between the reflectance
of different bands. At present, we have not distinguished
between these two possibilities in our study yet, but we are
preparing a study focused on the correlation between different
bands based on ground validation.
V. C ONCLUSION
This study investigated the performance of six VI-based
DPMs for estimating FVC using a 3-D RT model. We first
calculated the theoretical uncertainty caused by reflectance
uncertainties based on the LPU. Then, we tested the per-
formance of the different VI-based models over simulated
3-D forest scenes. The multiband BRFs of these scenes were
calculated by computer simulation which helps to separate the
model uncertainty and other factors. Both MODIS and OLI
SR at their corresponding spatial resolutions were simulated.
In the analysis of theoretical uncertainty, we found that RVI-
and EVI-based models were most affected by sensors (the-
oretical uncertainty can reach 1), followed by NDVI, DVI,
RDVI, and EVI2. In the scene simulation analysis, we found
that the DVI-based model performed best, the FVC absolute
difference was less than 0.1, while the NDVI-based model
showed poor accuracy (the absolute difference can reach a
maximum of about 0.35). At the same time, SR uncertainties
further increased the uncertainty of FVC estimation, and RVI
was more sensitive to them. From the simulation experiments,
MODIS and OLI showed approximately equal uncertainties;
however, from the LPU calculation, MODIS showed larger
uncertainty than OLI. At the same time, the incidence angle
of the sun increases the shadow area, leading to an increase in
the uncertainty of the DPM. When the SZA is 70, the max-
imum bias can reach 95.2% for NDVI-based model, and the
minimum was also as large as 73.2% for RVI-based model, but
it was least affected by the change of the SZA. We also found
that different vegetation distributions affect FVC retrieved
by the DPM. In the case of uniform vegetation distribution,
the absolute difference was greatly affected by the SZA, and
did not vary with the SAA. The absolute differences under the
“half-half” and row structure distributions were approximately
symmetric. From the study of endmember purity, we found
Authorized licensed use limited to: China University of Geosciences. Downloaded on February 08,2021 at 15:17:42 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
12 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
TABLE III
PARTIAL DERIVATIVES OF RVI, DVI, EVI, EVI2, AND RDVI
that the accuracy of DPM does not increase with the purity
of the endmember. This might be attributed to the DPM that
its theoretical assumption is unreasonable. Overall, our results
show that based on the propagation of input uncertainty and
computer simulation experiments, DVI performs best both in
accuracy and stability, and RVI performs the worst. These
analyses provide a reference for the selection of the optimal
VI for FVC retrieval based on DPM.
APPENDIX
See Table III.
ACKNOWLEDGMENT
The authors thank the help from Prof. Youtang Hong for his
fruitful discussion about the law of propagation of uncertainty.
REFERENCES
[1] T. Purevdorj, R. Tateishi, T. Ishiyama, and Y. Honda, “Relationships
between percent vegetation cover and vegetation indices,” Int. J. Remote
Sens., vol. 19, no. 18, pp. 3519–3535, Dec. 1998.
[2] J. R. Etherington, P. D. Moore, and S. B. Chapman, “Methods in plant
ecology,J. E col., vol. 75, no. 1, p. 283, Mar. 1987.
[3] Y. Ding, H. Zhang, K. Zhao, and X. Zheng, “Investigating the accuracy
of vegetation index-based models for estimating the fractional vegetation
cover and the effects of varying soil backgrounds using in situ measure-
ments and the PROSAIL model,” Int. J. Remote Sens., vol. 38, no. 14,
pp. 4206–4223, Jul. 2017.
[4] A. A. Gitelson, Y. J. Kaufman, R. Stark, and D. Rundquist, “Novel
algorithms for remote estimation of vegetation fraction,” Remote Sens.
Environ., vol. 80, no. 1, pp. 76–87, Apr. 2002.
[5] D. S. Schimel, “Terrestrial ecosystems and the carbon cycle,” Global
Change Biol., vol. 1, no. 1, pp. 77–91, Feb. 1995.
[6] P. J. Sellers et al., “A revised land surface parameterization (SiB2) for
atmospheric GCMS. Part I: Model formulation,” J. Climate, vol. 9, no. 4,
pp. 676–705, Apr. 1996.
[7] R. K. Kaufmann et al., “The effect of vegetation on surface temperature:
A statistical analysis of NDVI and climate data,” Geophys. Res. Lett.,
vol. 30, no. 22, pp. 3–6, Nov. 2003.
[8] J. R. Dymond, P. R. Stephens, P. F. Newsome, and R. H.
Wilde, “Percentage vegetation cover of a degrading rangeland from
SPOT,” Int. J. Remote Sens., vol. 13, no. 11, pp. 1999–2007,
Jul. 1992.
[9] G. Wang, S. Wente, G. Z. Gertner, and A. Anderson, “Improve-
ment in mapping vegetation cover factor for the universal soil
loss equation by geostatistical methods with Landsat thematic map-
per images,” Int. J. Remote Sens., vol. 23, no. 18, pp. 3649–3667,
Jan. 2002.
[10] A. W. Kuchler, D. Mueller-Dombois, and H. Ellenberg, “Aims and
methods of vegetation ecology,Geograph. Rev., vol. 66, no. 1, p. 114,
Jan. 1976.
[11] A. Wilson et al., “Evaluation of methods of assessing vegetation change
in the semi-arid rangelands of Southern Australia,” Rangeland J.,vol.9,
no. 1, p. 5, 1987.
[12] C. Leprieur, Y. H. Kerr, S. Mastorchio, and J. C. Meunier, “Monitor-
ing vegetation cover across semi-arid regions: Comparison of remote
observations from various scales,” Int. J. Remote Sens., vol. 21, no. 2,
pp. 281–300, Jan. 2000.
[13] Q. Zhou and M. Robson, “Automated rangeland vegetation cover and
density estimation using ground digital images and a spectral-contextual
classifier,” Int. J. Remote Sens., vol. 22, no. 17, pp. 3457–3470,
Jan. 2001.
Authorized licensed use limited to: China University of Geosciences. Downloaded on February 08,2021 at 15:17:42 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
YA N et al.: EVALUATION OF THE VI-BASED DPM FOR FVC ESTIMATION 13
[14] J. M. Fernández-Guisuraga, L. Calvo, and S. Suárez-Seoane, “Com-
parison of pixel unmixing models in the evaluation of post-fire forest
resilience based on temporal series of satellite imagery at moderate
and very high spatial resolution,” ISPRS J. Photogramm. Remote Sens.,
vol. 164, pp. 217–228, Jun. 2020.
[15] J. Xiao and A. Moody, “A comparison of methods for estimating
fractional green vegetation cover within a desert-to-upland transition
zone in central New Mexico, USA,” Remote Sens. Environ., vol. 98,
nos. 2–3, pp. 237–250, Oct. 2005.
[16] J. M. Bioucas-Dias et al., “Hyperspectral unmixing overview: Geomet-
rical, statistical, and sparse regression-based approaches,” IEEE J. Sel.
Topics Appl. Earth Observ. Remote Sens., vol. 5, no. 2, pp. 354–379,
Apr. 2012.
[17] K. Jia, Y. Li, S. Liang, X. Wei, and Y. Yao, “Combining estimation of
green vegetation fraction in an arid region from Landsat 7 ETM+data,”
Remote Sens., vol. 9, no. 11, p. 1121, Nov. 2017.
