Agricultural and Forest Meteorology 218–219 (2016) 243–249
Contents lists available at ScienceDirect
Agricultural and Forest Meteorology
journal homepage: www.elsevier.com/locate/agrformet
Informative spectral bands for remote green LAI estimation
in C3 and C4 crops
Oz Kira a , Anthony L. Nguy-Robertson b , Timothy J. Arkebauer c , Raphael Linker a ,
Anatoly A. Gitelson a,b,∗
a
Faculty of Civil and Environmental Engineering, Israel Institute of Technology, Technion, Technion City, Haifa, Israel
School of Natural Resources, University of Nebraska, Lincoln, NE, USA
c
Department of Agronomy and Horticulture, University of Nebraska-Lincoln, Lincoln, NE, USA
b
a r t i c l e
i n f o
Article history:
Received 7 September 2015
Received in revised form
15 December 2015
Accepted 28 December 2015
Keywords:
Remote sensing
Reflectance
Neural network
Partial least squares
Vegetation index
Maize
Soybean
a b s t r a c t
Green leaf area index (LAI) provides insight into the productivity, physiological and phenological status
of vegetation. Measurement of spectral reflectance offers a fast and nondestructive estimation of green
LAI. A number of methods have been used for the estimation of green LAI; however, the specific spectral bands employed varied widely among the methods and data used. Our objectives were (i) to find
informative spectral bands retained in three types of methods, neural network (NN), partial least squares
(PLS) regression and vegetation indices (VI), for estimating green LAI in maize (a C4 species) and soybean
(a C3 species); (ii) to assess the accuracy of the algorithms estimating green LAI using a minimal number
of bands for each crop and generic algorithms for the two crops combined.
Hyperspectral reflectance and green LAI of irrigated and rainfed maize and soybean were taken during
eight years of observations (altogether 24 field-years) in very different weather conditions. The bands
retained in the best NN, PLS and VI methods were in close agreement. The validity of these bands was further confirmed via the uninformative variable elimination PLS technique. The red edge and the NIR bands
were selected in all models and were found the most informative. Identifying informative spectral bands
across all four techniques provided insight into spectral features of reflectance specific for each species as
well as those that are common to species with different leaf structures, canopy architectures and photosynthetic pathways. The analyses allowed development of algorithms for estimating green LAI in soybean
and maize with no re-parameterization. These findings lay a strong foundation for the development of
generic algorithms which is crucial for remote sensing of vegetation biophysical parameters.
© 2015 Elsevier B.V. All rights reserved.
1. Introduction
Green leaf area index (LAI) provides insight into the productivity, physiological and phenological status of vegetation as it is
a measure of the photosynthetically active leaf area over ground
area (Daughtry et al., 1992; Watson, 1947). Green LAI is a useful metric for ecological (e.g. Evans et al., 2013; Hardwick et al.,
2015), crop (Casa et al., 2012; Mishara et al., 2015), and carbon (e.g.
Anderson et al., 2015; Gonsamo and Chen, 2014) models, among
many others. Measurement of spectral reflectance offers a fast,
nondestructive method for LAI estimation (e.g., Asrar et al., 1984;
Delegido et al., 2015). A number of methods such as radiative transfer (e.g., Atzberger and Richter, 2012), look-up tables (e.g. Duan
∗ Corresponding author.
E-mail address: [email protected] (A.A. Gitelson).
http://dx.doi.org/10.1016/j.agrformet.2015.12.064
0168-1923/© 2015 Elsevier B.V. All rights reserved.
et al., 2014), vegetation indices (VI; e.g., Liu et al., 2012), partial
least squares (PLS; e.g. Darvishzadeh et al., 2011), neural networks
(NN; e.g., Chen et al., 2015), or a combination of thereof (e.g. Liang
et al., 2015) have been used previously for the estimating green LAI
(see Verrelst et al., 2015 for a review).
Early studies identified the red and NIR regions as sensitive to
leaf area index, thus, reflectances in these regions are commonly
used in VIs such as simple ratio (Jordan, 1969) and normalized
difference (Rouse et al., 1974, Huete et al., 1997) among others.
While these indices were found to be sensitive to low LAI, they
lose sensitivity as LAI increases above 2–3 (e.g. Tucker, 1977). Thus,
applicability of other spectral regions to estimate moderate-to-high
LAI was investigated.
Sensitivity of reflectance in the red edge region to greenness/chlorophyll content of vegetation has been recognized for
decades (Curran et al., 1991, 1990; Danson and Plummer, 1995;
Gitelson and Merzlyak, 1994; Horler et al., 1983; Vogelmann et al.,
244
O. Kira et al. / Agricultural and Forest Meteorology 218–219 (2016) 243–249
1993) and this spectral region was widely used for estimating vegetation biophysical parameters at leaf and canopy levels. Recently
published reviews provide a summary of the multispectral and
hyperspectral remote sensing efforts in more detail and the reader
is referred to these articles for a more thorough understanding
(Gitelson, 2011a,b; Hatfield et al., 2008; Le Maire et al., 2008; Ustin
et al., 2009; Verrelst et al., 2015).
While the utility of the red-edge spectral region has been
demonstrated in various studies, positions of red edge bands has
varied widely from 690 to 740 nm (e.g., Hatfield et al., 2008; Le
Maire et al., 2008). Most of the models developed for green LAI estimation are based on data taken for one species or in one growth
stage (see review of Verrelst et al., 2015). Only a few studies have
investigated the applicability of algorithms for several species with
no algorithm re-parameterization (e.g., Delegido et al., 2013; NguyRobertson et al., 2014). Performance of different techniques for
green LAI estimation has been compared, to our knowledge, only
for one species (e.g., Siegmann and Jarmer, 2015).
