Rapid detection of anthocyanin content in lychee pericarp duringstorage using hyperspectral imaging coupled with model fusion

Yi-Chao Yang a, Da-Wen Sun a,b,*, Hongbin Pu a, Nan-Nan Wang a, Zhiwei Zhu a

aCollege of Light Industry and Food Sciences, South China University of Technology, Guangzhou 510641, Chinab Food Refrigeration and Computerized Food Technology, Agriculture and Food Science Centre, University College Dublin, National University of Ireland,Belfield, Dublin 4, Ireland

A R T I C L E I N F O

Article history:Received 6 October 2014Received in revised form 22 January 2015Accepted 23 February 2015Available online 11 March 2015

Keywords:Hyperspectral imaging (HSI)LitchiAnthocyaninRadial basis function support vectorregression (RBF-SVR)Model fusionRadial basis function neural network (RBF-NN)

A B S T R A C T

A quantitative approach was proposed to evaluate anthocyanin content of lychee pericarp usinghyperspectral imaging (HSI) technique. A HSI system working in the range of 350–1050nmwas used toacquire a 3-D lychee image. Successive projection algorithm (SPA) and stepwise regression (SWR)algorithmwere utilized to reduce data dimensionality and search for optimal wavelengths related withanthocyanin content in pericarp. Radial basis function support vector regression (RBF-SVR) was adoptedto establish quantitative relationship between hyperspectral image information in two sets of optimalwavelengths and anthocyanin content of pericarp. Finally, in order to improve prediction accuracy, SPA-RBF-SVR and SWR-RBF-SVR models were fused into a single model by radial basis function neuralnetwork (RBF-NN) algorithm. The results revealed that the fused model possessed a better performancethan either SPA-RBF-SVR or SWR-RBF-SVR models alone, as the fused model showed higher coefficientsof determination (R2) of 0.891 and 0.872, and lower root mean square errors (RMSEs) of 0.567% and0.610% for the training and the testing sets, respectively. Visualization maps based on the fused modelwere generated to display the anthocyanin distribution within lychee pericarp. This study demonstratesthat HSI is capable of predicting and visualizing anthocyanin evolution in the pericarp of lychee duringstorage.

ã 2015 Elsevier B.V. All rights reserved.

1. Introduction

Quality is a very important factor for the development of theagricultural industry, Therefore methods and techniques such asdrying (Sun and Byrne, 1998; Sun and Woods, 1997; Delgado andSun, 2002a,b), refrigeration (Sun, 1997a,b; Sun et al., 1996;McDonald and Sun, 2001; McDonald et al., 2001; Kiani and Sun,2011) and edible coating (Xu et al., 2001) are often used to ensureproduct quality. Lychee or litchi (Litchi chinensis Sonn.), as a non-climacteric subtropical and tropical fruit, is one of the importantagricultural products, especially in China, Vietnam and the rest oftropical Southeast Asia. At full maturity, its pericarp is bright redand white flesh is sweet and juicy. After harvest, the pericarprapidly desiccates and turns brownness within 3 days at ambienttemperature (Zhang and Quantick, 1997). Pericarp browning

decreases commercial value of lychee and has been regarded asthe major problem of postharvest lychee (Jiang et al., 2006). Therate of lychee browning is closely related to degradation of redpigments (identified as anthocyanins) and the formation of brown-colored by-products (Huang et al., 1990; Lee and Wicker, 1991).Therefore, anthocyanins are considered as one of the mostsignificant parameters for evaluating lychee quality and areincreasingly used to grade fruits into different quality levels.Therefore besides utilizing technologies for keeping its quality, it isequally important to develop methods for evaluating its quality.The available methods for assessing anthocyanin content arecommonly based on ultraviolet-visible spectrophotometry (Joaset al., 2005) and chromatographic analyses (Zhang et al., 2003).These methods are functional but inefficient and difficult to beapplied to on-line inspection of anthocyanin content. Consequent-ly, there is a need for rapid and non-contact techniques to inspectthe fruit quality.

Hyperspectral imaging (HSI) is a cutting-edge optical techniquethat combines imaging or computer vision (Sun and Brosnan,2003; Valous et al., 2009; Jackman et al., 2008; Sun, 2004; Patricket al., 2009; Wang and Sun, 2002) with spectroscopic techniquesinto a single system (Wu and Sun, 2013a,b). It can continuously and

* Corresponding author at: South China University of Technology, College of LightIndustry and Food Sciences, Guangzhou 510641, China. Tel.: +353 1 7167342; fax:+353 1 7167493.

E-mail addresses: dawen.sun@ucd.ie, http://www.ucd.ie/refrig,http://www.ucd.ie/sun (D.-W. Sun).

http://dx.doi.org/10.1016/j.postharvbio.2015.02.0080925-5214/ã 2015 Elsevier B.V. All rights reserved.

Postharvest Biology and Technology 103 (2015) 55–65

Contents lists available at ScienceDirect

Postharvest Biology and Technology

journal homepage: www.elsev ier .com/ locate /postharvbio

rapidly scan fruit samples and simultaneously provide spectral andspatial information for each pixel, making it carry a great potentialin non-destructively detecting food quality attributes and visual-izing their spatial distribution (Cubero et al., 2011; Kamruzzamanet al., 2011, 2012; ElMasry et al., 2011a,b, 2012; Barbin et al., 2012;Wu et al. 2012c). The feasibility of using HSI for evaluatingpigments of fruits has been demonstrated by Fernandes et al.(2011), in which a visible-near infrared (Vis-NIR) HSI (400–1000nm) system was developed to determine anthocyaninconcentration of grape (Cabernet Sauvigon variety) skins. Adaptiveboosting neural network algorithm was adopted to establishcalibration model of anthocyanin content. The principal compo-nents of grapes’ spectra were the inputs of the adaptive boostingneural network, and anthocyanin concentrations measured usingconventional laboratory techniques were the outputs. A squaredcorrelation coefficient of 0.65 was achieved, which revealed howbeneficial the development of a neural network performance couldbe. In another study, Qin and Lu (2008) measured opticalproperties of fruits (apples, peaches, pears, kiwifruits and plums)and vegetables (cucumbers, zucchini squash and tomatoes) by Vis-

NIR (500–1000nm) HSI. Three major pigments; i.e., chlorophyll,anthocyanin and carotenoid, were investigated. A good correlationwas found between diffuse reflectance spectra of samples and thethree pigments concentrations.

The above several studies were the only papers available ondetermining pigments of fruits by HSI. At present, no research hasbeen done on application of HSI for evaluating pigment content oflychee fruits, especially anthocyanin content of pericarp. The lackof such studies may be because it is quite difficult to estimatepigment content of lychee by HSI due to its short shelf life, shapeand bumpy surface. Lychee is a seasonal fruit with the marketingtime of 2–3 months and the shelf life of only 3–5 days, leading tothat the quality of lychee is ever-changing. This increases thedifficulty of measuring quality attributes of many lychee samplesin the same conditions using traditional methods for developingprediction models. Furthermore, the curvature of lychee fruit cancause an uneven distribution of light on its surface when it isirradiated by a light source, causing the peripheral area of lycheedarker than the central area in the measured hyperspectral image.In addition, as noted by Huang et al. (2011), bumpy surface of

[(Fig._1)TD$FIG]

Fig. 1. Key steps of the experimental procedure and data processing.ROI: region of interest; SPA: successive projection algorithm; SWR: stepwise regression; RBF-SVR: radial basis function support vector regression; RBF-NN: radial basisfunction neural network.

56 Y.-C. Yang et al. / Postharvest Biology and Technology 103 (2015) 55–65

lychee also makes the nondestructive evaluation of lychee qualityby HSI more difficult.