[18] K. Jia et al., “Fractional vegetation cover estimation algorithm for
Chinese GF-1 wide field view data,” Remote Sens. Environ., vol. 177,
pp. 184–191, May 2016.
[19] M. J. Hill, Q. Zhou, Q. Sun, C. B. Schaaf, and M. Palace, “Relationships
between vegetation indices, fractional cover retrievals and the structure
and composition of Brazilian Cerrado natural vegetation,” Int. J. Remote
Sens., vol. 38, no. 3, pp. 874–905, Feb. 2017.
[20] L. J. Zhang, H. Z. Ma, Z. B. Zhou, Z. L. Ren, X. B. Zhu, and L. Sun,
“Vegetation canopy coverage estimation using physical models,” Adv.
Mater. Res., vols. 726–731, pp. 4709–4713, Aug. 2013.
[21] M. Chopping, L. Su, A. Rango, J. Martonchik, D. Peters, and
A. Laliberte, “Remote sensing of woody shrub cover in desert grasslands
using MISR with a geometric-optical canopy reflectance model,” Remote
Sens. Environ., vol. 112, no. 1, pp. 19–34, Jan. 2008.
[22] Y. Ding, H. Zhang, Z. Li, X. Xin, X. Zheng, and K. Zhao, “Comparison
of fractional vegetation cover estimations using dimidiate pixel models
and look-up table inversions of the PROSAIL model from Landsat 8 OLI
data,” J. Appl. Remote Sens., vol. 10, no. 3, Sep. 2016, Art. no. 036022.
[23] S. Jacquemoud and F. Baret, “PROSPECT: A model of leaf optical
properties spectra,” Remote Sens. Environ., vol. 34, no. 2, pp. 75–91,
Nov. 1990.
[24] S. Jacquemoud et al., “PROSPECT+SAIL models: A review of use
for vegetation characterization,” Remote Sens. Environ., vol. 113,
pp. S56–S66, Sep. 2009.
[25] X. Zhang, C. Liao, J. Li, and Q. Sun, “Fractional vegetation cover
estimation in arid and semi-arid environments using HJ-1 satellite hyper-
spectral data,” Int. J. Appl. Earth Observ. Geoinf., vol. 21, pp. 506–512,
Apr. 2013.
[26] P. J. Sellers et al., “A global 1by 1NDVI data set for climate
studies. Part 2: The generation of global fields of terrestrial biophysical
parameters from the NDVI,” Int. J. Remote Sens., vol. 15, no. 17,
pp. 3519–3545, Nov. 1994.
[27] G. Jiapaer, X. Chen, and A. Bao, “A comparison of methods for
estimating fractional vegetation cover in arid regions,” Agricult. Forest
Meteorol., vol. 151, no. 12, pp. 1698–1710, Dec. 2011.
[28] G. Gutman and A. Ignatov, “The derivation of the green vegetation frac-
tion from NOAA/AVHRR data for use in numerical weather prediction
models,” Int. J. Remote Sens., vol. 19, no. 8, pp. 1533–1543, Jan. 1998.
[29] Z. Jiang et al., “Analysis of NDVI and scaled difference vegetation index
retrievals of vegetation fraction,” Remote Sens. Environ., vol. 101, no. 3,
pp. 366–378, Apr. 2006.
[30] S. J. Goetz, R. K. Wright, A. J. Smith, E. Zinecker, and E. Schaub,
“IKONOS imagery for resource management: Tree cover, impervious
surfaces, and riparian buffer analyses in the mid-atlantic region,” Remote
Sens. Environ., vol. 88, nos. 1–2, pp. 195–208, Nov. 2003.
[31] M. Weiss and F. Baret, “Evaluation of canopy biophysical variable
retrieval performances from the accumulation of large swath satellite
data,” Remote Sens. Environ., vol. 70, no. 3, pp. 293–306, Dec. 1999.
[32] L. Yang, K. Jia, S. Liang, J. Liu, and X. Wang, “Comparison of
four machine learning methods for generating the GLASS fractional
vegetation cover product from MODIS data, Remote Sens., vol. 8, no. 8,
p. 682, Aug. 2016.
[33] Y. Yang, T. Wu, S. Wang, and H. Li, “Fractional evergreen forest cover
mapping by MODIS time-series FEVC-CV methods at sub-pixel scales,
ISPRS J. Photogramm. Remote Sens., vol. 163, pp. 272–283, May 2020.
[34] J. Qi, “Spatial and temporal dynamics of vegetation in the san pedro river
basin area,” Agricult. Forest Meteorol., vol. 105, nos. 1–3, pp. 55–68,
Nov. 2000.
[35] A. R. Huete and R. D. Jackson, “Soil and atmosphere influences on
the spectra of partial canopies,” Remote Sens. Environ., vol. 25, no. 1,
pp. 89–105, Jun. 1988.
[36] J. Qi, A. Chehbouni, A. R. Huete, Y. H. Kerr, and S. Sorooshian,
“A modified soil adjusted vegetation index,” Remote Sens. Environ.,
vol. 48, no. 2, pp. 119–126, May 1994.
[37] B. Johnson, R. Tateishi, and T. Kobayashi, “Remote sensing of frac-
tional green vegetation cover using spatially-interpolated endmembers,”
Remote Sens., vol. 4, no. 9, pp. 2619–2634, Sep. 2012.
[38] F. Li, W. Chen, Y. Zeng, Q. Zhao, and B. Wu, “Improving estimates of
grassland fractional vegetation cover based on a pixel dichotomy model:
A case study in inner Mongolia, China,” Remote Sens., vol. 6, no. 6,
pp. 4705–4722, May 2014.
[39] K. Yan et al., “Evaluation of MODIS LAI/FPAR product collection 6.
Part 2: Validation and intercomparison,” Remote Sens., vol. 8, no. 6,
p. 460, May 2016.
[40] S. G. Leblanc and R. A. Fournier, “Hemispherical photography simula-
tions with an architectural model to assess retrieval of leaf area index,”
Agricult. Forest Meteorol., vol. 194, pp. 64–76, Aug. 2014.
[41] B. Brede et al., “Assessment of workflow feature selection on forest
LAI prediction with sentinel-2A MSI, Landsat 7 ETM+and Landsat 8
OLI,” Remote Sens., vol. 12, no. 6, p. 915, Mar. 2020.
[42] S. Chen et al., “Retrieval of the fraction of radiation absorbed by
photosynthetic components (FAPARgreen) for forest using a triple-
source leaf-wood-soil layer approach,” Remote Sens., vol. 11, no. 21,
p. 2471, Oct. 2019.
[43] S. Wu et al., “Characterization of remote sensing albedo over sloped sur-
faces based on DART simulations and in situ observations,” J. Geophys.
Res., Atmos., vol. 123, no. 16, pp. 8599–8622, Aug. 2018.
[44] G. Yan et al., “An operational method for validating the downward
shortwave radiation over rugged terrains,” IEEE Trans. Geosci. Remote
Sens., vol. 59, no. 1, pp. 714–731, Jan. 2021.
[45] J. Qi et al., “LESS: Large-scale remote sensing data and image simula-
tion framework over heterogeneous 3D scenes,” Remote Sens. Environ.,
vol. 221, pp. 695–706, Feb. 2019.
[46] Z.-Q. Hu et al., “Estimation of fractional vegetation cover based on
digital camera survey data and a remote sensing model,” J. China Univ.
Mining Technol., vol. 17, no. 1, pp. 116–120, Mar. 2007.
