Full Text

REVIEW PAPER

Recent Advances and Applications of Hyperspectral Imagingfor Fruit and Vegetable Quality Assessment

D. Lorente & N. Aleixos & J. Gómez-Sanchis & S. Cubero &

O. L. García-Navarrete & J. Blasco

Received: 23 June 2011 /Accepted: 3 November 2011 /Published online: 22 November 2011# Springer Science+Business Media, LLC 2011

Abstract Hyperspectral imaging systems are starting to beused as a scientific tool for food quality assessment. Atypical hyperspectral image is composed of a set of arelatively wide range of monochromatic imagescorresponding to continuous wavelengths that normallycontain redundant information or may exhibit a high degreeof correlation. In addition, computation of the classifiersused to deal with the data obtained from the images canbecome excessively complex and time-consuming for suchhigh-dimensional datasets, and this makes it difficult toincorporate such systems into an industry that demandsstandard protocols or high-speed processes. Therefore,recent works have focused on the development of new

systems based on this technology that are capable ofanalysing quality features that cannot be inspected usingvisible imaging. Many of those studies have also centred onfinding new statistical techniques to reduce the hyper-spectral images to multispectral ones, which are easier toimplement in automatic, non-destructive systems. Thisarticle reviews recent works that use hyperspectral imagingfor the inspection of fruit and vegetables. It explains thedifferent technologies available to acquire the images andtheir use for the non-destructive inspection of the internaland external features of these products. Particular attentionis paid to the works aimed at reducing the dimensionality ofthe images, with details of the statistical techniques mostcommonly used for this task.

Keywords Computer vision . Fruits . Vegetables . Quality .

Non-destructive inspection . Image analysis . Hyperspectralimaging .Multispectral imaging

NomenclatureANN Artificial neural networksANOVA Analysis of varianceAOTF Acousto-optic tunable filtersBMP Bitmap image formatBSQ Band sequentialCCD Charge-coupled deviceFLD Fisher’s linear discriminantFWHM Full width at half-maximumGALDA Genetic algorithm based on LDALCTF Liquid crystal tunable filtersLD Lorentzian distributionLDA Linear discriminant analysisMC Moisture contentMD Mahalanobis distanceNIR Near infrared

D. Lorente : S. Cubero :O. L. García-Navarrete : J. Blasco (*)Centro de Agroingeniería,Instituto Valenciano de Investigaciones Agrarias,Cra. Moncada-Náquera Km. 5,46113 Moncada, Spaine-mail: [email protected]

N. AleixosInstituto Interuniversitario de Investigación en Bioingeniería yTecnología Orientada al Ser Humano,Universitat Politècnica de València,Camino de Vera s/n,46022 Valencia, Spain

J. Gómez-SanchisIntelligent Data Analysis Laboratory, IDAL, ElectronicEngineering Department, Universitat de València,Dr. Moliner 50,46100 Burjassot (Valencia), Spain

O. L. García-NavarreteDepartamento de Ingeniería Civil y Agrícola,Universidad Nacional de Colombia-Sede Bogotá,Carrera 30 No. 45-03, Edificio 214, Oficina 206,Bogotá, Colombia

Food Bioprocess Technol (2012) 5:1121–1142DOI 10.1007/s11947-011-0725-1

PCA Principal component analysisPLS Partial least squarePLSDA PLS discriminant analysisPLSR PLS regressionRF RadiofrequencyRGB Red, green, blue colour spaceRGBI Red, green, blue, infraredSAM Spectral angle mapperSID Spectral information divergenceSSC Soluble solids contentTA Titratable acidTIFF Tagged image file formatUV Ultraviolet

Introduction

The application of machine vision to food analyses hasincreased considerably in recent years (Sun 2007) and hasbeen used with meat (Du and Sun 2009), fish (Quevedo etal. 2010; Quevedo and Aguilera 2010), grains (Manick-avasagan et al. 2010), bread (Farrera-Rebollo et al. 2011),fruits and vegetables (Cubero et al. 2011), among others.The breadth of applications depends, among many otherthings, on the fact that machine vision systems providesubstantial information about the nature and attributes ofthe objects present in a scene. Another important feature ofsuch systems is that they open up the possibility of studyingthese objects in regions of the electromagnetic spectrumwhere the human eye is unable to operate, such as in theultraviolet (UV), near-infrared (NIR) or infrared (IR)regions.

The high risk of human error in classification processeshas been underlined and is one of the most importantdrawbacks that machine vision can help prevent. In a studycarried out with different varieties of apples, where variousshape, size and colour parameters were compared, one ofthe conclusions reached was the limited human capacity toreproduce the estimation of quality, which is defined asinconsistency (Paulus et al. 1997). Moreover, as the numberof parameters considered in the decision-making processincreases, so does the rate of error in classification.Furthermore, it should also be mentioned that automaticinspection allows precise statistics to be generated onaspects related to the quality of the inspected product,which leads to greater control over it and facilitates itstraceability.

In this respect, the quality of a particular fresh orprocessed fruit or vegetable is defined by a series ofexternal characteristics that make it more or less attractiveto the consumer. Such attributes include its ripeness, size,weight, shape, colour, condition, or presence/absence of

defects, stems or seeds, as well as a series of internalproperties like sweetness, acidity, texture, hardness, amongothers. They are primary criterions in making purchasingdecisions of fresh fruits and vegetables (Kays 1999). Insum, they cover all of the factors that exert an influence onthe product’s appearance, on its nutritional and organolepticqualities, or on its suitability for preservation. Most of thesefactors have traditionally been assessed by visual inspectionor destructive sampling performed by trained operators, butcurrently, many of them, particularly the external ones, canbe estimated with commercial vision systems (Cubero et al.2011; Costa et al. 2011). These vision systems for fruitsorting are normally based on colour video cameras thatimitate the vision of the human eye by capturing imagesusing three filters centred on red, green and blue (RGB)wavelengths. Therefore, they are limited to observingscenes and are usually incapable of obtaining muchinformation about the external or internal composition ofthe products or to detect some defects or alterations whosecolour is similar to the colour of the sound skin.

An approach that enhances the possibilities if suchsystems are multispectral computer vision systems whichbasically can acquire a set of optimised monochromaticimages at few wavelengths that make it possible to estimateor discover features that are difficult with the traditionalsystems. These wavelengths normally belong to the visibleor near-infrared parts of the electromagnetic spectrum andin most cases are selected using hyperspectral imagingsystems. Beyond multispectral imaging, the use of hyper-spectral sensors makes it possible to conduct a moresophisticated analysis of the scene by acquiring a large setof monochromatic images corresponding to consecutivewavelengths. The reduction in the price of these systems,typically used for remote sensing and meteorology, allowsthem to be used now in laboratories for food quality, andthey are an emerging and promising tool for food qualityand safety control, as Gowen et al. (2007) and Elmasry etal. (2008a) stated in earlier reviews of the use of thistechnology in food inspection. The acquired multidimen-sional spectral signature (spectrum) characterising a pixelcan be used not only to analyse scenes like a standardcolour camera but also to obtain information about internalcompounds that can be related with the internal quality ofthe product. These systems work with a large number ofconsecutive monochromatic images of the same scene atdifferent wavelengths, thus enabling simultaneous analysisof the spatial and spectral information. The set ofmonochromatic images that are captured constitutes ahyperspectral image. As they are made up of a largecollection of images, hyperspectral images constitute a farmore extensive source of information than that provided bya single monochromatic image or a conventional RGBimage. The number of images depends on the spectral

1122 Food Bioprocess Technol (2012) 5:1121–1142

resolution of the system used, and they are combined byforming a 3D cube in which two dimensions are spatial(pixels) and the third is the spectrum of each pixel. Withoutadequate processing, such a large amount of data, despitebeing one of the main advantages of hyperspectral systems,can complicate the extraction of useful information sincemuch of the information obtained is redundant or, bynature, cannot be used to distinguish between regions withsimilar characteristics (Shaw 2003). The demanding indus-trial restrictions of working in real time often make itnecessary to reduce the dimensionality of the problem and toselect the greatest amount of non-redundant informationfrom the least number of wavelengths. Unsupervisedmethods such as principal component analysis or supervisedones such as linear discriminant analysis are commonlyemployed. In this respect, basic literature can be found inChang (2003) and Grahn and Geladi (2007).

This paper addresses the recent advances focused on theuse of this technology for the quality assessment of fruitsand vegetables. An explanation of multispectral systems,which expand the human eye capabilities and normally arethe final objective of studies using hyperspectral imaging, isdescribed in “Multispectral Imaging”. In the third section,different technologies for hyperspectral image acquisitionare explained and discussed. Hyperspectral images generatea large amount of information that can be processed usingdifferent statistical techniques. The techniques that are mostfrequently used in the scientific literature for processinghyperspectral images of fruits and vegetables are examinedin “Statistical Techniques”. Sometimes, the reduction of thedata is aimed at selecting some important wavelengths inorder to build multispectral systems that are easier toimplement using a more standard technology. Works withthis objective are revised in “Dimensionality Reduction andSelection of Spectral Features” followed by the applicationof hyperspectral technology to the inspection of the externaland internal quality of different species of fruits andvegetables. Finally, major problems of this technology arediscussed, and prospects of future trends and conclusionsare given.

Multispectral Imaging

One way to enhance the capability of traditional coloursystems that seek to imitate the human eye is the use ofmultispectral systems, but they have not to be confusedwith hyperspectral ones. A hyperspectral image is com-posed of a relatively wide range of continuous wavelengths,whereas a multispectral image consists of a few wave-lengths that do not necessarily have to be continuous. Themain advantages of multispectral imaging systems are therelatively low cost of the system in comparison with

hyperspectral systems and the fact that they can be morespecific for real applications. In fact, hyperspectral systemsare sometimes used just to select the particular set ofwavelengths that will finally be used. An RGB cameracould be considered a particular case of a multispectralsystem, although it is more common to include wavelengthsin frequencies outside the visible, like NIR. For instance,Aleixos et al. (2002) developed a multispectral camera forthe inspection of citrus fruits which was able to acquirevisible and near-infrared images (RGBI) from the samescene. The same authors also developed specific algorithmsfor inspecting the size, colour and presence of defects incitrus at a rate of between 5 and 10 fruits per second. Thecamera had two CCDs, one of which was a colour CCDthat provided RGB information and the other was mono-chromatic but coupled to an IR filter, centred on 750 nm,which provided IR information. For defect detection, aBayesian discriminant model was used to segment theimages at the pixel level, the independent variables beingthe grey levels of the RGBI bands and the classesbackground, defect and sound skin. The experiments werecarried out with oranges, mandarins and lemons. Compar-ing results with those obtained using human classificationshowed 94% coincidence in the worst case (when the fruitwas changing colour from green to orange). The systemwas also capable of correctly classifying lemons andmandarins and detected the external defects in 93% and94% of cases, respectively. One of the conclusions was thatthe B improved the detection of defects compared withusing only RGI, but its contribution was of little importance.Taghizadeh et al. (2011a) compared a conventional RGBimaging system based on a standard still camera with ahyperspectral imaging system (400–1,000 nm) to evaluatethe quality of mushrooms by estimating the hunter L value,which is the most commonly applied feature for mushroomquality grading. Different model performance indicatorsshowed the reasonably high potential of hyperspectralimaging models to predict the L value for mushroomsamples in comparison to RGB-based models.

A multispectral vision system developed by Kleynen etal. (2005) included four wavelength bands in the visible/NIR range for sorting apples cv. Jonagold based on thepresence of defects. They used interference filters centred at450, 500, 750 and 800 nm, but since the 500-nm spectralcomponent did not give any significant information fordiscriminating between defects and sound tissue, thisspectral component was finally not taken into account incomputing the frequency distributions. The 450-nm spectralband provided significant information with which toidentify slight surface defects like russet. The 750- and800-nm bands, on the other hand, offered good contrastbetween the defect and the sound tissue and were wellsuited to detecting internal tissue damage like hail damage,

Food Bioprocess Technol (2012) 5:1121–1142 1123

bruises and so forth. This system was recently used byUnay et al. (2011) to extract several features from defectiveskins of apples cv. Jonagold with the aim of classifying thefruit into different categories.

Bennedsen and Peterson (2005) also developed amultispectral machine vision system with the aim ofdetecting surface defects on apples. The system operatedon apples that were oriented with the stem/calyx axisperpendicular to the imaging camera. Images were acquiredthrough two optical filters at 740 and 950 nm. Due to thedifference in the detecting ability of the two wavebandsused, two training sets were constructed for each variety:one for 740-nm and one for 950-nm images. In order toevaluate the overall performance of the system, the binaryimages resulting from the segmentations of the 740- and950-nm images were combined. Apples of eight varietieswere used to test the combined performance of thesegmentation routines, with a success rate ranging from78% to 92%.