However, the above problems could be solved by somealgorithms from remote sensing and mathematical modelingfields. Spatial variation of light intensity could be corrected bygeometric correction factors from a 3-D model produced from a 2-D lychee image. A mask with the pixels from the 2-D image thatbelong to lychee fruit was acquired first to fix the exact position ofthe lychee fruit. Then the center of mask of the lychee fruit wascalculated based on distance between the points on the mask,which was considered to lie on the Z-axis at a height equivalent toaverage radius of the fruit. The 3-Dmodelwas developed accordingto the maximum height and the center point of the fruit. Moredetailed procedures could be found in the study reported byGómez-Sanchis et al. (2008). The performance of models forevaluating fruit quality can be improved by model fusion. Themodel fusion methods have been demonstrated to be a practicalapproach in improving performance of model (Wang et al., 2012).Themajority of previous studies have adopted a single algorithm toselect optimal wavelengths. Although such a method was feasible,some wavelengths with useful information might be rejected,thereby reducing the performance of the models. In order toeliminate the limitations associated with using a single algorithmof wavelength selection and to develop better valuation models, afew researchers have proposed somemodel fusion approaches thatused different wavelength selection algorithms to process the HSIdata. For example, Wang et al. (2012) proposed a method of modelfusion for predicting apple firmness. First, two sets of optimalwavelengths were selected by uninformative variable elimination(UVE) and supervised affinity propagation (SAP), respectively.Second, validation models based on the sets of optimal wave-lengths were developed using PLS algorithm. Third, apple firmnesswas predicted by UVE-PLS and SAP-PLS models and predictionvalues were imported as input layer of back propagation neuralnetwork (BP-NN), while the measured values of apple firmnesswere used as output data of the network. Finally, the network wastrained for achieving the best performance. The results indicatedthat the performance of validation models was considerablyimproved by fusing UVE-PLS and SAP-PLS models into a singlemodel, and the best prediction accuracy was reached by the fusedBP-NN model.

According to Nanyam et al. (2012), fusion methods includedata, feature, and decision fusions. Enlightened by Wang et al.(2012), the decision fusion was adopted in the current study. Thedecision fusion procedure was described as follows. Firstly,successive projection algorithm (SPA) and stepwise regression(SWR) algorithms were used to selected optimal wavelengths,respectively. Pericarp color and spectral reflectance in two sets ofselected optimal wavelengths were extracted and integrated forfurther processing. Secondly, radial basis function support vectorregression (RBF-SVR) was utilized to develop prediction model ofanthocyanin content of pericarp based on extracted image andspectral information. Finally, the two RBF-SVR models werefused into one model using radial basis function neural network(RBF-NN) algorithm for improving prediction accuracy ofanthocyanin content of pericarp and the performances of allmodels were compared. The specific objectives of the currentstudy were to:

(i) acquire hyperspectral images in the spectral range of 308–1105nm of lychees and remove the differences of lightintensity among different areas of the sample;

(ii) extract average spectra of region of interest (ROI) in the lycheeimage and select two sets of optimal wavelengths using SPAand SWR algorithms, respectively;

(iii) extract image feature of ROI from lychee image at the optimalwavelengths and build calibration models based on image andspectral information using RBF-SVR algorithm; and

(iv) fuse the SPA-RBF-SVR and SWR-RBF-SVR models into a singlemodel using RBF-NN and generate visualization map ofanthocyanin distribution by the fused model.

2. Materials and methods

2.1. Sample preparation

Lychee samples (Feizixiao variety, at commercial ripenessstage) were supplied in June 2013 from an orchard inGuangzhou, China. Uniform fruits free from physical damagewere selected as samples by visual inspection. 360 fruits with adiameter in the range of 31.47–33.59mm were divided into six

[(Fig._2)TD$FIG]

Fig. 2. Schematic diagram of hyperspectral imaging system.

Y.-C. Yang et al. / Postharvest Biology and Technology 103 (2015) 55–65 57

groups, containing 60 fruits in each. All fruits were put into aplastic bucket filled with 0.05% sporgon (the prochloraz-manganese chloride complex) (Hoechst Schering AgrEvo GmbH,Düsseldorf, Germany) for 2min and air-dried at room tempera-ture so as to inhibit the growth of bacteria on the surface of thelychee (Chen et al., 1997). These samples were stored in anenvironmental control chamber (SPX-300IC, Suzhou JiangdongPrecision Instruments Ltd., Suzhou, China) maintained at 27 �Cand 85% relative humidity (RH) (Underhill and Simons, 1993).Before being scanned by a HSI system, samples were removedfrom storage and left for about 30min to reach roomtemperature (25 �C). Finally, lychees stored for 0–5 days werescanned and the anthocyanin contents of their pericarps weredetermined by an ultraviolet spectrophotometer. Fig. 1 showsthe key steps of the whole experimental procedure.

2.2. Hyperspectral imaging system

Hyperspectral images were acquired using a line-scanning HSIsystem working in the spectral range of 308–1105 nm. As shownin Fig. 2, the HSI system consisted of four components: a spectralimaging system with a spectrograph, a charged couple device(CCD) camera and a lens; a lighting system; a translation stageand a computer. The spectrograph (ImSpector V10E, SpectralImaging Ltd., Oulu, Finland) with spectroscopic resolution of1.6 nm was connected to the CCD camera (DL-604M, Andor Co.,Chicago, USA) with effective resolution of 1004 pixels in thespatial dimension and 1002 pixels (bands) in the spectraldimension, which was equipped with a standard C-mount zoomlens. The lighting system was composed of two 150W halogenlamps (Olympus Optical Co., Tokyo, Japan) to illuminate thesample. The translation stage presented the sample to thespectral imaging system in different, non-overlapping positions,in order to inspect as much lychee surface as possible. Thecomputer, equipped with hyperspectral imaging analyzer soft-ware (V10E, Isuzu Optics Co., Taiwan, China), controlled themovement of the motor, conducted the scan and acquired thespectral image.

2.3. Image acquisition and calibration

In order to acquire accurate hyperspectral data, several systemparameters were set before scanning samples. The distancebetween the upper surface of samples and the CCD camera lenswas set as 515mm. The exposure time was 30ms, and the speed oftranslation stage was 1.0mm/s. During the test, the lychee samplewas placed on the black translation stage and scanned line-by-lineby a spectrograph with the movement of translation stage. Theactual size that a pixel corresponded to in the lychee sample wasclosely related to the distance between camera lens and samples.Under the condition of 515mmdistance, the lengths in both x and ydirections were 0.128mm. All the operations were carried out in adark chamber.

To remove effect of dark current, before scanning samples,white reference imagewas acquired with awhite Spectralon panel(�99% reflectance), followed by a dark image that was obtainedwith all light sources off and a black cap covering the camera lens.To obtain relative reflectance of lychee samples, flat-fieldcorrections were performed on original hyperspectral reflectanceimages using the equation below:

I ¼ I0 � IdIw � Id

� 100% (1)

where I is the corrected image, I0 is the raw hyperspectral image, Iwis the white reference image and Id is the dark image.

