biosystems engineering 106 (2010) 223–233
Available at www.sciencedirect.com
journal homepage: www.elsevier.com/locate/issn/15375110
Research Paper
Neural-network model for estimating leaf chlorophyll
concentration in rice under stress from heavy metals using
four spectral indices
Meiling Liu a, Xiangnan Liu a,*, Mi Li a, Meihong Fang a, Wenxue Chi b
a
b
School of Information Engineering, China University of Geosciences, Beijing 100083, China
Meteorological Observation Center of CMA, Beijing 100081, China
article info
Heavy metal stress in soils results in subtle changes in leaf chlorophyll concentration,
Article history:
which are related to crop growth and crop yield. Accurate estimation of the chlorophyll
Received 1 September 2009
concentration of a crop under heavy metal stress is essential for precision crop production.
Received in revised form
The objective of this paper is to create a back propagation (BP) neural-network model to
1 December 2009
estimate chlorophyll concentration in rice under heavy metal stress. Three experiment
Accepted 14 December 2009
farms located in Changchun, Jilin Province, China with level II pollution, with level I
Published online 15 May 2010
pollution and with safe level were selected, The assessment was based on the input
parameters normalised difference vegetation index (NDVI), optimized soil-adjusted vegetation index (OSAVI), modified triangle vegetation index/modified chlorophyll absorption
ratio index (MTVI/MCARI), MTVI/OSAVI and the output parameters of rice leaf chlorophyll
concentration. The output parameters were sensitive to heavy metal stress. The result
indicated that an optimum BP neural-network prediction model has 4-10-2-1 network
architecture with gradient descent learning algorithm and an activation function including
the sigmoid tangent function in the input layer, a hidden layer and sigmoid logistic
functions in the output layer. The correlation coefficient (R2) between the measured
chlorophyll concentration and the predicated chlorophyll concentration was 0.9014, and
the root mean square error (RMSE) was 2.58.
ª 2009 IAgrE. Published by Elsevier Ltd. All rights reserved.
1.
Introduction
With the development of industry, some pollutants have
affected farms. Heavy metals from industrial pollutants have
caused stress to crops adjacent to industrial areas. Heavy
metal stress is characterised by the reduction of the enzymes
involved in chlorophyll biosynthesis and induced photosynthetic inhibition (De Filipps & Pallaghy, 1994), membrane
disruption (Droppa & Horvath, 1990), substitution of metal
ions for the chlorophyll molecule and the reduction in leaf
chlorophyll concentration (Kastori et al., 1998). Heavy metal
stress results in crop growth restriction and crop yield loss. It
is therefore essential and urgent that measures are made to
detect crop heavy metal stress.
Remote sensing is a valid alternative to traditional groundbased methods to detect plant stress, especially with the
* Corresponding author.
E-mail addresses: [email protected] (M. Liu), [email protected] (X. Liu), [email protected] (M. Li), [email protected]
(M. Fang).
1537-5110/$ – see front matter ª 2009 IAgrE. Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.biosystemseng.2009.12.008
224
biosystems engineering 106 (2010) 223–233
Nomenclature
xi, xmin, xmax, xnorm real, minimum, maximum, normalised
value of an input variable
yai, ymin, ymax, ynorm, ymi real, minimum, maximum,
indexed value of an output variable, measured
output variable
emergence of hyperspectral data. Some research has been
conducted to establish a relationship between heavy metal
content in abandoned mines and reflectance characteristics,
biochemical composition, and pigment content of the plants
respectively (Choe et al., 2008; Gan, Liu, & Zhou, 2004; Liu, Gan,
& Wang, 2004). Studies that examined plant response to heavy
metal-contaminated soil have generally focused on mines
(Ellis & Eslick, 1997). Only a few studies have focused on
farmland under heavy metal stress. Recently, a few
researchers have demonstrated the correlation between crop
spectral characteristics and crop chlorophyll content under
copper or lead contamination stress (Chi, Liu, Liu, & Yang,
2006; Li, Yang, Zhang, Zhang, & Zhou, 2008; Ren, Zhuang, &
Pan, 2008). However, these experiments were performed
under laboratory conditions by adding copper, lead, chromium, or zinc, etc. The fact that most studies have been
conducted under laboratory conditions, rather than under
field conditions, is one of the primary stumbling blocks in
applying hyperspectral remote sensing to precision
agriculture.
The level of the pollutants found in farmlands is much
lower than the level found in abandoned mines or used in
laboratory-based experiments. Alterations in plant biochemistry and cellular composition imposed by heavy metal stress
are subtle, and the spectral differences derived from heavy
metal stress are not obvious; hence using conventional
methods utilising spectral characteristic analysis has presented a challenge with respect to detecting crop heavy metal
stress. Therefore, priority should be given to selecting effective methods for detecting variations in biochemical composition. It is well known that artificial neural-network models
have advantages in determining the input–output relationship for complex systems based on the strength of their
interconnections presented in a set of sample data (Howard &
Mark, 2000). Tumbo, Wagner, and Heinemann (2002) developed a back propagation (BP) neural-network model to estimate chlorophyll content based on spectrometry readings,
and verified that the estimated values from this neuralnetwork model correlated with chlorophyll measurements
reasonably. Wang and Thai (2003) created a generic algorithm
model to select an optimal second-order polynomial model to
assess crop nitrogen stress based on a ratio of leaf reflectance
measured at 700 nm and 740 nm. Noh, Zhang, Shin, Han, and
Feng (2006) used a BP neural-network model to assess crop
nitrogen stress based on maize canopy reflectance in three
channels (green, red, and near-infrared – NIR) of a multispectral image sensor.
