Scientia Agropecuaria 10(1): 143 – 161 (2019)

SCIENTIA AGROPECUARIA a. Facultad de Ciencias Agropecuarias Scientia Agropecuaria Universidad Nacional de Website: http://revistas.unitru.edu.pe/index.php/scientiaagrop Trujillo

REVIEW Trends in application of NIR and hyperspectral imaging for food authentication

Jeffrey Mendez1; Liz Mendoza1; J.P. Cruz-Tirado2, iD ; Roberto Quevedo3, iD ;

1,* iD Raúl Siche

1 Facultad de Ciencias Agropecuarias, Universidad Nacional de Trujillo, Av. Juan Pablo II s/n. Ciudad Universitaria, Trujillo, Perú. 2 Department of Food Engineering, University of Campinas, Brazil. 3 Departamento de Acuicultura y Recursos Agroalimentarios, Programa FITOGEN, Universidad de Los Lagos, Av. Fuchslocher 1305, Osorno, Chile.

Received December 10, 2018. Accepted March 2, 2019.

Abstract Food fraud can cause damage to consumer health and affect their confidence, destroy brands and generate large economic losses in the industry. Food authenticity allows to identify if food composition, geographical origin, genetic variety and farming system corresponds to what has been declared on the label. Although there are currently standardized methods to identify certain adulterants, the complexity of the food, the complexity of the supply chain and the appearance of new adulterants require the continuous development of analytical techniques to detect food fraud. NIR and Hyperspectral imaging (HSI) in tandem with chemometrics are non-destructive, non-invasive and accurate techniques for food authentication. This review focuses on NIR and HIS approaches to food authentication, including adulteration by substitution, geographical origin and farming system. In this context, the advances in NIR and HSI approaches reported since 2014 are discussed regarding their potential use in food authentication. Both techniques have shown to have efficiency, precision and selectivity to detect adulterants and identify geographic origin, genetic variety and farming system. Portability and remote access are shown as the next step for the industrialization of NIR and HSI devices.

Keywords: food fraud; spectroscopy; discrimination; regression.

1. Introduction (Branigan, 2008). Cases of food fraud Food authentication has become a growing cause, in addition to mistrust in the need throughout the world. Food consumer, large economic losses to authentication is aimed at detecting food companies or governments, destroying fraud, which is an illegal action carried out brands and devaluing the market value of for economic purposes through the the affected products. For instance, during adulteration of a food or false information the pork crisis in Ireland due to dioxin on the label (Barreto et al., 2018; Danezis et contamination, 1800 jobs were lost, and the al., 2016). This may include genetic variety, cost was estimated at US$138 million geographical origin, processing technology (Kennedy et al., 2009), while it is estimated and food composition (Esteki et al., 2018). that the adulteration of olive oil with Also, food authentication is important hazelnut oil causes a loss of 4 million euros because food fraud sometimes has per year for the European Union (Ozen and unfortunate consequences, for instance, Mauer, 2002). the adulteration of milk powder reported in In order to identify food fraud in food, China in 2008 (Gossner et al., 2009), which various techniques have been developed caused the death of six children and the using sophisticated and efficient technolo- hospitalization of thousands of others gies, which includes chromatographic

How to cite this article: Mendez, J.; Mendoza, L.; Cruz-Tirado, J.P; Quevedo, R.; Siche, R. 2019. Trends in application of NIR and hyperspectral imaging for food authentication Scientia Agropecuaria 10(1): 143-161.

------* Corresponding author © 2019 All rights reserved E-mail: [email protected] (R. Siche). DOI: 10.17268/sci.agropecu.2018.01.16

-143- J. Mendez et al. / Scientia Agropecuaria 10(1): 143 – 161 (2019) methods (Esteki et al., 2018), proteomics, This review shows a summary of the most metabolomics and genomics-based recent applications of NIR and NIR-HIS for methods (Böhme et al., 2019; Ortea et al., food authentication, which includes: geo- 2016) and spectroscopy techniques (Abbas graphical origin, genotype, genetic variety et al., 2018). This last group includes and adulteration by addition/substitution. techniques such as near-infrared spectros- copy (NIR), hyperspectral imaging (HIS), 2. 2. Recent applications of NIR and Fourier-transform infrared spectroscopy NIR-HIS for food authentication (FT-IR) and Raman (Figure 1). 2.1 Vegetable oil NIR is based on the absorption of Vegetable oils are one of the foods most electromagnetic radiation (light) in a susceptible to adulteration, both by wavelength range between 780 - 2500 nm addition and by false information regarding (Esteki et al., 2018). Each food presents a their geographical origin (Table 1). For characteristic spectrum, a fingerprint that instance, The Rapid Alert System for Food allows its identification and differentiation and Feed (RASFF) exposed a food fraud (Barbin et al., 2015). Variations in the case in sunflower oil with high levels of absorption of radiation at each wavelength mineral oil on 23 April 2008 (Picouet et al., are related to the chemical composition of 2018). This led to Picouet et al. (2018) to the food (Rady and Adedeji, 2018). There- develop a technique to detect “at-line” fore, factors such as the crop/process mineral oil in sunflower oil, using a portable conditions, geographical origin, variety or NIR coupled with a reflection probe, genotype that affect the chemical coupled with an immersion probe and a composition of foods are related with prototype of a multichannel Quasi Imaging different levels of absorption of radiation at Visible NIR spectrometer coupled to an a certain wavelength (Cozzolino, 2016). immersion probe. The prototype has been This allows us, with the help of supervised useful to detect adulteration at levels above and unsupervised chemometric methods, 2.5%. However, portable NIR coupled to to differentiate and classify a wide variety reflection probe to be more efficient (R2cal of foods based on the NIR spectral better than 0.99), achieving pure samples information. of sunflower oil with a probability of 98.5%, On the other hand, for certain food analysis and samples adulterated with mineral oil cases, NIR has not allowed to achieve the with a probability of 95%. Nevertheless, the expected efficiency, mainly because it is a authors conclude that detection of punctual technique. Hyperspectral imaging adulterated samples below 0.5% is difficult (HIS) is a variation of NIR (NIR-HIS), which using NIR, therefore, chromatographic combines the spectral information obtained techniques should be use to confirm the from the absorption of the radiation with fraud. the spatial information obtained from the On the other hand, NIR in transmission and image (Kamruzzaman et al., 2016). NIR-HIS transflectance modes combined with provides a greater amount of information SIMCA classifier allowed detect lard in than NIR, so one of the key approaches for palm oil with accuracy > 0.95 (Basri et al., its possible industrial application is the 2017). Also, the quantitative analyses were variable selection and the creation of performed by PLS coupled with variable multispectral models that spend less selection based on cumulative adaptive processing time. reweighted sampling (CARS) (Li et al., 2009). CARS allowed identify important wavelength interacted with the fat and oil chemical structure, being used for building robust prediction model (R2p > 0.99). CARS- PLS and CARS-ECR (elastic component regression) also showed good results for identify adulterated samples of sesame oil with different oil (see Table 1) (Chen et al., 2018). Therefore, this selection variable method should be tested in other oil type with different adulterants or different geographical origin. One of the vegetable oils with the highest risk of adulteration is olive oil (van Ruth et al., 2018) established that after the spice