2.4. Measurement of anthocyanin content of lychee pericarp

Anthocyanin content of lychee pericarp with different storagedays were measured using the method described by Zhang andQuantick (1997). First, five pieces of pericarp were acquired fromeach sample using a sampling device (Hurun Instrument Ltd.,Guangzhou, China) with a diameter of 5.0mm. Second, the fivepieces pericarp were placed in a mortar and triturated by a pestleusing liquid nitrogen freezing method (Liu et al., 2011). Acquiredpowder was immersed in 1% HCl–methanol (4ml) for 2h tocompletely extract anthocyanin pigment of the lychee pericarp.Then all turbid solution was centrifuged at 5000 rpm and 4 �C for15min in a cold centrifuge (JW-3021HR, Anhui Jiawen Instrumentsand Equipment Co., Hefei, China). The supernate was collected anddiluted two-fold with 1% HCl–methanol. Third, absorption of thediluent at 530 and 600nm was measured using an ultraviolet–visible spectrophotometer (UV-1800, Shimadzu Co., Kyoto, Japan).The accuracy of the UV-1800 measurement for measuringanthocyanin content was �0.004A. Finally, anthocyanin contentof the sample was calculated by the following equation:

Cmgg

� �¼ A530nm � A600nmð Þ � V � n�M

e�m

� �(2)

where A530nm and A600 nm are the absorbances at the correspond-ingwavelength;V is the total volume of extracting solution; n is thedilution ratio (it is two in this study);M is the molecular weight ofcyanidin-3-glucoside (i.e., 449.4); e is amolar extinction coefficientof 29,600 (M�1 cm�1), and m is the mass of the sample.

2.5. Image preprocessing and extraction of spectral information

In this study, image preprocessing used included resizinghyperspectral image of lychee, removing image background andcorrecting differences of light intensity in lychee image. Initially, auniform size of 300�300 pixels (lines� samples) was selectedfrom the original hyperspectral image of lychee for eliminatingirrelevant data of background. This step was achieved bydetermining the center of lychee sample of each hyperspectralimage and segmenting the image at the center of the sample to a300�300 size. However, because the 300�300 area was muchlarger than the largest lychee sample, the resized image(300�300 size) could contain the whole lychee sample. Thepurpose of this step was to uniform lychee image, removeredundant and irrelevant background information and reducecomputation load. In addition, in order to remove much noisesexisting in the spectral ranges of 308–350nm and 1050–1105nm,the spectral range of uniform size image was resized to 350–1050nm with a total of 431 bands. Afterwards, a template wasobtained by subtracting the 489nm image from the 914nm imageand used to build a mask by using a simple threshold value of0.024 to the template. Image background was then removed byapplying the mask to corresponding hyperspectral image forreducing stochastic noises. After being processed by the aboveoperations, the image only contained image and spectralinformation of lychee, while background values were set as zero.

Because the shape of lychee is round, light intensity is notevenly distributed on the surface of lychee when they areirradiated by parallel light sources. This causes central areas oflychee sample brighter than the peripheral area, making not allvariance of spectral reflectance be due to changes of fruit quality. Inorder to overcome the problem, geometric correction factors wereused to eliminate the spatial variance of light intensity in lycheeimage. Lycheewas a spherical fruit and it could be approximativelyconsidered as a Lambertian sphere. A 3-D Lambertian spheremodel was developed from a 2-D lychee image. Geometriccorrection factors were calculated from the 3-D model to describe

58 Y.-C. Yang et al. / Postharvest Biology and Technology 103 (2015) 55–65

the spatial variation of each pixel in lychee image. Finally,hyperspectral images of lychee were corrected by Eq. (3)(Gómez-Sanchis et al., 2008):

r lð Þ ¼ Ixy lð ÞaDcos fð Þ þ 1� aDð Þ½ � ¼

Ixy lð Þeg

(3)

where r(l) is the result after correcting Ixy(l); Ixy(l) is thecalibrated monochrome image at wavelength l and point (x,y), aD

is the ratio of direct light to total average light; f is the incidentangle between the beam of direct light and the direction of thenormal on the surface of lychee; and eg is the dimensionlessgeometric correction factor.

Before extracting spectral information of lychee image, roundareas containing about 2000 pixels were manually selected asregions of interest (ROIs). Average values of spectral reflectance ofthe ROIs were calculated and regarded as spectral information ofthe sample. All these operations were performed by the Environ-ment for Visualizing Images software (ENVI 4.7, Research SystemsInc., Boulder, CO, USA). In addition, multiplicative scatter correc-tion (MSC) and standard normal variate (SNV) were employed toeliminate scatter associated with curved surface of lychee. In theMSC method, it is assumed that scattering coefficients at allwavelengths are equal and that scattering spectra and chemicalabsorption information can be mathematically differentiated. Themethod makes it possible to compensate scatter light of differentareas and to correct uneven light intensity on the surface of roundfruit. Therefore, MSC is frequently used to improve predictionresults related to Vis-NIR data (Isaksson and Næs, 1988; Wu et al.,2012a,b). SNV is an ideal method in solving the problem of opticallength changes. Themethod is also able to eliminatemultiplicativeinterferences of scatter and achieve spectral vector normalization

(Candolfi et al., 1999). To remove slope variations on an individualspectrum basis, each original spectrum is transformed indepen-dently by subtracting the mean of all spectrum and dividing bytheir standard deviation.

2.6. Optimal wavelengths selection

In order to reduce computational load and improve theperformance of prediction model, various algorithms are oftentaken to remove redundant HSI information for selecting severaloptimal wavelengths (Wu et al., 2012a,b). In this study, SPA andSWRwere used to reduce the dimension of hyperspectral data andto find several important wavelengths closely related to anthocy-anin content of pericarp. SPA is a variable selection method thatselect optimal wavelengths by minimizing co-linearity amongwavelengths (Araújo et al., 2001), and has been used forwavelength selection in many cases (Liu et al., 2015; Wu andHe, 2014; Wu et al., 2013). It is conducted based on a sequence ofprojection operations that involve the columns of instrumentalresponse matrix. SPA calculation begins with one wavelengthrandomly and maps the wavelength to a vector space. If thewavelength can be represented by some vectors in the space, itwould be considered as collinear wavelengths. Then SPA incorpo-rates the collinear wavelengths into a new and independentwavelength at each iteration until a specified number N ofwavelengths is accomplished (Galvão et al., 2008). The objective ofSPA is to seek a set of representative variables that contain theminimum redundant information. In terms of SWR, it is awavelength selection method combining both forward andbackward selection procedures. It adds or removes variablesaccording to their significance for performance of a regression

[(Fig._3)TD$FIG]

Fig. 3. Main procedures of image feature extraction. (For interpretation of the references to color in the text, the reader is referred to the web version of this article.)

Y.-C. Yang et al. / Postharvest Biology and Technology 103 (2015) 55–65 59

model (Xu and Zhang, 2001). The significance is determined by thecorrelation coefficient between independent variables and depen-dent variable. The larger correlation coefficient of a variable has,the higher its significance is. The optimal variables are selected asthose with the highest significance. All operations of SPA and SWRalgorithms were performed using MATLAB Version 2010a (TheMathworks Inc., Natick, MA, USA).

2.7. Extraction of skin color feature

Besides spectral information of lychee pericarp, skin color isalso very significant information closely related to anthocyanincontent of pericarp (Lee and Wicker, 1991). Thus, pericarp colorwas also extracted from ROIs of lychee image in this study. Fig. 3illustrates the main procedures of extracting image feature oflychee. After selecting ROIs of lychee images, color values of eachpixel in the ROIs were extracted in three components of RGBspaces by MATLAB software. The mean values of red channel rm,green channel gm and blue channel bm were computed as colorfeatures of lychee (Karimi et al., 2012). Finally, spectralreflectance and image feature of ROIs were combined fordeveloping prediction model of anthocyanin content of lycheepericarp.

2.8. Modeling algorithms

RBF-SVR algorithm was utilized to establish quantitativerelationships between image and spectral information andanthocyanin content of pericarp. Two prediction models weredeveloped based on image and spectral information in two sets ofoptimal wavelengths. Finally, RBF-NN was applied to fuse twomodels into a single model for improving prediction accuracy ofanthocyanin content. All themodeling procedureswere carried outusing MATLAB 2010a.