Significant progress has been made in this field. However,
there has been little research on the variations in leaf chlorophyll
concentration under various levels of heavy metal stress,
RSME
root mean square error between real output
variable and measured output variable
wjk, wij, h, a connection weights of hidden layer and output
layer, input layer and hidden layer, learning rate
and momentum coefficient
especially for crops growing in field situations rather than under
laboratory conditions. In this paper, we propose a neuralnetwork model to retrieve leaf chlorophyll concentration accurately from a rice crop grown under heavy metal stress using
performance spectral indices from hyperspectral data.
2.
Materials and methods
2.1.
Field experiment design
The city of Changchun, Jilin Province in China is an important
industrial district and agricultural district. Some areas have
been contaminated by industrial pollutants; among the most
important of these are heavy metals. Suburban farms have
soils with copper and cadmium at concentrations above those
Fig. 1 – Location map for experimental sites in Changchun,
Jilin Province, China.
225
biosystems engineering 106 (2010) 223–233
Table 1 – The information of the experiment sites.
Geographical location
Copper content
(mg kg1)
Cadmium content
(mg kg1)
Pollution level
Soil quality
standarda (mg kg1)
68.2
45.5
20.4
0.465
0.182
0.093
II
I
Safe
II (50 Cu 400; 0.3 Cd 0.6)
I (35 Cu < 50; 0.2 Cd < 0.3)
Safe (Cu < 20.8; Cd < 0.097)
43 52.20 N, 125 10.20 E
43 54.60 N, 125 10.40 E
44 06.30 N, 125 10.20 E
a Soil quality standard according to the Environment Monitoring Centre of China.
considered to be normal for the area. Three experimental sites
(43 510 34.800 N–43 510 37.000 N, 125 090 07.200 E–125 100 25.300 E) adjacent to the China First Automobile Factory (i.e., the contamination source) in Changchun were selected (Fig. 1). Heavy
metal contamination stress levels in the soil of the three field
experiments (labelled A, B, and C) varied. The soil and the
stress rates were determined as safe level, level I pollution and
level II pollution according to a soil sample analysis (Table 1).
The site is within the temperate continental climate zone with
a mean annual rainfall of 522–615 mm, where soils are
dominated by black soils with sufficient organic matter (2–4%).
The crop selected in this site was rice which is one of the most
important crops in China.
2.2.
Data preparation
2.2.1.
Spectral data measurement
The data collection was carried out during four days during
a typical rice growth season: 8 July, 4 August, 29 August and 18
September 2008, which corresponded to the seeding, tillering,
booting and mature growth stages of rice. All spectral
measurements were taken under cloudless or near cloudless
conditions between 10:00 and 14:00, using an ASD FieldSpec
Pro spectrometer (Analytical Spectral Devices, Boulder, CO.
USA). The spectrometer was fitted with 10 field of view fibre
optics, operated in the 350–2500 nm spectral region with
sampling intervals of 2 nm. A BaSO4 calibration panel was
used for determining the black and baseline reflectance. A
panel radiance measurement was taken before and after rice
measurement using two scans each time. Rice radiance
measurements were made at 20 sites in each plot and each
site was scanned 10 times. These measurements were then
averaged for the particular plot. The average reflectance
curves for the three field experiments are shown in Fig. 2.
2.2.2.
Corporation, NJ, USA). The chlorophyll content in rice leaves
was measured at six randomly chosen times and the average
value for each sample was calculated. The chlorophyll
concentration was calculated from the SPAD-502 chlorophyll
readings (Wood, Reeves,& Himelrick, 1993), using the
following equation:
y ¼ 0:996x 1:52
where y is the chlorophyll concentration and x is the SPAD-502
chlorophyll readings (mg cm2).
2.3.
The effect of Cu and Cd contamination on rice
Cd is a non-essential element that can be easily absorbed by
plants. While Cu is an essential element, high doses can
adversely affect plant growth (Pahlsson, 1989). The interaction
between Cu and Cd is complex and has an effect on their
individual functions. Huang, Hu, and Liu (2009) demonstrated
that soil application of Cu greatly enhanced Cd accumulation,
but the application of Cd had a negligible effect on Cu uptake
by rice plants. Biochemical, physiological, and structural
aspects affected by excess Cu and Cd in rice included
following: (1) Cd and Cu can inhibit the activities of several
antioxidative enzymes including superoxide dismutase (SOD),
ascorbate peroxidase (APOD), and glutathione reductase (GR)
(Fernandes & Henriques, 1991); (2) they can inhibit the
stomatal opening (Barcelo & Poschenrieder, 1990); (3) they
damage the photosynthetic apparatus (Chien & Kao, 2000;
Krupa, 1988; Sidlecka & Baszynsky, 1993); (4) they are
substituted for magnesium (Mg), the central atom of chlorophyll, and then reduce the chlorophyll content (Chien, Wang,
Metal content and leaf chlorophyll measurement
The spectral measurements were correlated with measurements of the chlorophyll content of the rice and heavy metal
concentrations. The metal content in the soil was analysed by
the Chinese Academy of Agricultural Sciences, Beijing, China.
The soils samples were firstly digested by reverse aqua-regia
(HNO3: HCl ¼ 3:1) (Stevens, 2002), then Cu was measured using
a flame atomic absorption spectrophotometer (AAS) (Spectr
AA 110/220, Varian, USA) and Cd with a graphite furnace AAS.
Soil samples were also analysed for diethylenetriaminepentaacetic acid (DTPA) extractable heavy metals
using the method of Lindsay and Norvell (1969) and the metal
concentrations were determined by AAS.
A simple method of determining the chlorophyll content is
the portable Chlorophyll Meter SPAD-502 (Minolta
Fig. 2 – Reflectance spectrum of rice under different levels
of heavy metal pollution.
