ArticlePDF Available

Abstract and Figures

Accurate estimation of leafchlorophyll content (Cab) from remote sensing is of tremendous significance to mon- itor the physiological status ofvegetation or to estimate primary production. Many vegetation indices (VIs) have been developed to retrieve Cab at the canopy level from meter- to decameter-scale reflectance observations. However, most of these VIs may be affected by the possible confounding influence of canopy structure. The objective of this study is to develop methods for Cab estimation using millimeter to centimeter spatial resolution reflectance imagery acquired at the field level. Hyperspectral images were acquired over sugar beet canopies from a ground-based platform in the 400– 1000 nmrange, concurrently to Cab, green fraction (GF), green area index (GAI) ground measurements. The orig- inal image spatial resolution was successively degraded from 1 mm to 35 cm, resulting in eleven sets of hyperspectral images. Vegetation and soil pixels were discriminated, and for each spatial resolution, measured Cab values were related to various VIs computed over four sets of reflectance spectra extracted from the images (soil and vegetation pixels, only vegetation pixels, 50% darkest and brightest vegetation pixels). The selected VIs included some classical VIs from the literature as well as optimal combinations of spectral bands, including sim- ple ratio (SR), modified normalized difference (mND) and structure insensitive pigment index (SIPI). In the case ofmND and SIPI, the use ofa blue reference band instead of the classical near-infrared one was also investigated. For the eleven spatial resolutions, the four pixel selections and the five VI formats, similar band combinations are obtainedwhen optimizing VI performances: themain bands ofinterest are generally located in the blue, red, red- edge and near-infrareddomains. Overall, mNDblue[728,850] defined as (R440−R728)/(R440+R850) and computed over the brightest green pixels obtains the best correlations with Cab for spatial resolutions finer than 8.8 cmwith a root mean square error of prediction better than 2.6 μg/cm2. Conversely, mNDblue[728,850] poorly correlates with variations in GF and GAI, thus reducing the risk of deriving non-causal relationships with Cab that would actually be due to the covariance between Cab and these canopy structure variables. As mNDblue[728,850] can be calculated frommost current multispectral sensors, it is therefore a promising VI to retrieve Cab frommillime- ter- to centimeter-scale reflectance imagery.
Content may be subject to copyright.
1
Estimating leaf chlorophyll content in sugar beet canopies
1
using millimeter- to centimeter-scale reflectance imagery
2
3
Sylvain Jaya,b,, Nathalie Gorrettaa, Julien Morela, Fabienne Maupasc, Ryad Bendoulaa, Gilles
4
Rabatela, Dan Dutartrec, Alexis Comard, Frédéric Barete
5
a Irstea, UMR ITAP, 361 rue J.F. Breton, 34196 Montpellier, France
6
b Aix Marseille Univ, CNRS, Centrale Marseille, Institut Fresnel, F-13013 Marseille, France
7
c Institut Technique de la Betterave, 45 rue de Naples, 75008 Paris, France
8
d HIPHEN SAS, 22b rue Charrue, 84000 Avignon, France
9
e INRA UMR 114 EMMAH, UMT CAPTE, Domaine Saint-Paul, Site Agroparc, F-84914 Avignon, France
10
Abstract
11
Accurate estimation of leaf chlorophyll content (Cab) from remote sensing is of tremendous significance
12
to monitor the physiological status of vegetation or to estimate primary production. Many vegetation
13
indices (VIs) have been developed to retrieve Cab at the canopy level from meter- to decameter-scale
14
reflectance observations. However, most of these VIs may be affected by the possible confounding
15
influence of canopy structure. The objective of this study is to develop methods for Cab estimation
16
using millimeter to centimeter spatial resolution reflectance imagery acquired at the field level.
17
Hyperspectral images were acquired over sugar beet canopies from a ground-based platform in the
18
400-1000 nm range, concurrently to Cab, green fraction (GF), green area index (GAI) ground
19
measurements. The original image spatial resolution was successively degraded from 1 mm to 35 cm,
20
resulting in eleven sets of hyperspectral images. Vegetation and soil pixels were discriminated, and
21
for each spatial resolution, measured Cab values were related to various VIs computed over four sets
22
of reflectance spectra extracted from the images (soil and vegetation pixels, only vegetation pixels,
23
50% darkest and brightest vegetation pixels). The selected VIs included some classical VIs from the
24
literature as well as optimal combinations of spectral bands, including simple ratio (), modified
25
2
normalized difference () and structure insensitive pigment index (). In the case of  and
26
, the use of a blue reference band instead of the classical near-infrared one was also investigated.
27
For the eleven spatial resolutions, the four pixel selections and the five VI formats, similar band
28
combinations are obtained when optimizing VI performances: the main bands of interest are generally
29
located in the blue, red, red-edge and near-infrared domains. Overall,  defined as
30
      and computed over the brightest green pixels obtains the best
31
correlations with Cab for spatial resolutions finer than 8.8 cm with a root mean square error of
32
prediction better than 2.6 µg/cm². Conversely, poorly correlates with variations
33
in GF and GAI, thus reducing the risk of deriving non-causal relationships with Cab that would actually
34
be due to the covariance between Cab and these canopy structure variables. As  
35
can be calculated from most current multispectral sensors, it is therefore a promising VI to retrieve Cab
36
from millimeter- to centimeter-scale reflectance imagery.
37
Keywords : Leaf chlorophyll content, Millimeter to centimeter spatial resolutions, mNDblue, Reflectance
38
imagery, Vegetation index.
39
1. Introduction
40
Photosynthesis is one of the most important biological processes, allowing life on Earth through
41
production of oxygen and organic matter (Ustin et al., 2009). Chlorophyll is one of the major plant
42
pigments that contribute to the absorption of photosynthetically active radiation. Quantifying
43
chlorophyll temporal dynamics is therefore critical to monitor the vegetation physiological status or to
44
estimate primary production (Blackburn, 2007, 1998). For this purpose, non-destructive estimation of
45
leaf chlorophyll content (denoted Cab hereafter) based on optical measurements has proven to be
46
effective since Cab drives most of the leaf reflectance and transmittance variabilities in the visible
47
domain. A high Cab retrieval accuracy is usually obtained at the leaf scale under controlled experimental
48
conditions, e.g., using dedicated leaf clips measuring transmittance at a few wavelengths (Cerovic et
49
3
al., 2012), or using hemispherical reflectance and/or transmittance measurements to invert physical
50
models such as PROSPECT (Jacquemoud and Baret, 1990) or to apply spectral indices (Gitelson et al.,
51
2003; Féret et al., 2011; le Maire et al., 2004; ). The estimation of Cab is more challenging at the canopy
52
scale: soil reflectance and canopy architecture interact with leaf scattering properties to generate
53
canopy reflectance. As a consequence, the effect of leaf composition may be confounded with those
54
of canopy structural properties, making the inversion of canopy reflectance models an ill-posed
55
problem (Baret and Buis, 2008; Combal et al., 2003): several combinations of green area index (GAI)
56
and Cab values may indeed correspond to similar canopy reflectance spectra in the visible domain,
57
which increases the uncertainty of Cab retrieval (Baret and Buis, 2008). Further, non-causal
58
relationships between canopy reflectance and the targeted variable may be observed when structural
59
and biochemical variables are correlated as reported by Knyazikhin et al. (2013). Effects of canopy
60
structure and leaf composition should therefore be disentangled with great care when relating
61
remote-sensing observations to foliar biochemistry (Knyazikhin et al., 2013; Latorre-Carmona et al.,
62
2014; Ustin, 2013).
63
A first approach has been proposed to improve the Cab estimation performance at the canopy
64
level by maximizing the spectral sensitivity to foliar biochemistry while minimizing the effects of soil
65
and vegetation structure. It consists in using a ratio vegetation index (VI), where the numerator is a
66
Cab-sensitive VI such the Modified Chlorophyll Absorption Reflectance Index (MCARI) (Daughtry et al.,
67
2000), and the denominator is a VI sensitive to canopy structure such as the Optimized Soil-Adjusted
68
Vegetation Index (OSAVI) (Rondeaux et al., 1996). MCARI/OSAVI (Daughtry et al., 2000), TCARI/OSAVI
69
(Haboudane et al., 2002) and derived versions of these two VIs (Wu et al., 2008) are examples of such
70
combined indices, which have been demonstrated to provide accurate Cab estimation results at the
71
canopy level (Kooistra and Clevers, 2016).
72
Alternatively, a second approach consists in increasing the sensitivity to foliar biochemistry by
73
optimizing the sun-sensor geometry: off-nadir measurements are generally more sensitive to leaf
74
4
properties than nadir measurements (Baret et al., 2010; Comar et al., 2012; Dorigo, 2012; Jacquemoud
75
et al., 2009; Jay et al., n.d.). This is not only due to the higher proportion of vegetation seen by the
76
sensor, but also to the large fraction of photons that have interacted with leaves before reaching the
77
sensor (Jacquemoud et al., 2009). Further, the relative viewing azimuth angle affects the canopy
78
reflectance sensitivity: measurements acquired in the backward direction, where shadows are
79
minimized, generally exhibit a higher sensitivity to leaf biochemistry (Dorigo, 2012; Jacquemoud et al.,
80
2009; Jay et al., n.d.).
81
Finally, a third approach consists in focusing on the illuminated vegetation pixels when the spatial
82
resolution is sufficient: this limits the detrimental influences of soil and canopy architecture and
83
consequently strengthens the sensitivity to Cab (Moorthy et al., 2008; Zarco-Tejada et al., 2004, 2001).
84
Multi- and hyperspectral cameras operated from ground-based or low-altitude platforms provide
85
a very high spatial resolution, ranging from a few millimeters to a few decimeters. However, most
86
current retrieval methods do not fully exploit the new possibilities offered by such high spatial
87
resolution imagery, thereby stimulating the need for new algorithms (Elarab et al., 2015; Houborg et
88
al., 2015). The large variability of leaf orientation and illumination conditions observed at this scale
89
induces strong variations in leaf radiance. For example, Jay et al. (2016) have proposed to invert the
90
PROSPECT+COSINE (ClOse-range Spectral ImagiNg of lEaves) model to map Cab over individual leaves
91
when the influence of surrounding elements is negligible. However, when individual leaves are
92
submitted to the radiative transfer conditions that prevail in the canopy, the problem was not yet
93
addressed. Most current VIs have been designed for leaf and canopy levels, and may therefore be
94
suboptimal to handle the above-mentioned variations in leaf reflectance (Bånkestad and Wik, 2016).
95
This study focuses on Cab estimation in sugar beet canopies using millimeter- to centimeter-resolution
96
reflectance imagery. Hyperspectral images were acquired from a ground-based platform and
97
concurrent measurements of GAI and Cab were completed. These data were used to design VIs
98
dedicated to Cab estimation that take advantage of such high spatial resolution imagery. Performances
99
5
were compared to those obtained with several VIs of the literature for the range of spatial resolutions
100
investigated.
101
2. Materials and methods
102
2.1. Field experiments
103
Figure 1: Locations of the three study sites and photographs of sugar beet canopies illustrating the
encountered soil types (loamy soil for site 1 and chalky soil for sites 2 and 3).
104
Field experiments were conducted in France in 2015 and 2016. Three study sites with different soil
105
properties were considered as illustrated in Fig. 1. A chalky soil was present at the ”Vaucogne”
106
(431′N, 4°21′E, denoted site 2) and ”Viapres” (48°35′N, 4°2′E, denoted site 3) sites, while the La
107
Selve” site (435’N, 01E, denoted site 1) was characterized by a loamy soil. The details of these
108
field experiments are summarized in Table 1. Seven sugar beet cultivars exhibiting differences in plant
109
structure were submitted to variable levels of nitrogen fertilization. Rows were spaced 45 cm apart
110
and plant population density was between 10 to 12 plants per square meter. So as to further increase
111
the representativeness and heterogeneity of the data set, various phenological stages were considered
112
during the 2015 and 2016 growing seasons, i.e., on June, 2-3 2015, June, 23-24 2015 and July, 26-27
113
6
2016. In particular, the crops considered in the 2016 experiment were carefully chosen so as to
114
decorrelate Cab and canopy structural properties. In total, the overall data set included 55 samples and
115
encompassed a large variability due to differences in cultivars, nitrogen fertilizations, development
116
stages, and soil and weather conditions.
117
118
Table 1: Characteristics of field experiments. s is the sun zenith angle at the time of measurements.
Date
Site
s (°)
Illumination
Cultivar No.
06/02/2015
2
36
Clear
1-3
06/03/2015
1
31
Cloudy
1-3
06/23/2015
1
29
Partly cloudy
1-3
06/24/2015
2
29
Clear
1-3
07/26/2016
3
33
Clear
4-6
07/27/2016
3
33
Partly cloudy
7
119
2.2. Reflectance measurements
120
Figure 2: Ground-based platform used for hyperspectral measurements
121
For each plot, an area corresponding to five consecutive plants along a row was imaged using a HySpex
122
VNIR-1600 hyperspectral camera (Norsk Elektro Optikk, Norway) set up on a ground-based platform
123
7
as shown in Fig. 2. The push-broom camera pointed vertically downward from a 1.15 m distance to the
124
bare soil. It measured the reflected radiation in 160 spectral bands ranging from 415 to 994 nm with
125
a 3.7 nm spectral sampling interval and 4.5 nm full width at half maximum, and acquired successive
126
scans of 1600 pixels along the row. The across-track field of view (FOV) was about 35 cm per scan at
127
the ground level, providing a 0.02 cm across-track sampling distance. A 40% diffuse reflectance
128
reference panel (Spectralon®, Labsphere) was used to measure the incoming solar irradiance while
129
limiting possible saturation of the sensor. The reference panel was placed horizontally above the
130
canopy to reduce the influence of possible vicinity effects. The HDRF (Hemispherical-directional
131
Reflectance Factor) was finally computed by dividing the signal measured for each band and each pixel
132
over the target by that measured over the reference panel and multiplying it by the reflectance of the
133
reference panel provided by Labsphere (assuming the panel to be Lambertian). Completion of the
134
scans over the 5 plants took a few seconds during which the incoming radiation was supposed to be
135
stable. Measurements were collected around solar noon with solar zenith angle always lower than 36°.
136
Illumination conditions differed between experiments, ranging from a clear blue sky to a fully overcast
137
sky (Table 1).
138
2.3. Cab and canopy structure measurements
139
The leaf chlorophyll content was estimated for each plot after image acquisition over the same five
140
plants. Six measurements per plant were made using a Dualex scientific+TM (Force-A, Orsay, France).
141
This leafclip measures leaf transmittance in a few wavebands from which Cab is estimated using the
142
relationship proposed by Cerovic et al. (2012) for dicotyledons, achieving an accuracy of around
143
4 µg.cm2. Measurements we re performed at different leaf levels to better consider the possible
144
Cab vertical gradient between leaves of different age or differently located in the canopy. These thirty
145
Cab values were averaged to provide a single Cab value per plot. After Cab measurements, the five plants
146
were collected and the area of each individual leaf measured using a photography-based technique.
147
The GAI was finally obtained by multiplying the average leaf area per plant by the plant population
148
8
density. In addition, the green fraction (GF, the fraction of green elements seen by the sensor in its
149
view direction) was estimated from hyperspectral images using the discrimination method detailed in
150
Section 2.4.2.
151
Inspection of the co-distributions between Cab and GAI, and between Cab and GF shows very poor
152
correlations in both cases (Fig. 3). Incidentally, note the importance of the 2016 data that enable these
153
correlations to be significantly reduced. The independency between Cab and GAI (resp. GF) distributions
154
prevents from obtaining spurious empirical relationships between VIs and Cab that may be inherited
155
from a more causal relationship between VIs and GAI (resp. GF).
156
Figure 3: Cab, GAI and GF reference measurements.
157
158
9
2.4. Processing of radiometric data
159
2.4.1. Image spatial subsampling
160
The performance of Cab estimation was evaluated over a range of spatial resolutions. The original
161
spatial resolution of 0.02 cm was degraded to 0.1, 0.2, 0.4, 0.9, 1.8, 3.5, 4.4, 7, 8.8, 17.5 and 35 cm
162
by averaging over cells of NxN pixels. This resulted in eleven sets of hyperspectral images as
163
illustrated in Fig. 4.
164
Figure 4: Gradual degradation of spatial resolution. Illustration for (a) 0.1 cm, (b) 1.8 cm, (c) 4.4 cm, (d) 8.8 cm,
(e) 17.5 cm and (f) 35 cm.
165
2.4.2. Discrimination of soil and vegetation
166
VIs were not only used to estimate Cab, but also to discriminate green vegetation from senescent
167
elements and soil. Preliminary tests (not further developed here for the sake of brevity, but illustrated
168
in the figure provided in supplementary material) demonstrated that the modified version of MCARI
169
10
proposed by Wu et al. (2008) and defined as    
170