2.8.1. RBF-SVR algorithmSupport vector regression (SVR) is a supervised learning

method that has been used as a powerful tool to developquantitative relationships between spectral reflectance and qualityattributes of fruits (Camps-Valls et al., 2004). The principle of SVRalgorithm is that input variables are first mapped into a high-dimensional feature space by a specific mapping function, i.e.,transfer function (Menesatti et al., 2009). In the high-dimensionalspace, the relationship between independent variables anddependent variable can be described by a linear function (Basaket al., 2007). The transfer function is the core of SVR algorithm anddifferent transfer functions can be adopted to construct differenttypes of SVRmodels. There are three common functions employed:polynomial function, sigmoid function and RBF (Zhao et al., 2006).Compared to polynomial and sigmoid functions, RBF function hasspecial advantages (e.g., fewer numerical difficulties) and has beenchosen in many studies (Buhmann, 2003). Therefore, in this study,RBF functionwas used as transfer function of SVR, which is definedas:

KRBF ¼ exp� k Xi

! �Xj! k

2s2

� �2

(4)

where KRBF is the RBF kernel function, Xi! is the training vectors,

Xj! is the mean of training vectors, and s is the width of Gaussian

function.The SVR regression equation has the following form in terms of

RBF kernel function (Huang et al., 2007):

f x!� �

¼Xli¼1

ai � a�i

� �KRBF þ b (5)

where l is the number of variables, ai and ai* are the dual variables,

and b is the error correction term.The two most important parameters in RBF-SVR model are

penalty parameter c and kernel parameter g. They were optimizedusing genetic algorithm in this study attributed to that geneticalgorithm was parallel operations starting with many points andthus it could effectively avoid trapping in local optimum (GoldbergandHolland,1988).When genetic algorithmwas conducted,firstly,it randomly produced some initial solutions (initial values of c andg) and fitness value of each initial solution was calculated by anobjective function. Then replication probability of each initialsolution was acquired according to corresponding fitness value.The higher the replication probability was, the more descendantswould be produced in the next generation. In this way, outstandingsolutions would be selected for replication, while those inferiorones would be weeded out. A new generation was produced bycrossover and mutation of these outstanding solutions. The newgeneration was superior to the last one in both gene andperformance. Finally, fitness function was applied to the newgeneration again. If ending conditions were met, the operationswould be stopped. Otherwise, the above all operations wererepeated until the ending conditions were fulfilled. The final c andg were utilized to train SVR models.

2.8.2. Model fusion using the RBF-NN algorithmModels developed by a single algorithm often have poor

generalization ability and robustness (Wang et al., 2012). Modelfusion methods can not only overcome these shortcomings butalso benefit improvement of performance and reliability ofcalibration models. Thus, it is increasingly applied to establishingthe relationship among different data sources as an effectivemeans (Wang et al., 2012). The SPA algorithm can remove the co-linearity among wavelengths, but the selected wavelengths mayhave a low signal-to-noise ratio (SNR) or be useless for calibratingthe model, thereby reducing prediction accuracy (Dai et al., 2014).SWR algorithm is a combination of forward and backwardmethods. It can identify variables with redundant informationwhen new variables are subsequently added to regressionequation, which can overcome the drawbacks of forward selectionmethod. However, extra random variables may lead SWR tobecome unstable during the process of wavelength selection,making slight changes of data cause significant differences ofresults (Mundry and Nunn, 2009). Consequently, both SPA andSWR fail to provide a valid means for selecting the mostrepresentative variables. However, model fusionmethods combineadvantages of the two algorithms and minimize their disadvan-tages. Therefore, prediction accuracy of anthocyanin content couldbe improved in fused model compared to performances ofcalibration models built by single algorithm alone. RBF-NN is anonlinear and feed-forward network that is trained bya supervisedalgorithm (Pulido et al., 1999). RBF-NN is not only available forclassification of sample varieties (Sandoval et al., 2014), but is alsosuitable for the quantitative analysis of mixtures (Dong et al.,2012). There are three layers in the topology structure: an inputlayer, a hidden layer and an output layer. For RBF-NN, the primary

Table 1Anthocyanin content in lychee pericarp stored for 0–5 days.

Storage time Max (mg/g) Min (mg/g) Mean� SD (mg/g) Range (mg/g)

0 d 2.9246 2.2492 2.7652�0.0908 0.67541d 2.4905 2.0571 2.2798�0.0963 0.43342d 2.0664 1.4725 1.8306�0.0462 0.59393d 0.9122 0.4674 0.7216�0.0870 0.44484d 0.4787 0.1339 0.3250�0.0509 0.34485d 0.2733 0.0993 0.1233�0.0586 0.17400–5d 2.9246 0.0993 1.3409�0.2154 2.8253

60 Y.-C. Yang et al. / Postharvest Biology and Technology 103 (2015) 55–65

function that conducts nonlinear transformation from input layerto hidden layer is usually a Gaussian function, such as that shownin the equation below (Shao et al., 2011):

Ø k X � ci kð Þ ¼ exp �k X � ci k2s2i

!(6)

where X is the input variables, ci and si are the central value andwidth of Gaussian function, respectively.

The main training procedure of RBF-NN is optimazation of thecentroids (i.e., ci and si) and determination of connection weightsfrom the hidden layer to the output (Lorente et al., 2013). Thereare a number of mathematical methods for selecting thecentroids, such as genetic algorithms (Sánchez et al., 1996) andorthogonal least squares learning algorithm (Walczak andMassart, 1996). After RBF connected to output layer neurons,the response of each output neuron is computed by a linearlearning function ‘learngdm’ to generate a new output. Therelationship between predicted value and the input variables isshown as follows:

f n Xð Þ ¼ w0 þXni¼1

wiØ k X � ci kð Þ (7)

where n is the number of the input variables, w0 is the bias termand wi is the weight ascribed to the ith input.

After the models being developed, some indicators wereadopted to evaluate the performance of the models. For thecalibration model, the indicators were the coefficients of

determination of calibration R2c

� �and the root mean square error

of calibration (RMSEC). For the prediction model, the indicators

were the coefficients of determination of prediction R2c

� �and the

root mean square error of prediction (RMSEP). Generally, the largerthe prediction set of R2 and smaller of RMSEs, the more desirable itis (Sakai, 2001).

2.9. Visualization of anthocyanin distribution

To monitor the changes of anthocyanin content of pericarpduring postharvest browning process of lychee, the distribution ofanthocyanin content was visualized based on the developedprediction model. The visualization of different physicochemicalattributes of fruits in a pixel-wise manner is one of the mainadvantages of HSI over traditional spectroscopy. Chemicalmapping is an available methodology for visualization, whichcan be useful for predicting anthocyanin distribution in the tested

[(Fig._4)TD$FIG]

Fig. 4. Comparison diagram of uncorrected and corrected lychee images at 874nm. (a) Uncorrected lychee image, (b) corrected lychee image and (c) Y-profiles of amonochromatic lychee image at x =356 pixels. (For interpretation of the references to color in the text, the reader is referred to the web version of this article.)

Y.-C. Yang et al. / Postharvest Biology and Technology 103 (2015) 55–65 61

lychee pericarp. Predicted values of anthocyanin content wasobtained according to the best model and used to generate adistribution map of anthocyanin content based on the spatialposition of each pixel. The magnitude of the anthocyanin contentwas shown in a color scale determined by the magnitude of thepredicted anthocyanin content in the lychee pericarp. Therefore, itis easy to observe the changes of anthocyanin content of pericarpstored for different periods byexamining the color changes evidentin the visualization maps. All steps generating the distributionmaps of anthocyanin content were carried out by MATLAB 2010a.

