prediction and classification of sugar content of sugarcane based on skin scanning using visible and...
DESCRIPTION
ThesisTRANSCRIPT
-
visible and shortwave near
Nazmi Mat Nawi a,c,*, GuangnaFaculty of Engineering and Surveying, UnivebNational Centre for Engineering in AgricultucDepartment of Biological and Agricultural En
y and M
2013 IAgrE. Published by Elsevier Ltd. All rights reserved.
to enhance farm profits through efficient application of crop
inputs by matching them with variations of crop yield and
quality in the field (Wendte, Skotnikov, & Thomas, 2001). The
primary requirements for PA application are yield monitoring
andmapping. To date, several studies have been carried out to
Viator, Larsen, & Peters, 2011). However, an extensive review
published by Bramley (2009) regarding PA application in sug-
arcane industry revealed that current PA technologies only
monitor the yield but do not have the ability to measure
product quality. This is a serious limitation because both yield
* Corresponding author. Department of Biological and Agricultural Engineering, Faculty of Engineering, Universiti Putra Malaysia, 43400Serdang, Selangor, Malaysia.
.com (N. Mat Nawi).
Available online at www.sciencedirect.com
vi
b i o s y s t em s e n g i n e e r i n g 1 1 5 ( 2 0 1 3 ) 1 5 4e1 6 1E-mail addresses: [email protected], naj_miey@hotmail1. Introduction
Precision agriculture (PA) is a valuable management strategy
produce a yield map in sugarcane industries around the
world, including Australia (Cox, Harris, Pax, & Dick, 1996),
Brazil (Magalhaes & Cerri, 2007) and the US (Price, Johnson,Received 2 January 2013
Received in revised form
27 February 2013
Accepted 8 March 2013
Published online 19 April 2013
different commercial sugarcane varieties were used. Each sample was scanned at four
scanning points to obtain the spectra data which was later correlated with its Brix (soluble
solids content) values. Partial least square (PLS) model was developed and applied to both
calibration and prediction samples. Using reflectance spectra data, the model had a coef-
ficient of determination (R2) of 0.91 and root means square error of predictions (RMSEP) of
0.721 Brix. The artificial neural network (ANN) was also applied to classify spectra data
into five Brix categories. The ANN has yielded good classification performance, ranging
from 50 to 100% accuracy with an average accuracy of 83.1%. These results demonstrated
that the Vis/SWNIR spectroscopy technique could be applied to predict sugarcane Brix in
the field based skin scanning method.Selangor, MalaysiadDepartment of Agricultural Machiner
Khuzestan, Iran
a r t i c l e i n f o
Article history:1537-5110/$ e see front matter 2013 IAgrEhttp://dx.doi.org/10.1016/j.biosystemseng.201an Chen a,b, Troy Jensen a,b, Saman Abdanan Mehdizadeh d
rsity of Southern Queensland, Toowoomba, QLD 4350, Australia
re (NCEA), University of Southern Queensland, Toowoomba, QLD 4350, Australia
gineering, Faculty of Engineering, Universiti Putra Malaysia, 43400 Serdang,
echanization, Ramin Khuzestan University of Agriculture and Natural Resources,
The potential application of a visible and shortwave near infrared (Vis/SWNIR) spectro-
scopic technique as a low cost alternative to predict sugar content based on skin scanning
was evaluated. Two hundred and ninety one internode samples representing threeinfrared
based on skin scanning using
Prediction and classification of sugar content of sugarcaneResearch Paper
journal homepage: www.else. Published by Elsevier Lt3.03.005er.com/locate/ issn/15375110d. All rights reserved.
-
R Reflectance spectra
b i o s y s t em s e ng i n e e r i n g 1 1 5 ( 2 0 1 3 ) 1 5 4e1 6 1 155and quality components vary significantly across the field
(Bramley, Panitz, Jensen, & Baillie, 2012; Johnson & Richard,
2005). An accurate determination of both components is
important especially for a fair payment system to growers and
better farm management practices. Thus, it is desirable to
measure quality level along with yield data in the field during
harvesting to fully utilise the PA approach.