Figure 1. NIR and hyperspectral imaging to detect food supply chain, the olive oil supply chain is fraud. the most susceptible to food fraud, mainly

-144-

J. Mendez et al. / Scientia Agropecuaria 10(1): 143 – 161 (2019) related on fraud factors of (1) fraud NIR allowed in less than 5 min time to detectability in raw material and (2) fraud identify the samples and assign them to detectability in final product, (3) historical different groups based on their compo- evidence of fraud, (4) valuable components sition. Similar results were found to detect or attributes and (5) level of competition extra-virgin olive oil adulterated with edible branch of industry. For this reason, NIR oils using FT-NIR spectral data combined spectroscopy has been a useful tool for the with SIMCA classifier (Karunathilaka et al., detection of olive oil adulteration. Azizian et 2016) and for olive oil blended with edible al. (2016) developed PLS models using oil using NIR coupled by SVM (Wu et al., information from the FT-NIR to predict the 2016). Also, NIR was very efficient to detect content of various compositional parame- soybean oil (Mendes et al., 2015), different ters of olive oil and relate it to its edible oils (Mossoba et al., 2017) and authenticity. From 66 samples, 50 from refined and mild deodorized olive oils California (USA), 15 from different (Wójcicki et al., 2015) in extra-virgin olive European countries and 1 from Spain, only oil. 23 met the authenticity requirements. FT-

Table 1 NIR and HIS to vegetable oil authentication

Statistical Oil type Fraud control Instrument Chemometrics Reference parameters Back ground correction NIR on mode R2p=0.99; + baseline correction + reflectance RMSEP=0.23% 1st derivate + MSC/PLS Sunflower Picouet et al. Mineral oil Prototype of a oil (2018) multichannel Quasi Background division + R2p=0.91; Imaging Visible NIR 1st derivate/PLS RMSEC=1.0% spectrometer Precision=1.0 Specificity=1.0 NIR in transmission SIMCA Sensitivity=0.2 mode CARS-PLS R2p=0.99; RMSEC=0.33% Basri et al. Palma oil Lard Precision=1.0 (2017) NIR in Specificity=1.0 SIMCA transflectance Sensitivity=0.4 CARS-PLS mode R2p=0.99; RMSEP=0.41% RMSEC=0.0245; CARS-PLS Sesame NIR in absorbance RMSEP=0.055 Chen et al. Authenticity CARS-ECR (Elastic oil mode RMSEC=0.0188; (2018) component regression) RMSEP=0.039 SBS-PLS (Stimulated Shi et al. Olive oil Authenticity UV-VIS-NIR R2p=0.99 Brillouin scattering) (2019) Extra R2p=0.99 Mendes et al. virgin Soybean oil NIR PLS RMSEP=1.76 (2015) olive oil Refined and Extra R2p=0.98 mild Wójcicki et al. virgin NIR PCR RMSEP=2.7 deodorized (2015) olive oil RDP=8.5 olive oil Extra Jiménez- Geographical Overall virgin NIR PLS-DA Carvelo et al. origin classification=100% olive oil (2019) Extra R2p>0.95 Mossoba et al. virgin Authenticity FT-NIR PLS (2017) olive oil Extra Overall Karunathilaka virgin Authenticity FT-NIR SIMCA classification=100% et al. (2016) olive oil Overall Wu et al. Olive oil Authenticity NIR SVM classification=93 - (2016) 100% Extra Geographical R2c=0.93 Peršurić et al. virgin NIR PLS-DA origin RMSEC=0.128 (2018) olive oil Correct Extra Laroussi- Geographical classification= 89.55 virgin NIR SIMCA Mezghani et origin – 98.50% olive oil al. (2015)

Correct Geographical Forina et al. Olive oil NIR LDA classification= 98.5 – origin (2015) 100%