3. Results and discussion

3.1. Anthocyanin property of lychee pericarp

The change range and regularity of anthocyanin content inlychee pericarp stored for different days are of special significancefor prediction accuracy and robust of the calibration model. Theycan even indicate the performance of calibration model to someextent when anthocyanin content changes abnormally. Therelevant statistical data of the anthocyanin content of pericarpstored for 0–5 days are shown in Table 1. The maximum, minimumand mean declined consistently from the 0th day to the fifth day.These data were consistent with the fact that the pericarp ofpostharvest lychee gradually browned with the extension ofstorage time and its color started changing from red and green tobrown. Meanwhile, the changes shown in Table 1 also followed thedescription reported by Zhang et al. (2001). However, the changesin the standard deviation (SD) did not follow this rule or evenshowed some irregularity. This observation might be due to thediversity of lychee samples and the few sample number.

3.2. Lychee image correction

Due to spatial variation of light intensity on the surface of thelychee, the acquired hyperspectral images had to be corrected bygeometric correction factors before extracting average spectra ofROIs. An illustrative example of uncorrected and corrected lycheeimages at 874nm is shown in Fig. 4. In the uncorrected image (a),the central area is sharp and bright while the outer areas are blurryand dim. In the corrected image (b), the entire sample is uniformlysharp and bright. Fig. 4(c) shows the corrected and uncorrectedreflectance across the widest part of the fruit, which is denoted bythe vertical red line in (a) and (b). The corrected reflectance shows

essentially equal magnitudes across the entire width of the fruit,which is in stark contrast to the uncorrected reflectance, in whichthe edge values are much less than the central values. Given thesuccess of this correction methodology, before extracting imageand spectral information, all of the lychee HSI images were treatedin this manner prior to further processing.

3.3. Spectral features of pericarp

The main chemical components of lychee are pigments, water,ascorbic acid, sugar, other compounds, all of which contribute tospectral responses in the Vis-NIR region. During lychee browning,some of these compounds change, resulting in the variance ofreflectance curves of the lychee pericarp. It has been reported thatthe components that change themost during the browning processare pigments, moisture content, color, and pH (Holcroft andMitcham, 1996). Anthocyanins contribute significantly to the redcolor of the lychee pericarp, and its degradation is one of the mostimportant causes of lychee browning (Rivera-López et al., 1999).Therefore, this study focused on the changes of anthocyanincontent in the pericarp during the storage period. According to Qinand Lu (2008), anthocyanin pigments have their maximumabsorbance around 535nm. Fig. 5 shows the average spectracurves of lychees stored for 0–5 days. There were no valley or peakaround 535nm but the reflectance of lychee at 535nm slightlyincreasedwith the extension of storage time and the reflectance oflychee stored for three days increased the most. This wasattributed to the degradation of anthocyanin caused by thebrowning of lychee. Anthocyanin content in lychee pericarp is notso much and thus the reflectance does not change a lot. A minorreflectance valley can be seen in the region around 680nm that isassociated with the presence of chlorophyll in lychees (Qin and Lu,2008). The spectral reflectance around 680nm did not change inthe first three days but suddenly increased at the fourth day andkept the trend in the following two days. This might be due to thatlychee did not start browning in the first three days. From thefourth day on, lychee pericarp rapidly browned for losing waterand chlorophyll of pericarp began degrading. According to Nicolaïet al. (2007), the absorption around 840–960nm was related tohydroxyl group of moisture and sugar. Finally, the absorptionobserved around 960nm was attributed to moisture as describedby ElMasry et al. (2007). In short, reflectance was seen to increasewith the decrease of moisture content because the samplesbecame drier during storage.

3.4. Selection of optimal wavelengths and the calibration model

In order to implement application of HSI technique in aproduction line setting, several optimal wavelengths are com-monly selected to reduce computational burden and improve theprediction capability of the calibration model (Liu et al., 2014). Inthis study, SPA and SWR algorithms were adopted to removevariables with little useful information for modeling and to selectthe optimal wavelengths for predicting anthocyanin content in thelychee pericarp. SPA aims to find the wavelengths with the least

[(Fig._5)TD$FIG]

Fig. 5. Spectral characteristics of lychee stored for 0–5 days.

Table 2Performance of different models for predicting anthocyanin content of lycheepericarp.

Models Performances of models

R2c

RMSEC (%) R2p

RMSEP (%)

Full wavelengths models 0.928 0.409 0.916 0.511SPA-RBF-SVR 0.746 0.728 0.672 0.925SWR-RBF-SVR 0.753 0.701 0.712 0.841Fused model 0.891 0.567 0.872 0.610

62 Y.-C. Yang et al. / Postharvest Biology and Technology 103 (2015) 55–65

redundant information. It began with one wavelength randomly,and merges another wavelength at each iteration until a specificnumber of optimal wavelengths are acquired. The maximumnumber of optimal wavelengths was set to 15 and the minimumwas 4. A total of 180 samples were selected from the training set(240 samples) using Kennard-Stone method for finding theoptimal wavelengths, and the remaining samples were used tovalidate the performance of these wavelengths. The SPA algorithmselected nine optimal wavelengths (368, 458, 614, 678, 903, 988,1004, 1020 and 1042nm). SWR was also applied for wavelengthselection. The purpose of SWR was to find the wavelengths withthe largest significance for the performance of prediction model.Therefore, it selected the optimal wavelengths by correlationcoefficient between wavelengths and quality parameter. Thereflectance values over 1 was considered as singular values andremoved from the data list. Before SWR being conducted,collinearity diagnostics were performed first for discriminatingcollinearity among different wavelengths and determining wheth-er SWR algorithm could be conducted. Only eleven wavelengths(423, 453, 587, 630, 676, 707, 762, 805, 867, 911 and 977nm) weredetermined as the optimal wavelengths by SWR algorithm.

Inall sixgroupsofsamples,40 fruits ineachgroupwererandomlyselected as training set and the remaining 20 lychees were asprediction set. Therefore, therewere 240 fruits used for training and120 lychee samples were used for predicting. Firstly, calibrationmodel on full wavelengths was built based on the training set byestablishing correlation relationship between reflectance values oflychee and anthocyanin content using RBF-SVR algorithm. Thenpredictionvalues of anthocyanin contentwere obtained byapplyingthe calibration model to reflectance values of lychee samples ofprediction set. The performance of the calibration model was thenvalidatedbycalculating thevariancesbetweenpredictionvaluesandmeasurement values of anthocyanin content. In the same way, thesimplified models were developed using RBF-SVR algorithm basedon the twosetsofoptimalwavelengthsand theirperformanceswerecompared. Table 2 illustrates the detailed performance of each SVRmodel. The results of SPA-RBF-SVR and SWR-RBF-SVR modelsdemonstrated that SWRwas slightly superior to SPA in selecting theoptimal wavelengths. However, the performances of both SPA-RBF-SVR and SWR-RBF-SVRmodelsweremuchworse than that ofmodelbased on full wavelengths. This was likely due to that both SPA andSWR missed some useful information while massive redundantinformation was removed. As the optimal wavelengths selected by

SPA and SWRwere different, it could be supposed that the fusion ofSPA-RBF-SVR and SWR-RBF-SVRmodels might overcome the aboveshortcoming and improve the performance of SVR models.