Our objectives were (i) to find informative spectral bands
retained in three types of methods, PLS, NN, and VI, for estimating green LAI in maize (a C4 species) and soybean (a C3 species);
(ii) to assess the accuracy of the algorithms estimating green LAI
using a minimal number of bands in each crop and in two crops
combined.
2. Methods
2.1. Site summary
The study site was located at the University of NebraskaLincoln Agricultural Research and Development Center near Mead,
Nebraska, which is a part of the AmeriFlux network. This study
site consists of three approximately 65-ha fields. Each field was
managed differently as either continuous irrigated maize, irrigated
maize/soybean rotation, or rainfed maize/soybean rotation. There
were a total of 16 and 8 field-years for maize and soybean, respectively, which had maximal green LAI values ranging from 4.3 to
6.5 m2 m−2 for maize and 3.0 to 5.5 m2 m−2 for soybean. Specific
details of these three fields (AmeriFlux sites Ne-1, Ne-2 and Ne-3)
can be found in Verma et al. (2005) and Viña et al. (2011).
2.2. Field measurements
Located in each field were six 20 × 20 m plots that represented
major soil and crop production zones (Verma et al., 2005). For estimating LAI, 6 plants were selected from one or two rows totalling
1 m length within each plot every 10–14 days. Sampling rows were
alternated between collections to minimize edge effects. Samples
were transported on ice prior to LAI measurements using an area
meter (model LI-3100, LI-COR, Inc., Lincoln, NE). LAI measurements
were determined by multiplying the green leaf area or total leaf area
per plant by the plant population in the plot. The field-level LAI was
determined from the average of LAI measured at the six plots on a
given sampling date.
The hyperspectral data collection started in 2001 and ended
in 2008 using an all-terrain sensor platform (Rundquist et al.,
2014, 2004). Two radiometers (model USB2000, Ocean Optics, Inc.,
Dunedin, FL) with a dual-fiber system were used to collect canopy
reflectance in the spectral range from 400 to 1100 nm with an
approximate 2 nm resolution. The downwelling fiber was fitted
with a cosine diffuser to measure irradiance while the upwelling
fiber collected upwelling radiance. The upwelling fiber had a field
of view of approximately 2.4 m in diameter since the height of
the fiber was maintained at a constant height above the canopy
throughout the growing season. The median of 36 reflectance
measurements collected along access roads into each field was
used as the field-level reflectance measurement. A total of 278
spectra for maize and 145 for soybean were acquired over the
study period (details are in Nguy-Robertson et al., 2014, 2012 and
Viña et al., 2011). The reflectance measurements and LAI measurements were not always concurrent. Since LAI changes gradually,
spine interpolations were taken between destructive LAI sampling
dates for each field in each year using Matlab (V. 7.9.0.529, The
MathWorks, Massachusetts, USA; Nguy-Robertson et al., 2012).
2.3. Data analysis
A reference green LAI value was available for each reflectance
measurement. The errors of green LAI estimation were calculated
separately for maize and soybean as well as for maize and soybean
combined.
Two different non-parametric regression methods and VIs were
selected for LAI estimation – partial least square (PLS) regression
(Brereton, 2005) and neural networks (NN; Haykin, 2005). Both
methods are commonly used in remote sensing (e.g., Baret et al.,
2009; Siegmann and Jarmer, 2015). The NN method was used for
development of the LAI product of future space systems (Baret et al.,
2009). To ensure a fair comparison between the different methods the same calibration and validation sets were used in all the
methods.
Much of the information content within reflectance spectra may
be redundant and can be explained with fewer than 466 spectral
bands. We used uninformative variable elimination PLS (UVE PLS;
Cai et al., 2008; Centner et al., 1996) to identify the most informative
spectral regions of the hyperspectral data. This method has been
used previously to find informative spectral bands for leaf chlorophyll and carotenoid content estimation (Kira et al., 2015). UVE PLS
assists in reducing the data dimension by eliminating spectral data
which are uninformative or redundant. The reliability measure c
computed for each wavelength range allows the elimination of less
informative wavelengths:
c =
b
std(b )
where b is the wavelength dependent regression coefficient originated from the PLS regression, and std(b ) is the standard deviation
of the regression coefficient vector. A smaller absolute value of the
reliability measure indicates that the data are less informative and
can be removed at the user’s discretion. For the UVE-PLS computation 50% of the data was used. The reflectance measurements
at 2 nm resolution were grouped into non-overlapping spectral
bands which included five reflectances, e.g., the band centered
at 700 nm included the reflectance at 696 nm, 698 nm, 700 nm,
702 nm, and 704 nm. Information content of reflectance spectra of
both crops combined, for data with 2 nm and these data resampled to 10 nm resolution, is shown in Fig. 1. It can be seen that data
with 10 nm resolution adequately presented the information content of reflectance spectra. Thus, resampling to 10 nm resolution
minimized the loss of information content while greatly decreasing
the processing time for NN and PLS.
The training and validation process used a number of 10 nm
spectral bands which varied depending on the modeling method.
During the training process all possible combinations of the spectral
bands were tested. Unless otherwise stated, each band combination was tested 20 times using 50% of the measurements (selected
randomly), resulting in 20 models altogether. For each combination, the error measures reported below correspond to the average
error of all 20 models. Four error measures were chosen to assess
the performance of the models (Simon et al., 2012): normalized
25
245
10
NN
2nm
20
NRMSE (%)
Absolute reliability parameter
O. Kira et al. / Agricultural and Forest Meteorology 218–219 (2016) 243–249
10 nm
15
10
PLS
9
8
5
0
7
400
500
600
700
3
2
800
4
5
Number of bands
Wavelength (nm)
Fig. 1. Reliability parameter calculated using uninformative variable elimination
PLS technique for green LAI estimation in the two crops combined plotted versus
wavelength of reflectance taken with 2 nm and 10 nm resolution.