-145-

J. Mendez et al. / Scientia Agropecuaria 10(1): 143 – 161 (2019)

In contrast, UV-VIS-NIR (in wavelengths appreciated being the Arabica variety, with 670 and 455 nm) combined with stimulated a higher market price compared to the Brillouin scattering (SBS) achieved to robust variety. These variations in price are detect olive oil adulteration (Shi et al., related to the most delicate form of Arabica 2019). SBS is an inelastic dispersion coffee production, growing at heights process that is presented by the fluctuation between 600 - 2000 m (Barbin et al., 2014; of the density of the acoustic sources in the Caporaso et al., 2018). Therefore, the medium, which allows to relate it to the identification of the variety of coffee beans properties of the density and refractive represents a very important research index (Shi et al., 2012). These olive oil focus. In addition, because coffee is properties change with adulteration with consumed in powder, it can be adulterated other oils, and can be used as an efficient at various stages of its supply chain, using method to separate between pure and lower value coffee or other food waste (i.e. adulterated samples. The method must be barley). Table 2 summarized the last extended for analysis of oils of species with research in the NIR and HIS application for a high index of polyunsaturated fatty acids, coffee authentication. as well as their mixtures, because the Bona et al. (2017) used support vector instability of these oils constantly machine (SVM) to compare NIR and FT-IR complicates their identification. for classification Arabica coffee according The determination of geographical origin of geographical origin in Brazil. SVM in olive oil is essential for the traceability of tandem with NIR spectral data allowed to the products and because the composition classify correctly (100%) Arabica coffee of the product is affected by their process samples. Similar results (sensitivity and conditions (Nenadis and Tsimidou, 2017; specificity = 1.0) were found for identify Wang et al., 2016). NIR combined with PLS- geographical origin and genotype origin of DA was more efficient than fluorescence Arabica coffee using PLS-DA in tandem excitation-emission matrix spectroscopy with NIR (Marquetti et al., 2016). For its technology to classify Argentinian extra- part, for Robusta coffee beans, FT-NIR virgin olive oil (Jiménez-Carvelo et al., coupled with self-organizing map (SOM) 2019). Similar results were found for allowed to identify and classify Robusta NIR/SIMCA to classify Tunisian extra-virgin coffee genotype origin (100% correct olive oil (Laroussi-Mezghani et al., 2015), classification) (Luna et al., 2017). After, in NIR/PLSDA to classify Croatian extra-virgin Brazil, a new research compared proton olive oil (Peršurić et al., 2018) and NIR/LDA transfer reaction mass spectrometry (PTR- to classify Italian olive oil (Forina et al., MS) and NIR to classify Arabica and 2015) according geographical origin. Arabica/Robusta (rate no informed) coffee What is questionable, perhaps, is the need according geographical origin (5 cities) and for more representative and corroborative farming system (conventional and organic) sampling. The time storage of oils, the type (Monteiro et al., 2018). NIR coupled with of olive cultivar and the process kind PLS-DA obtained a correct classification > carried out for oil extraction must be 80% for identify farming system, which is an considered (Binetti et al., 2017). These advantage compared with PTR-MS due to factors affect the composition of the oil, so low price, rapid analysis and minimal the differences may be given not by the sample preparation. More recently, in culture conditions of a specific place inter-laboratory analyses was confirmed (geographical origin), but may be the FT-NIR potential for classify Arabica dependent on other factors. For instance, if and Robusta coffee from different different cultivars are used, it is very likely geographical origin (2 continents, 9 that the variations in composition are countries) (Giraudo et al., 2019). PLS-DA related to the type of cultivar, so the results was a chemometric used to develop cannot be conclusive to differences related efficient classification models (correct to geographical origin. classification > 93%) for classify coffee beans according continent origin and 2.2 Coffee country origin. At same time, the results Coffee is one of the most consumed food were confirmed for two laboratories and no species in the world. Among the varieties significant difference (p < 0.05) was found, with greater economic importance is the which allows to create more reliable and Arabica coffee (Coffea arabica) and robust analytical methods. Robusta coffee (Coffea canephora) Variety identification (Robusta or Arabica) varieties (Thorburn-Burns et al., 2017). in green or roasted beans is important for Both species have differentiated food fraud control. Roasting coffee is an organoleptic characteristics, the most important stage for its consumption,

-146-

J. Mendez et al. / Scientia Agropecuaria 10(1): 143 – 161 (2019) because roasting process intensifies the data coupled with PLS to detect Robusta organoleptic properties of coffee beans. coffee in Arabica coffee (Bertone et al., However, roasting also disguises the 2016). appearance of coffee beans, making Hyperspectral imaging (HIS) is powerful identification difficult. (De Luca et al., 2016) technique for coffee beans identification studied the possibility of using NIR and characterization (Caporaso et al., combined with the PLS-DA or SIMCA 2018; Zhang et al., 2013). Due to a big data classifiers to identify the variety (Arabica or provided by HIS, a variable selection to Robusta) of roasted coffee. Both classifiers build a multispectral imaging system is proved to be efficient in differentiating essential for industrial application (Amigo between coffee varieties regardless of et al., 2013). For Arabica and Robusta coffee origin (Table 2). On the other hand, coffee discrimination was tested sparse portable NIR was used successfully to methods (sPCA + KNN and sPLS-DA) using identify blends of Arabica coffee with HIS data (Calvini et al., 2015). The sparse Robusta coffee (in different roasting levels), methods allow to perform variable selection corn, peels and sticks (Correia et al., 2018). at the same time as classification, giving PCA and PLS were chemometrics parsimonious models. Both sPCA + KNN employed to identify the adulterated and sPLS-DA converged to the same samples, obtaining a quantification limit important wavelengths, however, the (LOQ) of 5 – 8% w/w. Similar results (R2p = analysis time for sPCA + KNN was higher 0.97) were found using FT-NIR spectral (13.5 s) than sPLS-DA (< 0.1 s).