Blasco et al. (2007) developed a multispectral inspectionsystem to detect and sort citrus fruits according to 11different types of external defects by combining theinformation obtained from four image acquisition systemsthat are sensitive to NIR, visible, UV and fluorescence.Compared with the results obtained using only colourimages, the multispectral system showed that the contribu-tion of non-visible information increased the rate of successin fruit classification by about 78%. This research was laterenhanced in Blasco et al. (2009) to include somemorphological features of defects. The overall success rateincreased to 86%. However, the greatest increment wasachieved in the identification of anthracnose and greenmould (95% and 97%, respectively), which are dangerousdefects that spread the disease to sound fruits, and their rateof identification must be as high as possible. A further stepwas taken by Qin et al. (2012) who developed a prototypefor real-time inspection of citrus canker based on twowavelengths cantered at 730 and 830 nm. To discriminatebetween the stem end and the actual defects, they combinedanother two bands to create the band ratio R830/R730. Singlewavebands can be combined to create spectral indexes.Some of these indexes were studied by Lleó et al. (2011) todetermine the ones that best fit the ripeness prediction ofRichlady peaches, with two new indexes being proposed.Different spectral indexes obtained from three wavelengths(450, 650 and 800 nm) were also studied by Lunadei et al.(2012) to monitor changes in quality during the storage ofspinach leaves.

The detection of contaminants can be one of theapplications of such multispectral systems. Kalkan et al.(2011) developed a 2D local discriminant bases algorithmto discriminate between aflatoxin-contaminated and uncon-taminated hazelnuts and red chilli peppers flakes. The

samples were screened with 12 different filters, some at400–510 nm with 10-nm full width at half-maximum(FWHM) and others at 550 and 600 nm with 70 and40 nm FWHM, respectively. The algorithm classified theflakes into aflatoxin-contaminated and uncontaminatedclasses with a 79.2% accuracy rate so that the level ofaflatoxin in the test set was decreased from 38.26 to22.85 ppb by removing the ones that were classified ascontaminated. The hazelnut kernels were independentlysubjected to two different classifications: first, on thedetection of contamination and, second, on the detectionof fungal infestation without considering their aflatoxinconcentrations. A correct classification accuracy of 92.3%was achieved for classifying the hazelnuts as aflatoxin-contaminated (>4 ppb) or not (<4 ppb).

Internal quality can also be predicted using these systems.For instance, Lleó et al. (2009) used a multispectral imagingsystem to classify peaches into different levels of maturity atharvest and to compare this classification with referencemeasurements such as firmness or reflectance at 680 nmachieved with a visible spectrometer. The proposed systemhad three band-pass filters centred at 800 nm IR, 675 nm red(R) and 450 nm blue (B), with a bandwidth of 20 nm. Twonon-supervised classifications based on the Ward methodwere applied on the histograms extracted from the region ofinterest, i.e. the skin of the peach. The first classificationconsidered the R channel image of each sample, whilst thesecond used the histograms of the R/IR images, whichachieved better results (90% agreement). The use of the R/IRratio avoided the effect of fruit shape on light reflectance andthus improved the definition of multispectral maturityclusters. In contrast, the contribution of the B component inthe classification was poor.

Hyperspectral Imaging Devices

The essential elements for constructing hyperspectralimaging systems include light sources, wavelength selec-tion devices and area detectors (Sun 2010). Depending onthe technology used, the selection of the wavelengths canbe performed by dispersing the incident radiation into itsindividual wavelength or blocking the radiation in such away that only the desired wavelength reaches the detector.The most frequently used are usually imaging spectro-graphs, liquid crystal tunable filters (LCTF) and, to a lesserextent, acousto-optic tunable filters (AOTF). Althoughthese types are the most commonly used, there are alsoother kinds of equipment that have been developed for theacquisition of reflectance hyperspectral images (Kim et al.2001). Figure 1 shows a hyperspectral image of an orangewith some external defects that are clear in particularwavelengths and practically invisible in others.


Some devices for acquiring hyperspectral images evenshare technology with spectrometry, although the twotechniques should not be confused. Hyperspectral imagingprovides spectral and spatial information (what and where),whilst spectrometry provides information about spectralinformation captured at a particular spot on the sample. Toknow more about this technology, good reviews aboutspectrometry were carried out by Nicolaï et al. (2007) andMagwaza et al. (2011), or other literature can be consulted(Ozaki et al. 2006; Sun 2009).

Acousto-optic Tunable Filters

In recent years, technology based on AOTF has grown,thereby providing an alternative to LCTF and to imagingspectrographs (Vila et al. 2005), and its use is starting to beintroduced for optimising agricultural and chemical pro-cesses (Bei et al. 2004). For instance, Jiménez et al. (2008)used an AOTF to obtain the spectrum of olive oil frominside a horizontal centrifugal decanter. This equipmentallows instantaneous scanning of the oil flowing through

the sensor. The imaging system was programmed to scanthree spectra for each olive residue sample at a rate of10 scans per second in the 1,100- to 2,250-nm range.Moreover, AOTF technology has also been used for thedetermination of physical features in food applications.Cayuela et al. (2009) described a portable AOTF-NIRspectrophotometer with a wide spectral range between1,100 and 2,300 nm which was equipped with a reflectancepost-dispersive optical configuration and an InGaAs detectorused for NIR prediction of fruit moisture content (MC), freeacidity and oil content in intact olives.

An AOTF is a solid-state device that works as anelectronically tunable band-pass filter based on light–soundinteractions in a crystal (Chang 1976). It can isolate a singlefrequency of light from a broadband source in response toan applied acoustic field. The main components of anAOTF are a suitably oriented birefringent uniaxial crystal towhich a piezoelectric transducer is bonded. The mostcommon crystal for constructing an AOTF is telluriumdioxide (TeO2). The application of a radio frequency (RF)signal to the transducer produces an acoustic wave that

Fig. 1 Hyperspectral image decomposed on their monochromatic images showing an orange with some external defects


propagates inside the crystal. The acoustic waves changethe refractive index of the crystal by compressing andrelaxing the crystal lattice. The changes in the refractiveindex make the crystal act like a transmission diffractiongrating. Unlike a classical diffraction grating, the AOTFonly diffracts one specific wavelength of light so that it actsmore like a band-pass filter with a narrow bandwidth than adiffraction grating. The diffracted light is divided into twobeams with orthogonal polarisations, i.e. horizontallypolarised and vertically polarised. The undiffracted beamand the undesired diffracted beam (e.g. vertically polarisedbeam) are blocked by a beam stop. The wavelength of lightselected is a function of the frequency of the RF applied tothe crystal. Therefore, the wavelength of the diffractedbeam is controlled by changing the frequency of the RFsource (Vila-Francés et al. 2011). Furthermore, the band-width and the intensity of the filtered light can also beadjusted by controlling the RF source. In addition, ifmultiple RF frequencies are launched into the crystal, thencombinations of frequencies can be diffracted simulta-neously, which makes it more flexible than LCTFs whichgenerate only a single band-pass at a time. Since AOTF isan advanced electronically tunable filter, it includesimportant features similar to those found in LCTF, suchas accessibility to random wavelengths, flexible controlla-bility, high spectral resolution, fast wavelength switching,wide spectral range, narrow bandwidth and a relativelylarge optical aperture. AOTF technology presents, as itsmain advantage, a tuning speed that is higher than thatobtained with LCTF technology, typically in tens ofmicroseconds, owing to the fact that it is only limited bythe speed of the sound propagation in the crystal.Furthermore, AOTF presents broader wavelength ranges(Vila-Francés et al. 2010). On the other hand, AOTFs havesmaller apertures compared with LCTFs, and in addition,the imaging quality is comparatively poor.

Liquid Crystal Tunable Filters

An LCTF is a solid-state instrument that uses electronicallycontrolled liquid crystal cells to transmit light with aselectable wavelength whilst excluding all others. TheLCTF is based on Lyot filters which are built from a seriesof optical stages each consisting of a combination of abirefringent retarder (an optical property of a material thatcauses the polarisations of light to travel at different speeds)and a liquid crystal layer sandwiched between two parallelpolarisers. As the incident linearly polarised light traversesthe retarder, it is split into two light rays, i.e. the ordinaryand the extraordinary rays, which have different opticalpaths though the retarder and emerge with a phase delaythat is dependent on the wavelength. After transmissionthrough the retarder, only those wavelengths that are in

phase are transmitted by the polariser to the next filter stage(Hetch 2001). To introduce tunability, a liquid crystal layeris used in each stage. Tunability is provided by the partialalignment of the liquid crystals along an applied electricfield between the two polarisers. Tuning time for randomlyaccessing wavelengths depends on the liquid crystalmaterial used and the number of stages in the filter.Typically, it takes tens of milliseconds to switch from onewavelength to another, which is far longer than theresponse time of the AOTF.

A typical LCTF-based system includes the electronicmodule and the tunable filter. Gómez-Sanchis et al. (2008a)described the main elements, configuration and a spectralcharacterisation of an LCTF-based system for citrusinspection. The main characteristics included two liquidcrystal filters, with spectral ranges of 400–720 and 650–1,100 nm respectively. The combined use of both filtersprovided an extended working range from 400 to 1,100 nm.A uniform global efficiency of the system across allfrequencies was achieved by varying the integration timefor each band using a certified white reference. Images inthose bands where the efficiency of the system was lowerneeded more acquisition time than others where the systemwas more efficient. In Peng and Lu (2006), an LCTF-basedhyperspectral imaging system was developed for measuringfruit firmness of apples. Spectral images from RedDelicious and Golden Delicious apples were acquired from650 to 1,000 nm in increments of 10 nm. Gómez-Sanchis etal. (2008a) studied the feasibility of an LCTF hyperspectralsystem for detecting decay in citrus fruits in the early stagesof infection using halogen lighting instead of the traditionalinspection using UV lighting. Two filters were used toachieve this: one that was sensitive to the visible (460–720 nm) and one that was sensitive to the NIR (730–1,020 nm). Figure 2 shows a possible arrangement for thesimultaneous use of two LCTF filters (i.e. sensible tovisible and NIR) with a single camera without altering thescene. Wang et al. (2011b, c) describe a methodology toimplement a whole system based on this kind of filters fromthe calibration to the integration in a complete imageacquisition system.

Imaging Spectrographs

An imaging spectrograph is an optical device that is capableof dispersing incident broadband light into different wave-lengths instantaneously on an area detector (e.g. a CCDdetector). Wavelength dispersion is carried out using aprism. The imaging spectrograph generally operates in aline scanning mode, i.e. the object is scanned line-by-line asthe entire field of view is acquired. The light from ascanning line is dispersed into different wavelengths andare projected onto the area detector, creating a special 2D


image: one dimension represents spatial information andthe other the spectral dimension. Each vertical line alongthe spectral axis of the 2D area detector forms a continuousspectrum from a fixed spatial point on the object surface.The object must be moved stepwise under the acquisitionsystem by means of a stepper table, whilst at each step aline is scanned, as in a push broom scanner. Thus, a fullspectral image can be obtained. One advantage of theimaging spectrograph is its high spectral resolution. On theother hand, the major drawback is the need to move theobject with respect to the spectrograph. Therefore, it is notpossible to acquire an entire image without properlysynchronising the image acquisition with the movement ofthe object. This fact makes hyperspectral image acquisitiondifficult and thus makes it necessary to use complexcalibration techniques in order to achieve quality hyper-spectral images (Sun 2010).

