Computers and Electronics in Agriculture 152 (2018) 281–289

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

Original papers

Assessment of spinach seedling health status and chlorophyll content by T

multivariate data analysis and multiple linear regression of leaf image
features
⁎
Avinash Agarwal , Snehasish Dutta Gupta
Agricultural and Food Engineering Department, Indian Institute of Technology Kharagpur, Kharagpur 721302, India

A R T I C LE I N FO A B S T R A C T

Keywords: Plant health and physiological status significantly influence chlorophyll content and photosynthetic capacity.
Leaf spectral reflectance Analysis of leaf reflectance information from digitized leaf images allows high-throughput, non-invasive and
Plant health status real-time estimation of chlorophyll content in a cost-effective manner. In the present study the application of
Image analysis multivariate data analysis tools, viz. principal component analysis (PCA) and agglomerative hierarchical clus-
Multivariate data analysis
tering analysis (AHCA), has been discussed for distinguishing between spinach seedlings having high and low
Multiple linear regression
chlorophyll contents by simultaneously using the information provided by various image features. Further, leaf
color information contained within different color spaces, viz. RGB (red, green and blue), rgb (normalized red,
green and blue), HSI (hue, saturation and intensity), CIE (Commission Internationale de l’Eclairage) L∗a∗b∗, CIE-
XYZ, and CIE-xyY color spaces, has been used to predict chlorophyll content in terms of SPAD (Soil Plant
Analysis Development) chlorophyll meter values by multiple linear regression. It was observed that the color
indices R, G, R + G, R−B, G−B, R + G−B, Y (luminance) and DGCI (dark-green color index) exhibited high
correlation (R2 > 0.8) with the SPAD values. Further, subjecting the leaf reflectance information provided by
these color indices to PCA and AHCA enabled a clear segregation of seedlings with high and low chlorophyll
contents. SPAD values predicted by the L∗a∗b∗ color space information yielded the lowest RMSE (root mean
square error) and the highest R2 (coefficient of determination) amongst the six color space features assessed. The
findings of the present study indicate that concatenation of leaf reflectance information provided by different
color indices may be more useful than individual color indices for assessing plant health status and predicting
chlorophyll content using machine vision.

1. Introduction
Environmental stress and nutrient deficiencies faced by plants are
manifested as changes in the pigment composition of leaves (Carter,
1993; Bacci et al., 1998). Decrease in chlorophyll content under con-
ditions of stress results in reduced light absorption and poor photo-
synthetic activity (Gitelson et al., 2003). Unhealthy and senescent
leaves typically exhibit reduced greenness due to lower chlorophyll
content (Carter and Knapp, 2001). Reduction in leaf water content is
also perceptible through changes in leaf spectral reflectance (Ahmad
and Reid, 1996). Since nitrogen is a key constituent of the chlorophyll
molecule, plant nitrogen status is also reflected through leaf coloration
(Pagola et al., 2009; Rorie et al., 2011). Hence, canopy reflectance is
considered as a reliable indicator of plant health status and physiolo-
gical stress (Kawashima and Nakatani, 1998; Carter and Knapp, 2001;
Raun et al., 2001). Assessment of leaf spectral reflectance yields in-
formation about the photosynthetic potential and allows timely detec-
tion of symptoms of stress and nutrient deficiencies in plants (Wood
et al., 1993; Carter and Knapp, 2001; Xu et al., 2011; Riccardi et al.,
2014).
Although the use of chlorophyll meters as well as chlorophyll
fluorometry, thermal imaging and hyperspectral imaging have proven
to be reliable for studying leaf spectral properties and for rapid and
non-destructive assessment of plant health status (Wood et al., 1993;
Mahlein et al., 2012), the techniques require expensive specialized
equipments and are thus out of reach for most cultivators. Advances in
digital image acquisition and processing techniques over the past dec-
ades have equipped us with tools for rapid, non-destructive, high-
throughput and cost-effective real-time quantification of leaf spectral
properties using just a digital camera and a computer (Dutta Gupta

⁎
Corresponding author.
E-mail addresses: avinash.agfe@iitkgp.ac.in, avinash.agarwal.1612@gmail.com (A. Agarwal), sdg@agfe.iitkgp.ac.in (S. Dutta Gupta).