226
biosystems engineering 106 (2010) 223–233
values of Cd and Cu always occurred in rice with level II pollution,
followed by level I pollution, and then safe level. A distinction
was found between the uptakes of Cd and Cu at different growth
stages of rice. The highest values of Cd and Cu uptake occurred in
the tillering and mature growth stages respectively.
2.4.
Fig. 3 – Chlorophyll concentration in rice leaves at different
growth stages.
Lin & Kao, 2001; Lagriffoul, Mocquot1, Mench1, & Vangronsveld, 1998; Larsson, Bordman, & Asp 1998). In summary,
the main toxic effect of Cu and Cd is oxidative stress.
Heavy metal toxicity in rice leaves was assessed by the
decrease in chlorophyll and protein contents (Huang et al., 2009).
An average value of chlorophyll concentration was obtained
from four critical growth stages of rice. The result showed that
Cd and Cu induced stress in rice had a close correlation with
chlorophyll concentration in rice at different growth stages
(Fig. 3). From Fig. 3, within seeding, tillering, booting and mature
growth stages, the chlorophyll concentration of rice with safe
level was higher than those with level I pollution, and higher
than level II pollution. In addition, the greatest difference in
chlorophyll concentration in rice with different pollution levels
occurred at the mature growth stage. Therefore, the difference in
chlorophyll concentration of rice proved credible and feasible
method for estimating Cu and Cd contamination levels. Low
values in the chlorophyll concentrations indicated that the rice
was suffering from very serious contamination.
The reduction of chlorophyll concentration in rice is caused
by Cd and Cu preferentially accumulating in the chloroplasts as
the rice grows. Cd and Cu levels in rice leaves with different
growth stages are displayed in Fig. 4. As seen in Fig. 4, the highest
Spectral indices used in the model
As illustrated in Fig. 2, in the NIR region, the leaf reflectance of
rice under heavy metal stress is somewhat lower than that of
the safe level. A number of studies have demonstrated that
the shifts in vegetation spectra of plant under heavy metal
stress occur in both the visible and the NIR part of the spectrum (Kooistra et al., 2004). With the previous analysis results,
the primary effect of Cu and Cd metal-induced stress on rice is
the corresponding reduction in chlorophyll. Therefore, our
goal was to select spectral indices sensitive to leaf chlorophyll
concentration in rice. Strong relationships exist between
chlorophyll concentrations and spectral indices formulated
with some specific narrow spectral bands including 550 nm,
670 nm, 700 nm, 800 nm (Haboudane et al., 2002; ZarcoTejada, Miller, Noland, Mohammed, & Sampson, 2001).
Therefore, several spectral indices formulated with the above
values were selected in this study.
2.4.1.
Normalised difference vegetation index (NDVI)
The most well known and widely used vegetation index is the
normalised difference vegetation index (NDVI) developed by
Rouse, Haas, Schell, Deering, and Harlan (1974). It is based on the
contrast between the maximum absorption in the red due to
chlorophyll pigments and the maximum reflection in the infrared
caused by leaf cellular structure. Using hyperspectral narrow
wavebands, this index is quantified by the following equation:
NDVI½670; 800 ¼
R800 R670
R800 þ R670
(1)
where Rx is the reflectance at the given wavelength (nm).
Our results showed that NDVI has a strong non-linear
relationship with chlorophyll concentration (Fig. 5), the
correlation coefficient (R2) being 0.6986. Gallagher, Pechmann,
Bogden, Grabosky, and Weis (2008) demonstrated that NDVI
can predict heavy metal stress and that it had a curvilinear
relationship with total heavy metal load (R2 ¼ 0.85).
Fig. 4 – Cu and Cd in rice leaves at different growth stages.
biosystems engineering 106 (2010) 223–233
227
Fig. 5 – Relationship between different vegetation indices and chlorophyll concentration.
2.4.2.
Modified triangle vegetation index (MTVI)
The MTVI belongs to the triangle vegetation index family
(Broge & Leblance, 2000). The general principle behind this
index is to describe the radiative energy absorbed by the
pigments as a function of the relative difference between red
and NIR reflectance in conjunction with the magnitude of
reflectance in the green region, where the light absorption by
chlorophylls is relatively insignificant. The index is calculated
as the area of the triangle defined by the green peak, the
chlorophyll absorption minimum, and the NIR shoulder in
spectral space. It based on the fact that chlorophyll absorption
causes a decrease in red reflectance. Some researchers have
tried to modify the triangle vegetation index by adjusting the
coefficient and the waveband of equation. Guang and Liu
(2009) developed the MTVI by as the followed equation:
MTVI ¼ 1:5½1:2ðR712 R550 Þ 2:1ðR670 R550 Þ
(2)
The results showed that MTVI was insensitive to chlorophyll concentration of rice under heavy metal stress. This was
attributed to the insignificant difference in the area of the
triangle derived from green peak and the chlorophyll
absorption maximum, especially in rice with level I pollution.
2.4.3.
Modified chlorophyll absorption ratio index (MCARI)
To reduce non-photosynthetic factors, Kim, Daughtry, Chappelle, McMurtrey, and Walthal (1994) developed the chlorophyll absorption ratio index (CARI) which measures the depth
of chlorophyll absorption at 670 nm relative to the green
reflectance peak at 550 nm and the reflectance at 700 nm. It
uses bands corresponding to the minimum absorption of the
photosynthetic pigments, centred at 550 nm and 700 nm, in
conjunction with the chlorophyll a maximum absorption
band, around 670 nm. Daughtry et al. (2000) developed the
MCARI which can be simplified as the following equation:
228
biosystems engineering 106 (2010) 223–233
Table 2 – Hyperspectral vegetation indices used in the sensitivity analysis.