 provided the best discrimination performance as compared to the other VIs investigated
171
(Table 2). MCARI[705,750] is designed to minimize the effects of soil and non-photosynthetic
172
materials, and its discrimination capacity appeared to be little dependent on soil properties and
173
illumination conditions. Further, as illustrated in Fig. 5b and Fig. 5g for strongly different soil properties
174
and illumination conditions, a single threshold value of   ensured accurate
175
discrimination performance at the 0.1 cm spatial resolution, for which the fraction of mixed pixels was
176
negligible (note that GF was estimated at this resolution). The fraction of mixed pixels significantly
177
increased when the resolution degraded, making the classification more difficult. The
178
MCARI[705,750] threshold value was thus adjusted for each resolution investigated to keep the GF
179
similar to that computed with the 0.1 cm spatial resolution. The threshold value was therefore
180
increased as the spatial resolution degraded in order to compensate for the increase in the number of
181
mixed pixels (Fig. 5).
182
Figure 5: Discrimination results obtained from two contrasted situations, i.e., low GF with chalky soil and under
sunny conditions (a), and large GF with loamy soil and under cloudy conditions (f). The same MCARI[705,750]
threshold value is used for the finest (0.1 cm) spatial resolution (b,g). This threshold is then adjusted for
11
coarser resolutions to keep the same GF (e). The resulting discrimination results are presented for 3.5 (c,h)
and 7 cm (d,i) spatial resolutions. Masked soil pixels are colored in cyan.
183
2.5. Cab estimation
184
2.5.1. Selection of vegetation indices
185
Several VIs were selected from the literature to estimate Cab. All of them are ratios based on two or
186
three wavebands located in the 400-900 nm spectral domain (Table 2). They are potentially sensitive
187
to Cab, GF, and GAI. VIs computed as ratios of linear combinations of bands present the advantage to
188
minimize the possible influence of multiplicative factors, including slope effects and variations in
189
illumination conditions when the radiance measurements on the reference panel and on the target
190
are not performed concurrently. These ratio VIs may be split into three different categories as
191
presented in Table 2.
192
The simplest one, SR (Simple Ratio), corresponds to the ratio of reflectance in two wavebands: a band,
193
hereafter denoted , that is sensitive to both absorption by chlorophyll and scattering by leaf, and a
194
reference band, denoted , that is only sensitive to scattering in order to correct from this effect
195
(Blackburn, 2007). For example,  and  (Gitelson et al., 2006, 2005, 2003) are based
196
on simple ratios, and combine a near-infrared reference band with a band located in medium
197
chlorophyll absorption domains to avoid saturation in Cab. Both  and  have been
198
found to provide accurate estimates of Cab at the leaf level (Gitelson and Merzlyak, 1994; Gitelson et
199
al., 2003, 2006; Schlemmer et al., 2013) and canopy chlorophyll content at the canopy level (Gitelson
200
et al., 2005; Clevers and Kooistra, 2012; Clevers and Gitelson, 2013; Schlemmer et al., 2013).
201
VIs from the second category are based on modified Normalized Difference ratios (), for which a
202
third waveband, denoted , can be introduced at the denominator. Note that SR and  are
203
functionally related when  , since      . Alternatively, taking  
204
may increase the sensitivity to Cab while reducing the impact of canopy structure and soil background
205
12
properties.  (Gitelson et al., 1996),  (Gitelson and Merzlyak, 1994) and  (Rouse et
206
al., 1973) are examples of  indices.
207
A simple modification of  indices, i.e.,  (Structure Insensitive Pigment Index) was proposed
208
by Rondeaux and Vanderbilt (1993) and Penuelas et al. (1995) to decrease the confounding influence
209
of leaf surface as well as canopy structure effects (Bousquet et al., 2005; Vigneau et al., 2011; Comar
210
et al., 2014; Jay et al., 2016). For -like indices, various reference bands have been proposed: for
211
example, Sims and Gamon (2002) have demonstrated that using a blue saturating waveband as
212
reference within  (modified Simple Ratio) may increase the Cab estimation performance as
213
compared to  indices.  (MERIS Terrestrial Chlorophyll Index) proposed by Dash and Curran
214
(2004) is also a SIPI-like index, where the three bands used are conveniently located in the red-edge
215
spectral domain, the reference band (709 nm) being close to the red one.
216
Because of the diversity of bands used to compute the three types of VIs as outlined in Table 2, the
217
data set presented in Section 2 was used to find optimal sets of wavebands in the case of millimeter-
218
to centimeter-scale reflectance imagery of sugar beet canopies. Similarly to Penuelas et al. (1995) and
219
Inoue et al. (2012), Cab prediction performance of these three VIs were evaluated for every possible
220
combination of and bands between 415 and 900 nm by 3.7 nm step. Cab prediction performances
221
were assessed using the Spearman’s rank correlation coefficient, denoted. As compared to using the
222
usual Pearson’s coefficient, , using reduces the influence of possible non-linearities between VIs
223
and Cab as well as between Dualex readings and actual Cab values. In the case of  that uses only two
224
wavebands, the reference band was also varied systematically from 415 to 900 nm, and  indices
225
were thus noted  . In the cases of  and  indices, as the reference waveband should
226
be insensitive to Cab variations over the considered Cab range, it was set either to near infrared (850
227
nm) or to blue (440 nm). The corresponding VIs were thus noted respectively   and
228
  if     nm, and   and   if   
229
 nm. Note that, although the atmosphere has lesser influence on the measured signal when using
230
13
low altitude sensors (e.g., ground- or tower-based, or embedded on unmanned aerial vehicles (UAVs))
231
instead of classical satellite- and airborne sensors, the reflected radiation in the blue domain may
232
strongly vary with changes in illumination conditions caused by atmospheric absorption and scattering.
233
This implies that the use of   and   requires a particular attention to
234
properly convert the measured radiance into reflectance. Also, if necessary due to the lower sensitivity
235
of CCD sensors in the blue region, the signal-to-noise ratio may be increased by aggregating a few
236
wavebands around 440 nm.
237
Table 2: Vegetation indices selected from the literature and their generic formulation.
Generic
VI name
VI
formulation