https://doi.org/10.1016/j.compag.2018.06.048
Received 5 June 2017; Received in revised form 14 August 2017; Accepted 28 June 2018
0168-1699/ © 2018 Elsevier B.V. All rights reserved.
A. Agarwal, S. Dutta Gupta
Fig. 1. Steps for acquisition of SPAD values and image features from spinach leaves.
et al., 2014). Leaf digital-image analysis aims at assessing the health of
plants via colorimetric characteristics of the leaves. The close associa-
tion of chlorophyll content with plant health, leaf greenness, photo-
synthetic efficiency and plant nitrogen status forms the basis of image-
based plant health assessment (Rorie et al., 2011). The technique en-
ables the quantification of leaf reflectance information in terms of
various color spaces and numerous chromatic transformations derived
thereof. The color spaces commonly used for digital quantification of
colorimetric information include RGB (red, green and blue), HSI (hue,
saturation and intensity), CIE (Commission Internationale de l’E-
clairage) L*a*b*, CIE-XYZ and CIE-xyY. Characteristic features of these
color spaces have been discussed at length in earlier publications
(Palus, 1998; Gonzalez and Woods, 2002; Ohta and Robertson, 2005;
Schanda, 2007).
RGB color space data contained within digital leaf images has been
used for assessing chlorophyll content (Kawashima and Nakatani, 1998;
Yadav et al., 2010; Hu et al., 2013), plant stress (Bacci et al., 1998;
Ahmad and Reid, 1996), nitrogen content (Pagola et al., 2009; Tewari
et al., 2013; Wang et al., 2013; Saberioon et al., 2014) and macro-
nutrient deficiencies (Wiwart et al., 2009; Xu et al., 2011). The rgb color
features derived from the RGB values have also been used in many such
studies (Ahmad and Reid, 1996; Dutta Gupta et al., 2013; Lee and Lee,
2013; Wang et al., 2013, 2014; Dey et al., 2016). Indicating leaf color
features in terms of hue (H) and saturation (S) in combination with
brightness (B), luminosity (L) and intensity (I) parameters has been
discussed in a few reports (Ahmad and Reid, 1996; Wiwart et al., 2009;
Mata-donjuan et al., 2012; Lee and Lee, 2013; Vesali et al., 2015; Rigon
et al., 2016). The application of L*a*b* color space values for assessing
plant nutrition status has also been demonstrated earlier (Graeff et al.,
2008; Hu et al., 2010; Wang et al., 2014). Assessment of plant stress
using xy chromaticity data obtained from spectrometer (Ruth et al.,
1991) and colorimeter (Bacci et al., 1998) readings has been reported
earlier. However, earlier reports have not addressed the use of CIE-XYZ
tristimulus color space or the derived CIE-xyY color model for evalu-
ating chlorophyll content and assessing plant health using digitized leaf
images.
282
Computers and Electronics in Agriculture 152 (2018) 281–289
Most of the previous reports have credibly correlated leaf chlor-
ophyll content with color indices such as ‘(R−B)/(R + B)’ (Kawashima
and Nakatani, 1998), ‘R + G’ (Hu et al., 2010) and ‘G’ (Yadav et al.,
2010). Dark green color index (DGCI), a colorimetric transformation of
the HSV based image features, has been applied for estimating leaf
chlorophyll content by many scientists (Rorie et al., 2011; Raper et al.,
2012; Saberioon et al., 2014; Vesali et al., 2015; Rigon et al., 2016). The
studies have focused on comparing the potential of individual color
indices in predicting plant health, chlorophyll content and nutrient
status from leaf images. We propose the concatenation of information
acquired from different color indices for segregating the healthy and
stressed plants by multivariate data analysis and predicting chlorophyll
content on the basis of various trichromatic leaf image features.
The objective of this study was to explore the potential of multi-
variate data analysis tools, viz. principal component analysis (PCA) and
agglomerative hierarchical cluster analysis (AHCA), for assessing
seedling health status by distinguishing between spinach seedlings with
high and low chlorophyll contents using digitized leaf image features.
Another aim of the study was to assess the capacity of different image
feature trichromatic values or triplets, viz. RGB, rgb, HSI, CIE-L*a*b*,
CIE-XYZ tristimulus values and the CIE-xyY color features, to predict
chlorophyll content in terms of SPAD values by a multiple linear re-
gression approach.
2. Materials and methods
2.1. Plant material and growth conditions
Seeds of spinach (Spinacia oleracea L. cv. “All Greens”) were pur-
chased from Sutton and Sons India Pvt. Ltd., Kolkata, India. The seeds
were sown in six plastic trays containing soilrite mix (perlite + ver-
miculite + sphagnum peat moss 1:1:1). The plastic trays were kept in a
controlled environment chamber for a period of 35 days under cool
white fluorescent lamps (FL), blue LEDs (BL, 470 nm), red LEDs (RL,
630 nm) and different combinations of blue + red LEDs (3:1, B3RL; 1:1,
BRL; 1:3, BR3L). Variations in leaf chlorophyll content under different
A. Agarwal, S. Dutta Gupta
LED treatments have been well documented for various plant species in
earlier reports (Agarwal and Dutta Gupta, 2016). A photoperiod of 16 h
at 40 µmol.m−2 s−1 photosynthetic photon flux density (PPFD) was
maintained for all light treatments. Trays were irrigated as per re-
quirement and a temperature of 25 ± 2 °C with relative humidity of
55 ± 5% was maintained within the chamber.
2.2. SPAD measurement and leaf image acquisition
Fifty fresh leaves were excised at 35 d from each of the six trays.
Leaf chlorophyll content was determined using a SPAD-502Plus chlor-
ophyll meter (Konica Minolta, Inc., Japan). Ten readings were taken
from each leaf avoiding the midrib and margins for calculating the
mean SPAD value. Each leaf was immediately scanned with a Scanjet