In Australia, sugarcane quality is determined based on its
sugar content, known as commercial cane sugar (CCS). CCS is
derived from Brix (soluble solids content), Pol (the sugarcontent measured using the property of optical activity which
causes polarised light to be rotated) and fibre content. The Brixand Pol are usually measured in the laboratory using juice
samples. Unfortunately, replicating this laboratory method in
the field is very difficult due to the absence of reliable sensor
and the difficulties obtaining sufficient juice samples. There-
fore, an ability to predict sugarcane quality by directly scan-
ning stalk skin using spectroscopic methods would be an
attractive alternative. Spectroscopic methods have widely
been applied to non-destructively predict internal crop quality
based on skin scanning of a number of crops, e.g. kiwifruit
(Moghimi, Aghkhani, Sazgarnia, & Sarmad, 2010), apples (Lu,
Nomenclature
A Absorbance spectra
ANN Artificial neural network
ASCII American Standard Code for Information
Interchange
ASD Analytical spectral device
BSES Bureau of Sugar Experimental Station
CCS Commercial cane sugar
FOV Field-of-view
FWHM Full width at half maximum
H High
L Low
LVs Latent variables
M Medium2004; Park, Abbott, Lee, Choi, & Choi, 2003) and pineapple
(Chia, Rahim, & Rahim, 2012; Guthrie & Walsh, 1997).
For field application, however, a portable, robust and low
cost spectrometer is preferable (Montes et al., 2006). Lately,
the development of a portable visible and shortwave near
infrared (Vis/SWNIR) instrument with a wavelength range
from 350 to 1100 nm appears promising because the bands are
ascribed to the third and fourth overtones of OeH and CeH
stretchingmodes and the instrument is low-cost and portable
enough for in-field measurements (Walsh, Guthrie, & Burney,
2000). The use of this spectrometer, coupled with the use of
multivariate statistical techniques could create the possibility
to intensively and preciselymap crop quality over large parcel
of cropping land (Menesatti et al., 2010). Several studies have
been successfully conducted in the field using a spectrometer
mounted on a harvester for different crops, including maize
(Montes et al., 2006), forage (Digman & Shinners, 2008) and
sugar beet (Panigrahi & Hofman, 2003).
Unfortunately, there are no reported studies regarding the
skin scanning method for quality prediction in sugarcaneindustry. This method works based on the principle that when
a light beam hits the stalk, a small fraction is reflected at the
surfaceas a specular reflectanceand the restwill penetrate into
fruit tissues. In the tissues, photons are absorbed or migrate in
different directionswhere radiationwill be scattered backward
to the surface as diffuse reflectance, while the remaining ra-
diation migrates forward into the tissues as absorbance (Qing,
Ji, & Zude, 2007). Light absorption is related to certain chemi-
cal constituents, such as sugar, acid, water, etc. (Williams &
Norris, 2001). A reflectance spectrometer measures the aggre-
gate amount of light reflected from a sample, from which light
absorption may be estimated and then related to certain
chemical constituents of the crop (Lu, 2004).
For mapping purposes, the crop quality data must also be
classified into suitable quality classes. The artificial neural
network (ANN) is a well-known non-linear method which
could provide a robust classification model (Lee et al., 2010).
ANN is amultivariate statistical data analysis technique which
often gives higher recognition and prediction probability than
statistical classification algorithms. Inspired by biological ner-
vous system, the neural networks are composed of a number of
elements operating in parallel. By adjusting the weights (con-
R2 coefficient of determination
RMSEP Root mean square error of prediction
SEC Standard error of calibration
SEP Standard error of prediction
SNV Standard normal variate
SSC Soluble solids content
Vis/SWNIR Visible/shortwave near infrared
VisNIR Visible near infraredMSC Multiplicative scatter correction
PA Precision agriculture
PCA Principal component analysis
PCs Principal components
PLS Partial least squarenections) between elements, a neural network can perform
several functions, such as prediction, data filtration, data
conceptualization, classification, and data association (Haykin,
1994). The applications of ANN to classify crop quality have
been reported by Wu, Feng, Zhang, and He (2008) and Xing,
Symons, Shahin, and Hatcher (2010). In this study, ANN
models were applied to classify sugarcane quality based on
several Brix values. Thus, the specific objectives of this studywere (1) to investigate the feasibility of using Vis/SWNIR to
predict the quality level based on skin scanning; (2) to classify
sugarcane quality based on five Brix classes using ANN.
2. Material and methods
2.1. Sample preparation
A total of 291 internodes were extracted from 22 sugarcane
stalk samples. The stalk samples were collected from the
research station of Bureau of Sugar Experimental Station
-
(BSES), Bundaberg, Queensland, inMay 2012. The stalk samples
were of a commercial variety cut fromapropagation block used
to source materials for plant breeding trials. They were a plant
crop that was planted in September 2011 (8 months old). The
crop was grown under commercial conditions with the fertil-
isation based on soil test and the six easy steps nutrition
guidelines. The stalks belong to commercial variety trials rep-
resenting three different maturity stages, namely early-
maturing (Q155), mid-maturing (Q208) and late-maturing
(Q190) crops. The selection of these three varieties was to
ensure that the models developed in this study cover a wide
range of Brix (7.6e22.2) which is representative during har-vesting. Whole stalk samples were first topped and cut into
b i o s y s t em s e n g i n e e r i n g 1 1 5 ( 2 0 1 3 ) 1 5 4e1 6 1156individual internode using a cutter. Each internode samplewas
scanned at four different scanning points from bottom to the
top (Fig. 1). Each section was imaginarily labelled as S1, S2, S3
and S4 following the sequence fromS1 to S4 (bottom to the top).
