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: nazmi@eng.upm.edu.my, 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