G3110 Photo Scanner (Hewlett Packard, USA) at 300 dpi resolution and
saved as 24-bit bitmap image (.bmp) for feature extraction and analysis
(Fig. 1). The spectral response characteristics and RGB color sensitivity
of the scanner have been provided in the supplementary data (Fig. S1,
Table S1).
Supplementary data associated with this article can be found in the
online version at https://doi.org/10.1016/j.compag.2018.06.048.
2.3. Image feature extraction and calculation of color indices
Leaf images were processed using MATLAB® (ver. R2013a) Image
Processing Toolbox for extracting the color features. The color in-
formation contained within the images was originally stored as RGB
values. These values were used to obtain the HSI, L*a*b* and XYZ color
features. Ten pixels were selected randomly from different regions of
the leaf image avoiding the midrib and margins, as depicted in Fig. 1,
for extracting the image features. The mean image feature values of the
ten pixels were used as the representative image feature values of the
entire leaf.
The expressions for obtaining H (hue), S (saturation) and V (value,
also known as brightness) from the RGB values are as follows (Rorie
et al., 2011):
⎧ 60{(G−B)/(max[R, G, B]−min[R, G, B])}; if max[R, G, B] = R
H = 60{2 + (B−R)/(max[R, G, B]−min[R, G, B])}; if max[R, G, B] = G
⎨
⎩ 60{4 + (R−G)/(max[R, G, B]−min[R, G, B])}; if max[R, G, B] = B
(1)
S = {max[R, G,B]−min[R, G,B]}/max[R, G,B] (2)
V = max[R, G,B] (3)
The values for I (intensity), CIE-XYZ and CIE-L*a*b* were obtained
from the RGB coordinates as follows (Wiwart et al., 2009):
I = (R + G + B)/3 (4)
⎡ X ⎤ ⎡ 0.412453 0.357580 0.180423 ⎤ ⎡ R ⎤
⎢ Y ⎥ = ⎢ 0.212671 0.715160 0.072169 ⎥ × ⎢G ⎥
⎣ Z ⎦ ⎣ 0.019334 0.119193 0.950227 ⎦ ⎣ B ⎦ (5)
for 0 ≤ (R, G, B) ≤ 1.
116.0 × f (Y/Yn)−16.0; if Y/Yn > 0.008856
L* = ⎧
⎨
⎩ 903.3 × (Y/Yn); if Y/Yn ⩽ 0.008856 (6)
a * = {500.0 × (f (X/Xn)−f (Y/Yn))} (7)
b* = {200.0 × (f (Y/Yn)−f (Z/Z n))} (8)
where, f(p) = p1/3, if p > 0.008856; f(p) = (7.787 × p) + (16.0/
116.0), if p ≤ 0.008856. Xn, Yn, Zn indicate the tristimulus values of the
reference white point.
The rgb and xyz features were obtained as follows:
r = R/(R + G + B) (9)
283
Computers and Electronics in Agriculture 152 (2018) 281–289
g = G/(R + G + B) (10)
b = B/(R + G + B) (11)
x= X/(X + Y + Z) (12)
y= Y/(X + Y + Z) (13)
z= Z/(X + Y + Z) (14)
Various color indices, viz. R + G, R + B, G + B, R−G, R−B, G−B,
R + G−B, G/R, G/B, (R−B)/(R + B), (R−G)/(R + G), (G−B)/
(G + B), (R−B)/(R + G + B), (R−G)/(R + G + B), (G−B)/
(R + G + B), were calculated from the obtained RGB values for corre-
lation with leaf SPAD values as reported earlier (Kawashima and
Nakatani, 1998; Yadav et al., 2010; Yuzhu et al., 2011; Hu et al., 2013).
Dark green color index (DGCI) was calculated using the expression
by Karcher and Richardson (2003):
DGCI = [{(H - 60) ÷ 60} + (1−S) + (1−V)]/3 (15)
where, H, S and V stand for hue, saturation and value obtained pre-
viously.
It may be noted that since the G (green) channel predominates the
RGB color space in green leaf images, the information contained by V
(value) is the same as G. The V color feature has only been extracted for
the calculation of DGCI. The term V has been used instead of B
(brightness) to avoid ambiguity with B (blue) from the RGB color space.
2.4. Plant health assessment by multivariate data analysis
Multivariate data analysis of color indices was performed to in-
tegrate the information possessed within the various image features and
provide a consolidated output by segregating the healthy and unhealthy
seedlings on the basis of the leaf image features. The color indices
having high correlation with SPAD values were chosen for the multi-
variate data analyses.
2.4.1. Selection of color indices by correlation with SPAD measurements
The SPAD values of all the 300 leaves were subjected to linear
correlation with the various color indices using Microsoft Excel 2007
Data Analysis ToolPak. The color indices having high correlation with
SPAD measurements (R2 ≥ 0.8) were selected for performing the mul-
tivariate data analyses.
2.4.2. Principal component analysis (PCA)
PCA is a dimensionality reduction tool that creates variables or
principal components (PCs) by the linear combination of all the input
variables, which in this case were the selected color indices. The values
of the selected color indices were normalized in Microsoft Excel 2007
using the “standardize” function. The normalized values were subjected
to PCA using MATLAB® (ver. R2013a) Statistics and Machine Learning
Toolbox. The values of the selected color indices for all the leaves
(n = 300) were uploaded into the software irrespective of the light
treatment to obtain the equations of the PCs in terms of the color in-
dices and the PCA score of each leaf in terms of individual PCs. The
coefficients of the color indices and the average scores of the individual
light treatments obtained by the PCA were visualized by creating a PCA
biplot with the PCs explaining more than ninety percent of the variance.
2.4.3. Agglomerative hierarchical cluster analysis (AHCA)
AHCA creates a relationship tree by calculating the Euclidean dis-
tance, a measure of dissimilarity, between two individual samples in a
pair-wise manner (Day and Edelsbrunner, 1984). Means of color index
values from each light treatment were normalized in Microsoft Excel
2007 and subjected to AHCA using MATLAB® (ver. R2013a) Statistics
and Machine Learning Toolbox. The Euclidean distances between the
image features of individual light treatments were calculated by the
‘nearest neighbor’ method (Day and Edelsbrunner, 1984). The
A. Agarwal, S. Dutta Gupta Computers and Electronics in Agriculture 152 (2018) 281–289