Table 2 NIR and HIS to coffee beans authentication

Coffee variety Fraud control Instrument Chemometrics Statistical parameters Reference Geographical Sensitivity=1.0 Bona et al. Arabica NIR SVM origin Specificity=1.0 (2017) Portable R2p=0.96 – 0.99 Correia et Arabica Authenticity PCA and PLS NIR RMSEP=2.8 – 6.6% al. (2018) Blends with R2p=0.97 Bertone et Arabica FT-NIR PLS Robusta coffee RMSEP=4.3% al. (2016) Robusta and Geographical Giraudo et FT-NIR PLS-DA Correct classification > 93% Arabica origin al. (2019) Geographical Arabica and NIR PLS-DA Correct classification > 61% origin Monteiro et Arabica/Robusta Farming al. (2018) blend NIR PLS-DA Correct classification > 80% system PLS-DA NIR Correct classification = 100% Arabica and Variety De Luca et Robusta roasted identification al. (2016) NIR SIMCA Sensitivity and specificity > 90% Efficiency prediction (test set) = 100% NIR-HIS sPCA + kNN Efficiency prediction (test image) Robusta and Variety = 86.9% Calvini et Arabica identification Efficiency prediction (test set) = al. (2015) 100% NIR-HIS sPLSDA Efficiency prediction (test image) = 80.2% Efficiency prediction (test set) = PLS-DA (4 94.9% NIR-HIS variables) Efficiency prediction (test image) = 74 - 92.2% Efficiency prediction (test set) = Robusta and Variety PLS-DA (32 100% Calvini et NIR-HIS Arabica identification variables) Efficiency prediction (test image) al. (2017) = 71 – 92.1% Efficiency prediction (test set) = sPLS-DA (2 100% NIR-HIS variables) Efficiency prediction (test image) = 83.9 – 93.1% Extreme Robusta and Variety Bao et al. NIR-HIS Learning Correct classification = 93.5% Arabica identification (2015) Machine Robusta and Variety Zhang et al. NIR-HIS SVM Correct classification = 98% Arabica identification (2018) Geographical Sensitivity = 0.75 – 1.0 NIR PLS-DA origin Specificity = 1.0 Marquetti et Arabica Genotype Sensitivity = 0.75 – 1.0 al. (2016) NIR PLS-DA origin Specificity = 0.93 – 1.0 FT-NIR PLS-DA Correct classification = 82.9% Genotype FT-NIR SIMCA Correct classification = 99.6% Luna et al. Robusta origin FT-NIR SOM Correct classification = 100% (2017) FT-NIR SVM Correct classification = 99.6%

-147-

J. Mendez et al. / Scientia Agropecuaria 10(1): 143 – 161 (2019)

After, the same research group used four process control method (MSPC) (Verdú et commercial filters: 1150 nm related to C-H al., 2016) and PLS-DA (Ziegler et al., 2016) aromatic second overtones, 1200 and 1250 allowed to discriminate between pure nm C-H aliphatic second overtone and 1400 wheat samples and adulterated samples. nm O-H first overtone of aliphatic, to build In many countries, including China, Japan multispectral model from the data obtained and South American countries, rice (Oryza from a hyperspectral image (Calvini et al., sativa L.) is one of the staple foods of their 2017). daily diet (Maione and Barbosa, 2018). Rice The results (see Table 2) suggest that the is a source of vitamins, minerals, fiber and multispectral model (4 variables) in tandem many essential elements (da Silva et al., with filters (individual) + PLS-DA or filters 2018; Dang and Vasanthan, 2019). The (combined) + sPLS-DA are as efficient as determination of the geographical origin the models built using full spectrum (150 and variety of rice has been an increasing variables). However, the analysis time is research line in recent years, especially in considerably reduced, which is important Asia region (Maione and Barbosa, 2018). for its online application. In the pixel-to- And, although the analysis of isotopes and pixel image classification, there is a shape minerals have been quite decisive for the effect of the coffee beans that are most authentication of rice (Mo et al., 2017), NIR noticeable in the prediction set. Although, spectroscopy is shown as a promising in general, the classification was correct technique (Table 3). (100%), new investigations can be carried NIR coupled with support vector data out to correct the shape effects, allowing a description (SVDD) was used to verify the better prediction in each pixel of the image. black rice authenticity (Chen et al., 2018). These results were better to those found in SVDD showed a best performance (100% previous works using NIR-HIS combined specificity and 94.2% sensitivity to identify with Extreme Learning Machine (correct authentic black rice) compared with k- classification = 93.5%) (Bao et al., 2015) nearest neighbor data description (KNNDD) and NIR-HIS coupled with SVM (correct and GAUSS method. This study should be classification = 98%) (Zhang et al., 2018) extended for all rice varieties and more using full spectra. chemometrics as PLS-DA and SIMCA must be used. On the other hand, diverse 2.3 Cereals wavelength selection methods should be Grains of cereals, wheat, rice, barley, oats; tested in order to reduce analysis time by and its derivatives, such as derived flour, creating multispectral models. are important ingredients for the most In other research, the NIR potential was important staple foods throughout the tested to identify and classify rice world, since they make foods such as according farming system: organic or breads, pastas, cakes and cookies conventional (Xiao et al., 2018). PCA and (Murniece and Straumite, 2014). For this, PLS were used to analyze NIR spectral the risk for the illicit activity associated with data. PLS showed a good performance cereals and derived products, are part of a (R2cv=0.8430 and RMSECV=0.1992) to big problem because of the impact for identify organic and conventional rice. everyone. Appropriate prevention However, the PLS was constructed using measures can only be implemented if the pure samples and no wavelength selections nature and type of illegal activity is was performed. Then, PLS-DA, KNN or understood (Tähkäpää et al., 2015). SIMCA probably present a better Therefore, this research aims to determine discrimination capacity to classify rice and analyze the extent of reported cases of according farming system, and new food adulteration in cereals and products to research should be drive in this sense. identify potential trends and frame the Hyperspectral imaging allowed identify and development of future empirical research in classify rice from different China’s regions this area. In this case, optical techniques (Sun et al., 2017). The spectral, texture and arise, such as the use of NIR and morphological features obtained of rice´s hyperspectral images, as efficient methods hyperspectral imaging were combined with to control the quality of these foods. SVM for classify rice according The adulteration of wheat flour using geographical origin. The features were cheaper flours is a common fraud case. HSI tested individually and combined, being the overcomes the difficulties of the spectral-texture-morphological model heterogeneity of the sample, allowing to (based on 9 important wavelengths) better efficiently detect flours of different grains. (correct classification = 91.67%) than other Methods based on HSI with multivariate models. analysis such as multivariate statistical