This technology is cited in the scientific literature morefrequently than LCTF and AOTF technologies for agricul-tural applications probably because of its higher spectralrange and resolution. Some examples where this technologyis well described include ElMasry et al. (2008b), where ahyperspectral imaging system based on a spectrograph wasused in the spectral region between 400 and 1,000 nm forthe early detection of bruises on different backgroundcolours of apples cv. McIntosh. Three effective wave-lengths in the NIR region (750, 820, 960 nm) were foundthat could potentially be implemented in multispectralimaging systems for the detection of bruises on apples. Inthe study conducted by Al-Mallahi et al. (2008), a hyper-spectral imaging system was used to distinguish potatotubers from soil clods. An imaging spectrograph wasemployed to take hyperspectral images of 60 wavebandsin the 321- to 1,044-nm range. In order to improve success

rates, hyperspectral images were taken and analysed to findthe most relevant wavebands to perform an optimumdiscrimination. This paper highlights the usefulness ofemploying an imaging spectrograph to discriminatebetween tubers and clods and shows the significance ofadding a waveband from the NIR to accomplish a highsuccess rate of discrimination under any moisture con-ditions. Such systems have also been successfully used forthe evaluation of physical properties in food science. Forinstance, Polder et al. (2003) used an imaging spectrograph(393–710 nm) to estimate lycopene and chlorophyllcontents, which play a role in the ripening of tomatoes.They also used an imaging spectrograph to measure theconcentration of the main constitutive compounds intomatoes (including lycopene, lutein, β-carotene, chloro-phyll a and chlorophyll b) at different stages of ripening(Polder et al. 2004). Spectral images from 396 to 736 nmwith a resolution of 1.3 nm were recorded and used todetermine the presence of these compounds. Figure 3shows a possible arrangement of a hyperspectral systembased on image spectrograph. The sample has to be placedover a conveyor belt or a similar mechanism that displacesit under the camera as the image is acquired. A possiblealternative is the use of a mirror scan, a device that capturesthe image reflected on a moving mirror.

Comparison

Apart from the technology used to acquire images, the maindifferences among the different systems are related to theway they organize the images. Whilst LCTF and AOTFgive directly images (2D data arrays) with a specific size inthe spatial dimension that can be saved in standard graphicfile formats such as BMP or TIFF, imaging spectrographsacquire images where one dimension corresponds to asingle spatial line whilst the other represents the spectral

Fig. 3 Possible arrangement for hyperspectral imaging spectrograph

Fig. 2 System for the acquisition of hyperspectral images based ontwo LCTF


values of each pixel in this line. For all systems, the fullhyperspectral image obtained from these 2D imagescompose a 3D cube that contains both spatial (x,y) andspectral (λ) dimensions. What is different is how theyacquire this image cube. LCTF and AOTF acquire imageswith a particular size and high spatial resolution, but limitedspectral resolution. On the contrary, imaging spectrometerscan acquire images with high spectral resolution but lowspatial resolution. With respect to the size of images, it islimited in one dimension of the plane whilst the other isvirtually unlimited when using a pushbroom scanningsystem, which presents an advantage when acquiringimages of elongated objects. However, the images acquiredwith these systems are normally saved in RAW or in BSQgraphic file formats and need to be opened with specificapplications such as ENVI©, Matlab©, LabView© orothers which are similar.

It is important to note that for covering a relatively largeregion of the spectrum, for instance by including the visibleand NIR, usually more than one device is needed regardlessof the technology used. Therefore, some mechanicalarrangements have to be done if we want to acquire theconsistent images of the same scene in the entire spectrumunder study (Erives and Fitzgerald 2005), although thespecific ranges are different for each manufacturer andchange as technology evolves. However, it is veryimportant to take into account that all systems have lowsensitivity near the limits of the working spectral range, sosome wavelengths near these limits must be discarded inthe experiments due to the high noise introduced. Anotherimportant parameter to consider when choosing an appro-priate system is the FWHM parameter, measured as thespectral separation between the two points where thetransmission of a particular wavelength attains 50% of thepeak value, which determines the spectral resolution of thedevice and gives an idea of the purity of the wavelength.

Between the tunable filters, one advantage of the LCTFis a high rejection ratio for out-of-band transmission, whichimplies high image quality. Another important advantage isthat LCTFs can be manufactured with larger apertures thanAOTFs. On the other hand, their major drawback is theirlonger tuning time relative to AOTFs, as previouslycommented.

Statistical Techniques

Hyperspectral imaging systems have, on the one hand,problems similar to those presented by traditional computervision systems based on colour or monochromatic images.These problems are related with bad illumination, thepresence of artefacts or noise in the images, the selectionof regions of interest and so forth, which need to be

corrected in pre-processing steps. Several techniques havebeen developed for these tasks (Gonzalez and Woods 2008;Russ 2011). On the other hand, having a large number ofbands is of great interest, but also increases the complexityof the analysis of the information. The statistical problemsthat arise when multicollinearity in a set of variables to beanalysed is present makes the selection of importantwavelengths or the classification of pixels for hyperspectralimage segmentation difficult (Mather 1998). This sectionanalyses the most widely used statistical techniques foundin the scientific literature for this purpose: principalcomponent analysis (PCA), partial least squares (PLS),linear discriminant analysis (LDA) and artificial neuralnetworks (ANN). It should be noted, however, that most ofthese techniques are focused just on the spectral raw datawithout including information that is spatially adjacent inthe images; hyperspectral data are not treated as images butas simple values stored in a spectral array. However, spatialinformation can be included afterwards in the imagesobtained from the most important sources of variability(score images).

Other statistical techniques can also be applied forprocessing hyperspectral images. In this work, we focusonly on the ones that are most widely cited for the analysisof fruits and vegetables. General multivariate imageanalysis is reviewed in Prats-Montalbán et al. (2011).

Linear Discriminant Analysis

Discriminant analysis is a statistical technique for classify-ing objects into mutually exclusive groups (classes) basedon a set of measurable features of the objects, which, in thecase of hyperspectral images, are normally spectral features.This supervised method is focused on maximizing the ratioof the variance between groups and variance within groups(Jobson 1992). The class membership of a sample can bepredicted by calculating the distance to the centroid of eachclass in the transformed space and then assigning thesample to the class with the smallest distance to it. LDA hasno free parameters to be tuned and the extracted featurespotentially interpretable under linearity assumptions. Thistechnique is related to ANOVA and regression analysis,which also attempt to express one dependent variable as alinear combination of other features or measurements(Fisher 1936; McLachlan 2004). Furthermore, it is alsoclosely related to PC when dealing with multidimensionaldata arrays since LDA needs the X variables to beindependent and normally distributed, which can be easilyperformed by PCA feature extraction (working on thescores or latent variables obtained from PCA). However,this technique does not take into account any differences inclass (Martinez and Kak 2004). These capabilities have ledto its extensive use and practical exploitation in many fields


of application, such as hyperspectral imaging, whichinvolves a large number of features.

Polder et al. (2002) compared a hyperspectral systemwith a standard RGB for measuring the stages of ripenessof tomatoes. The spectrograph had a spectral range of 396–736 nm with a spectral resolution of 1.3 nm, which enabledit to obtain spectral images with 257 bands. LDA wasperformed on both hyperspectral and RGB images as wellfor classifying pixels into five different classes of ripeness.The classification of tomatoes was performed by a simplemajority vote regarding individual pixel classification, thespectral images being more discriminating than standardRGB images. On the other hand, Al-Mallahi et al. (2008)used linear discriminant analysis to distinguish potatotubers from soil clods. Five effective bands from hyperspectralimages were analysed. The discriminant functions from LDAwere able to separate the pixels and classify the objects astuber or clod under wet and dry conditions with higher ratesthan with just colour information. In this case, the hyper-spectral images analysed using LDA also offered better resultsthan the traditional RGB systems.

Gómez-Sanchis et al. (2008b) analysed the feasibilityof detecting decay in citrus fruits caused by Penicilliumdigitatum by finding out a reduced set of optimallyselected bands. Four feature selection methods wereevaluated in order to select the most discriminant bandsfor distinguishing between early fungal damage and soundskin, such as correlation analysis, mutual information,stepwise multivariate regression and genetic algorithmsbased on LDA. Gowen et al. (2009a) used LDA to analysehyperspectral images for the early detection of freezedamage in white button mushrooms. For each image, themean reflectance spectra from the central part wereobtained and PCA was applied to the calibration set ofdata in order to concentrate that spectral information intothe two first PC scores. LDA was then applied in order toestimate a boundary to separate the clusters of undamagedand freeze-damaged spectra. Using this method, freeze-damaged mushrooms could be correctly classified with anaccuracy >95% after 45 min thawing, even when freeze–thaw damage was not visibly evident.

Principal Component Analysis

PCA is one of the most popular multivariate statisticaltechniques commonly used to reduce the dimensionality ofdata and to solve the multicollinearity problem. Thismethod simplifies the task of obtaining an overview of allthe information in the dataset because it is an unsupervisedprojection method which summarises data by forming newvariables as uncorrelated and linear combinations of theoriginal variables. Normally, a few of these new variables,called principal components (PCs), should explain most of

the common variations in all the data. Therefore, by onlyconsidering several PCs, high-dimensional data can bereduced to a lower dimensionality with a minimal loss ofinformation. A drawback of PCA is that it does notguarantee the class separability of data due to its unsuper-vised nature related to the fact that PCA summarises thevariance in the data, which may not be related to thesegregation of the classes (Jolliffe 2002). This technique iswidely used in hyperspectral imaging as it is considered apowerful and robust tool for obtaining an overview of suchcomplex data and for reducing the large dimension of thedata provided by the hyperspectral images.

Mehl et al. (2004) used this technique on hyperspectralimages (628–900 nm) to detect different defects andcontamination on the surfaces of Red Delicious, GoldenDelicious, Gala and Fuji apples. Differences in spectralresponses were analysed using PCA and second differencemethods for sorting wholesome and contaminated apples.The study showed that both methods gave very similarresults for the detection of damages, the main differencesbeing related to the computing time. Whilst PCA is morecomplex to use and requires more data processing time, theasymmetric second difference method only required threewavelengths and much less computation time to process theimages. Zhao et al. (2010) applied PCA to the spectralregion ranging from 526 to 824 nm in order to get the bestscore image representing the bruised region of pears cv.Cristal. The score image was used prior to severalclassification algorithms to detect the bruises, finding thatthe bruised region could be clearly identified from the PC2score image. Then, several classification algorithms wereused in a comparative manner to classify pixels into bruisedor sound classes, including maximum likelihood, Euclideandistance, Mahalanobis distance (MD) and spectral anglemapper (SAM). The results showed that MD and SAMoffered the best performance, with detection accuracies of93.8% and 95.0%, respectively. Compared with the otherclassification algorithms, MDC and SAM were able toovercome the effects of uneven illumination in detectingbruising of pears, which is an important question to takeinto account.

Partial Least Squares

PLS regression is an unsupervised statistical method usedwhen not only a data array coming from X data is availablebut also a Y array that we want to predict from our X data.The aim of PLS analysis is to find a latent variables linearregression model by projecting the X variables and the Yvariables into a new latent space, where the covariancebetween these latent variables is maximized. In otherwords, a PLS model tries to find the latent multidimen-sional direction in the X space that explains the direction of


the maximum multidimensional covariance in the Y space.The PLS algorithm determines a set of orthogonalprojection axes, called PLS weights. Typically, most ofthe variance could be captured with the first few latentvariables, whilst the remaining latent variables describedrandom noise or linear dependencies between the depen-dent variables and the independent variables. Moreover,PLS, as PCA, is used to convert a set of highly correlatedvariables into a set of independent latent variables (t scores)variables by using linear transformations and can beemployed as a dimensionality reduction technique. Due toall these benefits, PLS models are widely used in hyper-spectral imaging to extract and summarise spectral infor-mation from hyperspectral images, to reduce the highdimensionality of the spectral data and to overcome theproblem of multicollinearity when we want to infer some Yproperty from the former (hyperspectral images; Vinzi et al.2010).

In order to study ripening in tomatoes, Polder et al.(2004) analysed the concentrations of different compoundsusing HPLC and by analysing spectral images using PLSregression at the pixel and region (tomato) levels. PLSregression models at the tomato level were constructedeither by averaging the prediction or by training the modelon the mean of the spectra of all the pixels belonging totomato class (i.e. only one spectrum is used per tomato).Although regression on complete tomatoes gave a smallererror than regression on individual pixels, pixel-basedregression made it possible to construct concentrationimages of tomatoes with non-uniform ripening.

Partial least squares discriminant analysis (PLS-DA) is asupervised variant of PLS regression where the independentvariable is a categorical one expressing the class member-ship of the samples. The usual PLS discriminant analysiswas proposed to overcome the multicollinearity problem ofLDA, and it can be defined as a straightforward extensionof the PLS regression (Sjöström et al. 1986). Using a PLS-DA calibration model, Nicolaï et al. (2006) discriminatedbetween pixels of sound apple skin and bitter pit lesions.Leave-one-out cross-validation was used to determine thenumber of latent variables based on the minimisation of theroot mean squared error of cross-validation prior tosegmentation. The number of latent variables in the PLScalibration was 2 and the system was able to identify bitterpit lesions, even when they were not visible to the nakedeye, such as just after harvest. Menesatti et al. (2009)studied the spectral region between 1,000 and 1,700 nmusing PLS-DA with the aim of developing an objectivemethod to assess the starch index in apples cv. GoldenDelicious. The starch classification of each apple was basedon the percentage of classified starch-free areas over thetotal (starch-free and starch) classified pixels. The potentialof hyperspectral imaging (445–945 nm) to discriminate

between casing soil, enzymatic browning and undamagedtissue on mushroom surfaces was investigated by Taghizadehet al. (2011b). Damage-free mushrooms, mushrooms artifi-cially smeared with casing soil and mechanically damagedmushrooms, resulting in enzymatic browning, were tested.PLS-DA models were developed to classify mushroomtissue as one of the three classes investigated using pixelspectra from each class.