2.2. Reflectance measurement
The reflectance measurement method was chosen for this
study because this method is the easiest to perform as it re-
quires no contact with the fruit and light levels are relatively
high (Schaare & Fraser, 2000). Reflectance measurement
methods have been successfully used to measure crops
quality by Chia et al. (2012) andDigman and Shinners (2008). In
this study, a handheld Vis/NIR spectroradiometer (FieldSpec
HandHeld and FieldSpec Pro FR, 325e1075 nm, Analytical
Spectral Devices (ASD), Inc., Boulder, CO, USA) was used to
collect spectra data from the skin surface over the wavelength
range of 325e1075 nm in 1.5 nm intervals. The equipment
used has a spectral resolution (full width at half maximum
(FWHM) of a single emission line) of approximately 3 nm at
around 700 nm. The scanning was undertaken using the 25
field-of-view (FOV) of the spectroradiometer. Two halogen
lamps (Lowell Pro-Lam 14.5 V tungsten bulb, Ushio Lighting,
Inc., Japan) were used to provide illumination to the sensor.
The lamps were placed at a distance of 800 mm and an angle
at 45 away from the sample. The distance between sampleand sensor was set at 70 mm, resulting in a measured spot of
0.031m diameter as calculated based on formula given by ASD
(2005). Distances between sensor and samples were main-
tained by fixing the sensor to a tripod and the samples were
held by a fixed sample holder.Fig. 1 e Scanning positions along individual internodes.The equipment was set to record the average of 20 scans
for each spectrum. All spectral data were stored in a computer
and processed using the RS3 software forWindows (Analytical
Spectral Devices, Boulder, CO, USA) designed with a graphical
user interface. Firstly, the reflectance spectra (R) were trans-
formed into ASCII format by using the ASD ViewSpec Pro
software (Analytical Spectral Devices, Boulder, CO, USA).
Then, the absorbance spectra (A) were calculated as Log (1/R)
using R data. Four spectra from each internode (S1eS4) were
then averaged and later used for calibration against Brixvalues obtained by the refractometer method.
The experiment was conducted inside a light-proof mea-
surement box (900 600 400 mm). The box was built toenclose the artificial light source, detecting probes and the
samples from ambient light. Relative reflectance spectra were
calculated by dividing stalk radiance with reference radiance
from a spectralon white reference panel for each wavelength.
In order to reduce signal-to-noise ratio, the first and last 75 nm
data points were removed from the spectral data as recom-
mended by (Liu, He, &Wang, 2008). Therefore, only the regions
between 400 and 1000 nm were used for the calculations.
2.3. Brix measurement
After the spectra acquisition, each internode was squeezed
using a clamp to extract juice samples for Brix measurement.Juice from each internode was collected in a small container,
properly shaken and poured onto handheld Brix refractom-eter (Model: RHB-32ATC, from Huake Instrument Co., Ltd,
Baoan, Shenzhen, China; the Brix range is from 0 to 32% withautomatic temperature compensation) to measure the Brix.The spectra scanning and Brix measurement were carriedout in the same day. Although the standard sugarcane quality
is measured using CCS which is derived from Brix, Pol andfibre content, only Brix value was used in this study becauseit is a relevant industrial parameter for sugarcane industry
which is the easiest, least expensive quality parameter to be
measured with little preparation. The correlation betweenBrix and CCS has been reported by Staunton, Donald, andPope (2011).
2.4. Pretreatment of spectral data
Both reflectance and absorbance spectra were pre-processed
for optimal performance. Pretreatment of spectral data is a
key part of spectral analysis to improve the accuracy. Several
common chemometrics methods were investigated including
smoothing technique of moving average, multiplicative scat-
ter correction (MSC), first and second derivatives, standard
normal variate (SNV) transformation and mean normal-
isation. After some trials and computations, MSCwas found to
be the best pretreatment method for this study. In fact, MSC
technique is the most popular normalisation technique
offered by most chemometrics software packages (Ns,
Isaksson, Fearn, & Davies, 2004). MSC is used to compensate
for additive (baseline shift) and multiplicative (tilt) effects in
the spectral data, which are induced by physical effects, such
as the non-uniform scattering throughout the spectrumas thedegree of scattering is dependent on the wavelength of the
radiation, the particle size and the refractive index. The
-
was selected because it is an excellent pattern classifier
(Torrecilla, Otero, & Sanz, 2004). The ANN selected was a feed-
forward network with supervised learning. The ANN model
consists of two layers with connections to the outside world
(an input layer where data are presented to the network and
an output layer which holds the networks response to given
inputs), and one hidden layers with ten neurons. More hidden
layers may cause over fitting, since the network focuses
The summary of statistical characteristics for calibration and
b i o s y s t em s e ng i n e e r i n g 1 1 5 ( 2 0 1 3 ) 1 5 4e1 6 1 157pretreatment processes were carried out using The Un-
scrambler V 9.6 software (Camo Process, AS, Oslo, Norway).
