prediction and classification of sugar content of sugarcane based on skin scanning using visible and...

Upload: aleine-leilanie-oro

Post on 14-Oct-2015

16 views

Category:

Documents


3 download

DESCRIPTION

Thesis

TRANSCRIPT

  • 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