-148-

J. Mendez et al. / Scientia Agropecuaria 10(1): 143 – 161 (2019)

Table 3 NIR and HIS applied to cereal products authentication

Statistical Product Fraud control Instrument Chemometrics Reference parameters Sorghum Control of R2p=0.95 Multivariate 2 Oat R p=0.97 Verdú et al. Wheat HSI Statistical (2016) Corn Process R2p=0.99 (MSPC) Barley Erkinbaev Correct Oat Wheat HIS PCA-PLSDA et al. classification =0.96 Rye (2017) R2p=0.97 Common wheat RMSEP=0.038 Avatar R2p=0.98 Su and Sun Cassava flour HIS FMCIA-PLSR Wheat RMSEP=0.026 (2017) R2p=0.97 Corn flour RMSEP=0.036 Rye R2p=0.96 Organic Organic wheat PLSR R2p=0.97 spelt Su and Sun Common spelt HIS R2p=0.93 (Triticum (2016) Rye R2p=0.96 spelta L.) MLR Organic wheat R2p=0.97 Sun et al. Rice Geographical origin HIS SVM R2p=0.9167 (2017) Correct Mo et al. Rice Geographical origin HIS PLS-DA classification = 99% (2017) Correct LDA Wheat classification = 82% Zhao et al. Geographical origin NIR grains Correct (2013) PLS-DA classification = 92% Wheat Correct Zhao et al. Geographical origin NIR PCA-PLS grains classification = 85% (2014) Correct Zhou et al. Corn Geographical origin NIR PLS-DA classification = (2015) 87.9% Barley, Canola, Maize Wheat Flaxseed / Oats / Rye / Soybean SNV / K- Canada Correct Ravikanth Broken Kernels / Buckwheat / Nearest Western NIR - HIS classification = 100 et al. Chaff / Stones / Wheat Spikelets / Neighbors / Red Spring % (2015) Wild Oats Naive Bayes (CWRS) Deer Droppings / Rabbit Droppings Vermeulen Durum Common wheat HIS - NIR PLS - DA Sensibilidad = 88.3 et al. wheat (2018) Sensibilidad = 90 Ziegler et Wheat Authenticity NIR PLS - DA Especificidad = 100 al. (2016) Exactitud = 95 Correct Cooked Shao et al. Soybean flour HIS - NIR LS - SVM classification = millet flour (2018) 98.30% RSQ = 0.96 Revilla et Lentils Geographical origin FT - NIR PLS - DA RMSECV = 0.40 al. (2019)

The correct sample collection is essential 2012). Perhaps the greatest likelihood of to develop an analysis method based on food fraud is found in meat products. spectral information, which should be Because the meat processing removes represent the variability studied (i.e. external morphological features of a geographical origin). For instance, if the muscle, which makes it difficult to identify aim is to differentiate by geographical the meat species in the processed product origin, factors such as genetic variety, (Sentandreu and Sentandreu, 2014). Due to genotypes (cultivars) and year of harvest the dynamism of the meat industry, fast, must be controlled. reliable and robust methods are needed that allow the authentication of meat and 2.4 Meat and fish meat products (Amaral et al., 2016; Kumar Adulteration in meat and meat products and Chandrakant-Karne, 2017). NIR and with undeclared animal species has HIS are presented as powerful, non- generated concern among consumers, destructive, non-invasive, fast and reliable because these meats are considered an technologies for meat authentication edible (i.e. horse meat in USA) or due to (Aredo et al., 2017; Cheng et al., 2017). social taboos (i.e. pork meat in Mulsim and Table 4 shows the last research focused in Jewish communities) (Boyacı et al., 2014; meat and fish authentication. Monahan et al., 2018; Nakyinsige et al.,