Actual VI name
References


780
550
-
   
Gitelson et al. (2005,
2003, 2006)
780
710
-
   

  
  
800
670
670

Rouse et al. (1973)
750
550
550

Gitelson et al. (1996)
750
705
705

Gitelson and Merzlyak
(1994)

  
  
850
445
680

Penuelas et al. (1995)
445
750
705

Sims and Gamon (2002)
709
754
681

Dash and Curran (2004)
238
2.5.2. Estimation procedure
239
Every tested VI was related to the measured Cab values using the 55 images, the strength of the
240
relationship being quantified based on the squared Spearman’s correlation coefficient, . This process
241
was applied for each spatial resolution using four subsets of reflectance spectra. (1) For the first subset,
242
all pixels were used, i.e., including both soil and vegetation parts. (2) In the second subset, only green
243
14
pixels were used. These vegetation pixels were then sorted according to their brightness level
244
computed as the average reflectance value in the 770-900 nm spectral domain. This allowed us to
245
define the two last subsets, corresponding to (3) the 50% darkest green pixels, and (4) the 50%
246
brightest green pixels. Such pixel selections influence the impact of canopy structure on VIs as reported
247
by Zarco-Tejada et al. (2001). Furthermore, two strategies were considered to compute the VI average
248
value over selected pixels: either (1) reflectance spectra were first averaged over all pixels of the subset
249
and the VI was then calculated, or (2) the VI was first computed for each pixel of the subset and the
250
resulting VI values were then averaged. As VIs are generally non-linear functions of reflectance, these
251
two strategies may lead to different results if the images are heterogeneous (Steven et al., 2015). The
252
performance of Cab estimation were thus evaluated for the eleven spatial resolutions, the four subsets
253
of pixels and the two VI averaging strategies, resulting into 88 regressions for each VI. Ultimately, linear
254
and best non-linear relationships between Cab and best VIs were determined. Prediction performances
255
for each relationship were then quantified based on the coefficient of determination (R²) and the root
256
mean square error of prediction (RMSEP), both being estimated using a leave-one-out cross-validation
257
process because of the relatively small number of images available.
258
In addition, the same procedure was applied for GF and GAI estimations since the considered VIs are
259
also potentially sensitive to these structural variables.
260
261
3. Results and discussion
262
In the following, the selection of optimal sets of wavebands to compute the VIs is first investigated.
263
These optimized VIs are then compared to the classical ones presented in Table 2 based on their
264
relationships with Cab, GF and GAI, especially emphasizing the effects of spatial resolution and pixel
265
selections.
266
3.1. Optimal band selection for Cab estimation from vegetation indices
267
15
Because of the multidimensional aspect of this study and the associated complexity for reporting the
268
results in an exhaustive way, emphasis is put on 0.9 cm and 17.5 cm spatial resolutions that illustrate
269
two contrasted situations. Further, only the results obtained by computing VIs from average
270
reflectance spectra of vegetation pixels are presented. Very similar results were observed for the other
271
spatial resolutions, subsets of pixels and when computing VIs by averaging pixel-level VI values. For
272
the sake of brevity, these results are not presented in this article.
273
For every VI, similar patterns of squared Spearman’s correlation coefficient are observed for the
274
two spatial resolutions investigated (Fig. 6). The maximum values are obtained for very similar
275
combinations of  wavebands. However, the 0.9 cm spatial resolution provides generally higher
276
values as well as slightly broader patterns of high correlations as compared to the 17.5 cm
277
resolution. The largest difference between the two spatial resolutions occurs for the  in the
278
yellow to red domains where soil and vegetation reflectances show the largest contrast. In this case,
279
the higher proportion of mixed pixels significantly degrades the correlations with Cab. Symmetrical
280
patterns are observed for  and  indices, since     and  
281
 . Conversely, non-symmetrical patterns are observed for  indices that do not verify
282
this property.
283
For  , the best performances are obtained taking in the red-edge (between 710 and 735
284
nm) and having a longer wavelength. The highest correlation is observed for , which
285
is close to the  proposed by Gitelson et al. (2005, 2003, 2006). Selecting a red-edge band
286
actually increases the sensitivity to Cab variation for high Cab values, i.e., it minimizes the saturation
287
effect. For  ,the best performances are obtained choosing λ1 in the red-edge and λ2 in
288
the red-edge and near infrared domains, i.e.,   . For such waveband combinations, we have
289
     . A convergence is thus reached between these two
290
generic VIs when optimizing the set of wavebands for Cab estimation. Anyway, the highest correlation
291
16
Figure 6: Squared Spearman’s correlation () obtained between VIs and Cab as a function of and
wavebands. VIs are computed from average reflectance spectra of vegetation pixels for the 0.9 cm (left-
hand column) and 17.5 cm (right-hand column) spatial resolutions. The five generic VIs are considered,
i.e.,  ,  ,  ,  , and   from top to bottom.
The color scale is the same for the five VIs and the two spatial resolutions.
17
is obtained for . For  , the best performances are observed for in
292
the red-edge and in the green to red domains, the optimal combination being .
293
This VI is similar to the original SIPI (Penuelas et al., 1995) with the exception of the use of a red-edge
294
band at the numerator instead of a blue band, which reduces possible saturation effects for the
295
considered Cab range. Note that no bands are selected in the near-infrared range since the reference
296
band is already in this domain. For  , the best performances are observed for in the
297
red-edge and in the red-edge and near-infrared domains. Note that taking in the green range
298
(around 533 nm) also provides strong correlations, especially at 0.9 cm spatial resolution. In addition
299
to the blue band used as a reference, the bands showing the largest contrast in chlorophyll absorption
300
coefficients are selected. Unlike the other VIs tested, the best Cab-sensitive waveband slightly differs
301
between the two spatial resolutions, ranging from 728 nm at 0.9 cm to 717 nm at 17.5 cm. Since the
302
best performances are obtained for  nm,  is selected for further analysis.
303
Finally, for  , the maximum correlations are obtained taking in the red-edge and
304
in the NIR plateau for the same reasons as those involved for . The optimal waveband
305
combination is , whose expression mainly differs from the  index proposed by
306
Sims and Gamon (2002) by the use of a longer Cab-sensitive waveband.
307
3.2. Sensitivity to Cab
308
The VI sensitivity to Cab (in terms of ) is shown in Fig. 7 for the classical VIs presented in Table 2 as
309
well as the five optimized VIs designed in Section 3.1. These VIs are computed for the eleven spatial
310
resolutions, the four subsets of pixels and the two strategies to compute the VIs.
311
18
Figure 7: Squared Spearman’s correlation () between VIs and Cab measured over the 55 plots as a function
of spatial resolution. In the left-hand column, VIs are computed from the average reflectance spectra of
considered pixels (strategy (1)), while in the right -hand column, VIs are computed by averaging pixel-level
VI values (strategy (2)). For both strategies, the four subsets of pixels are tested (vegetation and soil pixels,
vegetation pixels, 50% darkest and brightest vegetation pixels).
312
313
19
Comparing the two strategies for computing VI values from images shows that the differences mainly
314
depend on the level of heterogeneity within the considered pixels. The heterogeneity increases as the
315
spatial resolution increases and the pixel selection becomes less restrictive (in order, 50% darkest or
316
brightest green pixels, green pixels, all the pixels). When the level of heterogeneity is the largest, as
317
when considering all the pixels in the case of high resolution (Fig. 7, line 1), averaging first the
318
reflectance values over the pixels and then computing the VI (Strategy 1) provides better Cab estimation
319
performances than Strategy 2 (averaging the VI values computed for each individual pixel). This may
320
be due to the fact that averaging the reflectances put more weight on the brightest pixels that bear
321
more reliable information on Cab as demonstrated later. Conversely, computing first the VIs at the pixel
322
level may lead to unrealistic VI values when the pixels have low reflectance values: since VIs are
323
computed as ratios, low values in the denominator will provide unstable values. Note that, since the
324
average reflectance value in the image does not depend on spatial resolution, the performances do
325
not change with spatial resolution when computing VIs from the average reflectance spectra over all
326
the pixels (Fig. 7, top left-hand plot). When the distribution of pixel values is reduced as in the case of
327
medium to coarse spatial resolution for a restricted selection of pixels (green pixels, 50% darkest or
328
brightest green pixels), the two strategies lead to very similar results (Fig. 7, lines 2-4).
329
Only Strategy 1 will therefore be considered in the following since it provides either best or equal Cab
330
estimation performances as Strategy 2.
331
When all the pixels are considered (Fig. 7, line 1), the best-performing VI is  (
332
), followed by  and  ( ). A substantial improvement is observed
333
for every VI except  when considering only vegetation pixels, especially for spatial resolutions
334
finer than 8.8 cm.   obtains significantly higher correlations than other VIs with
335
ranging from 0.84 to 0.87 for resolutions finer than 8.8 cm. Again, the  VIs show good correlations,
336
with ,  and  achieving close to 0.74 for the finest
337
resolutions. Incidentally, it is worth noting that, even if  was originally designed for estimating
338
20
canopy chlorophyll content from low-resolution satellite sensors (Dash and Curran, 2004), the good
339
performances obtained here are consistent with those observed in previous studies dealing with leaf
340
chlorophyll content estimation from higher resolution remote sensing (Haboudane et al., 2008; Hunt
341
et al., 2012; Jay et al., n.d.). While the performances obtained with most VIs remain nearly stable when
342
the resolution degrades, those obtained with  drop down to   at 35 cm
343
spatial resolution. This indicates that   is very sensitive to the soil influence
344
observed in mixed pixels. In the case of coarse spatial resolutions for which vegetation and soil cannot
345
be accurately discriminated, using a shorter wavelength more sensitive to Cab for is expected to