Fig. 2. Normalized root mean square error (NRMSE) of green LAI estimation in maize
and soybean combined as function of number of spectral bands used by NN and PLS
regression.
10
root mean square error (NRMSE), coefficient of determination (R2 ),
ND =
CI =
(1 − 2 )
(1 + 2 )
Maize
NRMSE (%)
mean normalized bias (MNB), and normalized mean bias (NMB).
The NN method was trained with a single spectral band (40 models) or two (821 combinations), three (10,661 combinations), four
(101,271 combinations), and five (749,399 combinations) 10 nm
bands. Due to the processing time, the 5-band NN model was tested
5 times rather than 20 times for each combination.
The NN used in this study was a standard two-layer
feed-forward network with sigmoid transfer functions. The
Levenberg–Marquardt algorithm was selected for the training
process. The stopping criteria included the performance error
(1 × 10−5 ), minimal gradient (1 × 10−15 ), Marquardt adjustment
parameter (1 × 1010 ), and maximum number of iterations (100).
PLS models were calibrated with the same training and validation
sets as the NN models.
The VIs used for LAI estimation were normalized difference (ND)
and chlorophyll index (CI):
Soybean
Two crops
9
8
7
NN
PLS
CI red edge
Fig. 3. Normalized root mean square error (NRMSE) of green LAI estimation in
maize, soybean and in two crops combined by NN with four bands, PLS regression
with three bands and CIred edge using two bands.
3. Results and discussion
3.1. Accuracy of green LAI estimation
1
−1
2
where 1,2 are reflectances at 10 nm width spectral bands.
In this study all the possible combinations of bands (821) were
tested to find the optimal one for the specific dataset based on
NRMSE. The VIs were fitted to the corresponding green LAI values
by using a 2nd order polynomial, a function commonly used for
reflectance model vs. green LAI relationships (e.g., Hatfield et al.,
2008).
To find the optimal number of bands used by the NN and PLS
methods, the accuracy of green LAI estimation in maize and soybean separately and combined was assessed (Table 1). For NN, the
optimal number of 10 nm spectral bands (each contains 5 wavelengths) was four as the difference between four and five bands
was less than 0.2% (Fig. 2). For PLS, only three 10 nm bands were
necessary as the improvement from three to four spectral bands
was only 0.3% (Fig. 2).
The NN method had the highest accuracy of green LAI estimation
in terms of NRMSE and bias for each crop and both crops combined (Fig. 3, Table 1). The algorithms produced by any of the three
methods were able to estimate green LAI in both crops combined
Table 1
Descriptive statistics of the relationships between green LAI measured and estimated by NN, PLS regression, and VIs, CIred edge and red edge normalized difference (NDred edge ).
NRMSE is root mean square error normalized to range of green LAI variation (%), R2 is coefficient of determination, MNB is the mean normalized bias (%), NMB is the normalized
mean bias (%).
NN
Maize
Soybean
Two crops combined
PLS
NRMSE
R2
MNB
NMB
NRMSE
R2
MNB
NMB
7.6
8.6
7.7
0.92
0.9
0.92
7.5
−1.9
4.4
−0.29
−1.94
−0.73
8.7
9.0
8.5
0.89
0.89
0.90
5.9
−6.4
1.8
0.21
−4.12
−0.94
NRMSE
R2
MNB
NMB
NRMSE
R2
MNB
NMB
8.5
9.7
8.6
0.90
0.87
0.90
9.8
1.8
7.1
−0.77
0.44
0.45
8.3
9.5
8.5
0.90
0.87
0.90
10.9
3.3
8.4
−0.55
−0.23
−0.47
Red edge ND
Maize
Soybean
Two crops combined
Red edge CI
246
O. Kira et al. / Agricultural and Forest Meteorology 218–219 (2016) 243–249
NN
PLS
4
2
4
2
0
0
0
2
4
4
2
0
0
6
CI
6
LAI predicted
6
LAI predicted
LAI predicted
6
2
LAI measured
4
6
0
2
4
LAI measured
Fig. 4. Green LAI predicted in maize and soybean combined by (A) NN, (B) PLS regression and (C) CIred
without re-parameterization for each crop with NRMSE below 8.6%
(Figs. 3 and 4; Table 1).
3.2. Spectral bands selection
10
9
8
7
6
5
4
3
2
1
Band 1
NN
Band 2
Band 3
Band 4
450
500
550
600
650
700
750
Band 1
PLS
A
400
10
9
8
7
6
5
4
3
2
1
400
800
band1
band2
C
400
450
500
550
600
650
Wavelength (nm)
Band 3
500
600
700
800
Wavelength (nm)
Top combinations
Top combinations
CIred edge
Band 2
B
Wavelength (nm)
10
9
8
7
6
5
4
3
2
1
edge .
in the blue (400–500 nm) were amongst the least useful variables.
Notably, the reliability parameter spectra of the two highly contrasting crop species, maize and soybean, were remarkably similar
(not shown).