[47] I. Elizabeth et al., “Measurement uncertainty evaluation in vickers
hardness scale using law of propagation of uncertainty and Monte Carlo
simulation,” MAPAN, vol. 34, no. 3, pp. 317–323, Sep. 2019.
[48] B. N. Taylor and C. E. Kuyatt, “Guidelines for evaluating and expressing
the uncertainty of NIST measurement results,” Nat. Inst. Standards
Technol., Gaithersburg, MD, USA, NIST Tech. Note 1297, 1994.
[49] E. Vermote, C. Justice, M. Claverie, and B. Franch, “Preliminary analysis
of the performance of the Landsat 8/OLI land surface reflectance
product,” Remote Sens. Environ., vol. 185, pp. 46–56, Nov. 2016.
[50] 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.
[51] Y. M. Govaerts and M. M. Verstraete, “Raytran: A Monte Carlo ray-
tracing model to compute light scattering in three-dimensional hetero-
geneous media,” IEEE Trans. Geosci. Remote Sens., vol. 36, no. 2,
pp. 493–505, Mar. 1998.
[52] J. M. Chen, P. D. Blanken, T. A. Black, M. Guilbeault, and S. Chen,
“Radiation regime and canopy architecture in a boreal aspen forest,
Agricult. Forest Meteorol., vol. 86, nos. 1–2, pp. 107–125, Aug. 1997.
[53] S. G. Leblanc, P. Bicheron, J. M. Chen, M. Leroy, and J. Cihlar, “Investi-
gation of directional reflectance in boreal forests with an improved four-
scale model and airborne POLDER data,” IEEE Trans. Geosci. Remote
Sens., vol. 37, no. 3, pp. 1396–1414, May 1999.
[54] Z.-H. Jiao, H. Ren, X. Mu, J. Zhao, T. Wang, and J. Dong, “Eval-
uation of four sky view factor algorithms using digital surface and
elevation model data,” Earth Space Sci., vol. 6, no. 2, pp. 222–237,
Feb. 2019.
[55] L. Li et al., “A half-Gaussian fitting method for estimating fractional
vegetation cover of corn crops using unmanned aerial vehicle images,”
Agricult. Forest Meteorol., vol. 262, pp. 379–390, Nov. 2018.
[56] J. Pu et al., “Evaluation of the MODIS LAI/FPAR algorithm based on
3D-RTM simulations: A case study of grassland,” Remote Sens., vol. 12,
no. 20, p. 3391, Oct. 2020.
[57] Y. Guo and F. Zeng, “Atmospheric correction comparison of spot-5
image based on model FLAASH and model QUAC,ISPRS-Int. Arch.
Photogramm., Remote Sens. Spatial Inf. Sci., vol. XXXIX-B7, pp. 7–11,
Jul. 2012.
Authorized licensed use limited to: China University of Geosciences. Downloaded on February 08,2021 at 15:17:42 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
14 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
[58] P.-Y. Chen, G. Fedosejevs, M. Tiscareño-López, and J. G. Arnold,
“Assessment of MODIS-EVI, MODIS-NDVI and VEGETATION-NDVI
composite data using agricultural measurements: An example at corn
fields in Western Mexico,Environ. Monitor. Assessment, vol. 119,
nos. 1–3, pp. 69–82, Jul. 2006.
[59] X. Y. S. Xu and R. Shen, “A comparative study of different vegetation
indices for estimating vegetation coverage based on the dimidiate pixel
model,” Remote Sens. Land Resour., vol. 24, no. 4, pp. 95–100, 2012.
[60] L. Zhang, X. Sun, T. Wu, and H. Zhang, “An analysis of shadow effects
on spectral vegetation indexes using a ground-based imaging spectrom-
eter,IEEE Geosci. Remote Sens. Lett., vol. 12, no. 11, pp. 2188–2192 ,
Nov. 2015.
[61] D. A. Roberts, M. Gardner, R. Church, S. Ustin, G. Scheer, and
R. O. Green, “Mapping chaparral in the santa monica mountains using
multiple endmember spectral mixture models,” Remote Sens. Environ.,
vol. 65, no. 3, pp. 267–279, Sep. 1998.
[62] J. L. Silván-Cárdenas and L. Wang, “Retrieval of subpixel tamarix
canopy cover from Landsat data along the forgotten river using linear
and nonlinear spectral mixture models,Remote Sens. Environ., vol. 114,
no. 8, pp. 1777–1790, Aug. 2010.
[63] X. Chen, J. Chen, X. Jia, B. Somers, J. Wu, and P. Coppin, “A quantita-
tive analysis of virtual Endmembers’ increased impact on the collinearity
effect in spectral unmixing,” IEEE Trans. Geosci. Remote Sens., vol. 49,
no. 8, pp. 2945–2956, Aug. 2011.
[64] T. W. Ray and B. C. Murray, “Nonlinear spectral mixing in desert
vegetation,” Remote Sens. Environ., vol. 55, no. 1, pp. 59–64, Jan. 1996.
[65] A. Halimi, Y. Altmann, N. Dobigeon, and J.-Y. Tourneret, “Nonlinear
unmixing of hyperspectral images using a generalized bilinear model,”
IEEE Trans. Geosci. Remote Sens., vol. 49, no. 11, pp. 4153–4162,
Nov. 2011.
[66] J. M. P. Nascimento and J. M. Bioucas-Dias, “Nonlinear mixture
model for hyperspectral unmixing,” Proc. SPIE, vol. 7477, Sep. 2009,
Art. no. 74770I.
[67] R. A. Muller and A. Buffington, “Real-time correction of atmospheri-
cally degraded telescope images through image sharpening,J. Opt. Soc.
Amer., vol. 64, no. 9, p. 1200, Sep. 1974.
[68] R. L. Pearson, “Remote mapping of standing crop biomass for estimation
of the productivity of the shortgrass prairie,” in Proc. 8th Int. Symp.
Remote Sens. Enviroment, 1972, pp. 1357–1381.
[69] A. J. Richardson and C. L. Wiegand, “Distinguishing vegetation from
soil background information,” Photogramm. Eng. Remote Sens., vol. 43,
no. 12, pp. 1541–1552, Dec. 1977.
[70] A. Bannari, D. Morin, F. Bonn, and A. R. Huete, “A review of vegetation
indices,” Remote Sens. Rev., vol. 13, nos. 1–2, pp. 95–120, Aug. 1995.
[71] D. Haboudane, “Hyperspectral vegetation indices and novel algorithms
for predicting green LAI of crop canopies: Modeling and validation in
the context of precision agriculture,” Remote Sens. Environ., vol. 90,
no. 3, pp. 337–352, Apr. 2004.
Kai Yan received the B.S. degree in mapping and
surveying from the Beijing University of Civil Engi-
neering and Architecture, Beijing, China, in 2011,
and the Ph.D. degree in GIS/RS from Beijing Nor-
mal University, Beijing, in 2018.
He was a Visiting Scholar with the Department of
Earth and Environment, Boston University, Boston,
MA, USA, from 2014 to 2016. He is an Assistant
Professor with the School of Land Science and Tech-
niques, China University of Geosciences, Beijing.
He was involved in the generation and assessment
of official MODIS/VIIRS global leaf area index (LAI) and fraction of
photosynthetically active radiation absorbed by vegetation (fPAR) products.
His research interests include the bidirectional reflectance distribution func-
tion (BRDF) modeling and LAI&fPAR retrieval.