Indices
Wavebands (nm)
NDVI[670,800]
670,800
OSAVI[670,800]
MCARI[670,700]
MTVI[550,712]
MTVI/MCARI[550,712]
670,800
550,670,700
550,670,712
550,670,700,712
MTVI/OSAVI[550,800]
550,670,712,800
Formula
R800 R670
R800 þ R670
OSAVI ¼ ð1 þ 0:5ÞðR800 R670 Þ=ðR800 þ R670 þ 0:5Þ
MCARI ¼ ½ðR700 R670 Þ 0:2ðR700 R550 ÞðR700 =R670 Þ
MTVI ¼ 1:5½1:2ðR712 R550 Þ 2:1ðR670 R550 Þ
1:5½1:2ðR712 R550 Þ 2:1ðR670 R550 Þ
MTVI=MCARI ¼
½ðR700 R670 Þ 0:2ðR700 R550 ÞðR700 =R670 Þ
1:5½1:2ðR712 R550 Þ 2:1ðR670 R550 Þ
MTVI=OSAVI ¼
ð1 þ 0:5ÞðR800 R670 Þ=ðR800 þ R670 þ 0:5Þ
NDVI½670; 800 ¼
MCARI ¼ ½ðR700 R670 Þ 0:2ðR700 R550 ÞðR700 =R670 Þ
(3)
Reference
Rouse et al. 1974
Huete, 1988
Daughtry et al. 2000
Guang & Liu, 2009
_
In sensitivity analysis, bands at 550 nm, 670 nm, 700 nm,
712 nm and 800 nm were selected. Therefore, reflectance in
these bands was selected to formulate vegetation indices
which may have potential in estimating chlorophyll concentration. Six indices (Table 2) were tested in this study; the
result indicated that NDVI, OSAVI, MTVI/MCARI and MTVI/
OSAVI were sensitive to chlorophyll content variations of rice
under heavy metal stress.
Our results showed that MCARI was also insensitive to
chlorophyll concentration of rice under heavy metal stress;
while other researchers have reported that MCARI was
sensitive to chlorophyll concentration (Wu, Niu, Tang, &
Huang, 2008). The reason is that the green peak position
moved towards longer wavelengths as the heavy metal
content increased. Daughtry et al. (2000) have pointed out that
MCARI was still sensitive to non-photosynthetic element
effects, mainly at low chlorophyll concentrations.
2.5.
Neural-network model development
2.4.4.
2.5.1.
Data pre-processing phase
Optimized soil-adjusted vegetation index (OSAVI)
Daughtry et al. (2000) proved that when MCARI is combined
with a soil line vegetation index like OSAVI (Rondeaux, Steven, & Baret, 1996), the sensitivity to the underlying soil
reflectance properties can be reduced. OSAVI belongs to the
soil-adjusted vegetation index (SAVI; Huete, 1988) family and
is defined by the following equation:
OSAVI ¼ ð1 þ 0:5ÞðR800 R670 Þ=ðR800 þ R670 þ 0:5Þ
(4)
As shown in Fig. 5, rice leaf chlorophyll concentration was
strongly correlated with OSAVI (R2 ¼ 0.6990). This agrees well
with the results of Haboudane et al. (2002).
2.4.5.
Integrating two vegetation indices
Research has indicated that an integrated index modelling is
more sensitive to chlorophyll content variations and more
resistant to the variations of leaf area index (LAI) and solar
zenith angle than a single vegetation index (Haboudane et al.,
2002). Wu et al. (2008) proved that integrated indices transformed chlorophyll absorption ratio index (TCARI/OSAVI)
[705,750] and MCARI/OSAVI [705,750] have better linearity
with chlorophyll content together with high correlation: R2 of
0.8808 and 0.9406 respectively. To improve sensitivity of the
vegetation index to subtle variations in rice leaf chlorophyll
concentration under heavy metal stress, in this paper, two
integrated indices are proposed. They be defined by the
following equations:
MTVI=MCARI ¼
1:5½1:2ðR712 R550 Þ 2:1ðR670 R550 Þ
½ðR700 R670 Þ 0:2ðR700 R550 ÞðR700 =R670 Þ
(5)
MTVI=OSAVI ¼
1:5½1:2ðR712 R550 Þ 2:1ðR670 R550 Þ
ð1 þ 0:5ÞðR800 R670 Þ=ðR800 þ R670 þ 0:5Þ
(6)
Our results indicated that the two integrated indices MTVI/
MCARI, MTVI/OSAVI and leaf chlorophyll content were highly
correlated with R2 of 0.7280 and 0.7198 respectively (Fig. 5).
_
In this study, to avoid data saturation, the input variables in
this model were normalised, based on their possible ranges
using the following equation:
xnorm ¼
xi xmin
xmax xmin
(7)
where xi, xmin, xmax and xnorm are the real-valued input variable, the minimum input variable, maximum input variable
and its normalised value respectively. The output from
a neural-network model is an indexed value that corresponds
to the input variable. To get the real-valued output, the
indexed output value needs to be de-normalised according to
the following equation:
yai ¼ ymin þ ynorm ymax ymin
(8)
Where yai, ymin, ymax and ynorm are the real-valued output
variable, the minimum and maximum values of the realvalued output, and the indexed output value from the neuralnetwork model, respectively.
2.5.2.
Model building phase
Artificial neural network (ANN) models have found wide
applications, including prediction, classification,system
modelling, signal processing, noise filtering (Pachepsky, Timlin, & Varallyay, 1996; Pal, Pal, Das, & Majumdar, 2003; Plate,
Bert, Grace, & Band, 2000; Tumbo et al., 2002). BP neutral
networks are popular neutral network architectures (Bishop,
1995; Cichocki & Unbehauen, 1993; Foody, 1995; Pachepsky
et al., 1996; Pal et al., 2003; Plate et al., 2000; Tumbo et al., 2002).