-36-

J. Mendez et al. / Scientia Agropecuaria 10(1): 143 – 161 (2019)

The taste and texture of the meat varies 78.95 - 100% in validation set. The adulte- depending on the type of muscle con- ration of lamb and beef with chicken meat sumed, and many times various undeclared was more difficult to access. However, the muscles are mixed with fraudulent econo- Lidia breed cattle and foal in minced beef mic motivations. Hyperspectral images of could be detected at levels of adulteration Lamb muscles: Longissimus dorsi, Psoas above 2% and 1%, respectively. These major, Semimembranosus and Semiten- results were better than those found to dinosus, were acquired in order to classify identify and quantify turkey meat in minced according muscle type (Sanz et al., 2016). beef (fresh, frozen-thawed and cooked) After testing different machine learning using FT-NIR with PLS-DA (Sensitivity = techniques, the linear Least Mean Squares 0.84; specificity = 0.76) (Alamprese et al., (LMS) classifier showed the best result 2016). On the other hand, it was possible (96.67% correct classification). Nolasco et identify and classify (Correct classification al. (2018) studied the use of NIR to classify = 94.2 – 100%) beef steaks according their different chicken parts: drumstick, breast ageing time (3, 7, 14 and 21 days post and thigh. The linear discriminant analysis mortem) using Vis-NIR spectral data (LDA), random forest (RF), and support combined with PLS-DA (Moran et al., 2018). vector machine (SVM) algorithms were Also, the authenticity of Italian Valle used and a correct classification of 97.5% d’Aosta Arnad Protected Designation of was achieved. It should be noted that this Origin (PDO) lard can be determine using research was conducted using a portable NIR spectral data in tandem with PLS-DA NIR, which is more attractive and less (Sensitivity and Specificity = 94.4%), expensive industrially. although with less precision than using Rady and Adedeji (2018) studied the volatile compound (VOC) or fatty acid (FA) adulteration of beef with chicken or pork analyses (Sensitivity and Specificity = and with vegetable protein using Vis-NIR- 100%) (Chiesa et al., 2016). For its part, FT- HIS and NIR-HIS. Both sensors (Vis-NIR-HIS NIR in tandem with SVM allowed to analyze and NIR-HIS) showed to be efficient to and classify veal sausage adulterated with classify pure samples of beef, chicken, pork and pork fat (10 – 50% w/w) pork, texturized vegetable protein (TVP) (Schmutzler et al., 2015). The results found and wheat gluten (WG) (100% correct in this work are quite relevant, since the classification) and adulterated samples method developed in the laboratory was (96% using selected wavelengths). The tested in industrial and on-site instrumental adulterant identification was more difficult, setups, analyzing them successfully but acceptable (69 – 100% correct classifi- through their plastic packaging (75 - 100% cation). Vis-NIR-HIS in selected wave- correct classification). lengths was more efficient [r(RDP)] for Currently, the fish production and predict adulterant concentration in minced marketing chain has been internationalized, beef: 0.85 (1.77) for pork, 0.86 (1.95) for with fish exported from developing to TVP, 0.86 (1.98) for chicken, 0.86 (1.87) and developed countries. NIR spectroscopy in 0.87 (1.64) for WG. In addition, Vis-NIR-HIS tandem with SIMCA was used to classify coupled with PLS regression allowed to tilapia fillets according to their quantify duck meat concentration in min- geographical origin (China) (Liu et al., ced lamb (R2p = 0.98) (Zheng et al., 2019). 2015). On the other hand, NIR coupled with one- The results showed that the NIR spectral class classifier variant of the partial least information is able to classify, acceptably squares method (OCPLS) and the soft (Correct classification = 75 - 85%), between independent modeling of class analogy tilapia fillets from different regions of China (SIMCA) allowed to classify ground meat (Guangdong Province, Hainan Province, according specie origin (lamb, beef and Guangxi Province and Fujian Province). On pork) (Pieszczek et al., 2018). Therefore, it the other hand, Vis-NIR-HIS was tested for is possible that NIR coupled with any of classify fresh from cold-stored (4 °C for 7 these two classification techniques (SIMCA days) and frozen-thawed (−20 °C and −40 °C or OCPLS) is appropriate and feasible to for 30 days) grass carp fish fillets (Cheng et identify ground beef according to its al., 2015). SIMCA, PLS-DA, least squares- species at industrial level. Later, the support vector machine (LS-SVM) and potential of NIR to identify and classify beef probabilistic neural network (PNN) classi- and lamb meat adulterated with pork, fiers were tested using full and important chicken, Lidia breed cattle or foal was wavelengths (446, 528, 541, 596, 660, 759 studied (López-Maestresalas et al., 2019). and 970 nm). NIR spectral data in tandem with PLS-DA achieved a correct classification between