calculation of scaled Euclidean distances, clustering of color index data Table 1

and construction of dendrogram for the resultant hierarchical cluster Correlation coefficients (R) and coefficients of determination (R2) for color
tree were performed using MATLAB® commands. indices and SPAD values.
Color index Correlation coefficient Coefficient of determination
2.5. SPAD prediction using color space triplets (R) (R2)

R −0.90276* 0.81498
The SPAD values of 300 leaves and their corresponding image fea-
G −0.90269* 0.81486
tures were randomly divided into two groups of 60 and 240 (hereafter B 0.00784 0.00006
referred to as input data and target data, respectively) irrespective of r −0.86911* 0.75536
the light treatment. SPAD values of the input data were subjected to g −0.48537* 0.23559
b 0.85514* 0.73127
multiple linear regression with the image feature triplets to obtain the
H 0.85843* 0.73691
model equations as described earlier (Su et al., 2008; Yadav et al., 2010; S −0.66486* 0.44204
Dutta Gupta et al., 2013). The SPAD readings and triplets of input I −0.87297* 0.76209
(n = 60) color space values were subjected to multiple linear regression L* −0.82243* 0.67639
(α = 0.05) using the Microsoft Excel 2007 Data Analysis ToolPak to a* 0.34403* 0.11836
b* −0.69644* 0.48503
obtain the model equation for SPAD in terms of the respective tri-
X −0.89279* 0.79708
chromatic color coordinates. The target (n = 240) image feature values Y −0.89857* 0.80743
were plugged into the corresponding equations to obtain the predicted Z −0.54701* 0.29922
SPAD values. The predicted target SPAD values were correlated with x −0.87976* 0.77398
y −0.79833* 0.63734
the known target SPAD values to assess the accuracy of the prediction
z 0.85115* 0.72445
by calculating the root mean square error (RMSE) and coefficient of R+G −0.90994* 0.82799
determination (R2) values using Microsoft Excel 2007. R+B −0.81402* 0.66263
G+B −0.81553* 0.66510
3. Results and discussion R−G 0.01132 0.00012
R−B −0.90382* 0.81691
G−B −0.90230* 0.81414
3.1. Correlation of color indices and SPAD values R + G−B −0.92337* 0.85262
G/R 0.77496* 0.60056
Correlation analysis of color indices with SPAD values yielded high G/B −0.81075* 0.65732
correlation coefficient values for most of the color indices (Table 1). (R−B)/(R + B) −0.87451* 0.76477
(R−G)/(R + G) −0.776* 0.60217
The indices R, G, r, I, L*, X, Y, x, R + G, R + B, G + B, R−B, G−B, (G−B)/(G + B) −0.82435* 0.67955
R + G−B, G/B, (R−B)/(R + B), (G−B)/(G + B) and (R-B)/ (R−B)/(R + G + B) −0.87565* 0.76677
(R + G + B) exhibit strong negative correlation (R ≤ −0.8) with the (R−G)/(R + G + B) −0.67838* 0.46021
SPAD values. Conversely, b, H, z and DGCI exhibit strong positive (G−B)/(R + G + B) −0.79543* 0.63272
DGCI 0.89549* 0.80191
correlation (R ≥ 0.8) with the SPAD readings. However, only R, G,
R + G, R−B, G−B, R + G−B, Y and DGCI have high coefficients of Values have been calculated for n = 300.
determination (R2 ≥ 0.8) and have been used for the multivariate data * Indicates P < 0.01.
analysis.
As observed in the current study, negative correlation of R and G video-camera (Kawashima and Nakatani, 1998), image scanner (Yadav
with leaf chlorophyll content is also evident from the findings of Yadav et al., 2010) and mobile phone camera (Vesali et al., 2015; Rigon et al.,
et al. (2010) and Hu et al. (2010). Strong correlation between leaf R 2016), under a variety of lighting conditions including natural light
values and nitrogen content, a crucial determinant of chlorophyll bio- (Kawashima and Nakatani, 1998; Riccardi et al., 2014) and artificial
synthesis, has also been reported earlier (Mercado-Luna et al., 2010). lighting via halogen bulb (Hu et al., 2010), fluorescent lamp (Mercado-
Findings of Vollmann et al. (2011) indicate strong negative correlation Luna et al., 2010) and light-emitting diodes (Saberioon et al., 2014).
between G and chlorophyll content. Image analysis experiments con- Variations in lighting conditions during image acquisition significantly
ducted with barley leaves revealed the strong negative correlation of impact the information contained within the image and thus have to be
R + G and R−B with leaf chlorophyll content (Hu et al., 2010). The compensated for by white balancing (Saberioon et al., 2014; Wang
color index G-B has also been shown to correlate negatively with et al., 2014). The present study involves the use of an image scanner
chlorophyll content in wheat and rye (Kawashima and Nakatani, 1998). equipped with a cold-cathode fluorescent lamp which provides constant
The inverse relationship of these RGB color space derived color indices illumination during image acquisition, thus eliminating the need to
with the chlorophyll content may be attributed to the increased re- perform any such adjustments to the leaf images.
flectance in leaves with lower chlorophyll content leading to higher
RGB values. The observations are in concurrence with the findings of
Carter (1993) where stressed leaves exhibited increased reflectance due 3.2. SPAD values and selected color index values
to reduction in chlorophyll content. The color index Y belonging to the
CIE-XYZ color space indicates the luminance of a color (Palus, 1998). The SPAD measurements clearly indicate that seedlings raised under
Since higher chlorophyll content leads to increased light absorbance FL, B3RL, BRL and BR3L have significantly higher chlorophyll contents
and reduced luminance, Y also exhibits a negative correlation with the than the seedlings treated with BL and RL (Table 2). This indicates that
SPAD measurements. Positive correlation of DGCI with leaf chlorophyll both monochromatic blue and monochromatic red LED treatments re-
content observed in the present study is in agreement with earlier re- sult in significant reduction in chlorophyll formation, whereas red and
ports (Rorie et al., 2011; Raper et al., 2012; Rigon et al., 2016). High blue mixed LED treatments as well as white light promote chlorophyll
correlation of these color indices with SPAD measurements makes them biosynthesis in spinach seedlings. The color indices having high nega-
a reliable indicator of leaf chlorophyll content and plant health status. tive correlation with SPAD, viz. R, G, Y, R + G, R−B, G−B and
The studies conducted to assess the correlation of chlorophyll con- R + G−B, have lowest and highest mean values for BR3L and BL
tent with leaf image features have been performed using leaf images treatments, respectively (Table 2). Conversely, the only selected color
acquired through various devices including digital camera (Karcher and index having positive correlation with SPAD, i.e. DGCI, is highest for
Richardson, 2003; Pagola et al., 2009; Dutta Gupta et al., 2013), digital BR3L and lowest for BL treated seedlings. This suggests that the

284
A. Agarwal, S. Dutta Gupta Computers and Electronics in Agriculture 152 (2018) 281–289

Table 2
SPAD and selected color index values of leaves from different treatments.
Light treatment FL BL B3RL BRL BR3L RL

SPAD 36.21 ± 4.73 23.48 ± 3.84 34.27 ± 4.28 37.91 ± 5.36 39.08 ± 5.24 26.53 ± 4.21
R 89.89 ± 10.61 124.29 ± 12.25 81.37 ± 10.94 88.11 ± 14.09 70.07 ± 10.23 121.19 ± 13.13
G 139.46 ± 11.99 175.89 ± 9.39 126.66 ± 10.83 135.21 ± 14.63 115.84 ± 10.21 161.39 ± 11.66
Y 21.61 ± 3.75 36.09 ± 4.71 17.57 ± 3.31 20.34 ± 4.51 14.51 ± 2.66 30.58 ± 4.97
R+G 229.36 ± 22.24 300.18 ± 21.41 208.04 ± 21.25 223.32 ± 28.1 185.92 ± 20.11 282.57 ± 24.36
R−B 7.14 ± 12.48 43.26 ± 12.23 5.35 ± 13.33 11.25 ± 16.95 −10.24 ± 13.32 39.02 ± 13.12
G−B 56.71 ± 13.79 94.86 ± 9.49 50.64 ± 13.05 58.35 ± 17.9 35.53 ± 13.42 79.21 ± 11.44
R + G−B 146.61 ± 23.89 219.15 ± 20.59 132.02 ± 22.9 146.46 ± 30.38 105.61 ± 22.61 200.41 ± 21.04
DGCI 0.64 ± 0.09 0.44 ± 0.05 0.67 ± 0.09 0.63 ± 0.12 0.78 ± 0.1 0.46 ± 0.06

Values indicate mean ± SD, n = 50.

magnitudes of R, G, Y, R + G, R−B, G−B and R + G−B have an in-

verse relationship with chlorophyll content and plant health, whereas
DGCI is directly related to plant health status.

3.3. Plant health assessment by multivariate data analysis of color indices

The color indices exhibiting very high correlation with SPAD values,
viz. R, G, R + G, R−B, G−B, R + G−B, Y and DGCI, were used for
performing the PCA and AHCA. The results obtained from the two
analyses have been compared to the information provided by the mean
SPAD values (Table 2).