Si Gao was born in Henan, China. She is a Senior
Student with the China University of Geosciences,
Beijing, China. Her major is surveying and mapping
engineering and her interest is vegetation quantita-
tive remote sensing. She will pursue her master’s
degree in remote sensing under the direction of
Dr. Kai Yan from 2021.
Haojing Chi was born in Henan, China. She is a
Senior Student with the China University of Geo-
sciences, Beijing, China. Her major is in surveying
and mapping engineering and her interest is vege-
tation quantitative remote sensing. She will pursue
her master’s degree in remote sensing with the
Aerospace Information Research Institute, Chinese
Academy of Sciences, Beijing.
Jianbo Qi received the Ph.D. degree from Beijing
Normal University, Beijing, China, in 2019. He was
a joint-Ph.D. student with Paul Sabatier University,
Toulouse, France, from 2016 to 2018.
He is an Assistant Professor with Beijing Forestry
University, Beijing. His research interests include
3-D radiative transfer modeling, realistic forest scene
simulation, and vegetation parameter retrieval.
Wanjuan Song received the B.S. degree in geo-
graphic information science and the Ph.D. degree
in cartography and geography information system
from Beijing Normal University, Beijing, China,
in 2013 and 2019, respectively.
She was a Visiting Scholar with the Department of
Earth and Environment, Boston University, Boston,
MA, USA, from 2016 to 2017. She is with the
State Key Laboratory of Remote Sensing Science,
Aerospace Information Research Institute, Chinese
Academy of Sciences, Beijing. Her research interests
include the vegetation structure parameter estimation, vegetation remote
sensing, and the application of DSCOVR EPIC data.
Yiyi Tong received the B.S. degree in physi-
cal geography and M.S. degree in remote sensing
from Beijing Normal University, Beijing, China,
in 2017 and 2020, respectively.
Her research interests include the estimation of
radiation budget and validation of remote sensing
models in rugged terrains.
Xihan Mu received the B.S. degree in computer
science and technology from the College of Informa-
tion Science and Technology, Beijing Normal Uni-
versity, Beijing, China, in 1999, and the Ph.D. degree
in remote sensing from the School of Geography,
Beijing Normal University, in 2009.
He was a Visiting Student with the Laboratoire
des Sciences de l’Images, de l’Informatique et de la
Télédétection, Louis Pasteur University, Strasbourg,
France, in 2007, and a Visiting Scientist with the
Commonwealth Scientific and Industrial Research
Organization, Canberra, ACT, Australia, in 2016. He is with the State Key
Laboratory of Remote Sensing Science, Faculty of Geographical Science,
Beijing Normal University, Beijing. His research interest includes multian-
gular remote sensing, particularly in the retrieval/measurement of vegetation
structural parameters.
Guangjian Yan (Senior Member, IEEE) received
the Ph.D. degree from the Institute of Remote Sens-
ing Applications, Chinese Academy of Sciences,
Beijing, China, in 1999.
He is a Professor with the State Key Laboratory
of Remote Sensing Science, Faculty of Geographical
Science, Beijing Normal University, Beijing. He has
published more than 200 articles. His research inter-
ests include multiangular remote sensing, vegetation
remote sensing, radiation budget, and scale correc-
tion of remote sensing.
Authorized licensed use limited to: China University of Geosciences. Downloaded on February 08,2021 at 15:17:42 UTC from IEEE Xplore. Restrictions apply.
... Currently, this estimation predominantly relies on satellite data, which provides extensive spatial coverage across substantial historical periods. Various methods for long-term FVC estimation using satellite data have been developed, including the dimidiate pixel method (DPM), which is based on vegetation indices (VIs) [10,11], empirical models (EMs) [12], and machine learning models (MLMs) [13]. Among these, the DPM is the most prevalent, based on the assumption that each pixel comprises pure vegetation and soil components. ...
... Despite the advancements in remote sensing technology, current methods for long-term FVC estimation face significant spatial and temporal limitations. The coarse spatial resolution of satellite imagery compromises the accuracy of FVC predictions in heterogeneous landscapes [11], while the lack of historical UAV data restricts its long-term applicability [23]. Furthermore, the integration of high-resolution UAV data with satellite imagery for long-term FVC monitoring remains challenging due to discrepancies in spatial and temporal scales between the two sources [24]. ...
... NDVI was chosen as the principal parameter due to its broad applicability in FVC model development and its demonstrated effectiveness as a key FVC indicator [48]. Using NDVI to construct the linear FVC model offers a direct and efficient approach, streamlining implementation while maintaining alignment with [11,49]. Although the inclusion of additional vegetation indices may improve model precision, it could also introduce greater complexity and potential challenges [30], such as increased computational load and the risk of multicollinearity between indices. ...
Article
Full-text available
Accurate long-term estimation of fractional vegetation cover (FVC) is crucial for monitoring vegetation dynamics. Satellite-based methods, such as the dimidiate pixel method (DPM), struggle with spatial heterogeneity due to coarse resolution. Existing methods using unmanned aerial vehicles (UAVs) combined with satellite data (UCS) inadequately leverage the high spatial resolution of UAV imagery to address spatial heterogeneity and are seldom applied to long-term FVC monitoring. To overcome spatial challenges, an improved dimidiate pixel method (IDPM) is proposed here, utilizing 2021 Landsat imagery to generate FVCDPM via DPM and upscaled UAV imagery for FVCUAV as ground references. The IDPM uses the pruned exact linear time (PELT) method to segment the normalized difference vegetation index (NDVI) into intervals, within which DPM performance is evaluated for potential improvements. Specifically, if the difference (D) between FVCDPM and FVCUAV is non-zero, NDVI-derived texture features are incorporated into FVCDPM through multiple linear regression to enhance accuracy. To address temporal challenges and ensure consistency across years, the 2021 NDVI serves as a reference for inter-year NDVI calibration, employing least squares regression (LSR) and histogram matching (HM) to identify the most effective method for extending the IDPM to other years. Results demonstrate that (1) the IDPM, by developing distinct DPM improvement models for different NDVI intervals, considerably improves UAV and satellite data integration, with a 48.51% increase in R² and a 56.47% reduction in root mean square error (RMSE) compared to the DPM and UCS and (2) HM is found to be more suitable for mining areas, increasing R² by 25.00% and reducing RMSE by 54.05% compared to LSR. This method provides an efficient, rapid solution for mitigating spatial heterogeneity and advancing long-term FVC estimation.
... NDVI is the normalized ratio between the surface reflectance of Red (R) and Near Infra-red (NIR) bands. In other words, NDVI typically represents the chlorophyll content of the plant (18). NDVI measures the greenness and density of vegetation from spatial data (19). ...
... ii. Dimidiate pixel model assumes that any image is composed of two components viz., vegetation and non-vegetation. This model considers the biophysical parameter relating morphological structure and physiological characteristics of the crop, which can be derived spatially from NDVI (18,25). The FVC value of each pixel can be calculated from NDVI using the formula: ...
... Water stress of a crop is a resultant character from soil and atmospheric parameters along with crop water requirements (37). An increased plant surface area indicates more stomatal count and more conductance for water vapor from leaves as transpiration (18). Though the spatial application of crop coefficients to wide ecosystem provides biased estimation of consumptive water use by more than 50%, optimization theory states that the predicted crop coefficient value from vegetative indices or any other sources like leaf area index or NDVI will be more like the actual crop coefficient values. ...