Artificial Neural Networks

An ANN is a nonlinear statistical data modelling tool thatattempts to mimic the fault tolerance and capacity to learnof biological neural systems by modelling the low-levelstructure of the brain. An ANN consists of aninterconnected group of artificial neurons that works likea parallel system capable of resolving paradigms that linearcomputing cannot. In most cases, an ANN is an adaptivesystem that can change and adjust its knowledge byadjusting its parameters according to the samples of datathat are presented in order to solve the problem at hand.This is called the learning phase. They are usually used tomodel complex relationships between inputs and outputs orto find patterns in data. The most popular ANN is themultilayer perceptron (MLP), which is a feedforward ANNmodel that maps sets of input data onto a set of appropriateoutputs and consists of multiple layers of nodes (neurons)on a directed graph that is fully connected from one layer tothe next. MLP can employ a large variety of learningtechniques, the most popular being backpropagation. Thebackpropagation algorithm is a supervised learning methodbased on gradient descent in error which propagatesclassification errors back through the network and usesthese errors to update parameters (Shih 2010). ANN is acommonly used pattern recognition tool in hyperspectralimage processing because of the fact that it is capable ofhandling a large amount of heterogeneous data withconsiderable flexibility and because of its nonlinearproperty (Plaza et al. 2009). Furthermore, ANN haveseveral advantages over conventional pattern recognitionmethods: firstly, they can learn the intrinsic relationship byexample; secondly, they are more fault-tolerant thanconventional computational methods; and, finally, in someapplications, ANN are preferred over statistical patternrecognition because they require less domain-relatedknowledge of a specific application (Egmont-Petersen etal. 2002).

Because of their flexibility and the possibility of workingwith unstructured and complex data like that obtained frombiological products, ANN have been applied in almostevery aspect of food science, and it is a useful tool forperforming food safety and quality analyses (Huang et al.2007). For instance, a combination of PCA and ANN was


also used by Bennedsen et al. (2007) to detect surfacedefects on apples cv. Golden Delicious. The images werearranged as datasets in which each individual column wasconsidered a sample; the pixel values were taken asvariables. By using PCA, the columns of pixels werereplaced by the columns of PCs, where each column wasassigned a value of 0 if it did not represent a defect and avalue of 1 if it did. This matrix was used to train thesubsequent neural networks.

Unay and Gosselin (2006) applied 18 thresholding andclassification-based techniques for the pixel-wise segmen-tation of surface defects on apples cv. Jonagold. The resultsshowed that MLP was the most promising for segmentingsurface defects in high-speed machine vision-based appleinspection systems. Ariana et al. (2006) presented anintegrated approach using multispectral imaging in reflec-tance and fluorescence modes to acquire images of threevarieties of apples (Honeycrisp, Redcort and Red Deli-cious) to distinguish various defects on apples. Back-propagation ANN classification models were developedfor two classification schemes (binary and multi-class).Seven classification models for each variety were builtbased on imaging mode combinations, the integratedimaging model of reflectance and fluorescence (FUV+R)being more effective on cv. Honeycrisp apples, whereassingle imaging models of visible and NIR reflectance orUV-induced fluorescence were effective on cv. Redcort andcv. Red Delicious.

Dimensionality Reduction and Selection of SpectralFeatures

A typical hyperspectral image is composed of dozens ofcorrelative monochromatic images that normally exhibit ahigh degree of correlation. In addition, the computationalcomplexity of the classifier can become excessive for high-dimensional datasets, which makes it almost impossible todevelop automatic inspection systems capable of workingin-line or in real time. The optimal reduction in dimension-ality allows the performance system to be optimised interms of computational performance and simplicity. Theseproblems are commonly alleviated by using techniques thatretain most of the original information in fewer bands, butconserve the greatest amount of variability and the mostsignificant information contained in the hyperspectralimage (Du and Sun 2006). Guyon and Elisseeff (2003)summarised the main benefits of variable selection asimproving the inference performance, providing faster andcost-effective predictors, and better understanding of theunderlying process that generates the data. Methods forreducing the dimensionality can be divided into featureselection and feature extraction. Feature selection

approaches try to find a subset of the original variables,i.e. frequency bands in the case of hyperspectral images. Insome cases, data analysis such as regression or classifica-tion can be performed in the reduced space more accuratelythan in the original space. Feature extraction, on the otherhand, transforms the data in the high-dimensional spaceinto a space with fewer dimensions. The data transforma-tion may be linear, as in PCA, but many nonlineardimensionality reduction techniques, such as ANN, arealso commonly used (Lee and Verleysen 2007). Moststatistical techniques used for this purpose coincide withthose described in “Statistical techniques” for hyperspectralimage processing. However, this section aims to presentspecific applications of dimensionality reduction because tofind a small set of important wavelengths or to simplify theproblem in order to develop systems that can be automatedare some of the common objectives of the use of hyper-spectral systems.

The PCA technique has been widely applied for datareduction. For instance, Xing et al. (2005) used it to reducedata from a hyperspectral imaging system (400–1,000 nm)for detecting bruises on ‘Golden Delicious’ apples. Fourwavebands centred at 558, 678, 728 and 892 nm wereselected for later use in a multispectral imaging test usingPCA. The classification results indicated that about 93% ofthe non-bruised apples were recognised as sound, and anaccuracy of about 86% was achieved in bruised apples.Later, Xing et al. (2007a) used PCA in the same spectralregion to reduce the number of bands for separating stemend/calyx regions from true bruises on apples cv. GoldenDelicious and cv. Jonagold. They found four (558, 678, 728and 892 nm) and six (571, 608, 671, 709, 798 and 867 nm)effective wavebands for identifying the stem end/calyx,respectively, in the two varieties. With the same configu-ration and for the same purposes, Xing et al. (2007b) usedthe chemometric tool PCA and PLS-DA to extract andsummarise the spectral information from the hyperspectralimages. PCA was then carried out on the same foureffective wavebands as before. Later, PC2 and PC3 scoreimages were combined to construct virtual images, andfinally, the ‘moments’ thresholding method was used toprocess those virtual images, this tool being found to besuitable for this application.

Lefcout et al. (2006) proposed a robust method forselecting one or two wavelengths from hyperspectral datawith the aim of detecting faeces in Golden and RedDelicious apples using reflectance (452–729 nm) andfluorescence images (465–900 nm). The segmentation ofthe images was done by applying a binary threshold toimages of a single wavelength and to images constructedusing ratios or differences between images at two differentwavelengths. PCA was also used by Liu et al. (2005, 2006)to obtain spectral features for the detection of chilling


injury in cucumbers imaged using a hyperspectral systembased on a spectrograph with a spectral range of 447–951 nm. Both simple band ratio algorithms and PCA wereused in an attempt to discriminate the region of interest(ROI) spectra of good cucumber skins from those injuredby chilling. The PCA classification model was establishedusing two classes, ‘good’ and ‘injury’, based on SIMCA(Soft Independent Modelling of Class Analogy; Sjöström etal. 1986) of PCA with a Mahalanobis distance and aresidual spectral measurement. In the work by Gowen et al.(2008), the PCA method was used on a hyperspectralimaging system in the spectral region between 400 and1,000 nm with a resolution of 5 nm in order to detect bruisedamage on white mushrooms. Two data reduction methodswere investigated: in the first one, PCA was applied to thehypercube of each sample, and the second PC (PC2) scoreimage was used to identify bruise-damaged regions on themushroom surface. To detect various common skin defectson oranges, Li et al. (2011) used PCA to select mostdiscriminant wavelengths in the range 400–1,000 nm.Better detection results were achieved with the thirdprincipal component images using six wavelengths (630,691, 769, 786, 810 and 875 nm) and the second principalcomponent images using two wavelengths (691 and769 nm).

As already mentioned, PLS is one of the mostcommonly used methods for this purpose. This techniquewas used by ElMasry et al. (2007) to analyse hyperspectralimages in the visible and NIR (400–1,000 nm) regions forthe non-destructive determination of quality attributes instrawberry. PLS models were developed between theaverage reflectance spectra and the measured qualityparameters in order to predict quality parameters. Further-more, multiple linear regression models were establishedusing only the optimal wavelengths to predict the qualityattributes. Using PLSR, Gowen et al. (2009b) studiedspectral bands related with water in order to investigate thespectral behaviour of white mushrooms under differentlevels of mechanical vibration using a hyperspectralimaging system based on a spectrograph and operating inthe NIR wavelength of 950–1,700 nm. PLSR models werebuilt for the prediction of vibration time using meanabsorbance spectra as inputs. Changes in sample spectraarising from perturbation were examined by an observationof the PLSR coefficients. Candidate water matrix coor-dinates were proposed at 950, 1,174, 1,398, 1,433, 1,454,1,496 and 1,510 nm.

Qin and Lu (2005) developed a feedforward back-propagation ANN classifier to sort cherries with andwithout pits using a transmission hyperspectral imagingsystem. Single spectra obtained from a specified spatiallocation of the image and selected ROIs, both covering thespectral region between 692 and 856 nm, were compared as

inputs for the ANN. To reduce data dimensionality, PCAwas applied to the ROI data and only the score spectra ofthe first PC were used as inputs of the ANN. Neuralnetworks were also used by ElMasry et al. (2009) by meansof feedforward backpropagation ANN models, which weredeveloped to classify apples into injured and normal classesand to detect changes in firmness due to chilling injury.Five optimal wavelengths (717, 751, 875, 960 and 980 nm)were selected by an ANN model based on the maximumweight assigned to the input nodes. With the aim ofselecting an optimal set of wavelengths in the range 460-1020 nm for detecting decay in citrus fruits, Gómez-Sanchis et al. (2012) used a method call minimumredundancy maximal relevance (MRMR). Then, methodsbased on ANN and classification and regression trees(CART) were used to segment the images composed bythe 10 more relevant bands. Results showed an overallsuccess rate of 97% in discriminating pixels belonging todecay class from sound skin and other common defects.

A different approach was taken by Ariana and Lu(2010a) who used hyperspectral imaging under the trans-mittance mode to select important wavebands that can beused in the further development of an in-line inspectionsystem to detect internal defects in pickling cucumbers andwhole pickles. Hyperspectral images were acquired fromnormal and defective cucumbers and whole pickles using ahyperspectral reflectance system (400–740 nm) and atransmittance imaging system (740–1,000 nm). A total offour subsets of wavebands were determined by a branch-and-bound algorithm combined with the k-nearest neigh-bour classifier. The method used by Lorente et al. (2011) toselect most spectral relevant features to detect decay incitrus was based in the area under the receiver operatingcharacteristic (ROC) curve, which measures the quality ofthe classifier by optimising the true positive rate versus thefalse positive rate in binary problems, but they extend itsuse to multiclass problems to discriminate two types ofdecay, sound skin and other common defects.

In some cases, different works proposed different wave-lengths to solve the same problems, such as the detection ofdefects in certain fruits. This fact seems to indicate thatmore attention is paid to the statistical techniques or to thecapabilities of the technology than to what is reallyhappening in the fruits or vegetables. The results may beinfluenced by the illumination, the calibration of thesystem, the variables taken in the statistical models and,in general, by the particular experimental conditions. In thisrespect, it would be desirable to enhance the experimentswith chemical analyses that give some meaning to thewavelengths found. Table 1 summarises important worksthat use the described statistical techniques for the process-ing and reduction of dimensionality of hyperspectralimages.