In the next step we analyzed the selection of the optimal spectral
bands by the various methods. For individual crops and the maize
and soybean combined, NN retained two bands in the red edge
(710–730 nm) and NIR (above 760 nm). All of the ten top combinations of spectral bands providing maximal accuracy retained red
edge and NIR bands and resulted in NRMSE for the two crops combined below 9.2%. When NN with three spectral bands was used, in
addition to the red edge and NIR bands, the third band was located
either in the blue or red range of the spectrum. This resulted in
reducing the NRMSE to 8.3%. When NN used four spectral bands,
an additional red edge band was retained (Fig. 5A). Thus, NN with
four spectral bands, NIR, two bands in the red edge range, one of
them at 700–710 nm and another at 720–730 nm, and either blue
or red, was able to estimate green LAI in two crops combined with
a NRMSE below 7.7%.
Similar to NN, PLS based on two spectral bands retained the red
edge (710–730 nm) and NIR (above 760 nm) bands and was able to
achieve a NRMSE for the two crops combined below 9.8%. PLS based
on three spectral bands, in addition to the red edge and NIR bands
retained a third band located in either the blue (410–470 nm) or
red (670–680 nm) range (Fig. 5B); it decreased the NRMSE to 8.5%.
For the two formulations of VIs tested, the red edge
(710–720 nm) and NIR (above 750 nm) bands were essential
for green LAI estimation in maize and soybean and both crops
Top combinations
Top combinations
All three methods, NN, PLS and VI, using a small number of
selected spectral bands, demonstrated high accuracy estimating
green LAI of both maize and soybean separately as well as of the
two crops combined. Identifying these spectral bands across three
methods provides insight into spectral features of reflectance specific to each species as well as those common to species with very
different leaf structure and canopy architecture such as soybean
and maize. Identifying these spectral bands also allows development of algorithms for estimating green LAI in these crops with no
re-parameterization.
Firstly, UVE PLS was applied to evaluate quantitatively the information content of reflectance data for green LAI estimation. Centner
et al. (1996) cautions that this approach is not for band selection,
but it is a way to eliminate variables that are useless. The magnitude
of the reliability parameter was used in our study as an indicator of
useful information contained in reflectance spectra. This parameter
was then compared with the most informative bands identified by
the other methods (NN, PLS and VIs).
The main features of the absolute reliability parameter spectra
for both crops combined were prominent peaks in the red edge
range, 700 through 740 nm, and NIR range, above 760 nm (Fig. 1).
Importantly, the reflectance in main absorption bands of chlorophylls in the red range of the spectrum (around 660–680 nm) and
6
LAI measured
700
750
800
10
9
8
7
6
5
4
3
2
1
band1
NDred edge
band2
D
400
450
500
550
600
650
700
750
800
Wavelength (nm)
Fig. 5. Ten top band (10 nm width) positions retained for estimating green LAI in maize and soybean combined by (A) NN with 4 spectral bands, (B) PLS regression with 3
spectral bands, (C) CIred-edge using two bands and (D) NDred-edge using two bands.
O. Kira et al. / Agricultural and Forest Meteorology 218–219 (2016) 243–249
25
Two crops
Maize
Soybean
NRMSE (%)
20
15
10
5
400
450
500
550
600
650
700
750
Wavelength (nm)
Fig. 6. NRMSE of green LAI estimation by CI = (NIR / ) − 1 plotted versus wavelength . The position of the NIR band in the numerator was fixed at 780 nm.
combined. These bands were in all ten top combinations of the
spectral bands bringing NRMSE below 8.5% for maize, 9.8% for
soybean and 8.7% for crops combined (Fig. 5C and D).
It is instructive that the spectral band selection for VIs coincided
consistently with bands selected by the two other methods, NN and
PLS as well as with informative spectral regions identified by UVE
PLS. Since reflectances in the NIR range between 760 and 800 nm
are closely related (R2 was above 0.9), the position of the NIR band
in the range 760–800 nm was not critical and did not affect the
accuracy of green LAI estimation. However, the position of the red
edge band does have an effect on accuracy of estimation (NguyRobertson et al., 2014).
To determine the effect of band position on the accuracy of green
LAI estimation by the CI = (NIR / ) − 1, the position of the NIR band
was fixed at 780 nm and the position of the band in the denominator was allowed to change from 400 to 750 nm. The NRMSE was
calculated for each position of the spectral band in the denominator
(Fig. 6). The spectra of NRMSE for each crop and the two crops combined had similar shapes with two pronounced minima in the green
range (around 550 nm) and in the red edge range around 720 nm.
Thus, CIred edge with a spectral band at 710–730 nm allowed green
LAI estimation in both crops combined with a NRMSE below 8.5%.
PLS and NN are by no means the ultimate modeling tools and
other regression approaches such as random forests or Gaussian
processes regression should also be investigated (Verrelst et al.,
2015). However, we believe it to be unlikely that such approaches
would result in the selection of other spectral ranges. The fact that
the same spectral bands bring the highest accuracy of green LAI
estimation by all three methods, and were confirmed by UVE PLS,
gives strong confidence that these band combinations are indeed
the most informative ones and the result is not merely an artifact
of the modeling approaches.
To understand the selection of optimal spectral bands by all four
methods, we analyzed relationships between reflectance at informative spectral bands and green LAI (Fig. 7). Reflectance in the blue
(not shown) and the red range of the spectrum was sensitive to
low-to-moderate green LAI below 3 m2 m−2 , widely used for green
LAI estimation (Fig. 7A). Note almost the same sensitivity of red
reflectance to green LAI below 3 m2 m−2 for both species which
likely explains the increase in accuracy of green LAI estimation
when either blue or red bands were used as the last informative
band by PLS and NN – (third and fourth respectively; Fig. 5A and
B). It also explains the increase in information content of blue and
red reflectance indicated by UVE PLS for the two crops combined
(Fig. 1).