In this paper, a BP neural-network model was established to
estimate the leaf chlorophyll concentration, which in turn,
indicates the stress levels of rice growing under heavy metal
stress. The input variables for the BP neural-network model
were NDVI, OSAVI, MTVI/OSAVI and MTVI/MCARI, while the
output variable of this model was chlorophyll concentration.
biosystems engineering 106 (2010) 223–233
Fig. 6 – A simple BP neural-network example flow chart
with 4-6-1 structure.
The topological structure of this BP neural-network model
consisted of four input neurons in the input layer, one output
neuron in the output layer and one or two hidden layers. The
flow chart of the 4-6-1 structure as an example is shown in Fig. 6.
The model building phase aims to produce a robust ANN
model that can accurately map outputs from inputs. A good
ANN model mainly relies on the choice of an optimum neutral
network architecture and network internal parameters (i.e.,
the number of hidden layers, number of neurons, type of
scaling and activation functions, learning and momentum
rates, and stopping criterion), and the selection of the training
algorithm (Gautam, Panigrahi, & Franzen, 2006).
In this study, different network architectures, activation
functions and training algorithms were tested. The ideal
229
activation function develops models using a sigmoid tangent
function that scales the input data in the open interval of 1–1,
a hidden layer utilising the sigmoid tangent function, and an
output layer utilising sigmoid logistic functions which scales
the input data to lie within the open interval of 0–1. The typical
gradient descent BP algorithm in learning algorithms proved to
be suitable for this study. Maier and Dandy (2000) and Zealand,
Burn, and Simonovic (1999) reported that over 80% of previous
neural-network models used a gradient descent BP training
algorithm, which is a supervised learning paradigm. By
supervising learning, a desired response is available to guide
the learning process. In a BP algorithm, the weights are initially
assigned arbitrary small values. As training progresses, the root
mean square error (RMSE) between the target output and the
network output is calculated. If the network output does not
produce the expected criterion, the training transfers to the BP.
The error signal carries through a reverse calculation by
a connecting pathway. The gradient descent BP algorithm
utilises a learning rate and momentum coefficient to adjust
weight values and to move towards a global minimum in the
error surface. Fig. 7 shows the gradient descent BP learning
algorithm in the neural-network architecture.
By trying different numbers of hidden layers and different
numbers of neurons in each hidden layer, the optimum model
architectures were 4-6-1, 4-8-1, 4-10-2-1 and 4-10-3-1. Compared
with one hidden layer, two hidden layers had higher accuracy in
estimating chlorophyll concentration. Table 3 summarised the
optimum model architecture and internal parameters utilising
in the different models developed in this study.
2.5.3.
Model evaluation phase
Generally, RMSE and average accuracy have been used to
measure the performance of neural-network models (Gautam
Fig. 7 – A process flow chart of a gradient descent BP learning algorithm.
230
biosystems engineering 106 (2010) 223–233
Table 3 – Summary table showing optimum ANN model’ architecture and internal parameters.
Model
Optimum
network
Activation function
Training
algorithm
1
2
3
4
5
6
7
8
4-6-1
4-6-1
4-8-1
4-8-1
4-10-3-1
4-10-3-1
4-10-2-1
4-10-2-1
{tansig,
{tansig,
{tansig,
{tansig,
{tansig,
{tansig,
{tansig,
{tansig,
Traingdm
Traingdm
Traingdm
Traingdm
Traingdm
Traingdm
Traingdm
Traingdm
logsig}
linear}
logsig}
linear}
tansig, logsig }
tansig, linear}
tansig, logsig }
tansig, linear}
Learning
rate
Momentum
coefficient
R2
RMSE
Pav
0.2
0.2
0.2
0.2
0.15
0.15
0.15
0.15
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.8040
0.7412
0.8014
0.7016
0.8483
0.7378
0.9014
0.7791
4.01
5.25
3.37
4.34
3.14
4.03
2.58
3.06
0.8516
0.8423
0.8664
0.8510
0.8662
0.8421
0.8922
0.8737
Notes: Tansig denotes sigmoid tangent function; logsig denotes sigmoid logistic functions; linear denotes linear function; traingdm denotes
gradient descent BP algorithm.
et al., 2006). In this study, the parameters were: (i)RMSE; (ii)
average accuracy (Pav); (iii) R2. The three parameters were
computed by:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2
PN i¼1 yai ymi
RMSE ¼
N1
(9)
Where yai, ymi, RMSE are the real-valued output variable,
measured output variable, RMSE between real-valued output
variable and measured output variable respectively, N is the
sample number. The lower RMSE, the better the performance
of the model.
PN
Pav ¼
i¼1
Pi
N
j yai ymi j
and Pi ¼ 1 yai
(10)
Where Pav and Pi are the average accuracy and the accuracy of
the ith result. Pav provides information on the accuracy that the
model can yield using a given data set. The nearer the value
approaches 1; the better is the performance of the model.
PN i¼1
R2 ¼ PN i¼1
yai ya
yai ya
PN i¼1
2 PN
ymi ym
i¼1 ðymi
ym
2
2
(11)
Where R2, ym and ya are the correlation coefficient, average
measured output variable and average real-valued output
variable respectively. The R2 represents the correlation
between predicted and measured variables. It is assumed that
the predicted and measured variables follow a normal distribution. Its value ranges from 0 to 1. The higher the value of the
correlation, the stronger is the indication of existing linear
relations between the actual and predicted variables.
3.