2.5. Principal component analysis (PCA)
PCA is a well-known multivariate data reduction method
which transforms original data into new orthogonal variables
referred to as principal components, or PCs (Purcell, Leonard,
OShea, & Kokot, 2005). PCA produces a reduced representa-
tion of the training data based on the maximum variation
between the spectra (Park et al., 2003). Usually, only a few PCs
or latent variables (LVs) are required to describe most of the
data variance with the first PC accounting for the greatest
amount of variance. A low number of PCs were normally
desirable in order to avoid inclusion of signal noise in the
modelling (Xiaobo, Jiewen, Xingyi, & Yanxiao, 2007). In this
paper, PCA for Partial least square (PLS) models was imple-
mented using The Unscrambler V 9.6. During the building of
calibration models, PCA was also used to detect outlier sam-
ples which could affect model performance. In this paper, one
outlier was identified and removed before PLS modelling.
2.6. Partial least square
PLS is a well-known factor analysis multivariate method prin-
cipally applied for predictive purposes (Purcell et al., 2005).
It requires a calibration step in which a model is constructed
from a number of significant factors, which are selected, for
example, by the well-known cross-validation leave-one-out
method. PLS analysis is widely utilized in near infrared spec-
troscopy analyses. PLS analysis can be performed to establish a
regression model to predict quality parameters of sugarcane.
PLS considers simultaneously the variable matrix Y (Brixvalues) and the variable matrix X (spectral data). Thus, PLS
method was used in this study to interpret the spectra and
develop both calibration and prediction models for sugarcaneBrix. In the development of the PLSmodel, full cross validation(leave-one-out) was used to evaluate the quality and prevent
over fitting of the calibration model (Arana, Jaren, & Arazuri,
2005). The maximum LVs number for an acceptable PLS model
isusually ten (Moghimietal., 2010). In thispaper, thePLSmodels
were run using ten LVs as identified by The Unscrambler V 9.6.
External validationmethodwas used in this paper to check
the PLS model performance. The external validation proce-
dure determines the predictive ability of an equation, based
on a sample set which has not been used in the calibration
development. Before the calibration, samples were divided
into two sets. One part (75% of samples) was used to develop a
prediction equation (calibration set) and another part (25% of
samples) was used to validate the predictive equation (vali-
dation set). Samples for validationwere selected by taking one
of every four samples from the entire sample set, taking care
to ensure that each set included samples that covered the
entire range of Brix values.
2.7. Description of classification model using theartificial neural networksThe ANN used in this paper was a perceptron model also
known as back-propagation perceptron. This type of networkprediction data sets of stalk samples is shown in Table 2. The
calibration and prediction data set showed similar means,
ranges and standard deviation. A relativelywide range of Brixvalues was obtained due to the inclusion of three different
varieties with different maturity stages. The range of Brixvalues for internode samples from the top to the bottom of the
Q155, Q208 andQ190 varieties were 7.6e22.2, 8e21.4 and 8e21,
respectively. The variation of Brix values along the stalk foreach variety as measured from each internode sample is
shown in Fig. 2. The downward trend of Brix values along thestalk is consistent for all varieties. The graph also showed that
different varieties have different internode numbers. It can be
seen that the early maturing variety (Q155) had higher Brixvalues than other varieties especially for the several in-
ternodes at the bottom of the stalks.
The example of raw reflectance spectra (R) and absorbance
spectra (A) of three samples having high (22 Brix), medium
Table 1 e Classification table for sugarcane Brix.
Class Brix range
High (H) 19.3e22.2
Medium high (MH) 16.5e19.2
Medium (M) 13.5e16.4
Medium low (ML) 10.5e13.4excessively on the idiosyncrasies of individual samples (Ruan,
Almaer, & Zhang, 1995). The input layer had five neurons
which corresponded to first five PCs as determined by PCA
method.