However, when LAI exceeded 3 m2 m−2 , reflectance in these
two regions became almost invariant and insensitive to increasing green LAI (Fig. 7A). The saturation of the reflectance vs. green
LAI relationship at moderate to high green LAI in the red range is
due to the high chlorophyll absorption coefficient in this spectral
range (e.g., Buschmann and Nagel, 1993; Merzlyak and Gitelson,
1995; Tucker, 1977). This resulted in saturation of leaf absorption
at low chlorophyll content (below 200 mg m−2 ), corresponding to
slightly green leaves (Gitelson and Merzlyak, 1994). This saturation of absorption and reflectance prevents using these spectral
25
Reflectance @710 nm, %
Reflectance @670 nm, %
20
30
Maize
25
Soybean
A
Maize
Soybean
15
10
5
0
B
20
15
10
5
0
0
1
2
3
4
5
6
0
7
1
2
Green LAI
45
4
5
6
7
70
Maize
C
Reflectance @780 nm, %
30
25
20
15
10
5
0
D
Maize
60
Soybean
35
3
Green LAI
40
Reflectance @730 nm, %
247
Soybean
50
40
30
20
10
0
0
1
2
3
4
Green LAI
5
6
7
0
1
2
3
4
5
6
7
Green LAI
Fig. 7. Relationships between reflectance and green LAI (in m2 m−2 ) for maize and soybean: red range at 670 ± 5 nm (A), red edge range at 710 ± 5 nm (B) and at 730 ± 5 nm
(C), NIR range at 780 ± 5 nm (D).
248
O. Kira et al. / Agricultural and Forest Meteorology 218–219 (2016) 243–249
4
3.5
CIred edge
3
2.5
2
1.5
1
y = 0.5036x
R² = 0.90
0.5
0
0
1
2
3
4
5
6
7
Green LAI
Fig. 8. Relationships between red edge chlorophyll index (CIred edge ) and green LAI
for maize and soybean combined.
bands for green LAI estimation above 3 m2 m−2 (Gitelson et al.,
2003; Nguy-Robertson et al., 2012; Tucker, 1979; Viña et al., 2011).
In the red edge range, the depth of light penetration into leaf and
canopy is higher than in the red (Fukshansky, 1981; Fukshansky
et al., 1993; Merzlyak and Gitelson, 1995) and canopy reflectance
becomes more sensitive to green LAI (Fig. 7C for wavelength
710 nm). Reflectance in the red edge is much higher than in the visible range especially for moderate to high green LAI (7–12% in red
edge range vs. 1–2% in the red and blue ranges). Thus the upwelling
radiance in the red edge range is higher and less noisy which adds
to the advantage of using red edge bands for green LAI estimation.
Importantly, with increases of green LAI, reflectance of both
crops in the range between 690 and 720 nm decreased (Fig. 7B) with
soybean reflectance remaining higher than in maize. In contrast,
at wavelengths above 730 nm, reflectance of both maize and soybean increased with soybean reflectance consistently higher than
that of maize. In the range around 730 nm reflectance of both crops
was almost invariant with consistently higher soybean reflectance
(Fig. 7C). This point of equilibrium is a result of the tradeoff between
chlorophyll absorption, which is a greater factor at shorter wavelengths, and scattering, which becomes the governing factor at
wavelengths beyond 730 nm.
For algorithm development, that does not require reparameterization for each crop, the relative positions of maize and
soybean reflectance vs. green LAI relationships are critically important (Fig. 7). In the red edge range, light penetrates deeply inside the
leaf and the different structures of maize and soybean leaves do not
affect leaf absorption. Thus, the main factor affecting reflectance in
this range is the starkly different canopy architectures with higher
scattering by soybean. In the red edge, soybean reflectance is higher
than that of maize (Fig. 7B). The same is true for NIR reflectance
(Fig. 7D). Such consistent relative positions of reflectance vs. green
LAI relationships in the red edge and in the NIR was the reason for
the accurate green LAI estimation in both crops by VIs (Fig. 8). For
the same reason, both NN and PLS techniques consistently used NIR
and red edge bands. This result agrees well with those reported in
other studies (e.g., Clevers and Kooistra, 2012; Delegido et al., 2013).
4. Conclusions
Over the last few decades, technological developments have
made it possible to remotely assess green LAI, one of the main
factors in vegetation productivity and physiological status. The
focus of this study was to identify informative spectral bands
for four types of methods, neural network, partial least squares
regression, uninformative variable elimination PLS technique, and
vegetation indices, and test their performance for estimating green
LAI in two contrasting C3 and C4 crop species, maize and soybean,
using multiyear observations. Hybrids tested had the greatest
architectural diversity available, providing dramatically different
canopy light environments.
The spectral intervals selected in all four types of methods, NN,
PLS, VIs and UVE PLS, correspond to fundamental spectral features
of pigment absorption, reflectance and leaf/canopy scatterings. The
consistency of the spectral bands retained by PLS and NN as well as
VIs is quite remarkable, indicating the robustness of the band selection. A very important result is that all techniques tested were not
species-specific for two contrasting crops with biochemical, structural, and abiotic differences. This finding allows for accurate green
LAI estimation in these crops without re-parameterization of the
algorithms. Several studies have shown that algorithms calibrated
over Nebraska sites have been applicable to accurately estimate
maize yield across the U.S.A. (Sakamoto et al., 2014), maize and
soybean gross primary production across the corn belt (Gitelson
et al., 2012), and green LAI of wheat and potato grown in Israel
(Nguy-Robertson et al., 2014). Thus, it is likely that the models
developed in this study may be applicable in other agroecosystems.