-150-

J. Mendez et al. / Scientia Agropecuaria 10(1): 143 – 161 (2019)

Table 4 NIR and HIS to meat and fish authentication

Meat product Fraud control Instrument Chemometrics Statistical parameters Reference Muscle Linear least Mean Correct classification = Sanz et al. Lamb Vis-NIR HIS discrimination Squares (LMS) 96.67% (2016) LDA Muscle Nolasco et al. Chicken NIR Random Forest Correct classification = 97.5% discrimination (2018) SVM Overall classification = 76 - SVM 95% VIS-NIR HIS PLS R2p = 0.53 – 0.86 Rady and RMSEP = 0.17 – 1.36 Beef Adulteration Adedeji Overall classification = 81 - (2018) SVM 95% NIR-HIS PLS R2p = 0.53 – 0.86% RMSEP = 0.26 – 0.55% Sensitivity = 96 - 100% NIR SIMCA Specificity = 80 – 99% Specie one-class classifier Pieszczek et Ground meat identification variant of the partial Sensitivity = 95 - 100% al. (2018) NIR least squares method Specificity = 73 – 99% (OCPLS) López- Correct classification = 78.95 Lamb and Beef Adulteration NIR PLS-DA Maestresalas – 100% et al. (2019) R2p = 0.98 Zheng et al. Lamb Adulteration Vis-NIR-HIS PLS RMSEP = 2.51% (2019) Correct classification = 75 – Schmutzler et Veal sausage Adulteration FT-NIR SVM 100% al. (2015) Correct classification = 94.2 – Moran et al. Beef steaks Ageing time Vis-NIR PLS-DA 100% (2018) Sensitivity and specificity = Chiesa et al. Lard Authenticity NIR PLS-DA 94.4% (2016) Fresh, frozen- Sensitivity = 0.84 thawed and PLS-DA Specificity = 0.76 Adulteration FT-NIR cooked PLS R2p = 0.884 minced beef RMSEP < 10.8% Gray-level-gradient Fresh and co-occurrence matrix Correct classification = 90.91 Pu et al. Pork muscle frozen- Vis-NIR-HIS (GLGCM) + – 93.14% (2015) thawed probabilistic neural network (PNN) PLS-DA based on fused variables Fresh and combining spectra at Correct classification = Ma et al. Pork muscle frozen- Vis-NIR-HIS the optimal 97.73% (2015) thawed wavelengths and textures Gray-level-gradient co-occurrence matrix Farming (GLGCM) + radial Correct classification = Xiong et al. Chicken Vis-NIR-HIS system basis function- 93.33% (2015) support vector machine (RBF-SVM) Geographical Correct classification = 75 – Liu et al. Tilapia fillets NIR SIMCA origin 85% (2015) Farming Vis-NIR-HIS SVM Correct classification = 98.2% Xu et al. Salmon system NIR-HIS SVM Correct classification = 92.7% (2017) Fresh, cold- stored, Correct classification = 90 – Cheng et al. Carp Vis-NIR-HIS LS-SVM frozen- 100% (2015) thawed

1st derivate pre-processing technique and 2.5. Honey LS-SVM showed the best performance Honey is probably one of the most complex using full wavelengths (94.29% correct and consumed natural food (Pita-Calvo et classification) and important wavelengths al., 2017). Bees (Apis mellifera) collect (91.43% correct classification). Later, (Xu nectar, plant secretions or excretions of et al., 2017) studied and compared the plant‐sucking insects to produce honey, potential of computational vision and Vis- after complex enzymatic process. Honey is NIR-HIS for classify salmon according a complex mixture of carbohydrates (70- farming system: organic and conventional. 80% w / w), water (10-20% w / w) and a Vis-NIR-HIS combined with SVM was more large number of minor components successful (98.2% correct classification) (Ouchemoukh et al., 2007). than using computational vision (83.6% Fructose/glucose and fructose/glucose correct classification using PLS-DA) or NIR- disaccharide are the main carbohydrates in HIS (92.7% correct classification using honey (65-80% w/w) (de la Fuente et al., SVM) to classify salmon. 2006). Therefore, a common adulteration

-36-

J. Mendez et al. / Scientia Agropecuaria 10(1): 143 – 161 (2019) practice is through mislabeling and mixing LDA allowed to identify high fructose corn sugar syrup and lower-quality honeys or syrup in honey (CCR = 100%) and to predict production honey using sugar instead of adulterant concentration (Ferreiro- nectar. González et al., 2018). For high fructose In this context, NIR is probably a most corns syrup and maltose syrup adulterants, sensitive spectroscopy technique to detect a multispectral model based on NIR data honey fraud (Naila et al., 2018). However, combined with CARS/PLS-Da achieved high some studies show the efficiency of hyper- precision (CCR = 88.3%) (Li et al., 2017). spectral images to identify adulteration and CARS allows to select important the origin of honey. Table 5 summarizes the wavelengths, to build multispectral model works that apply NIR and hyperspectral for on-line applications (Li et al., 2009). images to determine honey authenticity. However, multispectral models must be The chemical composition and nutritional equal or more accurate than models using properties of honey make it a healthy and whole spectrum. On the other hand, NIR desired food. Therefore, its market value is spectral data obtained using fiber optic more expensive than common sweeteners immersion probe (transmittance mode) was such as sugar beet, cane, inverted syrups used to build PLS models for detecting and syrups with high fructose content (Pita- accurately high fructose corn syrup in Calvo et al., 2017). But, the adulteration of honey (RMSECV = 1.48; R2CV = 0.987) honey with inverted syrups or high fructose (Bázár et al., 2016). Later, Başar and syrups is usually difficult to detect, because Özdemir (2018) built regression models to they mimic the sucrose-glucose-fructose detect beet sugar and corn syrup in honey. profile of honey (Paradkar and Irudayaraj, Two regression chemometrics was tested: 2002). NIR is presented as a potential tool (1) PLS and (2) Genetic‐algorithm‐based to identify adulterated honey with high inverse least squares (GILS) (Karaman et fructose syrups. Discriminant methods are al., 2009). Both showed similar accuracy to main chemometrics to classify unknown predict adulterant concentration, but honey samples into clusters on the basis of multispectral model can be constructing for similarities (Sivakesava and Irudayaraj, on-line applications. 2001). NIR spectral data in tandem with