After removing one outlier from the original training data
set (200), the PCA method was then run over the training set
consisting of 219 samples while the testing set had 71 sam-
ples. The output layer consists of five neurons corresponding
to each Brix range as defined in classification table (Table 1).The ANNprocess is carried out by adjusting theweight of each
node, using the transfer function. The ANN algorithm
including PCAmethod for classification purposes in this study
was programmed and executed in Matlab (Version 7, The
Mathworks Inc. Natick, MA, USA). The transfer function
employedwas the sigmoid function bounded between 0 and 1.
The numerical values of the input and output variables used
by the ANN were normalised values in the range of 0e1.
3. Results and discussions
3.1. Overview of the spectra and statistic characteristicsof BrixLow (L) 7.6e10.4
-
(18 Brix) and low (14.2 Brix) Brix values as measured by Vis/SWNIR are shown in Fig. 3(a) and (c) respectively. In each
Table 2 e Statistical characteristics of Brix in sugarcanestalks for calibration and validation.
Model Sample no Min Max Mean SD
Calibration 200 7.5 22.2 17.86 3.04
Prediction 71 8.2 22 17.83 2.93
SD Standard deviation.
b i o s y s t em s e n g i n e e r i n g 1 1 5 ( 2 0 1 3 ) 1 5 4e1 6 1158figure, no obvious difference could be seen in the shape of the
spectra for different Brix values for both figures. However,a baseline shift problem existing in raw reflectance and
absorbance spectra has been eliminated usingMSCmethod as
shown in Fig. 3(b) and (d) respectively.
3.2. The prediction of sugarcane Brix
PLS models were developed using the pretreated reflectance
spectra by MSC technique. The performance of the final PLS
calibrationmodels could be evaluated by the standard error of
calibration (SEC), the coefficient of determination for calibra-
tion (R2), the standard error of prediction (SEP), and the coef-
ficient of determination for prediction (R2). A proper model
should have a low SEC, SEP and root mean square error of
prediction (RMSEP) and a high coefficient of determination for
both prediction and calibration models. The values of these
indices for PLS models developed using reflectance and
absorbance spectra are shown in scatter plots in Fig. 4(a) and
(b) respectively. In both figures, the ordinate and abscissa
represent the predicted andmeasured values of the Brix. TheR2 and RMSEP values for reflectance spectra were 0.91 and 1.41
while for absorbance spectra were 0.89 and 1.51 respectively.
The prediction performance of PLS models for both reflec-
tance and absorbance spectra showed good agreement be-
tween the reference and estimated values. Shenk and
Westerhaus (1996) suggested that an R2 value greater than
0.9 indicates excellent quantitative information of themodels.
From Fig. 4, it can be seen that the PLS model developed
with reflectance spectra performed better than the model
developed with absorbance spectra as indicated by R2 and
RMSEP. This is reasonable finding since the absorbance spectra
is related to the presence of chemical components such as
sugars, while the reflectance spectra of fruit contain informa-
tion on both the absorption as well as scattering properties of
24Q155 Q208 Q19010
12
14
16
18
20
22
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
B
rix
Internode Number
Fig. 2 e Typical average Brix values of internodes fordifferent sugarcane varieties.the tissue (Nicola et al., 2008). Since this study was performed
on skin surface, scattering properties of the tissue would also
have contributed to the model performance. The ten LVs used
in PLS model were also found to be sufficient for predictingBrix value from skin surface of sugarcane stalks.
The R2 value of sugar content (Brix) obtained in this paperfrom reflectance spectra was slightly better than those ob-
tained by Peirs, Lammertyn, Ooms, and Nicola (2001) with R2
values ranged between 0.73 and 0.89 using different apple
varieties. The result of this study was also better than that
reported by Wu et al. (2008) who used Vis/NIR to detect an
infection on eggplant leaves with prediction ability with an
85% accuracy rate. It is also better than that reported by
Schaare and Fraser (2000) who obtained R2 of 0.86 for pre-
dicting soluble solids content (SSC) on kiwifruit. Chia et al.
(2012) also reported lower prediction value (R2 0.68) whenpredicting sugar content from pineapple. This result is the
same as obtained by Khuriyati and Matsuoka (2004) for non-
destructively scanning tomato (R2 0.91) on the skin to pre-dict sugar content. However, the result of this study is lower
than the result (R2 0.93) reported by Moghimi et al. (2010) forpredicting sugar content of kiwifruit. Overall, this proposed
technique has demonstrated that Vis/SWNIR can be used to
predict sugar content of sugarcane based on skin scanning.
This finding suggested that spectroscopic method has the
potential to be mounted on a chopper harvester for producing
a quality map across the field.
3.3. Sugarcane Brix classification by artificial neuralnetwork
Table 3 shows the accuracy of classifying sugarcane Brixusing ANN with five PCs. Five classes with their respective
threshold values were predefined from the distribution ofBrix as discussed in Table 1. In this model, the accuracy ofclassification ranged from 50% to 100% with the average ac-
curacy being 83.1%. The accuracy ranges obtained in this
studywere better than that (45.2e93.5%) reported by Park et al.