346
improve the performances by mitigating the soil influence as discussed earlier (Fig. 6).
347
Considering the 50% brightest green pixels (Fig. 7, line 4) generally improves the performances of Cab
348
estimation for every resolution as compared to using all the green pixels (Fig. 7, line 2). In particular,
349
  reaches    for resolutions finer than 4.4 cm, while  and
350
 led to   and   respectively. The benefit of considering only the
351
brightest green pixels for Cab estimation is in agreement with previous findings (Moorthy et al., 2008;
352
Zarco-Tejada et al., 2004, 2001). Conversely, using the 50% darkest pixels results in lower correlations,
353
especially for the highest spatial resolutions. For those pixels, the incoming radiation contains a higher
354
proportion of photons that have already interacted with the canopy before reaching the considered
355
leaves. This may make the illumination conditions locally very variable, leading to lower correlations
356
with Cab.
357
358
21
Figure 8: Relationships between Cab and (a) , (b) , (c) ,
and (d) . The spatial resolution is 3.5 cm, and VIs are computed from average reflectance spectra of
the 50% brightest vegetation pixels. For each VI, the prediction performances obtained using linear (in black)
and best non-linear (second-degree polynomials, in red) regressions are shown.
359
The best cases are analyzed in more detail to better quantify their performances (Fig. 8). They
360
correspond to  , ,  and  computed from
361
the average reflectance spectra of the 50% brightest green pixels observed at the 3.5 cm spatial
362
resolution.  achieves the best performances with   and   
363
µg/cm² (i.e., 9.6 % of Cab range) using a linear regression model, while the non-linear regression model
364
does not improve the results. The three other best VIs lead to significantly lower estimation
365
performances, ranging from   and    µg/cm² (13.4 %) for  to
366
  and    µg/cm² (14.0 %) for . These RMSEP values are about 40 % higher
367
than that obtained with  . Note that these RMSEP values should be reassessed by
368
using more accurate Cab measurements as obtained from a pigment extraction method instead from
369
transmittance-based Dualex measurements.
370
22
3.3. Sensitivity to GF and GAI
371
As mentioned earlier, VIs may be simultaneously sensitive to variations in Cab and canopy structural
372
variables such as GF or GAI. The effect of leaf biochemical composition must therefore be carefully
373
disentangled from that of structural variables.
374
For this purpose, the correlation between the tested VIs and GF (Fig. 9) or GAI (Fig. 10) was
375
investigated. For both structural variables, the performances of the two strategies to compute the VIs
376
similarly depend on the level of heterogeneity in the pixel selection. However, unlike for Cab, when the
377
heterogeneity is the largest, as when considering all the pixels at the highest resolutions, averaging the
378
pixel-level VI values (Strategy 2) generally provides slightly better correlations with both GF and GAI
379
than computing the VI from the average reflectance spectra (Strategy 1). As observed for Cab (Fig. 7),
380
Strategy 2 enhances the influence of the heterogeneity: the latter contains little information about Cab
381
variations while being strongly driven by the canopy structure, e.g., through the proportion of soil and
382
vegetation pixels as well as that of shaded and illuminated pixels. This explains why Strategy 2 provides
383
more accurate retrievals of these structural variables as compared to Strategy 1. However, when the
384
distribution of pixels is reduced, very similar results are obtained for the two strategies (Fig. 9 and
385
Fig. 10, lines 2-4), similarly to what is observed for Cab. Only Strategy 2 will therefore be considered in
386
the following since it provides either best or equal GF and GAI estimation performances as Strategy 1.
387
In the case of GF estimation, strong correlations ( ) are generally obtained with all tested VIs
388
except  and  when all the pixels are considered (Fig. 9, line 1).
389
 achieves the best performances for spatial resolutions close to 8 cm ( ). Unlike for Cab,
390
no further improvement is observed when using only vegetation pixels (Fig. 9, lines 2-4) since, by
391
definition, GF relates to the relative proportions of vegetation and soil.
392
In the case of GAI estimation when considering all the pixels, the best VIs are  ( )
393
followed by  and  ( ) for spatial resolutions finer than 1.8 cm because of the
394
beneficial influence of pure soil pixels (Fig. 10, line 1). For similar reasons as for GF estimation, focusing
395
23
on green pixels generally decreases the performances for most resolutions and tested VIs (Fig. 10,
396
lines 2-4), especially when considering the 50% brightest green pixels at the finest resolutions for which
397
the impact of canopy structure is minimum. The main exceptions occur for resolutions around 8 cm
398
with , ,  and , whose performances achieved when considering the 50%
399
darkest green pixels improve as compared to using all the pixels. The best correlation over all pixel
400
selections is actually obtained by  at the 7 cm spatial resolution ( ). This may be due to
401
the fact that this combination resolution/pixel selection offers the best compromise to maximize the
402
sensitivity to canopy structure variations (maximum for the finest resolutions) while minimizing 
403
saturation (minimum for the coarsest resolutions).
404
These results thus demonstrate the specificities associated with each VI: most tested VIs appear to be
405
mainly related to GF and GAI while   and to a lesser extent  are
406
strongly related to Cab and almost not sensitive to the structural variables. Such a result is of
407
tremendous importance since it demonstrates that the sensitivity of  to Cab will
408
not derive from non-causal relationships due to possible covariance between Cab and canopy structure
409
observed over the training dataset.
410
411
24
Figure 9: Squared Spearman’s correlation () between VIs and GF measured over the 55 plots as a function
of spatial resolution. In the left-hand column, VIs are computed from the average reflectance spectra of
considered pixels (strategy (1)), while in the right -hand column, VIs are computed by averaging pixel-level
VI values (strategy (2)). For both strategies, the four subsets of pixels are tested (vegetation and soil pixels,
vegetation pixels, 50% darkest and brightest vegetation pixels).
25
Figure 10: Squared Spearman’s correlation () between VIs and GAI measured over the 55 plots as a function
of spatial resolution. In the left-hand column, VIs are computed from the average reflectance spectra of
considered pixels (strategy (1)), while in the right -hand column, VIs are computed by averaging pixel-level
VI values (strategy (2)). For both strategies, the four subsets of pixels are tested (vegetation and soil pixels,
vegetation pixels, 50% darkest and brightest vegetation pixels).
412
413
26
Figure 11: Relationships between  and GF when considering all the pixels at the 8.8 cm resolution (a),
and between  and GAI when considering the 50% darkest green pixels at the 7 cm resolution (b). In both
cases, VIs are computed by averaging pixel-level VI values. Prediction performances obtained using linear (in
black) and best non-linear (in red, power function for GF and exponential function for GAI) regressions are
shown.
414
The best cases observed in Fig. 9 and Fig. 10 are further detailed in Fig. 11 to quantify their
415
performances. In the case of GF estimation, the best case corresponds to  computed from all the
416
pixels observed at the 8.8 cm resolution (Fig. 11.a). It obtains   and    ( of
417
the GF range) using either linear or non-linear regression models. In the case of GAI estimation, the
418
best case corresponds to  computed over the 50% darkest green pixels observed at the 7 cm
419
resolution (Fig. 11.b). While using a linear model leads to reasonable estimation results ( 
420
and    m²/m² corresponding to  of the GAI range), using an exponential model
421
significantly improves the performance to take into account the saturation occurring for high GAI
422
values, as suggested by Baret and Guyot (1991) (  and   /m² corresponding
423
to  of the GAI range).
424
4. Conclusions and perspectives
425
This study focuses on the estimation of leaf chlorophyll content from reflectance observations using
426
an empirical approach applied to sugar beet crops. It is clearly demonstrated that computing the
427
spectral indices over the well-illuminated green pixels of the image improves the sensitivity to
428
chlorophyll content and decreases the possible impact of variations in canopy structure. Further, the
429
27
  index performs significantly better than every other index tested. This index has
430
two interesting properties: (1) as a ratio index,  is by construction independent to
431
any factor that affects the radiance reflected by the canopy in a multiplicative way, including slope
432
effects and variations in illumination conditions due to slight cloud coverage; (2) the particular
433
selection of wavebands enhances the sensitivity of   to leaf chlorophyll content
434
while minimizing the impact of canopy structure. It is symptomatic to observe that the best-performing
435
index for chlorophyll content estimation (  ) is the poorest one for canopy structure
436
retrieval. Reciprocally, the best index for canopy structure retrieval () offers the poorest
437
performances for chlorophyll content estimation. The main drawback of this type of spectral index
438
based on a blue band is its sensitivity to variations in the spectral distribution of incident light: the
439
latter may be indeed strongly influenced by changing illumination conditions that affect the scattering
440
by aerosols, water droplets or crystals. Great attention should therefore be paid either to the stability
441
of irradiance conditions (using potentially artificial light sources), or to the radiometric calibration that
442
will require frequent measurements over reference panels or corrections from ancillary information.
443
Alternatively, the use of a red reference located in the domain of maximum chlorophyll absorption
444
(i.e., around 680 nm) may decrease the sensitivity to atmospheric conditions while providing
445
interesting results if the spatial resolution and Cab range are such that this band nearly saturates for
446
every sample. Incidentally, another drawback of using a blue reference is that poor estimation
447
performances may be expected for very low values of leaf chlorophyll content, as in this case, the blue
448
waveband does not reach saturation and varies with chlorophyll content (Sims and Gamon, 2002).
449
Studying the impact of spatial resolution on Cab estimation shows that  may be
450
quite sensitive to the soil influence. In the case of sugar beet crops, similar results are obtained for all
451
the spatial resolutions finer than 4.4 cm, but these results deteriorate for coarser resolutions. In the