Estimation of Product Quality

Hyperspectral imaging has recently emerged as a powerfulinspection tool for the quality assessment of fruits andvegetables. The quality of a piece of fruit or vegetable isdefined by several attributes that determine its marketabilityand shelf life. Quality assessment is therefore one of themost important goals of the highly competitive foodindustry. Product quality includes external appearance, suchas colour and the presence of skin diseases or bruises, andinternal quality features, including sugar content or firm-ness. External quality measurements of fruits are currently

being determined in many cases by computer visionsystems based on colour images. Even though suchtechniques offer important advantages like real-time oper-ation, lower cost or simulation of human processes, theyalso have some limitations, the main one being the fact thatthe human eye is restricted to the visible part of theelectromagnetic spectrum and misses important informationthat is outside these limits. Therefore, to expand qualityinspection beyond human limitations, it is necessary toemploy instrumental measurements such as hyperspectralimaging (Sun 2010). One of the applications of thesesystems is thus the enhancement of the traditional computer

Table 1 Studies related to statistical techniques applied to hyperspectral image processing and hyperspectral data reduction

Reference Fruit Statisticaltechnique

Application

Mehl et al. (2004) Red Delicious, Golden Delicious, Gala,and Fuji apples

PCA Detection of disease, fungal contamination, bruises andsoil contamination

Xing et al. (2005) Golden Delicious apples PCA Reduction in the number of hyperspectral bands forbruise detection

Qin and Lu (2005) Cherries PCA andANN

Detection of pits

Lefcout et al. (2006) Golden Delicious and Red Deliciousapples

PCA Detection of faeces

Xing et al. (2007a) Golden Delicious and Jonagold apples PCA Reduction in the number of hyperspectral bands forbruise detection

Xing et al. (2007b) Golden Delicious and Jonagold apples PCA & PLS Stem end/calyx end detection

Gowen et al. (2008) White mushrooms PCA Bruise detection

Zhao et al. (2010) Crystal pears PCA Bruise detection

Li et al. (2011) Oranges PCA Bruise detection

Wang et al. (2012) Onions PCA Sour skin damage

ElMasry et al. (2007) Strawberry PLS MC, SSC, and acidity (expressed as pH) qualityattributes

Polder et al. (2004) Tomatoes PLS Constitutive compounds at different ripening stages

Nicolaï et al. (2006) Apples PLS Identification of bitter pit lesions

Menesatti et al.(2009)

Golden Delicious apples PLS Assess starch index

Taghizadeh et al.(2011b)

White mushrooms PLS enzymatic browning

Polder et al. (2002) Tomatoes LDA Compare hyperspectral system to standard RGB forripeness stages

Al-Mallahi et al.(2008)

Potatoes LDA Distinguish potato tubers from soil clods

Gómez-Sanchis et al.(2008b)

Clemenules mandarins LDA Decay

Blasco et al. (2007) Citrus fruits LDA Segmentation of visible images

Gowen et al. (2009a) White mushrooms LDA Early detection of freeze damage

Qin and Lu (2005) Montmorency tart cherries ANN Sort cherries with and without pits

Bennedsen et al.(2007)

Golden Delicious apples ANN Identification of surface defects

Unay and Gosselin(2006)

Jonagold apples ANN Pixel-wise segmentation of surface defects

Ariana et al. (2006) Honeycrisp, Redcort and Red Deliciousapples

ANN Distinguish bitter pit, soft scald, black rot, decay andsuperficial scald defects

ElMasry et al. (2009) Red Delicious apples ANN Predict the firmness of the apples


vision techniques in the inspection of external quality bydeveloping more accurate systems for the estimation ofparticular quality features. However, this technology isbeing used largely because of the capability of hyper-spectral imaging to detect the presence of numerouschemical compositions that can be related with internalquality, as well as estimating their spatial distribution,which cannot be achieved by traditional systems.

Estimation of External Quality Parameters

Detection of skin defects is one of the most widespreaduses of machine vision in the inspection of fresh fruits andvegetables since the perceived quality is highly associatedwith a good appearance of the product (Kays 1999).Hyperspectral imaging provides a valuable tool not onlyto detect external defects but also to differentiate betweendifferent types of defects that have similar appearance oreven to detect some that are not clearly visible. Manyapplications aimed at such detection have been reported inapples. Bruising of apples is one of the major surfacedefects since they lower the quality of the fruit and entailsignificant economic losses; therefore, a great deal of efforthas been made to improve bruise detection. Lu (2003)developed a NIR hyperspectral imaging system for detect-ing both new and old bruises on apples in the 900- to1,700-nm spectral region. PCA and minimum noise fractiontransform were performed on the images acquired from RedDelicious and Golden Delicious apples over a period of47 days after bruising. The system was able to detect bothnew and old bruises, with a correct detection rate between62% and 88% for Red Delicious and from 59% to 94% forGolden Delicious in the spectral region between 1,000 and1,340 nm. Xing and Baerdemaeker (2005) used a hyper-spectral imaging system (400–1,000 nm) to detect bruiseson Jonagold apples. Six optimal wavelengths were selected(571, 608, 671, 709, 798 and 867 nm) and PCA was thenperformed on the resulting multispectral image. Thecontour plots for the first PC score images were used todistinguish between sound apples and bruised apples, theresult being a classification rate for sound apples of 84.6%and for bruised apples of 77.5%. It is difficult to establish acomparison since these works studied different spectralranges. However, depending on the type of defect, one ofthe other can be applied. For instance, for clearly visibledamages, probably a system based on visible could fitbetter, but for those defects that are difficult to distinguishwith the naked eye, probably the NIR would work better.

The citrus fruits are another important cultivar with greatcommercial. Martínez-Usó et al. (2005) used a hyper-spectral system (400–720 nm) to develop an unsupervisedalgorithm capable of segmenting images of oranges for thedetection of surface defects. The algorithm employed a

minimisation function which took into account the intensityof each band together with edge information. Anotherhyperspectral imaging approach to defect detection in citrusfruits was described by Qin et al. (2009) who detectedcanker lesions and other common peel diseases by means ofthe spectral information divergence (SID) classificationmethod based on quantifying the spectral similarities byusing a predetermined canker reference spectrum. SID wasperformed on the hyperspectral images of Ruby Redgrapefruits in the spectral region from 450 to 930 nm,achieving an overall classification accuracy of 96.2%.

Other works have also demonstrated the value ofhyperspectral imaging for the detection of skin defectsand damage in other species of fruits. A number ofhyperspectral imaging methodologies have been researchedfor the external quality assessment of cucumbers. Cheng etal. (2004) proposed a novel method for hyperspectralfeature extraction and applied it to the detection of chillinginjury on cucumbers using a hyperspectral imaging systemwith a spectral range of 447–951 nm. The results achieveda minimum defect recognition rate of 91% and a minimumsound cucumber recognition rate of 88.3% in the worstcase. Later, Liu et al. (2005) used the same hyperspectralimaging system for detecting chilling injury on cucumbers.Both simple band ratio algorithms and PCA were tested todiscriminate good cucumber skins from those of chilling-injured cucumbers. The results revealed that both the dual-band ratio algorithm using the 811- and 756-nm spectralreflectance and the PCA model from a narrow spectralregion of 733–848 nm were able to detect chilling-injuredskins of cucumbers within 3–7 days of storage at RT, with asuccess rate of over 90%.

Karimi et al. (2009) studied the changes in reflectance(350–2,500 nm) of avocados coated with different formu-lations. Absorbance images (1,000–1,600 nm) were used bySugiyama et al. (2010) to discriminate between the skin,stem and leaves of frozen blueberries. The optimalillumination wavelengths for distinguishing foreign materi-als were determined to be 1,268 and 1,317 nm, according tothe results of a discriminant analysis of absorbance spectra.LDA was used to select the wavelengths and to separatebetween classes. A threshold was then used to classify thepixels. Wang et al. (2011a) used a reflectance hyperspectralsystem (400–720 nm) for the detection of external insectdamage in jujube fruits. The peel conditions of jujubesamples were tested at different undamaged stem end/calyxend/cheek regions and at insect-damaged stem end/cheekregions. Over 98% of the intact jujubes and 94% of theinsect-infested jujubes represented in the images wererecognised correctly, and the overall classification accuracywas about 97%. Dangerous damage in onions caused bysour skin was detected at wavelengths 1,040 and 1,400 nmby Wang et al. (2012). They used PCA to select these two


wavelengths from the range 950–1,650 nm, and severalclassification methods including thresholding and Fisher’sdiscriminant analysis, to detect the disease, with an 87%success rate. Table 2 summarises some of the mostimportant works related with the estimation of externalquality using hyperspectral systems.

Estimation of Internal Quality Parameters

Hyperspectral imaging has also been widely used (mostlyon apples) to measure the internal quality attributes offruits, such as sugar or SSC, flesh and skin colour,firmness, acidity and starch index, and so forth. Concerninginternal quality, maturity is extremely important in thedetermination of harvest time and the marketing context.

Several parameters are normally used to assess fruitmaturity, for instance firmness. In recent years, many workson the determination of the maturity of apples have beenreported. An example of such studies is that of Peng and Lu(2005) who developed a method of predicting apple fruitfirmness using a multispectral imaging system. A Lorentziandistribution (LD) function with three parameters was pro-posed to characterise spatial scattering profiles from scatteringimages for Red Delicious apples at four wavelengths (680,880, 905 and 940 nm). A multi-linear regression analysis wasperformed to predict the relationship between parameters ofthe scattering profile and the firmness of apples. A similarapproach was used by Lu and Peng (2006) for assessingfirmness in Red Haven and Coral Star peach fruit, the resultbeing that a wavelength of 677 nm had the highest

Table 2 Works related with the estimation of external quality features of fruits

Reference Fruit Features Wavelengths(nm)

Lefcout et al.(2006)

Apples Golden and RedDelicious

Faeces 465–900

Xing andBaerdemaeker(2005)

Apples Jonagold Bruises 400–1,000

ElMasry et al.(2008b)

Apples McIntosh Bruises 400–1,000

Lu (2003) Apples Red Delicious andGolden Delicious

Bruises 900–1,700

Mehl et al.(2004)

Apples Red Delicious, GoldenDelicious, Gala, and Fuji

Side rots, bruises, flyspecks, scabs, moulds,fungal diseases and soil contaminations

430–900

Karimi et al.(2009)

Avocados Reflectance 350–2,500

Sugiyama et al.(2010)

Blueberries Foreign material 1,000–1,600

Gómez-Sanchiset al. (2008b)

Citrus fruits Decay 460–1,020

Cheng et al.(2004)

Cucumbers Chilling injury 447–951

Liu et al. (2006) Cucumbers Chilling injury 447–951

Qin et al. (2009) Grapefruits Ruby Red Canker lesions, greasy spot, insect damage,melanose, scab and wind scarring

450–930

Wang et al.(2011a)

Jujubes Insect damage 400–720

Wang et al.(2012)

Onions Sour skin damages 950–1,650

Martinez andKak (2004)

Oranges Skin defects 400–720

Qin et al. (2012) Oranges Canker lesions, thrips scarring, copper burn, insect damageheterochromatic stripe, scale infestation, phytotoxicity, and windscarring.

400–1,000

Blasco et al.(2009)

Oranges and mandarins Thrips scarring, phytotoxicity, wind scarring, scale infestation,chilling injury, sooty mould, oleocellosis, anthracnose,stem-end injury, medfly egg deposition and green mould

Visible, NIR

Zhao et al.(2010)

Pears Bruises 408–1,117

Gowen et al.(2008)

White mushrooms Bruises 400–1,000

Gowen et al.(2009a)

White mushrooms Freeze damage 400–1,000


correlation with firmness for a single wavelength,although at least 11 wavelengths were needed to obtainbetter results. The same group also used an LCTF-basedimaging system to measure fruit firmness of apples(Peng and Lu 2006). Spectral images from Red Deliciousand Golden Delicious apples were acquired from 650 to1,000 nm. Similar to their previous work, they used amodified LD function with four parameters, instead ofthree, to characterise the scattering profiles. The resultsindicated that whilst LD parameters at single wavelengthswere related to fruit firmness, they were insufficient forthe accurate prediction of fruit firmness.

Hyperspectral scattering is a promising technique for thenon-destructive sensing of multiple quality attributes ofapple fruit (Peng and Lu 2008; Mendoza et al. 2011). Nohet al. (2007) developed an integrated hyperspectral reflec-tance and fluorescence imaging technique for measuringapple maturity. Both fluorescence and reflectance scatteringimages were acquired using a hyperspectral imaging systemin the 500- to 1,000-nm region from Golden Deliciousapples to estimate multiple maturity parameters like fleshand skin colour, firmness, SSC, starch and titratable acid(TA). An approach similar to the one employed in theworks mentioned above was employed to relate LDparameters to individual maturity features for each sensingmode and their combined data. Similarly, Noh and Lu(2007) employed laser-induced fluorescence scatteringimages to predict multiple quality parameters (fruit skinand flesh colour, firmness, SSC and TA). Fluorescencescattering images were acquired from Golden Deliciousapples by a hyperspectral imaging system (500–1,000 nm).A hybrid method combining PCA and ANN modelling wasused to predict fruit quality parameters. Good predictionswere obtained for apple skin hue, with r=0.94, andrelatively good predictions were obtained for fruit firmness,skin chroma and flesh hue (r≥0.74). On the other hand,poorer correlations were found for SSC, TA and fleshchroma. Zhao et al. (2009) applied hyperspectral imaging(408–1,117 nm) to determine the sugar content of apple cv.Fuji. On applying the PLS method to the spectral profiles ofthe fruits, it was found that the optimal spectral range forsugar content was 704–805 nm.