3.3.1. Principal component analysis

The first two principal components are able to explain almost
ninety-nine percent of the variance (Table 3). Thus, we have ignored
the remaining principal components for preparing the PCA biplot. It is
noteworthy that the Eigenvalue of PC1 (7.69) and variance explained
by PC1 (96.133%) are very high as compared to PC2, implying that PC1 Fig. 2. Biplot of PC1-PC2 obtained by the PCA of selected color indices.
may be considered as the major principal component in current study.
The expressions for the first two principal components, i.e. PC1 and
(Eq. (16)) and thus lie in the positive direction of PC1 on the biplot.
PC2, are as follows:
Conversely, DGCI, which has a positive correlation with the SPAD va-
PC1 = 0.3538 × R + 0.3544 × G + 0.3515 × Y + 0.3569 × (R + G) + 0.3507 × (R - B) lues (Table 1), has a negative loading on PC1 (Eq. (16)) and thus lies in
+ 0.3508 × (G - B) + 0.3604 × (R + G - B)−0.3497 × DGCI (16) the negative direction of PC1 (Fig. 2). Thus, it may be inferred that the
PC1 axis possesses an anti-parallel orientation with respect to SPAD.
PC2 = 0.2538 × R + 0.3241 × G + 0.4443 × Y + 0.2913 × (R+ G)−0.4315 × (R - B) Since PC1 itself explains more than 95% of the variance in data
−0.3586 × (G - B)−0.0546 × (R + G - B) + 0.4802 × DGCI (17) (Table 3) the present discussion focuses on PC1 only. The biplot in-
dicates that the treatments having negative PC1 scores (Fig. 2) consist
The biplot of PC1 and PC2 obtained by the PCA of the selected color
of seedlings with higher SPAD values (Table 2) and better health status,
indices reveals two distinct groups of treatments (Fig. 2), one having BL
and vice versa. Thus, the treatments lying in the negative direction of
and RL, and the other having BR3L, B3RL, BRL and FL. The treatments
PC1, viz. FL, B3RL, BRL and BR3L, possess healthier seedlings with high
having high R, G, Y, R + G, R−B, G−B and R + G−B values (Table 2),
chlorophyll content, whereas BL and RL treatments, lying along the
viz. BL and RL, are present in the same half of the biplot as these color
positive direction of PC1, comprise of seedlings with low chlorophyll
indices, i.e. along the positive direction of PC1 axis. Similarly, treat-
content. This implies that the PCA-based methodology can reliably
ments with high DGCI values (Table 2), viz. FL, B3RL, BRL and BR3L,
segregate the healthy and unhealthy seedlings on the basis of leaf
are present along the negative PC1 axis along with DGCI itself.
spectral properties. The procedure allows a lucid visualization of leaf
A clear distinction in the spectral reflectance characteristics of FL,
image information contained within eight variables or color indices in
B3RL, BRL and BR3L from the BL and RL treated seedlings is evident
the form of two PCs. It must be noted that the technique described in
from the biplot (Fig. 2). It is noteworthy that the selected color indices
the present study employs linear PCA as the selected color indices ex-
having negative correlation with the SPAD values, viz. R, G, R + G,
hibited significant linear correlation with the leaf SPAD values. Non-
R−B, G−B, R + G−B and Y (Table 1), have a positive loading on PC1
linear PCA may be applied to datasets having non-linear and complex
interrelations (Linting et al., 2007). The use of PCA for correlating
Table 3
Eigenvalues of the principal components obtained by the PCA of selected color
disease resistance and grain yield characteristics using canopy image
indices and the percentage of variance explained. features of wheat has been discussed earlier (Zhou et al., 2015). Pagola
et al. (2009) have used PCA for deriving a new color index from RGB
Principal Eigenvalue Variance Cumulative variance
image features of barley leaves for assessing the nitrogen nutrition
component explained (%) explained (%)
status.
PC1 7.69 96.133 96.133
PC2 0.22 2.746 98.879 3.3.2. Agglomerative hierarchical cluster analysis
PC3 0.07 0.852 99.731
The scaled Euclidean distances, obtained by the AHCA, between the
PC4 0.02 0.238 99.969
PC5 0.00 0.031 100 spectral properties of the seedlings from different light treatments have
been presented in Table 4. The color indices of BL and RL exhibit high