452
context of UAV-based remote sensing, this means that a resolution of about 4 cm offers a good
453
compromise between accuracy and efficiency, the latter improving as spatial resolution decreases and
454
flight altitude increases. This compromise should, however, be defined for each situation: for example,
455
28
time constraints may be such that efficiency can be improved by reasonably decreasing spatial
456
resolution and, therefore, accuracy (note that this decrease can somehow be compensated by
457
choosing a shorter Cab-sensitive waveband so as to mitigate the soil influence).
458
These results are derived from a comprehensive set of experiments including several locations, years,
459
cultivars and levels of nitrogen fertilization. Good robustness properties are thus expected for sugar
460
beet crops. However, these results need to be further evaluated for other species. In particular, the
461
spatial resolution to be retained should be adapted to each crop to limit the fraction of mixed pixels in
462
the computation of the spectral index. Also, the influence of the pixel subset used to compute the VIs
463
should also be questioned for vegetation species exhibiting strong leaf surface effects, although this is
464
already the case for the glossy sugar beet leaves. Ultimately, the proposed principles to derive optimal
465
spectral indices could be extended to the estimation of other leaf biochemical constituents such as
466
leaf water content: for example, the Cab-sensitive waveband (728 nm) could be replaced by a water-
467
sensitive one.
468
The rapid emergence of UAVs has opened a new era of Earth observation for vegetation monitoring
469
(Houborg et al., 2015). Multi- and hyperspectral cameras are now becoming efficient and affordable
470
sensors, and their combination with UAVs enables the acquisition of spectral data with the required
471
high spatial resolution while providing a significant spatial coverage capacity in a reasonable time. This
472
is particularly interesting for agricultural applications (Duan et al., 2014; Verger et al., 2014; Zarco-
473
Tejada et al., 2013) that require a timely revisit capacity to sample the crop at the optimal development
474
stage (Inoue et al., 2012). The VIs developed in this study take advantage of such high spatial resolution
475
data to improve the estimation of leaf chlorophyll content in sugar beet canopies as compared to
476
classical VIs of the literature. In particular, the use of    based on UAV observations
477
is promising to accurately estimate leaf chlorophyll content over larger areas.
478
Acknowledgments
479
29
This study was funded by the French Ministry of Agriculture, Agrifood, and Forestry (PHENOBET
480
project), and by the French National Research Agency, within the program ”Investissements d’avenir”
481
with the reference ANR-11-BTBR- 0007 (AKER project). Thanks a lot to the three anonymous reviewers
482
for their valuable comments and suggestions, as well as to Daniel Moura, David Bastidon and Jean-
483
Francois Bonicel for their help in the experiments.
484
References
485
Bånkestad, D., Wik, T., 2016. Growth tracking of basil by proximal remote sensing of chlorophyll
486
fluorescence in growth chamber and greenhouse environments. Comput. Electron. Agric. 128,
487
7786. doi:10.1016/j.compag.2016.08.004
488
Baret, F., Buis, S., 2008. Estimating canopy characteristics from remote sensing observations: review
489
of methods and associated problems, in: Liang, S. (Ed.), Advances in Land Remote Sensing:
490
System, Modeling, Inversion and Application. pp. 173201. doi:10.1007/978-1-4020-6450-0_7
491
Baret, F., de Solan, B., Lopez-Lozano, R., Ma, K., Weiss, M., 2010. GAI estimates of row crops from
492
downward looking digital photos taken perpendicular to rows at 57.5° zenith angle: Theoretical
493
considerations based on 3D architecture models and application to wheat crops. Agric. For.
494
Meteorol. 150, 13931401. doi:10.1016/j.agrformet.2010.04.011
495
Baret, F., Guyot, G., 1991. Potential and limitations of vegetation indices for LAI and APAR
496
assessment. Remote Sens. Environ. 35, 161173.
497
Blackburn, G.A., 2007. Hyperspectral remote sensing of plant pigments. J. Exp. Bot. 58, 855867.
498
doi:10.1093/jxb/erl123
499
Blackburn, G.A., 1998. Quantifying Chlorophylls and Caroteniods at Leaf and Canopy Scales. Remote
500
Sens. Environ. 66, 273285. doi:10.1016/S0034-4257(98)00059-5
501
Bousquet, L., Lachérade, S., Jacquemoud, S., Moya, I., 2005. Leaf BRDF measurements and model for
502
30
specular and diffuse components differentiation. Remote Sens. Environ. 98, 201211.
503
doi:10.1016/j.rse.2005.07.005
504
Cerovic, Z.G., Masdoumier, G., Ghozlen, N. Ben, Latouche, G., 2012. A new optical leaf-clip meter for
505
simultaneous non-destructive assessment of leaf chlorophyll and epidermal flavonoids. Physiol.
506
Plant. 146, 251260. doi:10.1111/j.1399-3054.2012.01639.x
507
Clevers, J., Kooistra, L., 2012. Using Hyperspectral Remote Sensing Data for Retrieving Canopy
508
Chlorophyll and Nitrogen Content. Ieee J. Sel. Top. Appl. Earth Obs. Remote Sens. 5, 574583.
509
doi:10.1109/JSTARS.2011.2176468
510
Clevers, J.G.P.W., Gitelson, A.A., 2013. Remote estimation of crop and grass chlorophyll and nitrogen
511
content using red-edge bands on sentinel-2 and-3. Int. J. Appl. Earth Obs. Geoinf. 23, 344351.
512
doi:10.1016/j.jag.2012.10.008
513
Comar, A., Baret, F., Obein, G., Simonot, L., Meneveaux, D., Viénot, F., de Solan, B., 2014. ACT: A leaf
514
BRDF model taking into account the azimuthal anisotropy of monocotyledonous leaf surface.
515
Remote Sens. Environ. 143, 112121. doi:10.1016/j.rse.2013.12.006
516
Comar, A., Burger, P., de Solan, B., Baret, F., Daumard, F., Hanocq, J.-F., 2012. A semi-automatic
517
system for high throughput phenotyping wheat cultivars in-field conditions: description and
518
first results. Funct. Plant Biol. 39, 914924. doi:10.1071/FP12065
519
Combal, B., Baret, F., Weiss, M., Trubuil, A., Macé, D., Pragnère, A., Myneni, R., Knyazikhin, Y., Wang,
520
L., 2003. Retrieval of canopy biophysical variables from bidirectional reflectance: Using prior
521
information to solve the ill-posed inverse problem. Remote Sens. Environ. 84, 115.
522
doi:10.1016/S0034-4257(02)00035-4
523
Dash, J., Curran, P.J., 2004. The MERIS terrestrial chlorophyll index. Int. J. Remote Sens. 25, 5403
524
5413. doi:10.1080/0143116042000274015
525
Daughtry, C.S.., Walthall, C.., Kim, M.., de Colstoun, E.B., McMurtrey, J.., 2000. Estimating Corn Leaf
526
31
Chlorophyll Concentration from Leaf and Canopy Reflectance. Remote Sens. Environ. 74, 229
527
239. doi:10.1016/S0034-4257(00)00113-9
528
Dorigo, W. a., 2012. Improving the robustness of cotton status characterisation by radiative transfer
529
model inversion of multi-angular CHRIS/PROBA data. IEEE J. Sel. Top. Appl. Earth Obs. Remote
530
Sens. 5, 1829. doi:10.1109/JSTARS.2011.2171181
531
Duan, S.B., Li, Z.L., Wu, H., Tang, B.H., Ma, L., Zhao, E., Li, C., 2014. Inversion of the PROSAIL model to
532
estimate leaf area index of maize, potato, and sunflower fields from unmanned aerial vehicle
533
hyperspectral data. Int. J. Appl. Earth Obs. Geoinf. 26, 1220. doi:10.1016/j.jag.2013.05.007
534
Elarab, M., Ticlavilca, A.M., Torres-Rua, A.F., Maslova, I., McKee, M., 2015. Estimating chlorophyll
535
with thermal and broadband multispectral high resolution imagery from an unmanned aerial
536
system using relevance vector machines for precision agriculture. Int. J. Appl. Earth Obs. Geoinf.
537
43, 3242. doi:10.1016/j.jag.2015.03.017
538
Féret, J.-B., François, C., Gitelson, A., Asner, G.P., Barry, K.M., Panigada, C., Richardson, A.D.,
539
Jacquemoud, S., 2011. Optimizing spectral indices and chemometric analysis of leaf chemical
540
properties using radiative transfer modeling. Remote Sens. Environ. 115, 27422750.
541
doi:10.1016/j.rse.2011.06.016
542
Gitelson, A.A., Gritz, Y., Merzlyak, M.N., 2003. Relationships between leaf chlorophyll content and
543
spectral reflectance and algorithms for non-destructive chlorophyll assessment in higher plant
544
leaves. J. Plant Physiol. 160, 271282. doi:10.1078/0176-1617-00887
545
Gitelson, A.A., Kaufman, Y.J., Merzlyak, M.N., 1996. Use of a green channel in remote sensing of
546
global vegetation from EOS-MODIS. Remote Sens. Environ. 58, 289298. doi:10.1016/S0034-
547
4257(96)00072-7
548
Gitelson, A.A., Keydan, G.P., Merzlyak, M.N., 2006. Three-band model for noninvasive estimation of
549
chlorophyll, carotenoids, and anthocyanin contents in higher plant leaves. Geophys. Res. Lett.
550
32
33, 16. doi:10.1029/2006GL026457
551
Gitelson, A.A., Merzlyak, M.N., 1994. Spectral Reflectance Changes Associated with Autumn
552
Senescence of Aesculus-hippocastanum L. and Acer-platanoides L. Leaves - Spectral Features
553
and Relation to Chlorophyll Estimation. J. Plant Physiol. 143, 286292. doi:10.1016/S0176-
554
1617(11)81633-0
555
Gitelson, A.A., Vina, A., Ciganda, V., Rundquist, D.C., Arkebauer, T.J., 2005. Remote estimation of
556
canopy chlorophyll content in crops. Geophys. Res. Lett. 32, 14. doi:10.1029/2005GL022688
557
Haboudane, D., Miller, J.R., Tremblay, N., Zarco-Tejada, P.J., Dextraze, L., 2002. Integrated narrow-
558
band vegetation indices for prediction of crop chlorophyll content for application to precision
559
agriculture. Remote Sens. Environ. 81, 416426. doi:10.1016/S0034-4257(02)00018-4
560
Haboudane, D., Tremblay, N., Miller, J.R., Vigneault, P., 2008. Remote estimation of crop chlorophyll
561
content using spectral indices derived from hyperspectral data. IEEE Trans. Geosci. Remote
562
Sens. 46, 423436. doi:10.1109/TGRS.2007.904836
563
Houborg, R., Fisher, J.B., Skidmore, A.K., 2015. Advances in remote sensing of vegetation function
564
and traits. Int. J. Appl. Earth Obs. Geoinf. 43, 16. doi:10.1016/j.jag.2015.06.001
565
Hunt, E.R., Doraiswamy, P.C., McMurtrey, J.E., Daughtry, C.S.T., Perry, E.M., Akhmedov, B., 2012. A
566
visible band index for remote sensing leaf chlorophyll content at the canopy scale. Int. J. Appl.
567
Earth Obs. Geoinf. 21, 103112. doi:10.1016/j.jag.2012.07.020
568
Inoue, Y., Sakaiya, E., Zhu, Y., Takahashi, W., 2012. Diagnostic mapping of canopy nitrogen content in
569
rice based on hyperspectral measurements. Remote Sens. Environ. 126, 210221.
570
doi:10.1016/j.rse.2012.08.026
571
Jacquemoud, S., Baret, F., 1990. PROSPECT: A model of leaf optical properties spectra. Remote Sens.
572
Environ. 34, 7591. doi:10.1016/0034-4257(90)90100-Z
573
33
Jacquemoud, S., Verhoef, W., Baret, F., Bacour, C., Zarco-Tejada, P.J., Asner, G.P., François, C., Ustin,
574
S.L., 2009. PROSPECT+SAIL models: A review of use for vegetation characterization. Remote
575
Sens. Environ. 113, S56S66. doi:10.1016/j.rse.2008.01.026
576
Jay, S., Bendoula, R., Hadoux, X., Féret, J.-B., Gorretta, N., 2016. A physically-based model for
577
retrieving foliar biochemistry and leaf orientation using close-range imaging spectroscopy.
578
Remote Sens. Environ. 177, 220236. doi:http://dx.doi.org/10.1016/j.rse.2016.02.029
579