Result and discussion
Due to the large variation of chlorophyll concentration in rice
with different levels of pollution that occur in the mature growth
stage, the best stage for monitoring heavy metal stress is at the
mature growth stage of rice. Therefore, in this research, 138
training data sets and 69 test data sets from the mature growth
stage of rice were obtained for different levels of heavy metal
pollution. By trying different neutral network architectures,
network internal parameters and training algorithms, satisfactory neural-network models were developed (Table 3); namely,
the BP neural-network architectures were 4-6-1, 4-8-1, 4-10-2-1
and 4-10-3-1; the learning algorithm was a gradient descent BP
Table 4 – Quantitative relationships of leaf chlorophyll
concentration ( y) to individual spectral index (x) in rice.
Spectral index
NDVI
OSAVI
MTVI/OSAVI
MTVI/MCARI
Fig. 8 – Comparison of predicated chlorophyll
concentration using an artificial neural-network model
and multiple regression models.
Regression equation
R2
y ¼ 4.9705x 0.7648
y ¼ 5.6535x2 2.0623x þ 1.3419
y ¼ 0.5746e2.1683x
y ¼ 2.8959ln(x) þ 3.7551
y ¼ 7.8019x 0.7146
y ¼ 14.493x2 3.4562x þ 1.3885
y ¼ 0.5859e3.4098x
y ¼ 2.8214ln(x) þ 5.0382
y ¼ 0.6401x þ 4.871
y ¼ 0.1123x2 1.5414x þ 6.5817
y ¼ 6.699e0.2786x
y ¼ 2.4517ln(x) þ 5.6369
y ¼ 1.2729x þ 4.8087
y ¼ 0.9391x2 5.3069x þ 8.9348
y ¼ 6.7585e0.5721x
y ¼ 2.7379ln(x) þ 4.0951
0.6799
0.6986
0.6514
0.6441
0.6780
0.6990
0.6521
0.6384
0.6886
0.7198
0.6568
0.7179
0.6131
0.7280
0.6236
0.6794
biosystems engineering 106 (2010) 223–233
231
Fig. 9 – Measured and predicted chlorophyll concentration using BP neutral network using different network structures.
algorithm; with a the learning rate of 0.2 or 0.15; the momentum
coefficient was 0.2; the activation function was the sigmoid
tangent function in the input layer and hidden layer and there
were sigmoid logistic functions in the output layer. According to
the three parameters for assessing the performance of the
neural-network models, optimal network architecture should
have the lowest RMSE, and the value of Pav and R2 close to 1. As
seen in Table 3, the optimal network architecture was obtained
when the neural-network model had ten neurons in the first
hidden layer and two neurons in the second hidden layer. The R2
for the optimal structure BP network model was 0.9014; RMSE
was 2.58 and Pav was 0.8922.
To assess the performance of the neural-network model
more thoroughly, a comparison between the neural-network
model and a statistical regression model was made. The
baseline model was a best-fit regression model obtained from
four spectral indices analysis of the rice crop and estimated
values of the optimal structure BP network model. Applying
two methods to the results gave the following respectively:
y ¼ 3:209 þ 1:993 NDVI 0:6390 OSAVI 0:363
ðMTVI=OSAVIÞ 0:220 ðMTVI=MCARIÞ
4.
Where R2 ¼ 0.713, y is predicated chlorophyll concentration
from four spectral indices using multiple regression models.
yai ¼ 0:9873ymi 0:091
2
variable (measured chlorophyll concentration) using optimal
structure BP network model.
The quantitative relationships between leaf chlorophyll
concentration and individual spectral indices in rice are
summarised in Table 4. As shown in Fig. 8 and Table 4, the
neural-network model showed stronger correlation than the
statistical regression model.
Before a neural-network model can be confidently applied
to predict the leaf chlorophyll concentration to determine rice
heavy metal stress levels, it is necessary to validate the model
by checking the degree to which the model predicted leaf
chlorophyll concentration matches the measured leaf chlorophyll concentration obtained from field tests. Fig. 9 shows
that the predicted leaf chlorophyll concentration matched the
measured leaf chlorophyll concentration well. As well as
predicting that the rice leaf chlorophyll concentration
increases with decreasing heavy metal content, the highest
chlorophyll concentration values were predicted for the
experiment site with safe level, followed by that of level I
pollution and then level II pollution.
(13)
Where R ¼ 0.9014,yai, ymi are the real-valued output variable
(predicated chlorophyll concentration), measured output
Conclusion
Hyperspectral data can provide an effective means for estimating leaf chlorophyll concentration variation of crops
under heavy metal stress. Because of heavy metal stress, the
biochemical component of the crop changes and crop leaf
cellular structure is damaged, resulting in differences in
232
biosystems engineering 106 (2010) 223–233
spectral reflectance. Therefore, it is feasible to detect such
variations in the biochemical composition of crops, specifically using leaf chlorophyll concentration in the crop through
spectral reflectance. In this research, some spectral indices
were formulated by the reflectance of two or more bands. In
sensitivity analysis, it was found that NDVI, OSAVI, MTVI/
MCARI and MIVI/OSAVI are more sensitive to chlorophyll
concentration than the other indices, so they were selected to
estimate leaf concentration. The relationship between these
indices and chlorophyll concentration was developed by the
BP neural-network model. The input neurons in the model
were NDVI, OSAVI, MTVI/MCARI and MIVI/OSAVI, and the
output neuron from this model was chlorophyll concentration
which was then used to assess rice heavy metal stress levels.
After testing a training set, our results showed that the neuralnetwork model developed can accurately estimate chlorophyll concentration variation to assess heavy stress levels.
When crops were contaminated by heavy metal, the leaf
chlorophyll concentration in the crop decreased. The higher
the level of heavy metal content, the lower the leaf chlorophyll
concentration. This finding agrees with the results of Kastori
et al. (1998).