285
A. Agarwal, S. Dutta Gupta Computers and Electronics in Agriculture 152 (2018) 281–289

Table 4 It may thus be inferred that the results obtained by the multivariate
Scaled Euclidean distances between the spectral properties of the leaves from data analyses, PCA and AHCA, performed using the selected color in-
different light treatments. dices, viz. R, G, R + G, R−B, G−B, R + G−B, Y and DGCI, are well in
Case FL BL B3RL BRL BR3L RL agreement with the information obtained from the SPAD measure-
ments.
FL 0.000
BL 0.615 0.000
B3RL 0.101 0.759 0.000
3.4. SPAD prediction using image feature triplets
BRL 0.000 0.625 0.081 0.000
BR3L 0.333 1.000 0.197 0.328 0.000 Multiple linear regression of the six color space trichromatic values
RL 0.448 0.139 0.587 0.454 0.830 0.000 yielded the following model equations for predicting SPAD values:
SPAD = 54.96376 - (0.18009×R)−(0.12136×G) + (0.15399×B) (18)
dissimilarity from FL, B3RL, BRL and BR3L. The color indices of FL and
SPAD = (0 × r ) + (177.95266 × g ) + (212.39774 × b)−101.20482 (19)
BRL have the least dissimilarity as indicated by the scaled Euclidean
distance. SPAD = 51.61614 + (0.13298×H)−(19.63592×S)−(0.23208×I) (20)
The dendrogram obtained by the AHCA of the color indices exhibits
a clear segregation into two major clusters, one containing FL, B3RL, SPAD = 77.61534 - (0.51851 × L*)−(0.14011 × a *)−(0.57617 × b*) (21)
BRL and BR3L treatments and the other containing the BL and RL
SPAD = 41.46242 - (1.29096×X)−(0.07739×Y) + (1.3128×Z) (22)
treatments (Fig. 3). It is noteworthy that the two major clusters ex-
hibited by the dendrogram have the same light treatments as the two SPAD = 95.06977 - (144.94674×x)−(12.12719×y)−(0.44925×Y) (23)
groups obtained in the PCA biplot (Fig. 2). The close association of the
spectral properties of leaves from the FL, B3RL, BRL and BR3L is evi- The SPAD values predicted using these equations yielded the lowest
dent from the low (< 0.4) Euclidean distances (Table 4). The ob- RMSE for the L*a*b* model equation (Eq. (21)) whereas the rgb model
servation is well in agreement with the information provided by the (Eq. (19)) exhibited comparatively higher RMSE than the rest of the
SPAD values (Table 2). The differences in mean SPAD values of various image feature triplets (Table 5). The coefficient of determination (R2)
treatments generally exhibit a similar pattern as the Euclidean dis- values indicate high degree of accuracy in projecting SPAD values from
tances. The light treatments having the highest Euclidean distance, viz. RGB, HSI, L*a*b*, XYZ and xyY color space values. The plots of actual
BL and BR3L, also have the highest difference between their mean versus predicted SPAD values obtained by the multiple linear regression
SPAD values (Table 2). Further difference in the mean SPAD values is models are presented in Fig. 4(a–f).
comparatively lower for treatments having low Euclidean distances. The interrelationship between plant health status, chlorophyll con-
Thus seedlings with high chlorophyll content can be easily segregated tent and leaf spectral reflectance has been well established (Daughtry
from the seedlings with low chlorophyll content by the present clus- et al., 2000; Carter and Knapp, 2001; Gitelson et al., 2003). Predict-
tering algorithm. However, unlike PCA, this technique does not indicate ability of chlorophyll content and SPAD values from leaf color indices
which cluster contains the healthy seedlings. Consequently, the same has been explored in previous studies (Yadav et al., 2010; Hu et al.,
has to be determined from the absolute color index values (Table 2) of 2013; Vesali et al., 2015). The use of color feature triplets for predicting
the samples having the highest Euclidean distance, i.e. BR3L and BL. chlorophyll content in the current study has been adapted from the
The high DGCI and low R, G, R + G, R−B, G−B, R + G−B and Y three-color analysis methodology applied for evaluating the chlorophyll
values of BR3L indicate that the cluster containing the BR3L treatment content of algae (Su et al., 2008). A similar RGB-based chlorophyll
also contains treatments with healthier seedlings. Conversely, low DGCI prediction model using multiple linear regression has also been dis-
value and high R, G, R + G, R−B, G−B, R + G−B and Y values of BL cussed by Yadav et al. (2010) where the accuracy of RGB and rgb tri-
imply poorer health status and low chlorophyll content of the seedlings plets in predicting the chlorophyll content of potato leaves has been
present in the pertinent cluster (Fig. 3). A similar clustering algorithm assessed. Riccardi et al. (2014) have employed a non-linear multiple
was adopted by Wiwart et al. (2009) to assess the changes in leaf regression model for predicting chlorophyll content from leaf image
spectral characteristics of three legume species under macronutrient features. Their study further indicated that RGB color modeling could
deficiencies. predict leaf chlorophyll content more accurately than a SPAD meter at
high chlorophyll concentrations. Recent studies have also demonstrated
the use of artificial neural networks for predicting chlorophyll content
from leaf image RGB features via non-linear modeling (Dutta Gupta
et al., 2013; Vesali et al., 2015; Dutta Gupta and Pattanayak, 2017).
In the present study, the performance of model equations obtained
by multiple linear regression of various color space values with SPAD
readings has been evaluated. Amongst the six image feature triplets,
viz. RGB, rgb, HSI, L*a*b*, XYZ and xyY, the equation obtained by the
L*a*b* color space parameters (Eq. (21)) yielded lowest RMSE values
and highest correlation between the predicted and the actual SPAD

Table 5
RMSE and R2 values of actual SPAD measurements vs SPAD values predicted by
various image feature triplets.
Image features used RMSE R2

RGB 2.796 0.854

rgb 3.552 0.767
HSI 2.903 0.844
L*a*b* 2.684 0.865
XYZ 2.829 0.850
Fig. 3. Dendrogram obtained by the hierarchical cluster analysis of the selected
xyY 2.776 0.856
color indices.