Other applications of hyperspectral imaging systems tothe assessment of the internal quality of apples have beencited in recent literature, such as chilling injury detection.Damage in fruit cell membranes due to chilling injuryaffects normal firmness, and therefore changes in firmnesscould be used as an indication of possible chilling injury.ElMasry et al. (2009) detected chilling injury and predictedfirmness in Red Delicious apples using a hyperspectralimaging system (400–1,000 nm) and ANN techniques.Experimental results demonstrated that a spectral imagingsystem associated with ANN could successfully distinguish

between chilling-injured apples and normal apples (98.4%accuracy) as well as detect changes in firmness (r=0.92).

Despite the importance of the parameters mentioned abovein the measurement of fruit maturity, themost reliable maturityindex is the starch degradation pattern of the pulp. Apple fruitmaturation is characterised by an almost complete conversionof starch into sugars. The most usual way to assess the starchconversion stage is the iodine test in which cut fruit is dippedin iodine solution and stain patterns are visually evaluated byinspectors comparing them with reference charts. However, oflate, researchers have been focusing on developing non-destructive techniques to assess the starch index that avoidsubjective assessment and the usage of toxic iodine solution.Peirs et al. (2003) used a hyperspectral imaging system (868–1,789 nm) to measure the starch distribution and starch indexof apple fruit during maturation. A threshold value wasdefined to classify the pixels of the first PC score image of anapple into pixels containing a concentration higher or lowerthan the threshold value for staining. Hence, these imagesconfirmed the spatial starch degradation pattern as deter-mined by the traditional iodine technique.

Another hyperspectral imaging approach to assess thestarch index is the one described by Menesatti et al. (2009).They used visible and NIR images of Golden Deliciousapples in the spectral region between 1,000 and 1,700 nmto study the relationships between both types of images andvisually assessed starch/starch-free patterns, measuringstages of apple maturity. Pixels belonging to classes starchand starch-free were classified, achieving 66% accuracy.The starch index was also employed by Nguyen Do Tronget al. (2011) to estimate the optimal cooking time ofpotatoes. The changes in the microstructure during cookingaffect the interaction of light with the starch granules indifferent regions inside the potato. The potential of hyper-spectral imaging (400–1,000 nm) was studied for contactlessdetection of the cooking front in potatoes.

Hyperspectral imaging has also been used for determin-ing the internal quality of other large variety of fruits andvegetables, which demonstrated the flexibility and capabil-ity of this technology for the inspection of agriculturalproducts. Qin and Lu (2005) used hyperspectral transmis-sion images in the 400- to 1,000-nm spectral region todetect pits in tart cherries using ANN to classify cherrieswith and without pits. Experiments resulted in low averageclassification errors (about 3%), showing that sampleorientation and colour did not significantly affect classifi-cation accuracy, but the size of the fruit did. ElMasry et al.(2007) determined MC, SSC and acidity in strawberries bymeans of a visible/NIR hyperspectral imaging system (400–1,000 nm), with results that showed a good predictionperformance for moisture content (r=0.91), SSC (r=0.80)and pH (r=0.94). Fernandes et al. (2011) reported a systembased on neural networks (Ada Boost) for the estimation of


grape anthocyanin concentration using hyperspectral images(400–1,000 nm). The inputs from the NN were the PCs of thespectra of the grapes. The internal quality of whole pickleswas studied by Ariana and Lu (2010b) using transmittance(675–1,000 nm) and reflectance (400–675 nm) hyperspectralimaging systems. Images of pickles were obtained using aprototype in-line hyperspectral imaging system. An overalldefect classification accuracy of 86% was achieved com-pared with an accuracy of 70% by the human inspectors.Finally, banana fruit quality and maturity stages were studiedby Rajkumar et al. (2012) at three different temperaturesusing a hyperspectral imaging technique in the visible andNIR regions (400–1,000 nm). Some quality parameters likeMC, firmness and SSC were determined and correlated withthe spectral data using PLS. Eight wavelengths were requiredto predict the maturity stages of banana fruits representingthe quality attribute in terms of the features that were studied.Table 3 summarises some of the most important worksrelated with the estimation of the internal quality of fruitsusing hyperspectral systems

Major Problematic Aspects Related to HyperspectralImaging

Despite its potential as a technique for inspecting agricul-tural products with capabilities beyond those of the naked

human eye or traditional computer vision systems based oncolour images, hyperspectral imaging systems have a seriesof problems that must be considered before using ordeveloping one of them. For instance, they are stillexpensive and complex systems compared with standardcameras. On the other hand, the current acquisition times ofthe images do not allow them to be used for in-line or real-time inspection systems, although they are valuable toolsfor developing them, as stated above, by selecting certainimportant wavelengths.

One of the main aspects to be taken into account isrelated to the problems that appear when the products to beinspected have a spherical shape, like citrus, peaches orapples. In this case, two problems arise: the presence ofbright spots caused by the reflection of light and theprogressive darkness of the borders, both caused by theeffect of Lambert’s cosine law. The glares or bright spotscan be reduced by means of some techniques like cross-polarisation, but for the degradation of light intensity fromthe centre to the borders, software-based corrections areneeded. For instance, Gómez-Sanchis et al. (2008c)proposed a method for correcting hyperspectral images ofcitrus fruits by estimating the height of the fruit from theradius in order to build a 3D model of the fruit. Theluminosity of each pixel can then be tuned depending onthe angle between the surface normal at the relative positionof the pixel in the volume of the fruit (the incidence of the

Table 3 Works related to the estimation of internal quality features of fruits

Reference Fruit Features Wavelengths(nm)

Peirs et al. (2003) Apples Starch distribution and index 868–1,789

Zhao et al. (2009) Apples Fuji Sugar content 408–1,117

Menesatti et al.(2009)

Apples Golden Delicious Starch 1,000–1,700

Noh and Lu (2007) Apples Golden Delicious Flesh and skin colour, firmness, SSC and TA 500–1,000

Noh et al. (2007) Apples Golden Delicious Flesh and skin colour, firmness, SSC, starchand TA

500–1,000

Mendoza et al.(2011)

Apples Golden Delicious Red delicious andJonagold

Firmness and SSC 500–1,000

ElMasry et al.(2009)

Apples Red Delicious Chilling injury, firmness 400–1,000 nm

Peng and Lu (2005) Apples Red Delicious Firmness 680, 880, 905,940

Peng and Lu (2006) Apples Red Delicious and Golden Delicious Firmness 650–1,000

Rajkumar et al.(2012)

Banana MC, SSC and firmness 400–1,000

Fernandes et al.(2011)

Grapes Anthocyanin concentration 400–1,000

Ariana and Lu(2010b)

Pickles Bloated 675–1,000

ElMasry et al.(2007)

Strawberries MC, SSC and acidity 400–1,000

Qin and Lu (2005) Tart cherries Pits 400–1,000


ray of light) and camera’s line of sight. In other cases, someerrors can be assumed.

Another detail to be borne in mind in order to obtainreliable, precise and accurate measurements is the need fora good calibration. When a hyperspectral image is acquired,it is affected by several factors such as the spatialdistribution of the intensity due to bad illumination, lenserrors or sensor errors, and the fact that the capture deviceor the spectral emission of the lamps is not uniform acrossthe spectrum. Therefore, it is crucial to ensure thatmeasurements are independent of the spectral efficiencyof the different components of the system by means of goodsystem and image calibrations. A good example for thistask is described in Geladi (2007). In this work, the maintools for image calibration, such as white references ofcertified reflectance, are presented and discussed togetherwith some calibration methodologies.

One main problem is related to the construction ofrobust calibration models that can be used for extendedperiods of time and for different situations, which canhappen on such variable biological products as fruits orvegetables. The statistical models developed are oftentime-, size- and dataset-related, and therefore the resultsare dependent on the experimental conditions. Mostworks use fruits harvested in a particular orchard duringone season, but they have not been shown to work formore than the current season. Therefore, when the aim isto select important wavelengths, for instance, in order todetect certain defects on apples, some works proposecertain wavelengths whilst other works propose others,depending on the methodology used. But these wavelengthsare not usually explained from the point of view of the productand are not related with any particular compound that justifiesand makes the results more comprehensible, which is one ofthe future challenges.

Future Trends

The future of hyperspectral systems applied to food inspectionis promising since both the industry and consumers arebecoming increasingly aware of the need to ensure the qualityand safety of food, and this technology is an important tool forthe automatic inspection and monitoring of these parameters.The price of the equipment is constantly decreasing, whilst thetechnology allows more accurate imaging systems to bedeveloped that are capable of going further into theelectromagnetic spectrum. This would enable researchers tocreate new applications oriented towards the non-destructiveestimation of internal compounds related with the organolep-tic quality or shelf life of the products.

However, there are still two challenges to be overcomeusing this technology. On the one hand, the acquisition and

processing times of the images are still slow, whichprevents widespread implementation in an industry thatneeds real-time inspection. The partial solution is to searchfor a small set of important wavelengths that can be used todeal with each problem individually, but which sometimesmiss important information or limit the potential scope ofthe final application. On the other hand, most of theresearch being conducted is aimed at detecting these sets ofwavelengths or obtaining results to relate with particularobjectives. But in most cases, the results are dependent onthe laboratory conditions (lighting, calibration, etc.) or onthe statistical techniques used and are not truly related withinternal compounds or physical–chemical properties thatcould support these results from the product point of view.In these cases, for example, different studies can obtaindifferent sets of wavelengths for similar applications.

Conclusions

This paper has summarised the current state of the art on theapplication of hyperspectral imaging for fruit and vegetableinspection. Most works deal with statistical techniques toreduce the dimensionality of the problem, the most cited beingbased on ANN, PCA, PLS or LDA. Using the whole capturedspectrum or reducing the information to a few bands, theultimate aim is the inspection of quality beyond thepossibilities of traditional computer vision systems based oncolour images. However, there are still challenges in this topicthat have to be overcome by researchers.

Although nowadays these systems fit probably betterwith laboratory developments, many current works try toprovide the industry with important practical solutions.However, very few of them investigate the physical–chemical and biological phenomena that are evidenced inthe images. Thus, different works provide very differentresults for similar problems, for instance in the selection ofparticular wavelengths.

From the tables shown in this review, it can be deduced thatmost groups use equipment capable of working in the spectralrange between 400 and 1,000 nm, although precisely becauseof the lack of information about the meaning of the selectedwavelengths, it can be a fact derived from the high price of thesystems capable of inspecting beyond this range, although thenumerical results (success rates) obtained with all thetechnologies and ranges are similar.

The increasing interdisciplinary nature of researchgroups offers the possibility of combining genetic, biolog-ical and physiological knowledge with physics and com-puter vision research to take an important step towardsintegrated solutions for the fruit and vegetable industry.These solutions will not only allow problems to be detectedbut will also afford the generation of tools with which to


prevent their causes. In general, this is a technology whoseuse is beginning to extend to inspect the external andinternal quality of many horticultural products, mainlybecause of the constant price reduction of the componentsand the increment in computation capacity of moderncomputers. However, its potential as a non-destructiveanalytical tool is not fully exploited and much remains tobe investigated.

Acknowledgement This work was partially funded by the InstitutoNacional de Investigación y Tecnologia Agraria y Alimentaria deEspaña (INIA) through research project RTA2009-00118-C02-01 andby the Ministerio de Ciencia e Innovación de España (MICINN)through research project DPI2010-19457, both projects with thesupport of European FEDER funds. This work was also been partiallyfunded by the Universitat de València through project UV-INV-AE11-41271.

References

Aleixos, N., Blasco, J., Navarrón, F., & Moltó, E. (2002). Multispec-tral inspection of citrus in real-time using machine vision anddigital signal processors. Computers and Electronics in Agricul-ture, 33(2), 121–137.

Al-Mallahi, A., Kataoka, T., & Okamoto, H. (2008). Discriminationbetween potato tubers and clods by detecting the significantwavebands. Biosystems Engineering, 100(3), 329–337.