Article
Full-text available
Estimation of evapotranspiration spatially is mandatory for water budgeting and water resource monitoring based on crop water demand. The amount of evapotranspiration can be estimated spatially from reference evapotranspiration and fractional vegetation cover over the region. The incorporation of specific characteristics like spatial variability over the land cover produced highly reliable results. The AquaCrop model gives reference evapotranspiration based on climatic parameters like temperature, rainfall, solar radiation, relative humidity and elevation. Normalized Difference Vegetation Index (NDVI) depicts the health of vegetation spatially ranging from -1 to +1. The dimidiate pixel model converts NDVI to Fractional Vegetation Cover (FVC) which was then used as a substitute for crop coefficient value. The products of these two parameters produce actual evapotranspiration spatially over the region with high spatial resolution. The average amount of actual evapotranspiration varies for each Land Use Land Cover (LULC) type over different agroclimatic zones as the climatic parameters and water usage patterns vary for each land cover type.
... In this study, the grassland vegetation coverage classification in the QTP region by Yan, K et al. [33,32] was used to divide the grassland coverage in this study into five levels ( Table 2). ...
... The pixel binary model serves as a straightforward and effective approach for estimating the FVC. Utilizing this model [32], the FVC for grasslands in the QTP is determined through an inversion model that relies on NDVI data. (The formula for these calculations can be found in the Supplementary Materials.) ...
... (The formula for these calculations can be found in the Supplementary Materials.) In this study, the grassland vegetation coverage classification in the QTP region by Yan, K et al. [32,33] was used to divide the grassland coverage in this study into five levels ( Table 2). ...
Article
Full-text available
Changes in grassland fractional vegetation coverage (FVC) are important indicators of global climate change. Due to the unique characteristics of the Tibetan Plateau ecosystem, variations in grassland coverage are crucial to its ecological stability. This study utilizes the Google Earth Engine (GEE) platform to retrieve long-term MODIS data and analyzes the spatiotemporal distribution of grassland FVC across the Qinghai–Tibet Plateau (QTP) over 24 years (2000–2023). The grassland growth index (GI) is used to evaluate the annual grassland growth at the pixel level. GI is an important indicator for measuring grassland growth status, which can effectively measure the changes in grassland growth in each year relative to the base year. FVC trends are monitored using Sen-Mann-Kendall slope estimation, the coefficient of variation, and the Hurst exponent. Geographic detectors and partial correlation analysis are then applied to explore the contribution rates of key driving factors to FVC. The results show: (1) From 2000 to 2023, FVC exhibited an overall upward trend, with an annual growth rate of 0.0881%. The distribution of FVC on the QTP follows a pattern of higher values in the east and lower values in the west; (2) Over the past 24 years, 54.05% of the total grassland area has shown a significant increase, 23.88% has remained stable, and only a small portion has shown a significant decrease. The overall trend is expected to continue with minimal variability, covering 82.36% of the total grassland area. The overall grassland GI suggests a balanced state of growth; (3) precipitation (Pre) and soil moisture (SM) are the main single factors affecting FVC changes in grasslands on the Tibetan Plateau (q = 0.59 and 0.46). In the interaction detection, in addition to the highest interaction between Pre and other factors, the interaction between SM and other factors also showed a significant impact on the changes in FVC of the QTP grassland; partial correlation analysis of hydrothermal factors and FVC of the QTP grassland. It shows that precipitation has a stronger correlation with QTP grassland FVC changes than temperature. This study has enhanced our understanding of grassland vegetation change and its driving factors on the QTP and quantitatively described the relationship between vegetation change and driving factors, which is of great significance for maintaining the sustainable development of grassland ecosystems.
... The estimation of FVC is primarily accomplished through two methods: field measurement and remote sensing estimation [7]. Due to its extensive coverage and strong capability for continuous observation, remote sensing has become a highly effective tool for regional and global FVC estimation [8,9]. The primary approaches in remote sensing estimation include empirical models, pixel unmixing models, physical model-based methods [9], and deep learning methods. ...
... Due to its extensive coverage and strong capability for continuous observation, remote sensing has become a highly effective tool for regional and global FVC estimation [8,9]. The primary approaches in remote sensing estimation include empirical models, pixel unmixing models, physical model-based methods [9], and deep learning methods. Empirical models are straightforward and computationally efficient, offering high FVC estimation accuracy in localized areas. ...
Article
Full-text available
Spectral information plays a crucial role in fractional vegetation cover (FVC) estimation, and selecting the appropriate spectral information is essential for improving the accuracy of FVC estimation. Traditionally, spectral feature selection is primarily guided by physical mechanisms or empirical statistical models. This has led to the use of multispectral and hyperspectral images, which often result in missing or redundant information, thereby decreasing the efficiency and accuracy of FVC estimation. This study proposes a novel dual-attention network to select the feature bands of Sentinel-2 multispectral images for the accurate FVC estimation of winter wheat. In the first step, the importance of hyperspectral band reflectances was determined using simulated data from the PROSAIL model, by combining the dual-attention mechanism with the convolutional neural network (DAM-CNN). In the second step, the importance of Sentinel-2 multispectral bands was converted from the hyperspectral band importance identified in the previous stage, and subsequently ranked accordingly. Based on the feature ranking results, multispectral simulated data translated from hyperspectral simulated data were used for CNN training, and multispectral feature selection was conducted based on FVC accuracy. Finally, the selected features were assessed based on their performance in FVC estimation using a CNN model with real data. The experimental results indicate that during the key growth period of winter wheat, the combination of red, green, and red-edge bands significantly influences the FVC estimation accuracy. Band 3 (Green), band 4 (Red), band 5 (Red-edge 1), and band 6 (Red-edge 2) of Sentinel-2 satellite images contribute most significantly to winter wheat FVC estimation, achieving an accuracy comparable to that obtained using all bands, while reducing the training time by 19.1%, as confirmed by field survey data.
... However, this digital monitoring system has shown inaccuracy, much like traditional methods, and low spatial resolution and is time-consuming; thus, it is not considered feasible for large-scale canopy vegetation monitoring. Moreover, it does not consider multiple scattering between vegetation and soil, therefore creating uncertainties if applied on large and heterogeneous areas under mixed grassland [24]. Physical modelling approaches based on RTM have potential to change the physical relations between spectral values and grassland canopy cover [25]. ...
Article
Full-text available
The potential of precision agriculture (PA) in forage and grassland management should be more extensively exploited to meet the increasing global food demand on a sustainable basis. Monitoring biomass yield and quality traits directly impacts the fertilization and irrigation practises and frequency of utilization (cuts) in grasslands. Therefore, the main goal of the review is to examine the techniques for using PA applications to monitor productivity and quality in forage and grasslands. To achieve this, the authors discuss several monitoring technologies for biomass and plant stand characteristics (including quality) that make it possible to adopt digital farming in forages and grassland management. The review provides an overview about mass flow and impact sensors, moisture sensors, remote sensing-based approaches, near-infrared (NIR) spectroscopy, and mapping field heterogeneity and promotes decision support systems (DSSs) in this field. At a small scale, advanced sensors such as optical, thermal, and radar sensors mountable on drones; LiDAR (Light Detection and Ranging); and hyperspectral imaging techniques can be used for assessing plant and soil characteristics. At a larger scale, we discuss coupling of remote sensing with weather data (synergistic grassland yield modelling), Sentinel-2 data with radiative transfer modelling (RTM), Sentinel-1 backscatter, and Catboost–machine learning methods for digital mapping in terms of precision harvesting and site-specific farming decisions. It is known that the delineation of sward heterogeneity is more difficult in mixed grasslands due to spectral similarity among species. Thanks to Diversity-Interactions models, jointly assessing various species interactions under mixed grasslands is allowed. Further, understanding such complex sward heterogeneity might be feasible by integrating spectral un-mixing techniques such as the super-pixel segmentation technique, multi-level fusion procedure, and combined NIR spectroscopy with neural network models. This review offers a digital option for enhancing yield monitoring systems and implementing PA applications in forages and grassland management. The authors recommend a future research direction for the inclusion of costs and economic returns of digital technologies for precision grasslands and fodder production.