(2003) for classifying soluble solids of Gala apples. However,
the result of this study was lower than the result (99.5% ac-
curacy) reported by Mohan, Karunakaran, Jayas, and White
(2005) who used ANN to classify grain quality.
From Table 3, it can be seen that two dominant classes
representing high and low Brix level (H and L classes) hadachieved 100% accuracy. The M class was also good with an
accuracy of 90.9%. However, both inter-middle classes (MH
and ML) showed lower accuracy. This low accuracy (misclas-
sification) which only affected the boundary samples was
probably due to these two inter-middle classes had Brixvalues which were in a borderline between the main classes.
Thus, they are having similar reflectance spectra with the
major classes. However, this misclassification is considered
minor because it is just affected the inter-middle group of
quality classes. The accuracy of the classification could be
improved if only three classes were used. Overall, the accu-
racy of this study can be considered as an acceptable and
useful especially for mapping the quality regions in the field.
The ability to map the quality regions across the paddockcould help growers to better manage their fields thus improve
the crop production in both yield and quality.
-
10
20
30
40
50
400 500 600 700 800 900 1000
Re
fle
cta
nc
e (%
)
Wavelength (nm)
22 Brix 18 Brix 14.2 Brix(a) (b)
(d)
15
20
25
30
35
40
45
400 500 600 700 800 900 1000
Re
fle
cta
nc
e (%
)
Wavelength (nm)
22 Brix 18 Brix 14.2 Brix
2
3
4
5
6
7
8
400 500 600 700 800 900 1000
Ab
so
rb
an
ce
Lo
g (1
/R
)
Wavelength (nm)
22 Brix 18 Brix 14.2 Brix(c)
2
3
4
5
6
7
400 500 600 700 800 900 1000
Ab
so
rb
an
ce
Lo
g (1
/R
)
Wavelength (nm)
22 Brix 18 Brix 14.2 Brix
Fig. 3 e Typical sugarcane spectrum at different Brix values for (a) raw reflectance spectra; (b) raw reflectance spectrapretreated with MSC; (c) raw absorbance spectra; (d) absorbance data pretreated with MSC.
12
14
16
18
20
22
12 14 16 18 20 22
Pre
dic
te
d (
B
rix
)
Reference (Brix)
RRMSEP
n = 71slope = 0.66
2 = 0.89= 1.51
(a)
10
12
14
16
18
20
22
10 12 14 16 18 20 22
Pre
dic
te
d (
B
rix
)
Reference (Brix)
(b)
R2
n = 71Slope = 0.72
= 0.91RMSEP = 1.41
Fig. 4 e Prediction versus reference values for prediction models of sugarcane Brix using (a) absorbance spectra and (b)reflectance spectra.
b i o s y s t em s e ng i n e e r i n g 1 1 5 ( 2 0 1 3 ) 1 5 4e1 6 1 159
-
Medium 10.5e13.4 4 2 50
b i o s y s t em s e n g i n e e r i n g 1 1 5 ( 2 0 1 3 ) 1 5 4e1 6 11604. Conclusions
Vis/SWNIR is a portable, low cost and non-destructivemethod
which was applied in this study to predict sugarcane Brixbased on stalk scanning. The PLS models showed a good
agreement between the reference and estimated values for
both reflectance and absorbance spectrawith R2 were 0.91 and
0.89 respectively. The ANN used to classify Brix into severalquality classes had yielded good classification performance
ranging from 50 to 100% accuracy with overall accuracy of
83.1%. Overall, this study has demonstrated that the Vis/
SWNIR spectroscopy technique coupled with ANN has the
potential to be used for on-line quality measurement to fulfil
the requirement of PA. This study also demonstrated that
quality prediction based skin scanning using a Vis/SWNIR
spectroscopy is feasible. This technique will be a better
alternative to current quality measurement which is based on
juice samples. Thus, the next study will focus on how to apply
this technique to map sugarcane Brix during harvesting inthe field. However, an effort is required to design a proper
sampling mechanism and installation configuration before
this method can be applied on a harvester. Spectral variations
due to different varieties, planting regions, seasons, weather
and cultural conditions also need to be considered in the
future works.
Acknowledgements
The authors acknowledge the financial supports provided by
Ministry of Higher Education, Malaysia and National Center
Low (ML)
Low (L) 7.6e10.4 3 3 100
Total 71 59 83.1Table 3 e ANN classification results.