Therefore, a fused model was generated from the predictionvalues of the SPA-RBF-SVRandSWR-RBF-SVRmodels using theRBF-NN algorithm. The effect was to integrate the two sets of optimalwavelengths into a single model. To accomplish this task, thepredicted values of anthocyanin content based on SPA-RBF-SVR andSWR-RBF-SVRmodelswere imported into the input layerof theRBF-NN as independent variables, and measured values of anthocyanincontentweretreatedasdependentvariables.AsshowninTable2, theprediction accuracy of the fused model was significantly improved.Compared with SPA-RBF-SVR or SWR-RBF-SVR models, the fusionmethod possessed obvious advantages, which also proved thesupposal that both SPA and SWR missed some important informa-tionwhen redundant informationwas eliminated. The results showthat (a) HSI technique has a great potential for estimatinganthocyanin content of lychee pericarp and (b) model fusion, as auseful mathematical approach integrating different algorithms, canovercomemany shortcomings of individual algorithms and improveprediction accuracy of the models.

[(Fig._6)TD$FIG]

Fig. 6. Distributionmaps of anthocyanin content in lychee pericarp stored for different days. (For interpretation of the references to color in the text, the reader is referred tothe web version of this article.)

[(Fig._7)TD$FIG]

Fig. 7. Comparison betweenmeasured anthocyanin content andmean anthocyanincontent of each predicted lychee fruit.

Y.-C. Yang et al. / Postharvest Biology and Technology 103 (2015) 55–65 63

3.5. Visualization of anthocyanin distribution

As noted previously, it is important to develop a robust modelwith high prediction accuracy and precision for predictinganthocyanin content. Equally important is the creation of animage or distribution map for use in real-time monitoring of thechanges in anthocyanin content during storage, because the nakedeyes or traditional imaging techniques are incapable of determin-ing the differences in anthocyanin content from location tolocation in the same sample. At the same time, it is difficult, if notimpossible, to identify the changes that are truly important inmaking an accurate prediction. In this study, the fused model wasapplied to each pixel of a lychee image. Specifically, predictedvalues of anthocyanin were first obtained by multiplying thespectrum of all spots in a lychee by regression coefficients of thefused model. These predicted values were ranked from low to highand represented by different colors, from blue and brown (low) toorange and red (high). Pixels with same reflectance had the similarpredicted values and were represented by the same color. Fig. 6shows distributionmaps of anthocyanin content in lychee pericarpstored for 0–5 days. From Fig. 6, the changes of anthocyanincontent during storage can be easily discerned and the differencesof anthocyanin content within the same fruit are also readilydistinguished. The sample stored for 0 day contained abundantamounts of anthocyanin and appeared as many spots of bright redin the map. As the storage time increased, the anthocyanin contentdeclined, and the color in the distribution map became more blueand brown. In addition, as evidenced by the mottled seen in Fig. 6,the distribution of anthocyanin content was non-uniformwithin alychee sample. Thiswaswhy some areas of fresh lychee present redwhile other areas were green. Another phenomenon worthy ofattention was that a lot of spots could not be distinguished frombackground and presented blue in the visualization map of lycheefruit, making the map be mottling. This was mainly due to tworeasons. One was that these sunk regions acquired less light thanthose bulges when they were irradiated. The other was thatanthocyanin was mostly contained in bulges, while sunk regionscontained little. This was why bulges turned from green to redwhen lychee was ripe, while sunk regions were green all the timeuntil lychee started browning. In order to validate the accuracy ofthe fused model for anthocyanin distribution prediction, the meananthocyanin content for all pixels of each predicted fruit wascalculated and compared with the actual value measured by anultraviolet–visible spectrophotometer. The result from suchcomparison is illustrated in Fig. 7. It could be observed that thecoefficients of determination between the two reached 0.870,which demonstrated that the poor predictions of sunk regions oflychee fruit did not much affect the accuracy of the models. Theproposed method was thus effective in evaluating anthocyanincontent of lychee pericarp during storage.

The above results indicated that the visualization ofanthocyanin content could improve on-line monitoring of fruitquality in ways that are not possible by traditional imaging orspectroscopic technique alone. HSI is capable of predicting andvisualizing anthocyanin evolution in pericarp stored for differ-ent days, which facilitates the development of quality monitor-ing techniques. It provides a new targeted and quantitativeapproach for quality estimation of fruits, which is beneficial tothe fruit industry.

4. Conclusions

A model fusion method was proposed in this study and used todevelop a calibration model that integrated different mathematicalalgorithms for predicting anthocyanin content in lychee pericarpstoredfor0–5days.Theresults indicatedthatanthocyanincontent in

pericarp during storage could be measured effectively by HSItechnique. Good performances were not only found in the fusedmodel but also in the models based on full wavelengths. The modelusing all of the wavelengths was the best, with high coefficients ofdeterminationof0.928and0.916,aswellas lowRMSEsof0.409%and0.511% for the training and testing sets, respectively. However, themodel could not be successfully applied to on-line detection of fruitquality in a production line setting because of the heavycomputational burden and the associated low efficiency of analysis.Although SPA-RBR-SVR and SWR-RBF-SVR models possessed goodpredictive ability, themodel fusion approach that integrated the twomodels into one single model appeared to overcome each model’sweaknessesandachievedanimprovedperformance.Byapplying thefusedmodel toeachspotof the lychee image,distributionmapswerecreated for visualizing the changes of anthocyanin content inpericarp with increased storage time. Although the proposedmethod was effective in evaluating anthocyanin content of lycheepericarp during storage, it should be improved in many aspectsbefore applying to other quality attributes of lychee fruit. In additionmore efforts were needed for practical utilization of the proposedmethod in an industrial setting.

Acknowledgments

The authors gratefully acknowledge the Guangdong ProvinceGovernment (China) for its support through the program “LeadingTalent of Guangdong Province (Da-Wen Sun)”. This research wasalso supported by the National Key Technologies R&D Program(2014BAD08B09, 2015BAD19B03), the International S&T Coopera-tion Programme of China (2015DFA71150), and the InternationalS&T Cooperation Projects of Guangdong Province(2013B051000010).

References

Araújo, M.C.U., Saldanha, T.C.B., Galvão, R.K.H., Yoneyama, T., Chame, H.C., Visani, V.,2001. The successive projections algorithm for variable selection in spectroscopicmulticomponent analysis. Chemom. Intell. Lab. Syst. 57 (2), 65–73.

Barbin, D.F., ElMasry, G., Sun, D.-W., Allen, P., 2012. Predicting quality and sensoryattributes of pork using near-infrared hyperspectral imaging. Anal. Chim. Acta719, 30–42.

Basak, D., Pal, S., Patranabis, D.C., 2007. Support vector regression. Neural Inf.Process.-Lett. Rev. 11 (10), 203–224.

Buhmann, M.D., 2003. Radial Basis Functions: Theory and Implementations.Cambridge University Press, Cambridge, England, pp. 121–136 ISBN:521633389.

Camps-Valls, G., Gómez-Chova, L., Calpe-Maravilla, J., Martín-Guerrero, J.D., Soria-Olivas, E., Alonso-Chordá, L., Moreno, J., 2004. Robust support vectormethod forhyperspectral data classification and knowledge discovery. Geosci. RemoteSens. IEEE Trans. 42 (7), 1530–1542.

Candolfi, A., DeMaesschalck, R., Jouan-Rimbaud, D., Hailey, P., Massart, D., 1999. Theinfluence of data pre-processing in the pattern recognition of excipients near-infrared spectra. J. Pharm. Biomed. Anal. 21 (1), 115–132.

Chen, Q., Zhang, J., Shen, J., 1997. Recent trend for postharvest storage of tropical andsubtropical fruits in China. Kasetsart J. 32, 67–71.

Cubero, S., Aleixos, N., Moltó, E., Gómez-Sanchis, J., Blasco, J., 2011. Advances inmachine vision applications for automatic inspection and quality evaluation offruits and vegetables. Food Bioprocess Technol. 4 (4), 487–504.