Jay, S., Maupas, F., Bendoula, R., Gorretta, N., n.d. Retrieving LAI, chlorophyll and nitrogen contents
580
in sugar beet crops from multi-angular optical remote sensing: comparison of vegetation indices
581
and PROSAIL inversion for field phenotyping. F. Crop. Res. submitted.
582
Knyazikhin, Y., Schull, M. a, Stenberg, P., Mõttus, M., Rautiainen, M., Yang, Y., Marshak, A., Latorre
583
Carmona, P., Kaufmann, R.K., Lewis, P., Disney, M.I., Vanderbilt, V., Davis, A.B., Baret, F.,
584
Jacquemoud, S., Lyapustin, A., Myneni, R.B., 2013. Hyperspectral remote sensing of foliar
585
nitrogen content. Proc. Natl. Acad. Sci. U. S. A. 110, E185-92. doi:10.1073/pnas.1210196109
586
Kooistra, L., Clevers, J.G.P.W., 2016. Estimating potato leaf chlorophyll content using ratio vegetation
587
indices. Remote Sens. Lett. 7, 611620. doi:10.1080/2150704X.2016.1171925
588
Latorre-Carmona, P., Knyazikhin, Y., Alonso, L., Moreno, J.F., Pla, F., Yan, Y., 2014. On hyperspectral
589
remote sensing of leaf biophysical constituents: Decoupling vegetation structure and leaf optics
590
using CHRIS-PROBA data over crops in barrax. IEEE Geosci. Remote Sens. Lett. 11, 15791583.
591
doi:10.1109/LGRS.2014.2305168
592
le Maire, G., François, C., Dufrêne, E., 2004. Towards universal broad leaf chlorophyll indices using
593
PROSPECT simulated database and hyperspectral reflectance measurements. Remote Sens.
594
Environ. 89, 128. doi:10.1016/j.rse.2003.09.004
595
Moorthy, I., Miller, J.R., Noland, T.L., 2008. Estimating chlorophyll concentration in conifer needles
596
with hyperspectral data: An assessment at the needle and canopy level. Remote Sens. Environ.
597
34
112, 28242838. doi:10.1016/j.rse.2008.01.013
598
Penuelas, J., Baret, F., Filella, I., 1995. Semi-empirical indices to assess carotenoids/chlorophyll a ratio
599
from leaf spectral reflectance. Photosynthetica 31, 221230.
600
Rondeaux, G., Steven, M., Baret, F., 1996. Optimization of soil-adjusted vegetation indices. Remote
601
Sens. Environ. 55, 95107. doi:10.1016/0034-4257(95)00186-7
602
Rondeaux, G., Vanderbilt, V.C., 1993. Specularly modified vegetation indices to estimate
603
photosynthetic activity . Int. J. Remote Sens. doi:10.1080/01431169308954004
604
Rouse, J.W., Hass, R.H., Schell, J.A., Deering, D.W., 1973. Monitoring vegetation systems in the great
605
plains with ERTS. Third Earth Resour. Technol. Satell. Symp. 1, 309317. doi:citeulike-article-
606
id:12009708
607
Schlemmer, M., Gitelson, a., Schepers, J., Ferguson, R., Peng, Y., Shanahan, J., Rundquist, D., 2013.
608
Remote estimation of nitrogen and chlorophyll contents in maize at leaf and canopy levels. Int.
609
J. Appl. Earth Obs. Geoinf. 25, 4754. doi:10.1016/j.jag.2013.04.003
610
Sims, D.A., Gamon, J.A., 2002. Relationships between leaf pigment content and spectral reflectance
611
across a wide range of species, leaf structures and developmental stages. Remote Sens.
612
Environ. 81, 337354. doi:10.1016/S0034-4257(02)00010-X
613
Steven, M., Malthus, T., Baret, F., 2015. Toward Standardization of Vegetation Indices, in: Thenkabail,
614
P.S. (Ed.), Remotely Sensed Data Characterization, Classification, and Accuracies. CRC Press, pp.
615
175194. doi:doi:10.1201/b19294-13
616
Ustin, S.L., 2013. Remote sensing of canopy chemistry. Proc. Natl. Acad. Sci. U. S. A. 110, 8045.
617
doi:10.1073/pnas.1219393110
618
Ustin, S.L., Gitelson, a. a., Jacquemoud, S., Schaepman, M., Asner, G.P., Gamon, J. a., Zarco-Tejada,
619
P., 2009. Retrieval of foliar information about plant pigment systems from high resolution
620
35
spectroscopy. Remote Sens. Environ. 113, S67S77. doi:10.1016/j.rse.2008.10.019
621
Verger, A., Vigneau, N., Chéron, C., Gilliot, J.M., Comar, A., Baret, F., 2014. Green area index from an
622
unmanned aerial system over wheat and rapeseed crops. Remote Sens. Environ. 152, 654664.
623
doi:10.1016/j.rse.2014.06.006
624
Vigneau, N., Ecarnot, M., Rabatel, G., Roumet, P., 2011. Potential of field hyperspectral imaging as a
625
non destructive method to assess leaf nitrogen content in Wheat. F. Crop. Res. 122, 2531.
626
doi:10.1016/j.fcr.2011.02.003
627
Wu, C., Niu, Z., Tang, Q., Huang, W., 2008. Estimating chlorophyll content from hyperspectral
628
vegetation indices: Modeling and validation. Agric. For. Meteorol. 148, 12301241.
629
doi:10.1016/j.agrformet.2008.03.005
630
Zarco-Tejada, P.., Miller, J.., Morales, A., Berjón, A., Agüera, J., 2004. Hyperspectral indices and model
631
simulation for chlorophyll estimation in open-canopy tree crops. Remote Sens. Environ. 90,
632
463476. doi:10.1016/j.rse.2004.01.017
633
Zarco-Tejada, P.J., Guillén-Climent, M.L., Hernández-Clemente, R., Catalina, A., González, M.R.,
634
Martín, P., 2013. Estimating leaf carotenoid content in vineyards using high resolution
635
hyperspectral imagery acquired from an unmanned aerial vehicle (UAV). Agric. For. Meteorol.
636
171, 281294. doi:10.1016/j.agrformet.2012.12.013
637
Zarco-Tejada, P.J., Miller, J.R., Noland, T.L., Mohammed, G.H., Sampson, P.H., 2001. Scaling-up and
638
model inversion methods with narrowband optical indices for chlorophyll content estimation in
639
closed forest canopies with hyperspectral data. IEEE Trans. Geosci. Remote Sens. 39, 1491
640
1507. doi:10.1109/36.934080
641
642
... To define the optimal wavebands for each qualitative variable-specific ND, a grid search was performed by exhaustively combining the 60 available spectral bands, similarly to what was done by, for example, Inoue et al. (2012), Jay, Gorretta, et al. (2017), and Thenkabail et al. (2000). Each band combination was used to build a linear regression with each trait of interest and evaluate the capability of the adjusted spectral index to account for changes in quality. ...
Article
Full-text available
Forage crops are a cornerstone of the agricultural industry in Nordic countries. Economic and ecological performances are directly linked to adapted farming practices, which require timed and precise information on the nutritive value of the forage. Field spectrometers could offer an interesting alternative to time-consuming laboratory measurements, as they provide near real time information. We used a handheld version of a field spectrometer already commercialized for cereal adjustable rate fertilization, to evaluate its potential for grassland nutritive quality estimation. Spectral data and samples were acquired over experimental fields and plots in four locations in Northern Sweden; samples were analyzed using wet chemistry to determine the crude protein concentration, the in vitro true digestibility, the neutral detergent fiber and the neutral detergent fiber digestibility. Grid-based adjusted spectral indices, partial least squares, random forest and support vector machine were tested to link the spectral data to the nutritive traits. Partial least squares and support vector machine outperformed the adjusted spectral indices and random forest. Best predictions were obtained with partial least squares for in vitro true digestibility and neutral detergent fiber (R2 of 0.64 and 0.78 and normalized root mean square error [nRMSE] of 2.1 and 8.0%, respectively) and with support vector machine for crude protein and neutral detergent fiber digestibility (R2 of 0.49 and 0.65 and nRMSE of 13.0 and 3.8%, respectively). These results suggests that there is a potential for this affordable, industry-ready spectrometer to be used as a practical farming tool, although more comprehensive datasets are needed to ensure that robust models are developed.
... However, when spectral signatures are aggregated over a larger area to match the spatial resolution of the ground truth observations, they can also introduce noise and reduce the prediction performance [40]. Studies devoted to the spatial aggregation problem involve broadband measurements [40][41][42] or are focused on proximal sensing and vegetation indices [14,43], while the issue remains underexplored for machine learning models trained using hyperspectral data acquired with a UAV. ...
Article
Full-text available
The remote sensing of the biophysical and biochemical parameters of crops facilitates the preparation of application maps for variable-rate nitrogen fertilization. According to comparative studies of machine learning algorithms, Gaussian process regression (GPR) can outperform more popular methods in the prediction of crop status from hyperspectral data. The present study evaluates GPR model accuracy in the context of spring wheat dry matter, nitrogen content, and nitrogen uptake estimation. Models with the squared exponential covariance function were trained on images from two hyperspectral cameras (a frenchFabry–Pérot interferometer camera and a push-broom scanner). The most accurate predictions were obtained for nitrogen uptake (R2=0.75–0.85, RPDP=2.0–2.6). Modifications of the basic workflow were then evaluated: the removal of soil pixels from the images prior to the training, data fusion with apparent soil electrical conductivity measurements, and replacing the Euclidean distance in the GPR covariance function with the spectral angle distance. Of these, the data fusion improved the performance while predicting nitrogen uptake and nitrogen content. The estimation accuracy of the latter parameter varied considerably across the two hyperspectral cameras. Satisfactory nitrogen content predictions (R2>0.8, RPDP>2.4) were obtained only in the data-fusion scenario, and only with a high spectral resolution push-broom device capable of capturing longer wavelengths, up to 1000 nm, while the full-frame camera spectral limit was 790 nm. The prediction performance and uncertainty metrics indicated the suitability of the models for precision agriculture applications. Moreover, the spatial patterns that emerged in the generated crop parameter maps accurately reflected the fertilization levels applied across the experimental area as well as the background variation of the abiotic growth conditions, further corroborating this conclusion.
... Hyperspectral Imaging (HSI) is one of the most popular techniques for plant disease detection which make the use of various vegetation indices (VI). The advantage of these vegetation indices is theta it reductions the dimensionality of the data and scale factor which affect the variations in the illuminations condition [6]. Photochemical Reflectance Index (PRI) is a normalized vegetation index used for estimating photosynthetic light use efficiency (LUE) and also helpful in monitoring crop stress. ...
Article
Determining the spatial variation of different plant factors throughout growing season will help to resolve stress factors within a field in a timely basis. Whereas the spectral characterizes help to estimate the proper photosynthesis process. This research shows that the nitrogen reflectance index (NRI) help to predict the nitrogen level of healthy and diseased plants and photochemical reflectance index (PRI) affects the leaf spectral absorption. These indices are calibrated under the hyperspectral pushbroom camera Resonon PIKA-L (400-1000nm) which is non-destructive and less time consuming, it is available in RUSA lab in Dr. Babasaheb Ambedkar Marathawada University, Aurangabad, Maharashtra. The spectral bands considered for the calculation of NRI are 700nm, 670nm, 570nm and for PRI spectral bands considered were 531nm, 570nm. Statistical values for PRI were calculated like R-Square (0.727), RMSE (0.267), P-value (2.787), standard error (2.979) and the statistical values for NRI were R-Square (4.223), RMSE (0.512), P-value (0.968), standard error(2.648).Linear regression was calculated for finding the relation between the data.