... This study used a pixel dichotomy model that relied on the NDVI to quantify the vegetation covering in the Yongzhou manganese mining region. This approach is appropriate for monitoring the dynamic changes in vegetation on a regional perspective [58,59]. Based on the principles of the pixel dichotomy model, the vegetation coverage is obtained, and the calculation is shown in Equation (2). ...
Article
Full-text available
Manganese ore, as an important strategic metal resource for the country, was subject to unreasonable mining practices and outdated smelting technologies in early China, leading to severe ecological damage in mining areas. This study examines the trends in vegetation cover change in the historical manganese mining areas of Yongzhou under the influence of policy, providing technical references for mitigating the ecological impact of these legacy mining areas and offering a basis for adjusting mine restoration policies. This paper takes the manganese mining area in Yongzhou City, Hunan Province as a case study and selects multiple periods of Landsat satellite images from 2000 to 2023. By calculating the Normalized Difference Vegetation Index (NDVI) and the Fractional Vegetation Coverage (FVC), the spatiotemporal changes and driving factors of vegetation coverage in the Yongzhou manganese mining area from 2000 to 2023 were analyzed. The analysis results show that, in terms of time, from 2000 to 2012, the vegetation coverage in the manganese mining area decreased from 0.58 to 0.21, while from 2013 to 2023, it gradually recovered from 0.21 to 0.40. From a spatial perspective, in areas where artificial reclamation was conducted, the vegetation was mainly mildly and moderately degraded, while in areas where no artificial restoration was carried out, significant vegetation degradation was observed. Mining activities were the primary anthropogenic driving force behind the decrease in vegetation coverage, while effective ecological protection projects and proactive policy guidance were the main anthropogenic driving forces behind the increase in vegetation coverage in the mining area.
... Many studies have demonstrated the significant role of remote sensing technology in estimating FVC at regional or landscape scales [55][56][57]. However, current research typically focuses on community-level FVC estimation, with limited attention paid to precise species-level estimates [16,58,59]. Especially for coastal salt marsh wetlands, although species-level FVC estimation has gradually received attention in recent years, related research is still relatively scarce [60,7]. ...
Article
Full-text available
Coastal saltmarsh wetlands are vital “blue carbon” ecosystems. Fractional vegetation cover (FVC) is a key indicator revealing the spatial distribution and growth status of vegetation. Remote sensing has proven a vital tool for FVC estimation at regional or landscape scales. Establishing a species-level FVC estimation model usually requires sufficient field measurements as training/validation samples. However, field-based sample collection in wetlands is challenging because of the harsh environment. In this study, we proposed a Fractional Vegetation Cover Wasserstein Generative Adversarial Network (FVC-WGAN) model for FVC sample expansion. We chose the Yellow River Delta as the study area and utilized the time series Sentinel-2 imagery and random forest regression model for species-level FVC estimation with the assistance of FVC-WGAN-generated samples. To assess the efficacy of FVC-WGAN, we designed 13 experimental schemes using different combinations of real and generated samples. Our results show that the FVC-WGAN-generated samples had similar feature values to the real samples. Supplementing 500 real samples with generated samples can achieve good accuracy with an average RMSE < 0.1. As the number of real samples increased, the accuracies of FVC estimation improved. When the number of the generated samples was balanced with the real samples, the accuracy improved in terms of both R2, RMSE and the spatial consistency.
... According to the principle of the DPM, a pixel is composed of two parts. These are rocks and non-rocks or vegetation and non-vegetation [45]. The proportion of the pixel occupied by rocks represents the EBF of the pixel, whereas the proportion of pixels occupied by vegetation is indicative of the FVC of the pixel. ...
Article
Full-text available
Karst rocky desertification (KRD) is a significant issue that affects the ecological and economic sustainability of southwest China. Obtaining the accurate distribution of different levels of KRD can provide decision-making support for the effective management of KRD. The Sustainable Development Goals Science Satellite 1 (SDGSAT-1) is the world’s first scientific satellite serving the 2030 Agenda for Sustainable Development of the United Nations, and is dedicated to developing high-resolution, multi-scale, global public datasets to support policy and decision-making support systems for sustainable development. SDGSAT-1 multispectral data provide detailed ground information with a spatial resolution of 10 m and a rich spectral resolution. In this study, we combined the red-modified carbonate rock index (RCRI, an index that characterizes the degree of carbonate rock exposure) and the normalized difference red edge index (NDRE, an index that characterizes the degree of vegetation coverage) to propose a novel feature space method based on SDGSAT-1 multispectral data to classify the different levels of KRD in the Jinsha County of Guizhou Province, a representative region with significant KRD in southwest China. This method effectively identified different levels of KRD with an overall classification accuracy of 87%. This was 20% higher than that of the grading index method, indicating that SDGSAT-1 multispectral data have promising potential for KRD classification. In this study, we offer a new insight into the classification of KRD and a greater quantity of remote-sensing data to monitor KRD over a wider area and for a longer period of time, contributing to the economic development and environmental protection of KRD areas.
Article
Full-text available
High-precision digital soil mapping in complex terrain is challenging. This study proposed a new method using the partitioning around medoids (PAM) clustering algorithm to partition the study area into distinct habitat patches type. Utilizing multi-source data and three machine learning models, we estimated soil organic carbon (SOC) content in southwest China. Results showed higher SOC content (0-15 cm) in the southwestern mountains and the northwestern plateaus of Sichuan, while lower in the Sichuan Basin. The prediction uncertainty exhibited a similar pattern. Topographic and climatic variables played crucial roles in SOC estimation. Among the three machine learning models, RF and XGBoost demonstrated higher simulation accuracy than SVM (R2 increased by 2.86% to 82.35%). Using the RF feature selection (FS) method to select optimal factors as model input variables improved simulation accuracy compared to using all factors or selecting based on Pearson correlation analysis (PsCA) (R2 increased by 1.75% to 64.71%). The study found that a hybrid model based on different habitat patches achieved higher accuracy than the single model for the whole study area (for example, with RF FS method and modeling, R2 increased by 2.17% to 34.78%, and RMSE decreased by 2.19% to 28.80%). These findings enhanced the accuracy and refinement of existing mapping in southwest China (compared to SoilGrids 1km and SoilGrids 250m products, R2 increased by 20.45% and 317.73%, and RMSE decreased by 35.51% and 60.49%). Such improvements better characterized the spatial variability of SOC and provided important implications for future soil carbon stock account in complex terrain areas.