Dai, Q., Cheng, J.-H., Sun, D.-W., Zeng, X.-A., 2014. Advances in feature selectionmethods for hyperspectral image processing in food industry applications: areview. Crit. Rev. Food Sci. Nutr. doi:http://dx.doi.org/10.1080/10408398.2013.871692 (in press).

Delgado, A.E., Sun, D.-W., 2002a. Desorption isotherms for cooked and cured beefand pork. J. Food Eng. 51 (2), 163–170.

Delgado, A.E., Sun, D.-W., 2002b. Desorption isotherms and glass transitiontemperature for chicken meat. J. Food Eng. 55 (1), 1–8 PII S0206-8774(01)00222-9.

Dong, W.-J., Ni, Y.-N., Kokot, S., 2012. Quantitative analysis of two adulterants inCynanchum stauntonii by near-infrared spectroscopy combined with multi-variate calibrations. Chem. Pap. 66 (12), 1083–1091.

ElMasry, G., Wang, N., ElSayed, A., Ngadi, M., 2007. Hyperspectral imaging fornondestructive determination of some quality attributes for strawberry. J. FoodEng. 81 (1), 98–107.

64 Y.-C. Yang et al. / Postharvest Biology and Technology 103 (2015) 55–65

ElMasry, G., Sun, D.-W., Allen, P., 2011a. Non-destructive determination of water-holding capacity in fresh beef by using NIR hyperspectral imaging. Food Res. Int.44 (9), 2624–2633.

ElMasry, G., Iqbal, A., Sun, D.-W., Allen, P., 2011b. Quality classification of cooked,sliced turkey hams using NIR hyperspectral imaging system. J. Food Eng.103 (3),333–344.

ElMasry, G., Sun, D.-W., Allen, P., 2012. Near-infrared hyperspectral imaging forpredicting colour, pH and tenderness of fresh beef. J. Food Eng. 110 (1), 127–140.

Fernandes, A.M., Oliveira, P., Moura, J.P., Oliveira, A.A., Falco, V., Correia, M.J., Melo-Pinto, P., 2011. Determination of anthocyanin concentration in whole grapeskins using hyperspectral imaging and adaptive boosting neural networks. J.Food Eng. 105 (2), 216–226.

Galvão, R.K.H., Araújo,M.C.U., Fragoso,W.D., Silva, E.C., José, G.E., Soares, S.F.C., Paiva,H.M., 2008. A variable elimination method to improve the parsimony of MLRmodels using the successive projections algorithm. Chemom. Intell. Lab. Syst. 92(1), 83–91.

Goldberg, D.E., Holland, J.H., 1988. Genetic algorithms and machine learning. Mach.Learn. 3 (2), 95–99.

Gómez-Sanchis, J., Moltó, E., Camps-Valls, G., Gómez-Chova, L., Aleixos, N., Blasco, J.,2008. Automatic correction of the effects of the light source on sphericalobjects. An application to the analysis of hyperspectral images of citrus fruits. J.Food Eng. 85 (2), 191–200.

Holcroft, D.M., Mitcham, E.J., 1996. Postharvest physiology and handling of litchi(Litchi chinensis Sonn.). Postharvest Biol. Technol. 9 (3), 265–281.

Huang, C.C., Wu, X.D., Tong, W.Q., 2007. Infrared image simulation based onstatistical learning theory. Int. J. Infrared Millimeter Waves 28 (12), 1143–1153.

Huang, L.,Wu, D., Jin, H., Zhang, J., He, Y., Lou, C., 2011. Internal quality determinationof fruit with bumpy surface using visible and near infrared spectroscopy andchemometrics: a case study with mulberry fruit. Biosyst. Eng. 109 (4), 377–384.

Huang, S., Hart, H., Lee, H., Wicker, L., 1990. Enzymatic and color changes duringpost-harvest storage of lychee fruit. J. Food Sci. 55 (6), 1762–1763.

Isaksson, T., Næs, T., 1988. The effect of multiplicative scatter correction (MSC) andlinearity improvement in NIR spectroscopy. Appl. Spectrosc. 42 (7), 1273–1284.

Jackman, P., Sun, D.-W., Du, C.-J., Allen, P., 2008. Prediction of beef eating qualityfrom colour, marbling andwavelet texture features.Meat Sci. 80 (4),1273–1281.

Jiang, Y., Wang, Y., Song, L., Liu, H., Lichter, A., Kerdchoechuen, O., Joyce, D., Shi, J.,2006. Postharvest characteristics and handling of litchi fruit-an overview. Anim.Prod. Sci. 46 (12), 1541–1556.

Joas, J., Caro, Y., Ducamp, M.N., Reynes, M., 2005. Postharvest control of pericarpbrowning of litchi fruit (Litchi chinensis Sonn cv Kwaï Mi) by treatment withchitosan and organic acids: I. Effect of pH and pericarp dehydration. PostharvestBiol. Technol. 38 (2), 128–136.

Karimi, Y., Maftoonazad, N., Ramaswamy, H.S., Prasher, S.O., Marcotte, M., 2012.Application of hyperspectral technique for color classification avocadossubjected to different treatments. Food Bioprocess Technol. 5 (1), 252–264.

Kamruzzaman, M., ElMasry, G., Sun, D.-W., Allen, P., 2011. Application of NIRhyperspectral imaging for discrimination of lamb muscles. J. Food Eng. 104 (3),332–340.

Kamruzzaman,M., ElMasry, G., Sun, D.-W., Allen, P., 2012. Prediction of some qualityattributes of lamb meat using near-infrared hyperspectral imaging andmultivariate analysis. Anal. Chim. Acta 714, 57–67.

Kiani, H., Sun, D.-W., 2011. Water crystallization and its importance to freezing offoods: a review. Trends Food Sci. Technol. 22 (8), 407–426.

Lee, H., Wicker, L., 1991. Anthocyanin pigments in the skin of lychee fruit. J. Food Sci.56 (2), 466–468.

Liu, H., Song, L., You, Y., Li, Y., Duan, X., Jiang, Y., Joyce, D.C., Ashraf, M., Lu, W., 2011.Cold storage duration affects litchi fruit quality, membrane permeability,enzyme activities and energy charge during shelf time at ambient temperature.Postharvest Biol. Technol. 60 (1), 24–30.

Liu, D., Sun, D.-W., Zeng, X.-A., 2014. Recent advances in wavelength selectiontechniques for hyperspectral image processing in the food industry. FoodBioprocess Technol. 7 (2), 307–323.

Liu, K., Chen, X., Li, L., Chen, H., Ruan, X., Liu, W., 2015. A consensus successiveprojections algorithm–multiple linear regression method for analyzing nearinfrared spectra. Anal. Chim. Acta doi:http://dx.doi.org/10.1016/j.aca.2014.12.033 (in press).

Lorente, D., Aleixos, N., Gomez-Sanchis, J., Cubero, S., Blasco, J., 2013. Selection ofoptimal wavelength features for decay detection in citrus fruit using the ROCcurve and neural networks. Food Bioprocess Technol. 6 (2), 530–541.

McDonald, K., Sun, D.-W., 2001. The formation of pores and their effects in a cookedbeef product on the efficiency of vacuum cooling. J. Food Eng. 47 (3), 175–183.

McDonald, K., Sun, D.-W., Kenny, T., 2001. The effect of injection level on the qualityof a rapid vacuum cooled cooked beef product. J. Food Eng. 47 (2), 139–147.

Menesatti, P., Zanella, A., D’Andrea, S., Costa, C., Paglia, G., Pallottino, F., 2009.Supervised multivariate analysis of hyper-spectral NIR images to evaluate thestarch index of apples. Food Bioprocess Technol. 2 (3), 308–314.

Mundry, R., Nunn, C.L., 2009. Stepwise model fitting and statistical inference:turning noise into signal pollution. Am. Nat. 173 (1), 119–123.