Ariana, D. P., & Lu, R. (2010a). Hyperspectral waveband selection forinternal defect detection of pickling cucumbers and wholepickles. Computers and Electronics in Agriculture, 74(1), 137–144.

Ariana, D. P., & Lu, R. (2010b). Evaluation of internal defect andsurface color of whole pickles using hyperspectral imaging.Journal of Food Engineering, 96(4), 583–590.

Ariana, D. P., Guyer, D. E., & Shrestha, B. (2006). Integratingmultispectral reflectance and fluorescence imaging for defectdetection on apples. Computers and Electronics in Agriculture,50, 148–161.

Bei, L., Dennis, G. I., Miller, H. M., Spaine, T. W., & Carnahan, J. W.(2004). Acousto-optic tunable filters: Fundamentals and applica-tions as applied to chemical analysis techniques. Progress inQuantum Electronics, 28(2), 67–87.

Bennedsen, B. S., & Peterson, D. L. (2005). Performance of a systemfor apple surface defect identification in near-infrared images.Biosystems Engineering, 90(4), 419–431.

Bennedsen, B. S., Peterson, D. L., & Tabb, A. (2007). Identifyingapple surface defects using principal components analysis andartificial neural networks. Transactions of the ASABE, 50(6),2257–2265.

Blasco, J., Aleixos, N., Gómez, J., & Moltó, E. (2007). Citrus sortingby identification of the most common defects using multispectralcomputer vision. Journal of Food Engineering, 83(3), 384–393.

Blasco, J., Aleixos, N., Gómez-Sanchis, J., & Moltó, E. (2009).Recognition and classification of external skin damage in citrusfruits using multispectral data and morphological features.Biosystems Engineering, 103, 137–145.

Cayuela, J. A., García-Martos, J. M., & Caliani, N. (2009). NIRprediction of fruit moisture, free acidity and oil content in intactolives. Grasas y Aceites, 60(2), 194–202.

Chang, C. (1976). Acousto-optic devices and applications. IEEETransactions on Sonics Ultrasound, 23(1), 2–22.

Chang, C. (2003). Hyperspectral imaging: Techniques for spectraldetection and classification. New York: Springer.

Cheng, X., Chen, Y., Tao, Y., Wang, C., Kim, M. S., & Lefcourt, A.(2004). A novel integrated PCA and FLD method on hyper-spectral image feature extraction for cucumber chilling damageinspection. Transactions of ASAE, 47(4), 1313–1320.

Costa, C., Antonucci, F., Pallottino, F., Aguzzi, J., Sun, D.-W., &Menesatti, P. (2011). Shape analysis of agricultural products: Areview of recent research advances and potential application tocomputer vision. Food and Bioprocess Technology, 4, 673–692.

Cubero, S., Aleixos, N., Moltó, E., Gómez-Sanchis, J., & Blasco, J.(2011). Advances in machine vision applications for automaticinspection and quality evaluation of fruits and vegetables. Foodand Bioprocess Technology, 4(4), 487–504.

Du, C.-J., & Sun, D.-W. (2006). Learning techniques used in computervision for food quality evaluation: a review. Journal of FoodEngineering, 72, 39–55.

Du, C.-J., & Sun, D.-W. (2009). Retrospective shading correlation of con-focal laser scanning microscopy beef images for three-dimensionalvisualization. Food and Bioprocess Technology, 2, 167–176.

Egmont-Petersen, M., de Ridder, D., & Handels, H. (2002). Imageprocessing with neural networks—A review. Pattern Recogni-tion, 35(10), 2279–2301.

ElMasry, G., Wang, N., ElSayed, A., & Ngadi, M. (2007). Hyperspectralimaging for nondestructive determination of some quality attributesfor strawberry. Journal of Food Engineering, 81, 98–107.

ElMasry, G., Nassar, A., Wang, N., & Vigneault, C. (2008a). Spectralmethods for measuring quality changes of fresh fruits and vegetables.Stewart Postharvest Review, 4, 1–13.

ElMasry, G., Wang, N., Vigneault, C., Qiao, J., & ElSayed, A.(2008b). Early detection of apple bruises on different backgroundcolors using hyperspectral imaging. LWT, 41, 337–345.

ElMasry, G., Wang, N., & Vigneault, C. (2009). Detecting chillinginjury in Red Delicious apple using hyperspectral imaging andneural networks. Postharvest Biology and Technology, 52, 1–8.

Erives, H., & Fitzgerald, G. J. (2005). Automated registration ofhyperspectral images for precision agriculture. Computers andElectronics in Agriculture, 47(2), 103–119.

Farrera-Rebollo, R. R., Salgado-Cruz, M. P., Chanona-Pérez, J.,Gutiérrez-López, G. F., Alamilla-Beltrán, L., & Calderón-Domínguez, G. (2011). Evaluation of image analysis tools forcharacterization of sweet bread crumb structure. Food andBioprocess Technology. doi:10.1007/s11947-011-0513-y.

Fernandes, A. M., Oliveira, P., Moura, J. P., Oliveira, A. A., Falco, V.,Correia, M. J., et al. (2011). Determination of anthocyaninconcentration in whole grape skins using hyperspectral imagingand adaptive boosting neural networks. Journal of FoodEngineering, 105(2), 216–226.

Fisher, R. (1936). The use of multiple measurements in taxonomicproblems. Annals of Eugenics, 7, 179–188.

Geladi, P. L. M. (2007). Calibration standards and image calibration.In H. F. Grahn & P. Geladi (Eds.), Techniques and applications ofhyperspectral image analysis (pp. 203–220). Chichester: Wiley.

Gómez-Sanchis, J., Camps-Valls, G., Moltó, E., Gómez-Chova, L.,Aleixos, N., & Blasco, J. (2008a). Segmentation of hyperspectralimages for the detection of rotten mandarins. Lecture Notes inComputer Science, 5112, 1071–1080.

Gómez-Sanchis, J., Gómez-Chova, L., Aleixos, N., Camps-Valls, G.,Montesinos-Herrero, C., Moltó, E., et al. (2008b). Hyperspectralsystem for early detection of rottenness caused by Penicilliumdigitatum in mandarins. Journal of Food Engineering, 89(1), 80–86.

Gómez-Sanchis, J., Moltó, E., Camps-Valls, G., Gómez-Chova, L.,Aleixos, N., & Blasco, J. (2008c). Automatic correction of theeffects of the light source on spherical objects. An application tothe analysis of hyperspectral images of citrus fruits. Journal ofFood Engineering, 85(2), 191–200.


http://dx.doi.org/10.1007/s11947-011-0513-y

Gómez-Sanchis, J., Martín-Guerrero, J. D., Soria-Olivas, E., Martínez-Sober, M., Magdalena-Benedito, R., & Blasco, J. (2012).Detecting rottenness caused by Penicillium in citrus fruits usingmachine learning techniques. Expert Systems with Applications,39(1), 780–785.

Gonzalez, R. C., & Woods, R. E. (2008). Digital image processing(3rd ed.). Upper Saddle River: Prentice Hall.

Gowen, A. A., O’Donnell, C. P., Cullen, P. J., Downey, G., & Frias, J. M.(2007). Hyperspectral imaging—An emerging process analyticaltool. Trends in Food Science & Technology, 18(12), 590–598.

Gowen, A. A., O’Donnell, C. P., Taghizadeh, M., Cullen, P. J., Frias,J. M., & Downey, G. (2008). Hyperspectral imaging combinedwith principal component analysis for bruise damage detectionon white mushrooms (Agaricus bisporus). Journal of Chemo-metrics, 22(3–4), 259–267.

Gowen, A. A., Taghizadeh, M., & O’Donnell, C. P. (2009a).Identification of mushrooms subjected to freeze damage usinghyperspectral imaging. Journal of Food Engineering, 93, 7–12.

Gowen, A. A., Tsenkova, R., Esquerre, C., Downey, G., & O’Donnell,P. D. (2009b). Use of near infrared hyperspectral imaging toidentify water matrix co-ordinates in mushrooms (Agaricusbisporus) subjected to mechanical vibration. Journal of NearInfrared Spectroscopy, 17(6), 363–371.

Grahn, H. F., & Geladi, P. (2007). Techniques and applications ofhyperspectral image analysis. Chichester: Wiley.

Guyon, I., & Elisseeff, A. (2003). An introduction to variable andfeature selection. Journal of Machine Learning Research, 3,1157–1182.

Hetch, E. (2001). Optics (4th ed.). Reading: Addison Wesley.Huang, Y., Kangas, L. J., & Rasco, B. A. (2007). Applications of

artificial neural networks (ANNs) in food science. CriticalReviews in Food Science and Nutrition, 47(2), 113–126.

Jiménez, A., Beltrán, G., Aguilera, M. P., & Uceda, M. (2008). Asensor-software based on artificial neural network for theoptimization of olive oil elaboration process. Sensors andActuators B, 129, 985–990.

Jobson, J. D. (1992). Applied multivariate data analysis: Categoricaland multivariate methods, vol. Berlin: Springer.

Jolliffe, I. T. (2002). Principal component analysis (2nd ed.). NewYork: Springer.

Kalkan, H., Beriat, P., Yardimci, Y., & Pearson, T. C. (2011).Detection of contaminated hazelnuts and ground red chili pepperflakes by multispectral imaging. Computers and Electronics inAgriculture, 77(1), 28–34.

Karimi, Y., Maftoonazad, N., Ramaswamy, H. S., Prasher, S. O., &Marcotte, M. (2009). Application of hyperspectral technique forcolor classification avocados subjected to different treatments. Foodand Bioprocess Technology. doi:10.1007/s11947-009-0292-x.

Kays, S. J. (1999). Preharvest factors affecting appearance. PostharvestBiology and Technology, 15, 233–247.

Kim, M. S., Chen, Y. R., & Mehl, P. M. (2001). Hyperspectralreflectance and fluorescence imaging system for food quality andsafety. Transactions of ASAE, 44(3), 721–729.

Kleynen, O., Leemans, V., & Destain, M. F. (2005). Development of amulti-spectral vision system for the detection of defects onapples. Journal of Food Engineering, 69, 41–49.

Lee, J. A., & Verleysen, M. (2007). Nonlinear dimensionalityreduction. New York: Springer.

Lefcout, A. M., Kim, M. S., Chen, Y.-R., & Kang, B. (2006).Systematic approach for using hyperspectral imaging data todevelop multispectral imagining systems: Detection of feces onapples. Computers and Electronics in Agriculture, 54, 22–35.

Li, J., Rao, X., & Ying, Y. (2011). Detection of common defects onoranges using hyperspectral reflectance imaging. Computers andElectronics in Agriculture, 78(1), 38–48.

Liu, Y., Chen, Y. R., Wang, C. Y., Chan, D. E., & Kim, M. S. (2005).Development of a simple algorithm for the detection of chillinginjury in cucumbers from visible/near-infrared hyperspectralimaging. Applied Spectroscopy, 59(1), 78–85.

Liu, Y., Chen, Y. R., Wang, C. Y., Chan, D. E., & Kim, M. S. (2006).Development of hyperspectral imaging technique for the detectionof chilling injury in cucumbers: Spectral and image analysis.Applied Engineering in Agriculture, 22(1), 101–111.

Lleó, L., Barreiro, P., Ruiz-Altisent, M., & Herrero, A. (2009).Multispectral images of peach related to firmness and maturity atharvest. Journal of Food Engineering, 93(2), 229–235.

Lleó, L., Roger, J. M., Herrero-Langreo, A., Diezma-Iglesias, B., &Barreiro, P. (2011). Comparison of multispectral indexesextracted from hyperspectral images for the assessment of fruitripening. Journal of Food Engineering, 104(4), 612–620.

Lorente, D., Aleixos, N., Gómez-Sanchis, J., Cubero, S., & Blasco, J.(2011). Selection of optimal wavelength features for decaydetection in citrus fruit using the ROC curve and neuralnetworks. Food and Bioprocess Technology. doi:10.1007/s11947-011-0737-x.

Lu, R. (2003). Detection of bruises on apples using near-infraredhyperspectral imaging. Transactions of the ASAE, 46, 523–530.

Lu, R., & Peng, Y. (2006). Hyperspectral scattering for assessingpeach fruit firmness. Biosystems Engineering, 93(2), 161–171.

Lunadei, L., Diezma, B., Lleó, L., Ruiz-Garcia, L., Cantalapiedra, S.,& Ruiz-Altisent, M. (2012). Monitoring of fresh-cut spinachleaves through a multispectral vision system. Postharvest Biologyand Technology, 63, 74–84.