... These innovative strategies can be feasible on ground or UAV platforms where the observation angle and time can be easily adjusted with great flexibility. In addition, higher spatial resolution reflectance imagery can benefit the discrimination of vegetation and soil pixels and thus reduce the fraction of soil (Zarco-Tejada et al., 2013;Jay et al., 2017;Wang et al., 2022). Coupling the 3SV algorithm into these observation strategies could further reduce the soil effects and hence improve the LCC estimation accuracy. ...
Article
Full-text available
Leaf chlorophyll content (LCC) is an important indicator of foliar nitrogen status and photosynthetic capacity. Compared to physical models, the generality of empirical models based on vegetation indices is often questioned when they are used to estimate LCC due to the influence from canopy structure, such as leaf area index (LAI). A recent study developed the LAI-insensitive chlorophyll index (LICI) and established a semi-empirical LICI-based LCC quantification model, which inherits both the robustness of physical models and the simplicity of empirical models. However, it is unclear whether such a simple model is as accurate and generic as physical models. Here, we adopted an innovative approach to disentangle the confounding effects of LAI and LCC on LICI and found that LICI was strongly correlated to LCC but only marginally sensitive to LAI. Moreover, we also found that LICI was sensitive to the soil background and thus proposed a spectral separation of soil and vegetation (3SV) algorithm, which is automatic and does not require prior information of soil background. After implementing the 3SV algorithm to remove the contributed reflectance of soil, we then obtained the contributed reflectance of vegetation (CRv). Model simulations showed that the soil background effect on the CRv-derived LICI was largely eliminated and hence this index was viewed to be soil-removed. As a result, the accuracy and generality of the soil-removed LICI-based model for LCC estimation was evaluated using comprehensive datasets from multiple vegetation types, years, sites, and observation platforms and compared to that of a MatrixVI-based physical model and a MERIS terrestrial chlorophyll index (MTCI)-based semi-empirical model. The root-mean-square error (RMSE) for LCC estimated by the soil-removed LICI-based model was 6.22–6.87 μg/cm² for the crop datasets and 10.68 μg/cm² for the multi-ecosystem dataset when the equivalent wet soil fraction was <0.7. Although further efforts are required to mitigate the effects of soil on the LICI-based model over sparse vegetation, this research is highly beneficial for extending its potential applications to the globe and advancing the development of an operational LCC monitoring system in the emerging satellite hyperspectral era.
Article
Leaf chlorophyll plays an important role in forest management and ecosystem balance. Hyperspectral images have been widely applied in leaf chlorophyll content (LCC) estimation. However, the complexity of canopy structures, such as leaf area index (LAI), weakens the performance of LCC estimation from hyperspectral images. The effect of LAI must be considered. Light detection and ranging (LiDAR) has advantages in LAI extraction. Therefore, this study theoretically and experimentally explored the potential of integrating the hyperspectral image and LiDAR data to improve LCC estimation. The PROSAIL model constrained by LiDAR-derived LAI was developed when the hyperspectral image was applied to LCC estimation. Theoretical evaluation and experimental validation were conducted to determine the performance of constrained PROSAIL models with four minimization algorithms. Four minimization algorithms included the fminsearchbnd, simulated annealing, genetic algorithm and Pareto sets. Three datasets were used: the synthetic dataset followed a uniform distribution, the synthetic dataset followed a normal distribution, and the airborne sensor dataset (airborne hyperspectral and LiDAR data). The non-constrained PROSAIL model was as a comparison. Results showed the effectiveness of the constrained PROSAIL model through integrating hyperspectral and LiDAR data for the improvement of LCC estimation. The constrained PROSAIL model always had better performance than non-constrained PROSAIL model among all datasets and minimization algorithms. It demonstrated that constrained PROSAIL model with LAI constraint could improve the accuracy of LCC estimation, where the R² could be increased by 10% to 29%, and 9% to 62% for the synthetic datasets and airborne sensor dataset, respectively. The constrained PROSAIL model with Pareto sets had the best accuracy with an R² of 0.81 from airborne sensor dataset. This study provides a new strategy for the improvement of LCC estimation and has great potential to serve precision forestry.
Article
The important period of wheat grain accumulation is from the flowering stage to the filling stage, and the nitrogen content of wheat in this period is of great significance to the yield accumulation. With the rapid development of sensor technology, different sensors have been increasingly used for crop nitrogen status estimation due to their flexibility. This study aimed to investigate the use of a combination of image information from two proximal sensors (RGB and thermal sensors) to assess the nitrogen status of wheat at the reproductive growth stage. Previous studies have focused on estimating leaf N status at the nutritional growth stage of wheat, and the precision of N estimation is not high at the later stages. Considering that the canopy was composed of leaves and spikes in the reproductive stage, we integrated leaf N content and spike N content as plant N content for assessment. A two-year field trial was conducted, and this study used a Sony camera to acquire RGB images from flowering to maturity and obtained thermal images using the handle thermal infrared camera during the same period. Then, these images were further processed to extract the color features (17), the texture features (5) and temperature values (2). Based on these 24 indices, this study used three machine learning algorithms (i.e., Back-Propagation neural network (BP), Random Forest (RF) and Support Vector Regression (SVR)), resulted in nine estimation models based on a single dataset (i.e., c-based BP, te-based BP, t-based BP, c-based RF, te-based RF, t-based RF, c-based SVR, te-based SVR, t-based SVR) and 12 models based on data fusions (i.e., c+te-based BP, c+t-based BP, te+t-based BP, c+te+t-based BP, c+te-based RF, c+t-based RF, te+t-based RF, c+te+t-based RF, c+te-based SVR, c+t-based SVR, te+t-based SVR, c+te+t-based SVR). The performance of the 21 models was evaluated and compared with each other according to the coefficient of determination (R²), root mean square error (RMSE) and residual prediction deviation (RPD) in nitrogen content estimation. The results show that the best model was the c+te+t-based RF, which was a model based on the combination of color features, texture features and temperature values. It achieved high accuracy in estimating plant N content (R² = 0.89, RMSE = 3.23 mg g⁻¹, RPD = 1.90). In conclusion, the combination of information from RGB and thermal images has good potential for application in monitoring crop N content at late reproductive stages, and plant temperature values can be used as effective indicators for assessing crop growth and nitrogen nutrient status.
Chapter
Globalization, modern cultivation techniques, climate change, and human activities have promoted the distribution of plant pathogens, resulting in frequent host–pathogen interactions and disease incidences. These causative factors are impossible to control as even the most rigorous quarantine system could not completely avoid the movement of plant pathogens and germplasm across countries and continents. Susceptibility of cultivated varieties against plant pathogens, especially fungi, has significantly increased because varieties are developed focusing on higher yield. Additionally, plant pathogens undergo frequent mutations and genetic changes to adapt to climate changes, overcome pesticide resistance, and infect plant germplasm previously resistant. Plant disease identification, quantification, and estimation of subsequent yield losses are crucial in modern-day agriculture to ensure food safety and security for the increasing global population. Phytopathometry utilizes systematized and specialized approaches for plant disease assessment and presents qualitative and quantitative data. Phytopathometry underpins all activities in plant pathology and extends into other related disciplines such as agronomy, plant breeding, and horticulture. Digital and biotechnology underpinned by contemporary artificial intelligence efficiently process sensory data for plant disease measurement. Modern phytopathometry tools aided with detailed knowledge of pathogen–host system biology are poised to become an integral part of precision agriculture.Use of machine learning, deep learning, digital technology, biotechnology, engineering, and nanotechnology in phytopathometry is exposing plant pathologists to new terminology, concepts, and ideas which were not even thinkable a few decades ago. Moreover, innovations in robotics have provided flexibility and precision in deploying these sensors for accurate disease assessment, even in large field areas. In this chapter, we have discussed various phytopathometry tools and approaches and their transdisciplinary uses.KeywordsDisease indexVisual assessmentRemote sensingDigital imageryArtificial intelligenceAdvanced technology
Article
Full-text available