If these results are confirmed for other species, spatial resolution
requirements of aircraft and satellite observations would be less
restrictive by tolerating mixed pixels containing different crops.
The analysis presented in this study is an important step in the
development of generic algorithms crucial for remote sensing of
vegetation biophysical parameters.
Acknowledgements
This research was supported by the NASA NACP program, Lady
Davis and Marie Curie International Incoming Fellowship to AG and
partially supported by the U.S. Department of Energy. We sincerely
appreciate the support and the use of facilities and equipment provided by the Center for Advanced Land Management Information
Technologies (CALMIT) and data from the Carbon Sequestration
Program, University of Nebraska-Lincoln. We are thankful to two
anonymous reviewers for their constructive and helpful criticism
and suggestions.
References
Anderson, R.G., Tirado-Corbalá, R., Wang, D., Ayars, J.E., 2015. Long-rotation
sugarcane in Hawaii sustains high carbon accumulation and radiation use
efficiency in 2nd year of growth. Agric. Ecosyst. Environ. 199, 216–224, http://
dx.doi.org/10.1016/j.agee.2014.09.012.
Asrar, G., Fuchs, M., Kanemasu, E.T., Hatfield, J.L., 1984. Estimating absorbed
photosynthetic radiation and leaf area index from spectral reflectance in
wheat. Agron. J. 76 (2), 300–306, http://dx.doi.org/10.2134/agronj1984.
00021962007600020029x.
Atzberger, C., Richter, K., 2012. Spatially constrained inversion of radiative transfer
models for improved LAI mapping from future Sentinel-2 imagery. Remote
Sens. Environ. 120, 208–218, http://dx.doi.org/10.1016/j.rse.2011.10.035.
Baret, F., Weiss, M., Berthelot, B., 2009. Sentinel-2 MSI Products – WP1152
Algorithm Theoretical Basis Document for Product Group. INRA-EMMAH,
Avignon, France.
Brereton, R.G., 2005. Chemometrics, Data Analysis for the Laboratory and Chemical
Plant. John Wiley & Sons Ltd, Chichester, UK, http://dx.doi.org/10.1002/
0470863242.
Buschmann, C., Nagel, E., 1993. In vivo spectroscopy and internal optics of leaves as
basis for remote sensing of vegetation. Int. J. Remote Sens. 14, 711–722.
Cai, W., Yankun, Y., Shao, X., 2008. A variable selection method based on
uninformative variable elimination for multivariate calibration of
near-infrared spectra. Chemometr. Intell. Lab. 90, 188–194.
Casa, R., Varella, H., Buis, S., Guérif, M., De Solan, B., Baret, F., 2012. Forcing a wheat
crop model with LAI data to access agronomic variables: evaluation of the
impact of model and LAI uncertainties and comparison with an empirical
approach. Eur. J. Agron. 37 (1), 1–10, http://dx.doi.org/10.1016/j.eja.2011.09.
004.
Centner, V., Massart, D.-L., de Noord, O.E., de Jong, S., Vandeginste, B.M., Sterna, C.,
1996. Elimination of uninformative variables for multivariate calibration. Anal.
Chem. 68, 3851–3858, http://dx.doi.org/10.1021/ac960321m.
Chen, B., Wu, Z., Wang, J., Dong, J., Guan, L., Chen, J., Yang, K., Xie, G., 2015.
Spatio-temporal prediction of leaf area index of rubber plantation using
HJ-1A/1B CCD images and recurrent neural network. ISPRS J. Photogramm.
Remote Sens. 102, 148–160, http://dx.doi.org/10.1016/j.isprsjprs.2014.12.011.
O. Kira et al. / Agricultural and Forest Meteorology 218–219 (2016) 243–249
Clevers, J.G.P.W., Kooistra, L., 2012. Using hyperspectral remote sensing data for
retrieving canopy chlorophyll and nitrogen content. IEEE J. Sel. Top. Appl. Earth
Obs. Remote Sens. 5, 574–583, http://dx.doi.org/10.1016/j.jag.2012.10.008.
Curran, P.J., Dungan, J.L., Gholz, H.L., 1990. Exploring the relationship between
reflectance red-edge and chlorophyll content in slash pine. Tree Physiol. 7,
33–48.
Curran, P.J., Dungan, J.L., Macler, B.A., Plummer, S.E., 1991. The effect of a red leaf
pigment on the relationship between red edge and chlorophyll concentration.
Remote Sens. Environ. 35, 69–76.
Danson, F.M., Plummer, S.E., 1995. Red-edge response to forest leaf area index. Int.
J. Remote Sens. 16 (1), 183–188.
Darvishzadeh, R., Atzberger, C., Skidmore, A., Schlerf, M., 2011. Mapping grassland
leaf area index with airborne hyperspectral imagery: a comparison study of
statistical approaches and inversion of radiative transfer models. ISPRS J.
Photogramm. Remote Sens. 66 (6), 894–906, http://dx.doi.org/10.1016/j.
isprsjprs.2011.09.013.
Daughtry, C.S.T., Gallo, K.P., Goward, S.N., Prince, S.D., Kustas, W.P., 1992. Spectral
estimates of absorbed radiation and phytomass production in corn and
soybean canopies. Remote Sens. Environ. 39 (2), 141–152, http://dx.doi.org/10.
1016/0034-4257(92)90132-4.
Delegido, J., Verrelst, J., Meza, C.M., River, J.P., Alonso, L., Moreno, J., 2013. A
red-edge spectral index for remote sensing estimation of green LAI over
agroecosystems. Eur. J. Agron. 46, 45–52.