Magwaza, L. S., Opara, U. L., Nieuwoudt, H., Cronje, P. J. R., Saeys, W.,&Nicolaï, B. (2011). NIR spectroscopy applications for internal andexternal quality analysis of citrus fruit—A review. Food andBioprocess Technology. doi:10.1007/s11947-011-0697-1.

Manickavasagan, A., Jayas, D. S., White, N. D. G., & Paliwal, J.(2010). Wheat class identification using thermal imaging. Foodand Bioprocess Technology, 3(3), 450–460.

Martinez, A. M., & Kak, A. C. (2004). PCA versus LDA. IEEETransactions on Pattern Analysis and Machine Intelligence, 23(2),228–233.

Martínez-Usó, A., Pla, F., & García-Sevilla, P. (2005). MultispectraliSegmentation by energy minimization for fruit quality estimation.In: Pattern Recognition and Image Analysis: Second IberianConference (IbPRIA 2005), Estoril, Portugal, June 7–9, 2005.LNCS, 3523, 689–696.

Mather, P. M. (1998). Computer processing of remotely sensedimages. Chichester: Wiley.

McLachlan, G. J. (2004). Discriminant analysis and statistical patternrecognition. New Jersey: Wiley-Interscience.

Mehl, P. M., Chen, Y. R., Kim, M. S., & Chan, D. E. (2004).Development of hyperspectral imaging technique for detection ofapple surface defects and contaminations. Journal of FoodEngineering, 61, 67–81.

Mendoza, F., Lu, R., Ariana, D., Cen, H., & Bailey, B. (2011).Integrated spectral and image analysis of hyperspectral scatteringdata for prediction of apple fruit firmness and soluble solidscontent. Postharvest Biology and Technology, 62(2), 149–160.

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

Nguyen Do Trong, N., Tsuta, M., Nicolaï, B. M., De Baerdemaeker,J., & Saeys, W. (2011). Prediction of optimal cooking time forboiled potatoes by hyperspectral imaging. Journal of FoodEngineering, 105(4), 617–624.

Nicolaï, B. M., Lötze, E., Peirs, A., Scheerlinck, N., & Theron, K. I.(2006). Non-destructive measurement of bitter pit in apple fruit


http://dx.doi.org/10.1007/s11947-009-0292-x

http://dx.doi.org/10.1007/s11947-011-0737-x

http://dx.doi.org/10.1007/s11947-011-0737-x

http://dx.doi.org/10.1007/s11947-011-0697-1

using NIR hyperspectral imaging. Postharvest Biology andTechnology, 40, 1–6.

Nicolaï, B. M., Beullens, K., Bobelyn, E., Peirs, A., Saeys, W.,Theron, K. I., et al. (2007). Nondestructive measurement of fruitand vegetable quality by means of NIR spectroscopy: A review.Postharvest Biology and Technology, 46(2), 99–118.

Noh, H. K., & Lu, R. (2007). Hyperspectral laser-induced fluores-cence imaging for assessing apple fruit quality. PostharvestBiology and Technology, 43, 193–201.

Noh, H., Peng, Y., & Lu, R. (2007). Integration of hyperspectralreflectance and fluorescence imaging for assessing apple matu-rity. Transactions of the ASABE, 50(3), 963–971.

Ozaki, Y., McClure, W. F., & Christy, A. A. (Eds.). (2006). Near-infrared spectroscopy in food science and technology. NewJersey: Wiley-Interscience.

Paulus, I., De Busscher, R., & Schrevens, E. (1997). Use of imageanalysis to investigate human quality classification of apples.Journal of Agricultural Engineering Research, 68, 341–353.

Peirs, A., Scheerlinck, N., De Baerdemaeker, J., & Nicolaï, B. M.(2003). Starch index determination of apple fruit by means of ahyperspectral near infrared reflectance imaging system. Journalof near infrared spectroscopy, 11(5), 379–389.

Peng, Y., & Lu, R. (2005). Modeling multispectral scattering profilesfor prediction of apple fruit firmness. Transactions of ASAE, 48(1), 235–242.

Peng, Y., & Lu, R. (2006). An LCTF-based multispectral imagingsystem for estimation of apple fruit firmness: Part I. Acquisitionand characterization of scattering images. Transactions of ASAE,49(1), 259–267.

Peng, Y., & Lu, R. (2008). Analysis of spatially resolvedhyperspectral scattering images for assessing apple fruitfirmness and soluble solids content. Postharvest Biology andTechnology, 48, 52–56.

Plaza, A., Benediktsson, J. A., Boardman, J. W., Brazile, J., Bruzzone,L., Camps-Valls, G., et al. (2009). Recent advances in techniquesfor hyperspectral image processing. Remote Sensing of Environ-ment, 113(1), S110–S122.

Polder, G., van der Heijden, G. W. A. M., & Young, I. T. (2002).Spectral image analysis for measuring ripeness of tomatoes.Transactions of ASAE, 45, 1155–1161.

Polder, G., van der Heijden, G. W. A. M., & Young, I. T. (2003).Tomato sorting using independent component analysis onspectral images. Real-Time Imaging, 9, 253–259.

Polder, G., van der Heijden, G. W. A. M., van der Voet, H., &Young, I. T. (2004). Measuring surface distribution ofcarotenes and chlorophyll in ripening tomatoes using imagingspectrometry. Postharvest Biology and Technology, 34, 117–129.

Prats-Montalbán, J. M., de Juan, A., & Ferrer, A. (2011). Multivariateimage analysis: A review with applications. Chemometrics andIntelligent Laboratory Systems, 107, 1–23.

Qin, J., & Lu, R. (2005). Detection of pits in tart cherries byhyperspectral transmission imaging. Transactions of ASAE, 48(5), 1963–1970.

Qin, J., Burks, T. F., Ritenour, M. A., & Bonn, W. G. (2009).Detection of citrus canker using hyperspectral reflectanceimaging with spectral information divergence. Journal of FoodEngineering, 93, 183–191.

Qin, J., Burks, T. F., Zhao, X., Niphadkar, N., & Ritenour, M. A.(2012). Development of a two-band spectral imaging system forreal-time citrus canker detection. Journal of Food Engineering,108(1), 87–93.

Quevedo, R., & Aguilera. (2010). Color computer vision andstereoscopy for estimating firmness in the salmon (Salmon salar)fillets. Food and Bioprocess Technology, 3(4), 561–567.

Quevedo, R., Aguilera, J. M., & Pedreschi, F. (2010). Color of salmonfillets by computer vision and sensory panel. Food andBioprocess Technology, 3(5), 637–643.

Rajkumar, P., Wang, N., EImasry, G., Raghavan, G. S. V., & Gariepy,Y. (2012). Studies on banana fruit quality and maturity stagesusing hyperspectral imaging. Journal of Food Engineering, 108(1), 194–200.

Russ, J. C. (2011). The image processing handbook (6th ed.). BocaRaton: CRC.

Shaw, P. J. A. (2003). Multivariate statistics for the environmentalsciences. New York: Hodder-Arnold.

Shih, F. Y. (2010). Image processing and pattern recognition:Fundamentals and techniques. New York: Wiley-IEEE.

Sjöström, M., Wold, S., & Söderström, B. (1986). PLS discriminantplots. In E. S. Gelsema & L. N. Kanal (Eds.), Pattern recognitionin practice I (pp. 461–470). Amsterdam: Elsevier.

Sugiyama, T., Sugiyama, J., Tsuta, M., Fujita, K., Shibata, M., Kokawa,M., et al. (2010). NIR spectral imaging with discriminant analysisfor detecting foreign materials among blueberries. Journal of FoodEngineering, 101(3), 244–252.

Sun, D.-W. (Ed.). (2007). Computer vision technology for food qualityevaluation. London: Academic.

Sun, D.-W. (Ed.). (2009). Infrared spectroscopy for food qualityanalysis and control. London: Academic.

Sun, D.-W. (Ed.). (2010). Hyperspectral imaging for food qualityanalysis and control. London: Academic.

Taghizadeh, M., Gowen, A. A., & O’Donnell, C. P. (2011a).Comparison of hyperspectral imaging with conventional RGBimaging for quality evaluation of Agaricus bisporus mushrooms.Biosystems Engineering, 108(2), 191–194.

Taghizadeh, M., Gowen, A. A., & O’Donnell, C. P. (2011b). Thepotential of visible-near infrared hyperspectral imaging todiscriminate between casing soil, enzymatic browning andundamaged tissue on mushroom (Agaricus bisporus) surfaces.Computers and Electronics in Agriculture, 77(1), 74–80.

Unay, D., & Gosselin, B. (2006). Automatic defect segmentation of‘Jonagold’ apples on multi-spectral images: A comparative study.Postharvest Biology and Technology, 42, 271–279.

Unay, D., Gosselin, B., Kleynen, O., Leemans, V., Destain, M. F., &Debeir, O. (2011). Automatic grading of bi-colored apples bymultispectral machine vision. Computers and Electronics inAgriculture, 75(1), 204–212.

Vila, J., Calpe, J., Pla, F., Gómez, L., Connell, J., Marchant, J. A., etal. (2005). SmartSpectra: Applying multispectral imaging toindustrial environments. Real-Time Imaging, 11, 85–98.

Vila-Francés, J., Calpe-Maravilla, J., Gómez-Chova, L., & Amorós-López, J. (2010). Analysis of acousto-optic tunable filterperformance for imaging applications. Optical Engineering, 49(11), 113203–113203-9.

Vila-Francés, J., Calpe-Maravilla, J., Gómez-Chova, L., & Amorós-López, J. (2011). Design of a configurable multispectral imagingsystem based on an AOTF. IEEE Transactions on Ultrasonics,Ferroelectrics, and Frequency Control, 58(1), 259–262.

Vinzi, V., Chin, W. W., Henseler, J., & Wang, H. (Eds.). (2010).Handbook of partial least squares. Berlin: Springer.

Wang, J., Nakano, K., Ohashi, S., Kubota, Y., Takizawa, K., & Sasaki,Y. (2011a). Detection of external insect infestations in jujube fruitusing hyperspectral reflectance imaging. Biosystems Engineering,108(4), 345–351.

Wang, W., Li, C., Tollner, E. W., Rains, G. C., & Gitaitis, R. D. (2011b).A liquid crystal tunable filter based shortwave infrared spectralimaging system: Calibration and characterization. Computers andElectronics in Agriculture. doi:10.1016/j.compag.2011.09.003.

Wang, W., Ca, L., Tollner, E. W., Rains, G. C., & Gitaitis, R. D.(2011c). A liquid crystal tunable filter based shortwave infrared


http://dx.doi.org/10.1016/j.compag.2011.09.003

spectral imaging system: Design and integration. Computers andElectronics in Agriculture. doi:10.1016/j.compag.2011.07.012.

Wang, W., Li, C., Tollner, E. W., Gitaitis, R. D., & Rains, G. C.(2012). Shortwave infrared hyperspectral imaging for detectingsour skin (burkholderia cepacia)-infected onions. Journal ofFood Engineering, 109(1), 36–48.

Xing, J., & De Baerdemaeker, J. (2005). Bruise detection on‘Jonagold’ apples using hyperspectral imaging. PostharvestBiology and Technology, 37(2), 152–162.

Xing, J., Bravo, C., Jancsók, P. T., Ramon, H., & De Baerdemaeker, J.(2005). Detecting bruises on ‘Golden Delicious’ apples usinghyperspectral imaging with multiple wavebands. BiosystemsEngineering, 90(1), 27–36.

Xing, J., Jancsók, P. T., & De Baerdemaeker, J. (2007). Stem-end/calyxidentification on apples using contour analysis in multispectralimages. Biosystems Engineering, 96(2), 231–237.

Xing, J., Saeys, W., & De Baerdemaeker, J. (2007). Combination ofchemometric tools and image processing for bruise detection onapples. Computers and Electronics in Agriculture, 56(1), 1–13.

Zhao, J., Vittayapadung, S., Quansheng, C., Chaitep, S., & Chuaviroj,R. (2009). Nondestructive measurement of sugar content of appleusing hyperspectral imaging technique. Maejo InternationalJournal of Science and Technology, 3(1), 130–142.

Zhao, J., Ouyang, Q., Chen, Q., & Wang, J. (2010). Detection ofbruise on pear by hyperspectral imaging sensor with differentclassification algorithms. Sensor Letters, 8, 570–576.


http://dx.doi.org/10.1016/j.compag.2011.07.012

Full Text

Documents

van der heijden

universitat

soluble solids

principal

typical hyperspectral

artificial

principal

food bioprocess