Class Brixrange
No. of testsample ineach group
No. of correctclassification
Accuracy(%)
High (H) 19.3e22.2 30 30 100
Medium
High (MH)
16.5e19.2 23 14 60.9
Medium (M) 13.5e16.4 11 10 90.9for Engineering in Agriculture (NCEA), Toowoomba, Australia.
The authors also thank BSES Limited, Bundaberg, for
providing samples and equipment.
r e f e r e n c e s
Arana, I., Jaren, C., & Arazuri, S. (2005). Maturity, variety andorigin determination in white grapes (Vitis vinifera L.) usingNIRS. Journal of Near Infrared Spectroscopy, 13, 349e357.
ASD. (2005). Handheld spectroradiometer: Users guide version 4.05.Suite A Boulder, USA: Analytical Spectral Devices, Inc.
Bramley, R. G. V. (2009). Lessons from nearly 20 years of precisionagriculture research, development, and adoption as a guide toits appropriate application. Crop & Pasture Science, 60, 197e217.Bramley, R. G. V., Panitz, J. H., Jensen, T., & Baillie, C. (2012).Within block spatial variation in CCS e another potentiallyimportant consideration in the application of precisionagriculture to sugarcane production. Proceedings of theAustralian Society of Sugar Cane Technologists, 34, 1e8.
Chia, K. S., Rahim, H. A., & Rahim, R. A. (2012). Prediction ofsoluble solids content of pineapple via non-invasive low costvisible and shortwave near infrared spectroscopy and artificialneural network. Biosystems Engineering, 113, 158e165.
Cox, G., Harris, H., Pax, R., & Dick, R. (1996). Monitoring cane yieldby measuring mass flow rate through the harvester.Proceedings of the Australian Society of Sugar Cane Technologists,152e157.
Digman, M. F., & Shinners, K. J. (2008). Real time moisturemeasurement on a forage harvester using near infraredreflectance spectroscopy. Transactions of the ASABE, 51,1801e1810.
Guthrie, J., & Walsh, K. (1997). Non-invasive assessment ofpineapple and mango fruit quality using near infra-redspectroscopy. Australian Journal of Experimental Agriculture, 37,253e263.
Haykin, S. (1994). Neural network: A comprehensive foundation. NewYork, NY: Macmillan.
Johnson, R. M., & Richard, E. P., Jr. (2005). Sugarcane yield,sugarcane quality, and soil variability in Louisiana. AgronomyJournal, 97, 760e771.
Khuriyati, N., & Matsuoka, T. (2004). Near infrared transmittancemethod for nondestructive determination of soluble solidcontent in growing tomato fruits. Environmental Control inBiology, 42, 217e223.
Lee, W. S., Alchanatis, V., Yang, C., Hirafuji, M., Moshou, D., &Li, C. (2010). Sensing technologies for precision specialty cropproduction. Computers and Electronics in Agriculture, 74, 2e33.
Liu, F., He, Y., & Wang, L. (2008). Comparison of calibrations forthe determination of soluble solids content and pH of ricevinegars using visible and short-wave near infraredspectroscopy. Analytica Chimica Acta, 610, 196e204.
Lu, R. (2004). Multispectral imaging for predicting firmness andsoluble solids content of apple fruit. Postharvest Biology andTechnology, 31, 147e157.
Magalhaes, P. S. G., & Cerri, D. G. P. (2007). Yield monitoring ofsugar cane. Biosystems Engineering, 96(1), 1e6.
Menesatti, P., Antonucci, F., Pallottino, F., Roccuzzo, G.,Allegra, M., Stagno, F., et al. (2010). Estimation of plantnutritional status by Vis/NIR spectrophotometric analysis onorange leaves [Citrus sinensis (L) Osbeck cv Tarocco]. BiosystemsEngineering, 105, 448e454.
Moghimi, A., Aghkhani, M. H., Sazgarnia, A., & Sarmad, M. (2010).Vis/NIR spectroscopy and chemometrics for the prediction ofsoluble solids content and acidity pH of kiwifruit. BiosystemsEngineering, 106(3), 295e302.
Mohan, L. A., Karunakaran, C., Jayas, D. S., & White, N. D. G.(2005). Classification of bulk cereals using visible and NIRreflectance characteristics. Canadian Biosystems Engineering,47(7), 7e14.
Montes, J. M., Utz, H. F., Schipprack, W., Kusterer, B.,Muminovic, J., Paul, C., et al. (2006). Near-infraredspectroscopy on combine harvesters to measure maize graindry matter content and quality parameters. Plant Breeding, 125,591e595.
Nicola, B. M., Verlinden, B. E., Desmet, M., Saevels, S., Saeys, W.,Theron, K., et al. (2008). Time-resolved and continuous waveNIR reflectance spectroscopy to predict soluble solids contentand firmness of pear. Postharvest Biology and Technology, 47,68e74.