Delegido, J., Verrelst, J., Rivera, J.P., Ruiz-Verdú, A., Moreno, J., 2015. Brown and
green LAI mapping through spectral indices. Int. J. Appl. Earth Obs. Geoinf. 35,
350–358, http://dx.doi.org/10.1016/j.jag.2014.10.001.
Duan, S.-B., Li, Z.-L., Wu, H., Tang, B.-H., Ma, L., Zhao, E., Li, C., 2014. Inversion of the
PROSAIL model to estimate leaf area index of maize, potato, and sunflower
fields from unmanned aerial vehicle hyperspectral data. Int. J. Appl. Earth Obs.
Geoinf. 26, 12–20, http://dx.doi.org/10.1016/j.jag.2013.05.007.
Evans, D.M., Pocock, M.J.O., Memmott, J., 2013. The robustness of a network of
ecological networks to habitat loss. Ecol. Lett. 16 (7), 844–852, http://dx.doi.
org/10.1111/ele.12117.
Fukshansky, L., 1981. Optical properties of plant tissue. In: Smith, H. (Ed.), Plants
and the Daylight Spectrum. Springer, Berlin, pp. 253–303.
Fukshansky, L.A., Remisowsky, A.M., McClendon, J., Ritterbusch, A., Richter, T.,
Mohr, H., 1993. Absorption spectra of leaves corrected for scattering and
distributional error: a radiative transfer and absorption statistics treatment.
Photochem. Photobiol. 57, 538–555, http://dx.doi.org/10.1111/j.1751-1097.
1993.tb02332.x.
Gitelson, A.A., Merzlyak, M.N., 1994. Spectral reflectance changes associated with
autumn senescence of Aesculus hippocastanum L. and Acer platanoides L. leaves.
Spectral features and relation to chlorophyll estimation. J. Plant Physiol. 143,
286–292.
Gitelson, A.A., Viña, A., Arkebauer, T.J., Rundquist, D.C., Keydan, G., Leavitt, B., 2003.
Remote estimation of leaf area index and green leaf biomass in maize canopies.
Geophys. Res. Lett. 30 (5), 1248, http://dx.doi.org/10.1029/2002GL016450.
Gitelson, A.A., 2011a. Non-destructive estimation of foliar pigment (chlorophylls,
carotenoids and anthocyanins) contents: espousing a semi-analytical
three-band model. In: Thenkabail, P.S., Lyon, J.G., Huete, A. (Eds.),
Hyperspectral Remote Sensing of Vegetation. Taylor and Francis, pp. 141–165.
Gitelson, A.A., 2011b. Remote sensing estimation of crop biophysical
characteristics at various scales. In: Thenkabail, P.S., Lyon, J.G., Huete, A. (Eds.),
Hyperspectral Remote Sensing of Vegetation. Taylor and Francis,
pp. 329–358.
Gitelson, A.A., Peng, Y., Masek, J.G., Rundquist, D.C., Verma, S., Suyker, A., Baker,
J.M., Hatfield, J.L., Meyers, T., 2012. Remote estimation of crop gross primary
production with Landsat data. Remote Sens. Environ. 121, 404–414, http://dx.
doi.org/10.1016/j.rse.2012.02.017.
Gonsamo, A., Chen, J.M., 2014. Continuous observation of leaf area index at
Fluxnet-Canada sites. Agric. For. Meteorol. 189-190, 168–174, http://dx.doi.
org/10.1016/j.agrformet.2014.01.016.
Hardwick, S.R., Toumi, R., Pfeifer, M., Turner, E.C., Nilus, R., Ewers, R.M., 2015. The
relationship between leaf area index and microclimate in tropical forest and oil
palm plantation: forest disturbance drives changes in microclimate. Agric. For.
Meteorol. 201, 187–195, http://dx.doi.org/10.1016/j.agrformet.2014.11.010.
Hatfield, J.L., Gitelson, A.A., Schepers, J.S., Walthall, C.L., 2008. Application of
spectral remote sensing for agronomic decisions. Agron. J. 100, S-117–S-131,
http://dx.doi.org/10.2134/agronj2006.0370c.
Haykin, S., 2005. Neural Networks, A Comprehensive Foundation, 2nd ed. Pearson
Education, Inc., New Jersey.
Horler, D.N., Dockray, M., Barber, J., 1983. The red edge of plant leaf reflectance. Int.
J. Remote Sens. 4 (2), 273–288.
Huete, A.R., Liu, H.Q., Batchily, K., Van Leeuwen, W., 1997. A comparison of
vegetation indices over a global set of TM images for EOS-MODIS. Remote Sens.
Env. 59 (3), 440–451.
Jordan, C.F., 1969. Derivation of leaf-area index from quality of light on the forest
floor. Ecology 50 (4), 663–666.
249
Kira, O., Linker, R., Gitelson, A., 2015. Non-destructive estimation of foliar
chlorophyll and carotenoid contents: focus on informative spectral bands. Int.
J. Appl. Earth Obs. Geoinf. 38, 251–260, http://dx.doi.org/10.1016/j.jag.2015.01.
003.
Le Maire, G., François, C., Soudani, K., Berveiller, D., Pontailler, J.Y., Bréda, N., 2008.
Calibration and validation of hyperspectral indices for the estimation of
broadleaved forest leaf chlorophyll content, leaf mass per area, leaf area index
and leaf canopy biomass. Remote Sens. Environ. 112, 3846–3864.
Liang, L., Di, L., Zhang, L., Deng, M., Qin, Z., Zhao, S., Lin, H., 2015. Estimation of crop
LAI using hyperspectral vegetation indices and a hybrid inversion method.
Remote Sens. Environ. 165, 123–134, http://dx.doi.org/10.1016/j.rse.2015.04.