This research verified that the neural-network model
developed here could provide an effective and faithful estimation of leaf chlorophyll variation based on the four
performance spectral indices. In trials with the BP neuralnetwork model, the optimal model, with ten neurons in the
first hidden layer and two neurons in the second hidden layer,
performed better than a statistical model of multivariable
spectral indices and a single spectral index. Based on the
results obtained from this study, it can be concluded that
using a neural-network model and hyperspectral data can
provide sufficient information to efficiently detect crop
biochemistry components and assess the level of pollutants in
field operations.
Acknowledgements
This research was supported by the National Natural Science
Foundation of China (40771155) and National High-tech R&D
Program of China (863 Program) (2007AA12Z174).
references
Barcelo, J., & Poschenrieder, C. (1990). Plant water relations as
affected by heavy metal stress: review. Journal of Plant
Nutrition, 13, 1–37.
Bishop, C. M. (1995). Neural networks for pattern recognition. Oxford,
England: Clarendon Press.
Broge, N. H., & Leblance, E. (2000). Comparing prediction power
and stability of broadband and hyperspectral vegetation
indices for estimation of green leaf area index and canopy
chlorophyll density. Remote Sensing of Environment, 76, 156–172.
Chi, G. Y., Liu, X. H., Liu, S. H., & Yang, Z. F. (2006). Studies of
relationships between Cu pollution and spectral
characteristics of TritiZnm Aestivum L. Spectroscopy and Spectral
Analysis, 26(7), 1272–1276.
Chien, H. F., & Kao, C. H. (2000). Accumulation of ammonium in
rice leaves in response to excess cadmium. Plant Science, 156,
111–115.
Chien, H. F., Wang, J. W., Lin, C. C., & Kao, C. H. (2001). Cadmium
toxicity of rice leaves is mediated through lipid peroxidation.
Plant Growth Regulation, 33, 205–213.
Choe, E., Meer, F. V. D., Ruitenbeek, F. V., Werff, H. V. D.,
Smeth, B. D., & Kim, K. W. (2008). Mapping of heavy metal
pollution in stream sediments using combined geochemistry,
field spectroscopy, and hyperspectral remote sensing: a case
study of the Rodalquilar mining area, SE Spain. Remote Sensing
of Environment, 112, 3222–3233.
Cichocki, A., & Unbehauen, R. (1993). Neural networks for
optimization and signal processing. Chichester, NY, USA: John
Wiley & Sons.
Daughtry, C. S. T., Walthall, C. L., Kim, M. S., Brown de
Colstoun, E., & McMurtrey, J. E., III (2000). Estimating corn leaf
chlorophyll concentration from leaf and canopy reflectance.
Remote Sensing of Environment, 74, 229–239.
De Filipps, L. F., & Pallaghy, C. K. (1994). Heavy metals: sources and
biological effects. In L. C. Rai, J. P. Gaur, & C. J. Soeder (Eds.),
Advances in limnology series: algae and water pollution. Stuttgart,
Germany: E. Scheizerbartsche Press.
Droppa, M., & Horvath, G. (1990). The role of copper in
photosynthesis. CRC Critical Reviews in Plant Science, 9, 111–123.
Ellis, R. W., & Eslick, L. (1997). Variation and range of mercury uptake
into plants at a mercury-contaminated abandoned mine site.
Bulletin of Environmental Contamination and Toxicology, 59, 763–769.
Fernandes, J. C., & Henriques, F. S. (1991). Biochemical,
physiological, and structural effects of excess copper in
plants. The Botanical Reivew, 57(3), 246–266.
Foody, G. M. (1995). Using prior knowledge in artificial neural
network classification with a minimal training set.
International Journal of Remote Sensing, 16, 301–312.
Gallagher, F. J., Pechmann, I., Bogden, J. D., Grabosky, J., & Weis, P.
(2008). Soil metal concentrations and productivity of Betula
populifolia (gray birch) as measured by field spectrometry and
incremental annual growth in an abandoned urban
Brownfield in New Jersey. Environmental Pollution, 156, 699–706.
Gan, F. P., Liu, S. W., & Zhou, Q. (2004). Identification of mining
pollution using Hyperion data at Dexing copper mine in
Jiangxi province, China. Earth Science-Journal of China University
of Geosciences, 29(1), 119–126.
Gautam, R., Panigrahi, S., & Franzen, D. (2006). Neural network
optimisation of remotely sensed maize leaf nitrogen with
a genetic algorithm and linear programming using five
performance parameters. Biosystems Engineering, 95(3), 359–370.
Guang, L., & Liu, X. N. (2009). Two kinds of modified spectral
indices for retrieval of crop canopy chlorophyll content.
Advances in Earth Science, 29(5), 548–554.
Haboudane, D., John, R., Millera, J. R., Tremblay, N., ZarcoTejada, P. J., & Dextraze, L. (2002). Integrated narrow-band
vegetation indices for prediction of crop chlorophyll content
for application to precision agriculture. Remote Sensing of
Environment, 81, 416–426.
Howard, D., & Mark, B. (2000). Neural network toolbox, version4.
Natick, MA, USA: The MathWorks, Inc.
Huang, Y. Z., Hu, Y., & Liu, Y. X. (2009). Combined toxicity of
copper and cadmium to six rice genotypes (Oryza sativa L.).
Journal of Environmental Sciences, 21, 647–653.
Huete, A. R. (1988). A soil vegetation adjusted index (SAVI). Remote
Sensing of Environment, 25, 295–309.
Kastori, R., Plesnicar, M., Sakac, Z., Pancovic, D., & Arsenijevic
Maksimovic, I. (1998). Effect of excess lead on sunflower
growth and photosynthesis. Journal of Plant Nutrition, 21,
75–85.