Article
Full-text available
Uncertainty assessment of the moderate resolution imaging spectroradiometer (MODIS) leaf area index (LAI) and the fraction of photosynthetically active radiation absorbed by vegetation (FPAR) retrieval algorithm can provide a scientific basis for the usage and improvement of this widely-used product. Previous evaluations generally depended on the intercomparison with other datasets as well as direct validation using ground measurements, which mix the uncertainties from the model, inputs, and assessment method. In this study, we adopted the evaluation method based on three-dimensional radiative transfer model (3D RTM) simulations, which helps to separate model uncertainty and other factors. We used the well-validated 3D RTM LESS (large-scale remote sensing data and image simulation framework) for a grassland scene simulation and calculated bidirectional reflectance factors (BRFs) as inputs for the LAI/FPAR retrieval. The dependency between LAI/FPAR truth and model estimation serves as the algorithm uncertainty indicator. This paper analyzed the LAI/FPAR uncertainty caused by inherent model uncertainty, input uncertainty (BRF and biome classification), clumping effect, and scale dependency. We found that the uncertainties of different algorithm paths vary greatly (−6.61% and +84.85% bias for main and backup algorithm, respectively) and the “hotspot” geometry results in greatest retrieval uncertainty. For the input uncertainty, the BRF of the near-infrared (NIR) band has greater impacts than that of the red band, and the biome misclassification also leads to nonnegligible LAI/FPAR bias. Moreover, the clumping effect leads to a significant LAI underestimation (−0.846 and −0.525 LAI difference for two clumping types), but the scale dependency (pixel size ranges from 100 m to 1000 m) has little impact on LAI/FPAR uncertainty. Overall, this study provides a new perspective on the evaluation of LAI/FPAR retrieval algorithms.
Article
Full-text available
Estimation of downward shortwave radiation (DSR) is of great importance in global energy budget and climatic modeling. Although various algorithms have been proposed, effective validation methods are absent for rugged terrains due to the lack of rigorous methodology and reliable field measurements. We propose a two-step validation method for rugged terrains based on computer simulations. The first step is to perform point-to-point validation at local scale. Time-series measurements were applied to evaluate a three-dimensional (3-D) radiative transfer model. The second step is to validate the DSR at pixel-scale. A semiempirical model was built up to interpolate and upscale the DSR. Key terrain parameters were weighted by empirical coefficients retrieved from ground-based observations. The optimum number and locations of ground stations were designed by the 3-D radiative transfer model and Monte Carlo method. Four ground stations were selected to upscale the ground-based observations. Additional three ground stations were set up to validate the interpolated results. The upscaled DSR was finally applied to validate the satellite products provided by MODIS and Himawari-8. The results showed that the modeled and observed DSR exhibited good consistency at point scale with correlation coefficients exceeding 0.995. The average error was around 20 W/m² for the interpolated DSR and 10 W/m² for the upscaled DSR in theory. The accuracies of the satellite products were acceptable at most times, with correlation coefficients exceeding 0.94. From an operational point of view, our method has an advantage of using small amount of ground stations to upscale DSR with relatively high accuracy over rugged terrains.
Article
Full-text available
In Mediterranean fire-prone ecosystems, shifts in fire regime as a consequence of global change could modify the resilience of vegetation communities. In this paper, we aim to compare the efficiency of high and moderate spatial resolution satellite imagery in the evaluation of resilience in a fire-prone landscape under different fire regime categories using two pixel unmixing techniques. A time series of Landsat (ETM+ and OLI; spatial resolution of 30 m) and WorldView-2 (spatial resolution of 2 m) imagery collected between 2011 (pre-fire conditions) and 2016 were used to estimate the temporal variation of fractional vegetation cover (FVC) as a quantitative measure of forest resilience. For this time series, FVC was computed under four fire-regime categories of recurrence and severity using two approaches: dimidiate pixel model and multiple endmember spectral mixture analysis (MESMA). The dimidiate pixel model was computed using NDVI as spectral response for the case of Landsat imagery and NDVI and red-edge NDVI (RENDVI) for WorldView-2. MESMA was applied to unmix WorldView-2 and Landsat imagery into four fraction images: photosynthetic vegetation (PV), non-photosynthetic vegetation (NPV), soil and shade. The PV shade normalized fraction corresponds to the FVC. In summer of 2016 we established 85 30 × 30 m field plots and 360 2 × 2 m field plots to measure the percentage of total vegetation cover in order to validate the FVC estimates made from remote sensing data. The FVC time series showed the same general pattern with both spatial scales and modeling approaches, high fire recurrence categories registering the highest resilience. The accuracy of the dimidiate pixel model was significantly higher for WorldView-2 based estimates (RMSE: 5–10%) than for Landsat (RMSE: 10–15%). The dimidiate pixel model computed from NDVI for both Landsat and WorldView-2 underestimated FVC at high field-sampled vegetation cover, while MESMA estimations were accurate for the entire range of vegetation cover for both satellites. The fraction of photosynthetic vegetation calculated using WorldView-2 had a higher performance (RMSE: 4–6%) than that quantified from Landsat (RMSE: 6–8%). The linear relationships assumed for validation purposes were statistically significant for both sensors and modeling approaches. Our study demonstrates the highest performance of very high spatial resolution satellite imagery and MESMA models in the quantitative estimation of FVC as a measure of post-fire resilience.
Article
Full-text available
Evergreen forest plays a particular role in biodiversity, carbon sequestration, and soil and water conservation, and the spatial and temporal evolutions of evergreen forest complexly interact with climate change. However, it is difficult for the existing remote sensed data to meet both demands of large scale and high accuracy. To solve this problem, we developed a time-series FEVC-CV method by combining the fractional evergreen forest cover model (FEVC) with the coefficient of variation (CV) at sub-pixel scales. This method using the normalized difference vegetation index (NDVI) dataset from the Moderate Resolution Imaging Spectroradiometer (MODIS) product to meet the needs of large-scale mapping. Considering the severe mixed-pixel problem and spatial heterogeneity of the study area, the improved dimidiate pixel model was used to extract the fractional evergreen forest cover at sub-pixel scales by introducing a new variable the intra-annual NDVI minimum value (NDVIann-min) and dividing grid units. Meanwhile, the CV of the time-series NDVI highlighted the time series fluctuation stability of evergreen forest compared with other vegetation types, such as deciduous forest and continuous crop. Therefore, the intra-annual time-series CV (CVai) and the CV of the continuous crop key phenology (CVkp) period were used to eliminate the interferences from other vegetation types, which were extremely admixed with low-coverage evergreen forest. We then verified the accuracy of the algorithm using 2-m resolution Gaofen-1 images. The results revealed that the overall accuracy of our algorithm exceeded 90%, with a root mean square error (RMSE) of cover fraction of around 10%. In addition, the mean relative error (MRE) indicated that the extraction accuracy of evergreen forest in non-urban areas was superior to the accuracy in urban areas. The results show that the algorithm achieved fairly high accuracy in detecting evergreen forest, including evergreen trees in urban areas, at a large scale.