286
A. Agarwal, S. Dutta Gupta
Fig. 4. Plots of measured SPAD values vs predicted SPAD values for RGB (a), rgb (b), HSI (c), L*a*b* (d), XYZ (e) and xyY (f) color space models (n = 240).
values (Table 5). The L*a*b* color space is a perceptually uniform color
space having close semblance to the human vision and is derived from
the XYZ chromaticity coordinates (Palus, 1998). Here, L* takes into
consideration the lightness of the color, whereas a* and b* indicate the
redness-greenness and the yellowness-blueness features, respectively.
Although the L*, a* and b* parameters did not yield very high correla-
tion with the SPAD values individually (Table 1), the information
provided by their linear combination (Eq. (21)) is able to predict SPAD
values with high accuracy (Table 5). The observation clearly indicates
that the complete information about the leaf color defined in a parti-
cular color space may not be perceptible from the individual con-
stituent color indices.
The model equations derived from the RGB, HSI, XYZ and xyY color
spaces (Eqs. (18), (20), (22), (23)) were able to predict the chlorophyll
content reliably, as indicated by the RMSE and R2 values (Table 5).
287
Computers and Electronics in Agriculture 152 (2018) 281–289
Amongst the six color spaces considered in the present study, the model
equation obtained from the rgb coordinates (Eq. (19)) yielded the
highest RMSE and lowest R2 values (Table 5). It is noteworthy that in
the model equation obtained by the rgb color space coordinates (Eq.
(19)) the coefficient of r is zero. This may be due to the fact that since
r + g + b = 1, or r = 1 – g – b, only b and g values are sufficient for
providing the complete information about the color space (Palus,
1998). The performance of rgb models has been reported to be better
than the RGB model in previous studies (Yadav et al., 2010; Dutta
Gupta et al., 2013; Dutta Gupta and Pattanayak, 2017). The deviation
of the present findings from the earlier reports may be due to differ-
ences in the spectral properties of the plant species and the image
processing methodologies employed. It has been stated that the rgb
color space is more stable under changing illumination (Palus, 1998).
However, since the leaf images have been collected under constant
A. Agarwal, S. Dutta Gupta
illumination, the efficacy of the rgb color space information was not
fully perceptible in the present investigation.
To check the validity of the discussed method for distributions of
data apart from the 60–240 grouping shown above, combinations of
150–150 and 240–60 input and target values were also tested (data not
shown). As compared to the 60–240 distribution, RMSE values in-
creased < 3% for the 150–150, and < 10% for the 240–60 sample
groupings. A variation of < 2% for R2 was observed between the
60–240, 150–150 and 240–60 distributions. A similar trend in the
RMSE and R2 values was obtained for the six color spaces for all three
groupings. This indicates that the method is valid for different sample
groups as well.
4. Conclusion
Assessment of plant health status from the visual cues provided by
leaf spectral reflectance using image analysis techniques has been well
explored in the past two decades. The technology provides a tool for
rapid and cost-effective assessment of chlorophyll content. In the pre-
sent study multivariate data analysis tools were employed to distinguish
between healthy and stressed spinach seedlings on the basis of leaf
image features. Treatments were successfully segregated into two
groups having high and low chlorophyll contents, respectively, by the
PCA and AHCA of digital image features. The technique described in the
present study may be applied for quick and easy detection of stressed
seedlings or seedlings with low chlorophyll content from a large sample
size. To our knowledge, this is the first report demonstrating the ap-
plication of PCA for segregating leaves on the basis of chlorophyll using
digital color features. The multiple linear regression approach for pre-
dicting chlorophyll content from leaf images revealed that combining
the information provided by the three individual constituents of any
color space may yield information regarding the leaf chlorophyll con-
tent reliably even when the constituent color indices fail to do so in-
dependently. The techniques discussed in the present study provide an
alternative approach for simple, cost-effective and high-throughput
plant health assessment using machine vision. Application of these
procedures to different plant systems using portable imaging devices
having a fixed light source for on-site image acquisition under con-
trolled illumination shall broaden the scope of the study further.
References
Agarwal, A., Dutta Gupta, S., 2016. Impact of light-emitting diodes (LEDs) and its po-
tential on plant growth and development in controlled-environment plant production
system. Curr. Biotechnol. 5, 28–43.
Ahmad, I.S., Reid, J.F., 1996. Evaluation of colour representations for maize images. J.
Agric. Eng. Res. 63, 185–195.
Bacci, L., De Vincenzi, M., Rapi, B., Arca, B., Benincasa, F., 1998. Two methods for the
analysis of colorimetric components applied to plant stress monitoring. Comput.
Electron. Agric. 19, 167–186.
Carter, G.A., Knapp, A.K., 2001. Leaf optical properties in higher plants: linking spectral
characteristics to stress and chlorophyll concentration. Am. J. Bot. 88, 677–684.
Carter, G.A., 1993. Responses of leaf spectral reflectance to plant stress. Am. J. Bot. 80,
239–243.
Daughtry, C.S.T., Walthall, C.L., Kim, M.S., De Colstoun, E.B., McMurtrey, J.E., 2000.
Estimating corn leaf chlorophyll concentration from leaf and canopy reflectance.
Remote Sens. Environ. 74, 229–239.
Day, W.H., Edelsbrunner, H., 1984. Efficient algorithms for agglomerative hierarchical
clustering methods. J. Classif. 1, 7–24.
Dey, A.K., Sharma, M., Meshram, M.R., 2016. An analysis of leaf chlorophyll measure-
ment method using chlorophyll meter and image processing technique. Proced.
Comput. Sci. 85, 286–292.
Dutta Gupta, S., Ibaraki, Y., Pattanayak, A.K., 2013. Development of a digital image
analysis method for real-time estimation of chlorophyll content in micropropagated
potato plants. Plant Biotechnol. Rep. 7, 91–97.
Dutta Gupta, S., Ibaraki, Y., Trivedi, P., 2014. Applications of RGB color imaging in
plants. In: Dutta Gupta, S., Ibaraki, Y. (Eds.), Plant Image Analysis: Fundamentals and
Applications. CRC Press, Florida, USA, pp. 41–62.
Dutta Gupta, S., Pattanayak, A.K., 2017. Intelligent image analysis (IIA) using artificial
neural network (ANN) for non-invasive estimation of chlorophyll content in micro-
propagated plants of potato. In Vitro Cell Dev. Biol. – Plant. https://doi.org/10.1007/
s11627-017-9825-6.
Gitelson, A.A., Gritz, Y., Merzlyak, M.N., 2003. Relationships between leaf chlorophyll
288
Computers and Electronics in Agriculture 152 (2018) 281–289