032.
Liu, J., Pattey, E., Jégo, G., 2012. Assessment of vegetation indices for regional crop
green LAI estimation from Landsat images over multiple growing seasons.
Remote Sens. Environ. 123, 347–358, http://dx.doi.org/10.1016/j.rse.2012.04.
002.
Merzlyak, M., Gitelson, A., 1995. Why and what for the leaves are yellow in
autumn? On the interpretation of optical spectra of senescing leaves (Acer
platanoides L.). J. Plant Physiol. 145 (3), 315–320.
Mishara, A.K., Ines, A.V.M., Das, N.N., Khedun, C.P., Singh, V.P., Sivakumar, B.,
Hansen, J.W., 2015. Anatomy of a local-scale drought: application of
assimilated remote sensing products, crop model, and statistical methods to an
agricultural drought study. J. Hydrol. 526, 15–29.
Nguy-Robertson, A.L., Gitelson, A.A., Peng, Y., Viña, A., Arkebauer, T.J., Rundquist,
D.C., 2012. Green leaf area index estimation in maize and soybean: combining
vegetation indices to achieve maximum sensitivity. Agron. J. 104, 1336–1347,
http://dx.doi.org/10.2134/agronj2012.0065.
Nguy-Robertson, A.L., Peng, Y., Gitelson, A.A., Arkebauer, T.J., Pimstein, A.,
Herrmann, I., Karnieli, A., Rundquist, D.C., Bonfil, D.J., 2014. Estimating green
LAI in four crops: potential for determining optimal spectral bands for a
universal algorithm. Agric. For. Meteorol. 192–193, 140–148, http://dx.doi.org/
10.1016/j.agrformet.2014.03.004.
Rouse, J.W., Haas Jr., R.H., Schell, J.A., Deering, D.W., 1974. Monitoring vegetation
systems in the Great Plains with ERTS. In: NASA SP-351, Third ERTS-1
Symposium NASA, Washington DC. 1, pp. 309–317.
Rundquist, D., Perk, R., Leavitt, B., Keydan, G., Gitelson, A., 2004. Collecting spectral
data over cropland vegetation using machine-positioning versus
hand-positioning of the sensor. Comput. Electron. Agric. 43 (2), 173–178.
Rundquist, D.C., Gitelson, A.A., Leavitt, B., Zygielbaum, A., Perk, R., Keydan, G.P.,
2014. Elements of an integrated phenotyping system for monitoring crop
status at canopy level. Agron 4 (1), 108–123, http://dx.doi.org/10.3390/
agronomy4010108.
Sakamoto, T., Gitelson, A.A., Arkebauer, T.J., 2014. Near real-time prediction of U.S.
corn yields based on time-series MODIS data. Remote Sens. Environ. 147,
219–231.
Siegmann, B., Jarmer, T., 2015. Comparison of different regression models and
validation techniques for the assessment of wheat leaf area index from
hyperspectral data. Int. J. Remote Sens. 36 (18), 4519–4534, http://dx.doi.org/
10.1080/01431161.2015.1084438.
Simon, H., Baker, K.R., Phillips, S., 2012. Compilation and interpretation of
photochemical model performance statistics published between 2006 and
2012. Atmos. Environ. 61, 124–139.
Tucker, C.J., 1977. Asymptotic nature of grass canopy spectral reflectance. Appl.
Opt. 16 (5), 1151–1156, http://dx.doi.org/10.1364/AO.16.001151.
Tucker, C.J., 1979. Red and photographic infrared linear combinations for
monitoring vegetation. Remote Sens. Env. 8 (2), 127–150.
Ustin, S.L., Gitelson, A.A., Jacquemoud, S., Schaepman, M., Asner, G.P., Gamon, J.,
Zarco-Tejada, A.P., 2009. Retrieval of foliar information about plant pigment
systems from high resolution spectroscopy. Remote Sens. Environ. 113,
S67–S77, http://dx.doi.org/10.1016/j.rse.2008.10.019.
Verma, S.B., Dobermann, A., Cassman, K.G., Walters, D.T., Knops, J.M., Arkebauer,
T.J., Suyker, A.E., Burba, G.G., Amos, B., Yang, H., Ginting, D., Hubbard, K.G.,
Gitelson, A.A., Walter-Shea, E.A., 2005. Annual carbon dioxide exchange in
irrigated and rainfed maize-based agroecosystems. Agric. For. Meteorol. 131,
77–96, http://dx.doi.org/10.1016/j.agrformet.2005.05.003.
Verrelst, J., Camps-Valls, G., Muñoz-Mari, J., Rivera, J.P., Veroustraete, F., Clevers,
G.P.W., Moreno, J.J., 2015. Optical remote sensing and the retrieval of terrestrial
vegetation bio-geophysical properties – a review. ISPRS J. Photogramm.
Remote Sens. 109, http://dx.doi.org/10.1016/j.isprsjprs.2015.05.005.
Viña, A., Gitelson, A.A., Nguy-Robertson, A.L., Peng, Y., 2011. Comparison of
different vegetation indices for the remote assessment of green leaf area index
of crops. Remote Sens. Environ. 115, 3468–3478, http://dx.doi.org/10.1016/j.
rse.2011.08.010.
Vogelmann, T.C., Rock, B.N., Moss, D.M., 1993. Red edge spectral measurements
from sugar maple leaves. Int. J. Remote Sens. 14, 1563–1575.
Watson, D.J., 1947. Comparative physiological studies on the growth of field crops:
I. variation in net assimilation rate and leaf area between species and varieties,
and within and between years. Ann. Bot. 11 (1), 41–76.