Article
Full-text available
The European Space Agency (ESA)’s Sentinel-2A (S2A) mission is providing time series that allow the characterisation of dynamic vegetation, especially when combined with the National Aeronautics and Space Administration (NASA)/United States Geological Survey (USGS) Landsat 7 (L7) and Landsat 8 (L8) missions. Hybrid retrieval workflows combining non-parametric Machine Learning Regression Algorithms (MLRAs) and vegetation Radiative Transfer Models (RTMs) were proposed as fast and accurate methods to infer biophysical parameters such as Leaf Area Index (LAI) from these data streams. However, the exact design of optimal retrieval workflows is rarely discussed. In this study, the impact of five retrieval workflow features on LAI prediction performance of MultiSpectral Instrument (MSI), Enhanced Thematic Mapper Plus (ETM+) and Operational Land Imager (OLI) observations was analysed over a Dutch beech forest site for a one-year period. The retrieval workflow features were the (1) addition of prior knowledge of leaf chemistry (two alternatives), (2) the choice of RTM (two alternatives), (3) the addition of Gaussian noise to RTM produced training data (four and five alternatives), (4) possibility of using Sun Zenith Angle (SZA) as an additional MLRA training feature (two alternatives), and (5) the choice of MLRA (six alternatives). The features were varied in a full grid resulting in 960 inversion models in order to find the overall impact on performance as well as possible interactions among the features. A combination of a Terrestrial Laser Scanning (TLS) time series with litter-trap derived LAI served as independent validation. The addition of absolute noise had the most significant impact on prediction performance. It improved the median prediction Root Mean Square Error (RMSE) by 1.08 m2 m−2 when 5% noise was added compared to inversions with 0% absolute noise. The choice of the MLRA was second most important in terms of median prediction performance, which differed by 0.52 m2 m−2 between the best and worst model. The best inversion model achieved an RMSE of 0.91 m2 m−2 and explained 84.9% of the variance of the reference time series. The results underline the need to explicitly describe the used noise model in future studies. Similar studies should be conducted in other study areas, both forest and crop systems, in order to test the noise model as an integral part of hybrid retrieval workflows.
Article
Full-text available
The fraction of absorbed photosynthetically active radiation (FAPAR) is generally divided into the fraction of radiation absorbed by the photosynthetic components (FAPARgreen) and the fraction of radiation absorbed by the non-photosynthetic components (FAPARwoody) of the vegetation. However, most global FAPAR datasets do not take account of the woody components when considering the canopy radiation transfer. The objective of this study was to develop a generic algorithm for partitioning FAPARcanopy into FAPARgreen and FAPARwoody based on a triple-source leaf-wood-soil layer (TriLay) approach. The LargE-Scale remote sensing data and image simulation framework (LESS) model was used to validate the TriLay approach. The results showed that the TriLay FAPARgreen had higher retrieval accuracy, as well as a significantly lower bias (R2 = 0.937, Root Mean Square Error (RMSE) = 0.064, and bias = −6.02% for black-sky conditions; R2 = 0.997, RMSE = 0.025 and bias = −4.04% for white-sky conditions) compared to the traditional linear method (R2 = 0.979, RMSE = 0.114, and bias = −18.04% for black-sky conditions; R2 = 0.996, RMSE = 0.106 and bias = −16.93% for white-sky conditions). For FAPAR that did not take account of woody components (FAPARnoWAI), the corresponding results were R2 = 0.920, RMSE = 0.071, and bias = −7.14% for black-sky conditions, and R2 = 0.999, RMSE = 0.043, and bias = −6.41% for white-sky conditions. Finally, the dynamic FAPARgreen, FAPARwoody, FAPARcanopy and FAPARnoWAI products for a North America region were generated at a resolution of 500 m for every eight days in 2017. A comparison of the results for FAPARgreen against those for FAPARnoWAI and FAPARcanopy showed that the discrepancy between FAPARgreen and other FAPAR products for forest vegetation types could not be ignored. For deciduous needleleaf forest, in particular, the black-sky FAPARgreen was found to contribute only about 23.86% and 35.75% of FAPARcanopy at the beginning and end of the year (from January to March and October to December, JFM and OND), and 75.02% at the peak growth stage (from July to September, JAS); the black-sky FAPARnoWAI was found to be overestimated by 38.30% and 28.46% during the early (JFM) and late (OND) part of the year, respectively. Therefore, the TriLay approach performed well in separating FAPARgreen from FAPARcanopy, which is of great importance for a better understanding of the energy exchange within the canopy.
Article
Full-text available
The sky view factor (SVF) is a crucial variable widely used to quantify the characteristics of surface structures and estimate surface radiation budget. Many SVF models based on raster data have been developed but not yet evaluated in a more quantitative and uniform manner. In this paper, four typical SVF models (Dozier-Frew (D-F), Manners, Lindberg-Grimmond (L-G), and Helbig_h) are evaluated using the SVF derived from simulated fisheye images based on the digital surface model (DSM) and digital elevation model data. The SVF calculated by D-F method using DSM data has the best accuracy, with a mean bias error of −0.007, root-mean-square error of 0.069, and coefficient of determination (R ² ) of 0.914. For the SVF value derived from digital elevation model data, L-G method shows good performance, with an mean bias error of 0.013, root-mean-square error of 0.032, and R ² of 0.897. The pixels near the edges of buildings, within the valley or along ridgelines, have higher SVF deviations. In addition, the slope angle calculated using DSM data has some artificial defects that make the significant impact on the SVF biases due to their calculation method and the discontinuous surface in urban areas. Thus, L-G and Helbig_h methods are more applicable for the DSM data due to the difficulty in defining slope and aspect angles. Moreover, the high accordance of SVFs between Helbig_h and L-G methods implies that the Helbig_h method is an alternative in virtue of its simpler form and lower computation cost than L-G method.
Article
Full-text available
Three-dimensional (3D) radiative transfer modeling of the transport and interaction of radiation through earth surfaces is challenging due to the complexity of the landscapes as well as the intensive computational cost of 3D radiative transfer simulations. To reduce computation time, current models work with schematic landscapes or with small-scale realistic scenes. The computer graphics community provides the most accurate and efficient models (known as renderers) but they were not designed specifically for performing scientific radiative transfer simulations. In this study, we propose LESS, a new 3D radiative transfer modeling framework. LESS employs a weighted forward photon tracing method to simulate multispectral bidirectional reflectance factor (BRF) or flux-related data (e.g., downwelling radiation) and a backward path tracing method to generate sensor images (e.g., fisheye images) or large-scale (e.g. 1 km2) spectral images. The backward path tracing also has been extended to simulate thermal infrared radiation by using an on-the-fly computation of the sunlit and shaded scene components. This framework is achieved through the development of a user-friendly graphic user interface (GUI) and a set of tools to help construct the landscape and set parameters. The accuracy of LESS is evaluated with other models as well as field measurements in terms of directional BRFs and pixel-wise simulated image comparisons, which shows very good agreement. LESS has the potential in simulating datasets of realistically reconstructed landscapes. Such simulated datasets can be used as benchmarks for various applications in remote sensing, forestry investigation and photogrammetry.
Article
The Monte Carlo simulation (MCS) method for uncertainty evaluation of measurement results has gained popularity as an alternative to the method based on law of propagation of uncertainty (LPU) ever since the recommendation by JCGM through supplement JCGM 101:2008. In this paper, efforts have been made to compute the uncertainty in measurement of hardness in the Vickers hardness scale using MCS and LPU method. Three different hardness blocks of different hardness scales, namely HV1, HV10 and HV30, have been calibrated following the standard procedure in the primary Vickers hardness machine established in CSIR-National Physical Laboratory. The mean hardness values and the associated measurement uncertainty of the hardness blocks are computed using LPU and MCS methods. A comparison of the results obtained through LPU and MCS methods has been carried out. It is observed that there is a good agreement between findings from both the methods adopted, hence confirming that MCS method can be employed in the various fields of hardness metrology for evaluation of measurement uncertainty.