Nanyam, Y., Choudhary, R., Gupta, L., Paliwal, J., 2012. A decision-fusion strategy forfruitquality inspectionusinghyperspectral imaging.Biosyst.Eng.111 (1),118–125.

Nicolaï, B.M., Beullens, K., Bobelyn, E., Peirs, A., Saeys,W., Theron, K.I., Lammertyn, J.,2007. Nondestructive measurement of fruit and vegetable quality by means ofNIR spectroscopy: a review. Postharvest Biol. Technol. 46 (2), 99–118.

Patrick, J., Sun, D.-W., Du, C.-J., Allen, P., 2009. Prediction of beef eating qualitiesfrom colour, marbling and wavelet surface texture features using homogenouscarcass treatment. Pattern Recognit. 42 (5), 751–763.

Pulido, A., Ruisanchez, I., Rius, F., 1999. Radial basis functions applied to theclassification of UV–visible spectra. Anal. Chim. Acta 388 (3), 273–281.

Qin, J., Lu, R., 2008. Measurement of the optical properties of fruits and vegetablesusing spatially resolved hyperspectral diffuse reflectance imaging technique.Postharvest Biol. Technol. 49 (3), 355–365.

Rivera-López, J., Ordorica-Falomir, C., Wesche-Ebeling, P., 1999. Changes inanthocyanin concentration in Lychee (Litchi chinensis Sonn.) pericarp duringmaturation. Food Chem. 65 (2), 195–200.

Sánchez, M., Swierenga, H., Sarabia, L., Derks, E., Buydens, L., 1996. Performance ofmulti layer feedforward and radial base function neural networks inclassification and modelling. Chemom. Intell. Lab. Syst. 33 (2), 101–119.

Sakai, K., 2001. Nonlinear Dynamics and Chaos in Agricultural Systems. ElsevierScience Ltd. Amsterdam, Holland, pp. 107 ISBN: 444506462.

Sandoval, G., Vazquez, R.A., Garcia, P., Ambrosio, J., 2014. Crop classification usingdifferent color spaces and rbf neural networks. Artif. Intell. Soft Comput. 8467,598–609.

Shao, Y., Bao, Y., He, Y., 2011. Visible/near-infrared spectra for linear and nonlinearcalibrations: a case to predict soluble solids contents and pH value in peach.Food Bioprocess Technol. 4 (8), 1376–1383.

Sun, D.-W., 1997a. Thermodynamic design data and optimum design maps forabsorption refrigeration systems. Appl. Therm. Eng. 17 (3), 211–221.

Sun, D.-W.,1997b. Solar powered combined ejector vapour compression cycle for airconditioning and refrigeration. Energy Convers. Manage. 38 (5), 479–491.

Sun, D.-W., 2004. Computer vision – an objective, rapid and non-contact qualityevaluation tool for the food industry. J. Food Eng. 61 (1), 1–2.

Sun, D.-W., Byrne, C.,1998. Selection of EMC/ERH isothermequations for rapeseed. J.Agric. Eng. Res. 69 (4), 307–315.

Sun, D.-W., Woods, J.L., 1997. Simulation of the heat and moisture transfer processduring drying in deep grain beds. Drying Technol. 15 (10), 2479–2508.

Sun, D.-W., Brosnan, T., 2003. Pizza quality evaluation using computer vision – part 1– pizza base and sauce spread. J. Food Eng. 57 (1), 81–89 PII S0260-8774(02)00275-3.

Sun, D.-W., Eames, I.W., Aphornratana, S., 1996. Evaluation of a novel combinedejector-absorption refrigeration cycle. 1. Computer simulation. Int. J. Refrig.-Revue Internationale Du Froid 19 (3), 172–180.

Underhill, S.J., Simons, D.H., 1993. Lychee (Litchi chinensis Sonn.) pericarpdesiccation and the importance of postharvest micro-cracking. Sci. Hortic. 54(4), 287–294.

Valous, N.A., Mendoza, F., Sun, D.-W., Allen, P., 2009. Colour calibration of alaboratory computer vision system for quality evaluation of pre-sliced hams.Meat Sci. 81 (1), 132–141.

Walczak, B., Massart, D., 1996. The radial basis functions – partial least squaresapproach as a flexible non-linear regression technique. Anal. Chim. Acta 331 (3),177–185.

Wang, H.H., Sun, D.-W., 2002. Melting characteristics of cheese: analysis of effect ofcheese dimensions using computer vision techniques. J. Food Eng. 52 (3), 279–284 PII S0260-8774(01)00116-9.

Wang, S., Huang, M., Zhu, Q.B., 2012. Model fusion for prediction of apple firmnessusing hyperspectral scattering image. Comput. Electron. Agric. 80, 1–7.

Wu, D., Sun, D.-W., 2013a. Advanced applications of hyperspectral imagingtechnology for food quality and safety analysis and assessment: a review- Part I:fundamentals. Innovative Food Sci. Emerging Technol. 19, 1–14.

Wu, D., Sun, D.-W., 2013b. Advanced applications of hyperspectral imagingtechnology for food quality and safety analysis and assessment: a review- PartII: applications. Innovative Food Sci. Emerging Technol. 19, 15–28.

Wu, D., He, Y., 2014. Potential of spectroscopic techniques and chemometric analysisfor rapid measurement of docosahexaenoic acid and eicosapentaenoic acid inalgal oil. Food Chem. 158, 93–100.

Wu, D., Chen, J., Lu, B., Xiong, L., He, Y., Zhang, Y., 2012a. Application of near infraredspectroscopy for the rapid determination of antioxidant activity of bamboo leafextract. Food Chem. 135 (4), 2147–2156.

Wu, D., Shi, H., Wang, S., He, Y., Bao, Y., Liu, K., 2012b. Rapid prediction of moisturecontent of dehydrated prawns using online hyperspectral imaging system. Anal.Chim. Acta 726, 57–66.

Wu, D., Sun, D.W., He, Y., 2012c. Application of long-wave near infraredhyperspectral imaging for measurement of color distribution in salmon fillet.Innov. Food Sci. Emerg. Technol. 16, 361–372.

Wu, D., Shi, H., He, Y., Yu, X., Bao, Y., 2013. Potential of hyperspectral imaging andmultivariate analysis for rapid and non-invasive detection of gelatinadulteration in prawn. J. Food Eng. 119 (3), 680–686.

Xu, L., Zhang, W.-J., 2001. Comparison of different methods for variable selection.Anal. Chim. Acta 446 (1), 475–481.

Xu, S.Y., Chen, X.F., Sun, D.-W., 2001. Preservation of kiwifruit coated with an ediblefilm at ambient temperature. J. Food Eng. 50 (4), 211–216.

Zhang, D., Quantick, P.C., 1997. Effects of chitosan coating on enzymatic browningand decay during postharvest storage of litchi (Litchi chinensis Sonn.) fruit.Postharvest Biol. Technol. 12 (2), 195–202.

Zhang, Z., Pang, X., Ji, Z., Jiang, Y., 2001. Role of anthocyanin degradation in litchipericarp browning. Food Chem. 75 (2), 217–221.

Zhang, Z.Q., Pang, X.Q., Yang, C., Ji, Z.L., Jiang, Y.M., 2003. Purification and structuralanalysis of anthocyanins from litchi pericarp. Food Chem. 84 (4), 601–604.

Zhao, J., Chen, Q., Huang, X., Fang, C., 2006. Qualitative identification of teacategories by near infrared spectroscopy and support vector machine. J. Pharm.Biomed. Anal. 41 (4), 1198–1204.

Y.-C. Yang et al. / Postharvest Biology and Technology 103 (2015) 55–65 65