Rapid and accurate assessment of yield and nitrogen use efficiency (NUE) is essential for growth monitoring, efficient utilization of fertilizer and precision management. This study explored the potential of a consumer-grade DJI Phantom 4 Multispectral (P4M) camera for yield or NUE assessment in winter wheat by using the universal vegetation indices independent of growth period. Three vegetation indices having a strong correlation with yield or NUE during the entire growth season were determined through Pearson’s correlational analysis, while multiple linear regression (MLR), stepwise MLR (SMLR), and partial least-squares regression (PLSR) methods based on the aforementioned vegetation indices were adopted during different growth periods. The cumulative results showed that the reciprocal ratio vegetation index (repRVI) had a high potential for yield assessment throughout the growing season, and the late grain-filling stage was deemed as the optimal single stage with R ² , root mean square error (RMSE), and mean absolute error (MAE) of 0.85, 793.96 kg/ha, and 656.31 kg/ha, respectively. MERIS terrestrial chlorophyll index (MTCI) performed better in the vegetative period and provided the best prediction results for the N partial factor productivity (NPFP) at the jointing stage, with R ² , RMSE, and MAE of 0.65, 10.53 kg yield/kg N, and 8.90 kg yield/kg N, respectively. At the same time, the modified normalized difference blue index (mNDblue) was more accurate during the reproductive period, providing the best accuracy for agronomical NUE (aNUE) assessment at the late grain-filling stage, with R ² , RMSE, and MAE of 0.61, 7.48 kg yield/kg N, and 6.05 kg yield/kg N, respectively. Furthermore, the findings indicated that model accuracy cannot be improved by increasing the number of input features. Overall, these results indicate that the consumer-grade P4M camera is suitable for early and efficient monitoring of important crop traits, providing a cost-effective choice for the development of the precision agricultural system.
Article
Monitoring the quality attributes of grapes is a practice that allows the state of ripeness to be checked and the optimal harvest time to be identified. A non-destructive method based on hyperspectral imaging (HSI) technology was developed. Analyses were carried out directly in the field on a ‘Sangiovese’ (Vitis vinifera L.) vineyard destined for wine production, by using a Vis/NIR (400–1000 nm) hyperspectral camera. One vineyard row was analysed on 13 different days during the pre-harvest and harvest time. The soluble solids content (SSC) expressed in terms of °Brix was measured by a portable digital refractometer. Afterwards, the grape samples were split in two classes: the first one composed by the samples characterised by a °Brix lower than 20 (not-ripe), while the second one by the samples with a °Brix higher than 20 (ripe). Grape mean spectra were extracted from each hyperspectral image and used to predict the SSC by partial least squares regression (PLS), and to classify the samples into the two classes by PLS discriminant analysis (PLS-DA). SSC was predicted with a R² = 0.77 (RMSECV = 0.79 °Brix), and the samples were correctly classified with a percentage from 86 to 91%. Even if the number of wavelengths was limited, the percentages of correctly classified samples were again within the above-mentioned range. The present study shows the potential of the use of HSI technology directly in the field by proximal measurements under natural light conditions for the prediction of the harvest time of the ‘Sangiovese’ red grape.
Article
As a key phenolic pigment concentrated in the surface tissues of leaves, flavonoids (Flav) are the major bioactive ingredients in Ginkgo leaf extracts. Flav are also marked natural antioxidants and significant indicators of biotic and abiotic stresses, critical for determining cultivation quality and enhancing Flav yield. In particular, area-based Flav (Flavarea) is related to the shortwave-blue light interaction within leaves per unit leaf area, whereas mass-based Flav (Flavmass) is useful for the quantitative assessment of Flav yield. In order to accurately estimate the contents of Flavarea and Flavmass in leaves of Ginkgo plantations, in this study, we developed an advanced bidirectional reflectance factor (BRF) spectra-based approach by reducing the effects of specular reflection and enhancing the absorption signals of Flav (in the shortwave-blue region of spectrum), using a suite of new spectral indices (SIs) (i.e., flavonoid index (FI), modified flavonoid index (mFI) and double difference index (DD)) calculated from the leaf clip equipped spectrometers-collected data. The results demonstrated that most of the SIs derived from the developed BRF spectra-based approach obtained relatively high performance for Flav estimation by alleviating adverse effects of specular reflection to different extents (CV-R² = 0.60–0.76). In specific, DDnir434,421 selected from DD-type indices performed (CV-R² = 0.76 for Flavarea; CV-R² = 0.69 for Flavmass) better than other indices. These findings represent marked potentials of the developed BRF spectra-based approach for non-destructively estimating leaf Flav content, as well as improving the understanding of the mechanisms of specular effects on Flav estimations in leaves of Ginkgo plantations.
Article
Full-text available
Remote sensing has gained much attention for agronomic applications such as crop management or yield esti- mation. Crop phenotyping under field conditions has recently become another important application that re- quires specific needs: the considered remote-sensing method must be (1) as accurate as possible so that slight differences in phenotype can be detected and related to genotype, and (2) robust so that thousands of cultivars potentially quite different in terms of plant architecture can be characterized with a similar accuracy over different years and soil and weather conditions. In this study, the potential of nadir and off-nadir ground-based spectro-radiometric measurements to remotely sense five plant traits relevant for field phenotyping, namely, the leaf area index (LAI), leaf chlorophyll and nitrogen contents, and canopy chlorophyll and nitrogen contents, was evaluated over fourteen sugar beet (Beta vulgaris L.) cultivars, two years and three study sites. Among the di- versity of existing remote-sensing methods, two popular approaches based on various selected Vegetation Indices (VI) and PROSAIL inversion were compared, especially in the perspective of using them for phenotyping applications. Overall, both approaches are promising to remotely estimate LAI and canopy chlorophyll content (RMSE≤10%). In addition, VIs show a great potential to retrieve canopy nitrogen content (RMSE=10%). On the other hand, the estimation of leaf-level quantities is less accurate, the best accuracy being obtained for leaf chlorophyll content estimation based on VIs (RMSE=17%). As expected when observing the relationship be- tween leaf chlorophyll and nitrogen contents, poor correlations are found between VIs and mass-based or area- based leaf nitrogen content. Importantly, the estimation accuracy is strongly dependent on sun-sensor geometry, the structural and biochemical plant traits being generally better estimated based on nadir and off-nadir ob- servations, respectively. Ultimately, a preliminary comparison tends to indicate that, providing that enough samples are included in the calibration set, (1) VIs provide slightly more accurate performances than PROSAIL inversion, (2) VIs and PROSAIL inversion do not show significant differences in robustness across the different cultivars and years. Even if more data are still necessary to draw definitive conclusions, the results obtained with VIs are promising in the perspective of high-throughput phenotyping using UAV-embedded multispectral cameras, with which only a few wavebands are available.
Book
Toward Standardization of Vegetation Indices
Article
Remote sensing is a promising tool for plant phenotyping and precision farming, as it allows for non-invasive, fast and automated measurements of relevant plant traits with spatial and temporal resolution. The simplest and most used remote sensing application in the field is to use reflectance vegetation indices, based on the optical properties of chlorophyll, as indicators of variables of interest. However, the applicability is limited by their sensitivity to environmental conditions and canopy structure. Another remotely sensed signal related to chlorophyll is chlorophyll fluorescence. Compared to reflectance it is plant specific and directly linked to plant physiological processes; but it is also weak, which complicates its use for in-field applications. This study evaluates the performance of an active proximal remote sensing system utilizing the chlorophyll fluorescence ratio method, measuring the ratio of red fluorescence to far-red fluorescence (termed SFR), for the assessment of growth and biomass as an alternative or complement to reflectance vegetation indices.
Article
Keywords: Sorghum Wheat Leaf Surface roughness BRDF BRF Reflectance Conoscope Azimuthal anisotropy Optical properties Goniometer Physical model Refractive index Leaf reflectance of monocotyledons generally displays a strong azimuthal anisotropy due to the longitudinal orientation of the veins. The Cook and Torrance (CT) bidirectional reflectance distribution function model was adapted to account for this distinctive feature. The resulting ACT (Anisotropic Cook and Torrance) model is based on the decomposition of the roughness parameter into two perpendicular components. It is evaluated on sorghum (Sorghum halepense) and wheat (Triticum durum) leaf BRF (Bidirectional Reflectance Factor) measurements acquired using a conoscope system. Results show that the ACT model fits the measurements better than azimuthally isotropic surface models: the root mean square error computed over all the BRF measurements for both leaves decreases from ≈0.06 for the Lambertian model to ≈0.04 for the CT model and down to ≈0.03 for the ACT model. The adjusted value of the refraction index is plausible (n ≈ 1.32) for both leaves while the retrieved roughness values perpendicular to the veins (sorghum = 0.56; wheat = 0.46) is about two times larger than that parallel to the veins (sorghum = 0.27; wheat = 0.18). Nonetheless, the observed residual discrepancies between the ACT model simulations and the measurements may be explained mainly by the Lambertian assumption of the volume scattering.
Article
Chlorophyll content at leaf level is an important variable because of its crucial role in photosynthesis and in understanding plant functioning. In this study, we tested the hypothesis that the ratio of a vegetation index (VI) for estimating canopy chlorophyll content (CCC) and one for estimating leaf area index (LAI) can be used to derive chlorophyll content at the leaf level. This hypothesis for estimating chlorophyll content at the leaf level was tested using simulations with the PROSAIL radiative transfer model and field spectroradiometry measurements in five consecutive years (2010-2014) for potato crops on experimental fields. During the growing season, in-situ field measurements of LAI and leaf chlorophyll content (LCC) were performed. Results showed that good estimates of LCC were feasible using ratio vegetation indices (VIs). This was tested at satellite level using RapidEye images. This letter presents a proof of concept for estimating LCC using Sentinel-2 data. Results confirm the importance of the red-edge bands for agricultural applications, but also showed that indices using the red-edge bands may be replaced by indices using green bands. It should now be tested with real Sentinel-2 data whether its spectral bands at 10 m spatial resolution are suitable for estimating LCC, avoiding the need for red-edge bands that only are available at 20 m.