content and spectral reflectance and algorithms for non-destructive chlorophyll as-
sessment in higher plant leaves. J. Plant Physiol. 160, 271–282.
Gonzalez, R.C., Woods, R.E., 2002. Digital Image Processing, second ed. Prentice Hall,
New Jersey, USA.
Graeff, S., Pfenning, J., Claupein, W., Liebig, H.P., 2008. Evaluation of image analysis to
determine the N-fertilizer demand of broccoli plants (Brassica oleracea convar. bo-
trytis var. italica). Adv. Opt. Technol. https://doi.org/10.1155/2008/359760.
Hu, H., Liu, H.Q., Zhang, H., Zhu, J.H., Yao, X.G., Zhang, X.B., Zheng, K.F., 2010.
Assessment of chlorophyll content based on image color analysis, comparison with
SPAD-502. In: Proceedings of 2nd International Conference on Information
Engineering and Computer Science (ICIECS), Wuhan, China. doi: 10.1109/ICIECS.
2010.5678413.
Hu, H., Zhang, J., Sun, X., Zhang, X., 2013. Estimation of leaf chlorophyll content of rice
using image color analysis. Can. J. Remote Sens. 39, 185–190.
Karcher, D.E., Richardson, M.D., 2003. Quantifying turfgrass color using digital image
analysis. Crop Sci. 43, 943–951.
Kawashima, S., Nakatani, M., 1998. An algorithm for estimating chlorophyll content in
leaves using a video camera. Ann. Bot. 81, 49–54.
Lee, K.J., Lee, B.W., 2013. Estimation of rice growth and nitrogen nutrition status using
color digital camera image analysis. Eur. J. Agron. 48, 57–65.
Linting, M., Meulman, J.J., Groenen, P.J., van der Kooij, A.J., 2007. Nonlinear principal
components analysis: introduction and application. Psychol. Methods 12 (3),
336–358.
Mahlein, A.K., Oerke, E.C., Steiner, U., Dehne, H.W., 2012. Recent advances in sensing
plant diseases for precision crop protection. Eur. J. Plant Pathol. 133, 197–209.
Mata-Donjuan, G.F., Mercado-Luna, A., Rico-García, E., Herrera-Ruiz, G., 2012. Use of
improved hue, luminance and saturation (IHLS) color space in the estimation of ni-
trogen on tomato seedlings (Lycopersicon esculentum). Sci. Res. Essays 7, 2343–2349.
Mercado-Luna, A., Rico-García, E., Lara-Herrera, A., Soto-Zarazúa, G., Ocampo-
Velázquez, R., Guevara-González, R., Herrera-Ruiz, G., Torres-Pacheco, I., 2010.
Nitrogen determination on tomato (Lycopersicon esculentum Mill.) seedlings by color
image analysis (RGB). Afr. J. Biotechnol. 9, 5326–5332.
Ohta, N., Robertson, A. (Eds.), 2005. Colorimetry: Fundamentals and Applications. John
Wiley and Sons Ltd., West Sussex, England.
Pagola, M., Ortiz, R., Irigoyen, I., Bustince, H., Barrenechea, E., Aparicio-Tejo, P.,
Lamsfus, C., Lasa, B., 2009. New method to assess barley nitrogen nutrition status
based on image colour analysis: comparison with SPAD-502. Comput. Electron. Agric.
65, 213–218.
Palus, H., 1998. Representations of colour images in different colour spaces. In: Sangwine,
S.J., Horne, R.E.N. (Eds.), The Colour Image Processing Handbook. Chapman and
Hall, London, UK, pp. 67–90.
Raper, T.B., Oosterhuis, D.M., Siddons, U., Purcell, L.C., Mozaffari, M., 2012.
Effectiveness of the dark green color index in determining cotton nitrogen status from
multiple camera angles. Int. J. Appl. Sci. Technol. 2, 71–74.
Raun, W.R., Solie, J.B., Johnson, G.V., Stone, M.L., Lukina, E.V., Thomason, W.E.,
Schepers, J.S., 2001. In-season prediction of potential grain yield in winter wheat
using canopy reflectance. Agron. J. 93, 131–138.
Riccardi, M., Mele, G., Pulvento, C., Lavini, A., d’Andria, R., Jacobsen, S.E., 2014. Non-
destructive evaluation of chlorophyll content in quinoa and amaranth leaves by
simple and multiple regression analysis of RGB image components. Photosynth. Res.
120, 263–272.
Rigon, J.P.G., Capuani, S., Fernandes, D.M., Guimarães, T.M., 2016. A novel method for
the estimation of soybean chlorophyll content using a smartphone and image ana-
lysis. Photosynthetica 54, 559–566.
Rorie, R.L., Purcell, L.C., Mozaffari, M., Karcher, D.E., King, C.A., Marsh, M.C., Longer,
D.E., 2011. Association of “greenness” in corn with yield and leaf nitrogen con-
centration. Agron. J. 103, 529–535.
Ruth, B., Hoque, E., Weisel, B., Hutzler, P.J., 1991. Reflectance and fluorescence para-
meters of needles of Norway spruce affected by forest decline. Remote Sens. Environ.
38, 35–44.
Saberioon, M.M., Amin, M.S.M., Anuar, A.R., Gholizadeh, A., Wayayok, A., Khairunniza-
Bejo, S., 2014. Assessment of rice leaf chlorophyll content using visible bands at
different growth stages at both the leaf and canopy scale. Int. J. Appl. Earth Obs.
Geoinform. 32, 35–45.
Schanda, J. (Ed.), 2007. Colorimetry: Understanding the CIE System. John Wiley & Sons
Ltd., New Jersey, USA.
Su, C.H., Fu, C.C., Chang, Y.C., Nair, G.R., Ye, J.L., Chu, I., Wu, W.T., 2008. Simultaneous
estimation of chlorophyll a and lipid contents in microalgae by three-color analysis.
Biotechnol. Bioeng. 99, 1034–1039.
Tewari, V.K., Arudra, A.K., Kumar, S.P., Pandey, V., Chandel, N.S., 2013. Estimation of
plant nitrogen content using digital image processing. Agric. Eng. Int.: CIGR J. 15,
78–86.
Vesali, F., Omid, M., Kaleita, A., Mobli, H., 2015. Development of an android app to
estimate chlorophyll content of corn leaves based on contact imaging. Comput.
Electron. Agric. 116, 211–220.
Vollmann, J., Walter, H., Sato, T., Schweiger, P., 2011. Digital image analysis and
chlorophyll metering for phenotyping the effects of nodulation in soybean. Comput.
Electron. Agric. 75, 190–195.
Wang, Y., Wang, D., Shi, P., Omasa, K., 2014. Estimating rice chlorophyll content and leaf
nitrogen concentration with a digital still color camera under natural light. Plant
Methods 10, 36.
Wang, Y., Wang, D., Zhang, G., Wang, J., 2013. Estimating nitrogen status of rice using
the image segmentation of GR thresholding method. Field Crops Res. 149, 33–39.
Wiwart, M., Fordoński, G., Żuk-Gołaszewska, K., Suchowilska, E., 2009. Early diagnostics
of macronutrient deficiencies in three legume species by color image analysis.
Comput. Electron. Agric. 65, 125–132.
A. Agarwal, S. Dutta Gupta
Wood, C.W., Reeves, D.W., Himelrick, D.G., 1993. Relationships between chlorophyll
meter readings and leaf chlorophyll concentration, N status, and crop yield: a review.
Proc. Agron. Soc. N.Z. 23, 1–9.
Xu, G., Zhang, F., Shah, S.G., Ye, Y., Mao, H., 2011. Use of leaf color images to identify
nitrogen and potassium deficient tomatoes. Pattern Recognit. Lett. 32, 1584–1590.
Yadav, S.P., Ibaraki, Y., Dutta Gupta, S., 2010. Estimation of the chlorophyll content of
micropropagated potato plants using RGB based image analysis. Plant Cell Tissue
289
Computers and Electronics in Agriculture 152 (2018) 281–289
Organ Cult. 100, 183–188.
Yuzhu, H., Xiaomei, W., Shuyao, S., 2011. Nitrogen determination in pepper (Capsicum
frutescens L.) plants by color image analysis (RGB). Afr. J. Biotechnol. 10,
17737–17741.
Zhou, B., Elazab, A., Bort, J., Vergara, O., Serret, M.D., Araus, J.L., 2015. Low-cost as-
sessment of wheat resistance to yellow rust through conventional RGB images.
Comput. Electron. Agric. 116, 20–29.