Ns, T., Isaksson, T., Fearn, T., & Davies, T. (2004). A user-friendlyguide to multivariate calibration and classification. Charlton,
Chichester, UK: NIR Publications.
-
Panigrahi, S., & Hofman, V. (2003). On the go sugar sensor fordetermining sugar content during harvesting. U.S. Patent No.6,624,888.
Park, B., Abbott, J. A., Lee, K. J., Choi, C. H., & Choi, K. H. (2003).Near-infrared diffuse reflectance for quantitative andqualitative measurement of soluble solids and firmness ofdelicious and gala apples. Transactions of the ASAE, 46(6),1721e1731.
Peirs, A., Lammertyn, J., Ooms, K., & Nicola, B. M. (2001).Prediction of the optimal picking date of different applecultivars by means of VIS: NIR-spectroscopy. PostharvestBiology and Technology, 21, 189e199.
Price, R. R., Johnson, R. M., Viator, R. P., Larsen, J., & Peters, A.(2011). Fiber optic yield monitor for a sugarcane harvester.Transactions of the ASABE, 54(1), 31e39.
Purcell, D. E., Leonard, G. J., OShea, M. G., & Kokot, S. (2005).A chemometrics investigation of sugarcane plant propertiesbased on the molecular composition of epicuticular wax.Chemometrics and Intelligent Laboratory Systems, 76, 135e147.
Qing, Z., Ji, B., & Zude, M. (2007). Predicting soluble solid contentand firmness in apple fruit by means of laser lightbackscattering image analysis. Journal of Food Engineering, 82,58e67.
Ruan, R., Almaer, S., & Zhang, J. (1995). Prediction of doughrheological properties using neural networks. Cereal Chemistry,72(3), 308e311.
Schaare, P. N., & Fraser, D. G. (2000). Comparison of reflectance,interactance and transmission modes of visible-near infraredspectroscopy for measuring internal properties of kiwifruit(Actinidia chinensis). Postharvest Biology and Technology, 20,175e184.
Shenk, J. S., & Westerhaus, M. O. (1996). Calibration the ISI way. InA. M. C. Davies, & P. C. Williams (Eds.), Near infraredspectroscopy: The future waves (pp. 198e202). Chichester: NIRPublications.
Staunton, S., Donald, D., & Pope, G. (2011). Estimating sugarcanecomposition using ternary growth relationships. Proceedings ofthe Australian Society of Sugar Cane Technologists, 33.
Torrecilla, J. S., Otero, L., & Sanz, P. D. (2004). A neural networkapproach for thermal/pressure food processing. Journal of FoodEngineering, 62, 89e95.
Walsh, K. B., Guthrie, J. A., & Burney, J. W. (2000). Application ofcommercially available, low-cost, miniaturised NIRspectrometers to the assessment of the sugar content of intactfruit. Australian Journal of Plant Physiology, 27, 1175e1186.
Wendte, K. W., Skotnikov, A., & Thomas, K. K. (2001). Sugar caneyield monitor. U.S. Patent No. 6,272,819.
Williams, P. C., & Norris, K. H. (2001). Near-infrared technology in theagricultural and food industries (2nd ed.). St. Paul, MN: AACC.
Wu, D., Feng, L., Zhang, C., & He, Y. (2008). Early detection ofBotrytis cinerea on eggplant leaves based on visible and near-infrared spectroscopy. Transactions of the ASABE, 51(3),1133e1139.
Xiaobo, Z., Jiewen, Z., Xingyi, H., & Yanxiao, L. (2007). Use of FT-NIR spectrometry in non-invasive measurements of solublesolid contents SSC of Fuji apple based on different PLSmodels. Chemometrics and Intelligent Laboratory Systems, 87,43e51.
Xing, J., Symons, S., Shahin, M., & Hatcher, D. (2010). Detection ofsprout damage in Canada Western Red Spring wheat withmultiple wavebands using visible/near-infrared hyperspectralimaging. Biosystems Engineering, 106, 188e194.
b i o s y s t em s e ng i n e e r i n g 1 1 5 ( 2 0 1 3 ) 1 5 4e1 6 1 161
Prediction and classification of sugar content of sugarcane based on skin scanning using visible and shortwave near infrared1. Introduction2. Material and methods2.1. Sample preparation2.2. Reflectance measurement2.3. Brix measurement2.4. Pretreatment of spectral data2.5. Principal component analysis (PCA)2.6. Partial least square2.7. Description of classification model using the artificial neural networks
3. Results and discussions3.1. Overview of the spectra and statistic characteristics of Brix3.2. The prediction of sugarcane Brix3.3. Sugarcane Brix classification by artificial neural network
4. ConclusionsAcknowledgementsReferences