Kim M. S., Daughtry C. S. T., Chappelle E. W., McMurtrey III J. E., &
Walthal C. L. (1994). The use of high spectral resolution bands
biosystems engineering 106 (2010) 223–233
for estimating absorbed photosynthetically active radiation
(Apar). In Proceedings of the sixth symposium on physical
measurements and signatures in remote sensing (pp. 299–306). Val
D’Isere, France, January 17–21.
Kooistra, L., Salas, E. A. L., Clevers, J. G. P. W., Wehrens, R.,
Leuven, R. S. E. W., Nienhuis, P. H., et al. (2004). Exploring field
vegetation reflectance as an indicator of soil contamination in
river floodplains. Environmental Pollution, 127, 281–290.
Krupa, Z. (1988). Cadmium-induced changes in the composition
and structure of the light-harvesting complex II in radish
cotyledons. Physiologia Plantarum, 73, 518–524.
Lagriffoul, A., Mocquot1, B., Mench1, M., & Vangronsveld, J. (1998).
Cadmium toxicity effects on growth, mineral and chlorophyll
contents, and activities of stress related enzymes in young
maize plants (Zea mays L.). Plant and Soil, 200, 241–250.
Larsson, E. H., Bordman, J. F., & Asp, H. (1998). Influence of UV-B
radiation and Cd2þ on chlorophyll fluorescence, growth and
nutrient content in Brassica napus. Journal of Experimental
Botany, 49, 1031–1039.
Li, Q. T., Yang, F. J., Zhang, B., Zhang, X., & Zhou, G. Z. (2008).
Biogeochemistry responses and spectral characteristics of
Rhus Chinensis Mill under heavy metal stress. Journal of
Remote Sensing, 12(2), 284–290.
Lindsay, W. L., & Norvell, W. A. (1969). Development of a DTPA
micronutrients soil test. Agronomy Abstracts, 6, 84.
Liu, S. W., Gan, F. P., & Wang, R. S. (2004). The application of
Hyperion data to extracting contamination information of
vegetation in the DEXING copper mine Jiangxi province,
China. Remote Sensing for Land & Resources, 59, 7–10.
Maier, H. R., & Dandy, G. C. (2000). Neural networks for the
prediction and forecasting of water resources variables:
a review of modeling issues and application. Ecological
Modeling, 15, 101–124.
Noh, H., Zhang, Q., Shin, Han, S., & Feng, L. (2006). A neural
network model of maize crop nitrogen stress assessment for
a multi-spectral imaging sensor. Biosystems Engineering, 94(4),
477–485.
Pachepsky, V., Timlin, A. D., & Varallyay, G. (1996). Artificial neural
networks to estimate soil water retention from easily measurable
data. Soil Science Society of America Journal, 60(3), 727–733.
Pahlsson, A. M. B. (1989). Toxicity of heavy-metals (Zn, Cu, Cd, Pb)
to vascular plants. Water Air and Soil Pollution, 47, 287–319.
Pal, N. R., Pal, S., Das, J., & Majumdar, K. (2003). SOFM-MLP:
a hybrid neural network for atmospheric temperature
233
prediction. IEEE Transactions on Geoscience and Remote Sensing,
41(12), 2783–2791.
Plate, T., Bert, J., Grace, J., & Band, P. (2000). Visualizing the
function computed by a feed forward neural network. Neural
Computation, 12(6), 1355–1370.
Ren, H. Y., Zhuang, D. F., & Pan, J. J. (2008). Canopy hyperspectral
characteristics of paddy plants contaminated by lead. GeoInformation Science, 10(3). 314–319 (in Chinese).
Rondeaux, G., Steven, M., & Baret, F. (1996). Optimization of soil
adjusted vegetation indices. Remote Sensing of Environment, 55,
95–107.
Rouse J. W., Haas R. H., Schell J. A., Deering D. W., & Harlan J. C.
(1974). Monitoring the vernal advancements and
retrogradation of natural vegetation. In NASA/GSFC, Final
report (pp. 1–137). Greenbelt, MD, USA.
Sidlecka, A., & Baszynsky, T. (1993). Inhibition of electron flow
around photosystem I in chloroplasts of cadmium-treated
maize plants is due to cadmium-induced iron deficiency.
Physiologia Plantarum, 87, 199–202.
Stevens, D. (2002). CSIRO land and water’s methods manual. ACIAR
and CSIRO. pp. 15–16.
Tumbo, S. D., Wagner, D. G., & Heinemann, P. H. (2002).
Hyperspectral-based neural network for predicting chlorophyll
status in corn. Transactions of the ASAE, 45(3), 825–832.
Wang, T., & Thai, C. N. (2003). A genetic algorithm based tool for
analysis and modeling of multi-spectral data. ASAE Paper No.
03-3126. MI, USA: St. Joseph.
Wood, C. W., Reeves, D. W., & Himelrick, D. G. (1993). Relationship
between chlorophyll meter readings and leaf chlorophyll
concentration, N status, and crop yield: a review. Proceedings of
the Agronomy Society of New Zealand, 23, 1–9.
Wu, C. Y., Niu, Z., Tang, Q., & Huang, J. (2008). Estimating
chlorophyll content from hyperspectral vegetation indices:
modeling and validation. Agricultural and Forest Meteorology,
148, 1230–1241.
Zarco-Tejada, P. J., Miller, J. R., Noland, T. L., Mohammed, G. H., &
Sampson, P. H. (2001). Scaling-up and model inversion
methods with narrow-band optical indices for chlorophyll
content estimation inclosed forest canopies with
hyperspectral data. IEEE Transactions on Geoscience and Remote
Sensing, 39, 1491–1507.
Zealand, C. M., Burn, D. H., & Simonovic, S. P. (1999). Short term
stream flow forecasting using artificial neural networks.
Journal of Hydrology, 214, 32–48.