Table 5 NIR and HIS to Honey authentication

Fraud control Instrument Chemometrics Statistical parameters Reference Radial basis function Accuracy = 94% (RBF) Floral origin VIS/NIR-HSI Minaei et al. (2017) SVM Accuracy = 93% Random forest Accuracy = 93% Escuredo et al. Floral origin NIR PLS RPD = 2.1 – 3.5 (2015) Botanical origin NIR PLS-DA Accuracy = 85 – 100% Gan et al. (2016) Correct classification LDA High fructose corn = 100% Ferreiro-González et VIS/NIR syrup R2p = 0.98 al. (2018) PLS RMSEP = 4.71% High-fructose corn Correct classification NIR CARS/PLS-DA Li et al. (2017) syrup = 86.3% Correct classification CARS/PLS-DA = 96.1% Maltose syrup NIR Li et al. (2017) R2p = 0.90 – 0.98 PLS RMSEP = 1.78 - 4.04% ANN Accuracy = 95% SVM Accuracy = 92% Fructose-glucose VIS/NIR - HSI LDA Accuracy = 90% Shafiee et al. (2016) Fisher Accuracy = 89% Parzen Accuracy = 84% High fructose corn NIR in transflectance R2cv = 0.98 PLS Bázár et al. (2016) syrup mode RMSECV = 1.48 Benchtop NIR PLS-DA Accuracy = 96.9% Glucose and Portable NIR PLS-DA Accuracy = 93.7% Guelpa et al. (2017) fructose Mobile NIR PLS-DA Accuracy = 87.5% Genetic‐algorithm‐ RMSEP= 0.90 – 2.19% based inverse least Beet sugar and R2p = 0.99 Başar and Özdemir NIR squares (GILS) corn syrup (2018) Partial least squares RMSEP= 1.18 – 2.89% (PLS) R2p = 0.97 – 0.99 Jaggery RMSEC= 0.0075 Kumaravelu and NIR PLS adulterants R2c = 0.99 Gopal (2015) RMSECV= 4.52% Mouazen and Al- Glusose NIR PLS R2cv = 0.85 Walaan (2014) RPD = 2.53

-153-

For detecting glucose and fructose in image with adulterant or adulterant/honey adulterated honey sample, more one of features. discriminant analysis was tested accurately. PLS-DA was selected as 3. Data fusion in food authentication chemometrics tool to build discriminant The data fusion strategy has allowed models for authenticity of South African obtaining, in many cases, more precise and honey based on prediction of glucose and conclusive results than using several fructose concentrations (Guelpa et al., techniques separately (Borràs et al., 2015; 2017). This work shows that the loss of Li et al., 2019). Data fusion can be done in precision is minimal compared to portable three levels (Figure 2) (Biancolillo et al., NIR with table NIR, which industrially is 2014): quite important. Similar results was showed (1) Low-level: a single matrix is created that for glucose (Mouazen and Al-Walaan, 2014) includes all the raw data of the analyzed and jaggery (Kumaravelu and Gopal, 2015) sources, as long as the data of the sensors adulterants in honey using NIR spectral are proportional (same row and column data in tandem with PLS regression. On the number), being observations of similar other hand, VIS/NIR-HSI showed a good physical quantities (Banerjee et al., 2016; Di performance to detect honey adulterated Rosa et al., 2017). with fructose and glucose using SVM, LDA (2) mid-level: the data obtained from each and ANN (Shafiee et al., 2016). Chemical sensor are analyzed separately and maps based on ANN model were built for relevant characteristics are extracted from visual identification of sample adulteration. each information block. Then, the However, the large amount of information information is joined in a single matrix to offered by HSI makes its industrial perform the multivariate analysis (Borràs et application difficult to implement. New al., 2015). works should be directed to identify the (3) high-level: the information is analyzed wavelengths important for the identification separately and a model is generated for of each adulterant. Thus, the imple- each block of data, and then, the responses mentation of HSI in the honey supply chain are combined for a final fused response is less costly and easier to manage. (Doeswijk et al., 2011). Perhaps the Honey can be unifloral or multifloral. majority vote method is the easiest and Unifloral honey is more expensive, due to most widely used in food analysis (Di Rosa its low production, therefore its et al., 2017). susceptibility to adulteration is higher. For this, there is a protected designation of origin (PDO) and protected geographical indication (PGI) for honey (Cajka et al., 2009), and its identification is economically important. For more information on denomination of origin regulations, diverse analytical methods for authentication and analysis of honey composition you can access the reviews of Pita-Calvo et al. (2017), Trifković et al. (2017), Soares et al. (2017), Naila et al. (2018), Wu et al. (2017), among others. Pollen types (Castanea, Eucalyptus, Rubus and Erica) related with honey floral (or botanical) origin was successfully determinate using NIR in tandem with PLS (Escuredo et al., 2015). Similar results (CCR=100%) for botanical origin was reported using NIR and iPLS (Gan et al., 2016) (Table 4). Later, VIS/NIR-HIS combined with radial basis (RBF), SVM and

Random forest showed similar accuracy for Figure 2. Data fusion approach for detecting food fraud. detect honey floral origin (Minaei et al., 2017). Since hyperspectral imaging is a Data fusion has extended in various areas combination of machine vision and such as intelligence system design (Hong et spectroscopy, the adulterant distribution al., 2003), image processing (Zhu and can be observed in sample, and their use Basir, 2006) and food analysis (included could be expanded to associate chemical authentication) (Ballabio et al., 2018;

-153-

J. Mendez et al. / Scientia Agropecuaria 10(1): 143 – 161 (2019)

Sanaeifar et al., 2018). Initially, the data of the PDO Chianti Classico olive oil. fusion strategy was based on two data Individually, NIR spectral data (pre-treated blocks (i.e. spectral and textural data by 1st derivate) presented the best obtained from hyperspectral imaging) performance with 100% for sensitivity and (Xiong et al., 2015), but now, the literature specificity. Great results were achieved reports researches based on fusion data (only 3 false positives) using data fusion using three (Borràs et al., 2016; Di Rosa et combining 5 variables (3 NIR + 1 UV-visible al., 2017), four (Erich et al., 2015) and even + 1 artificial nose) selected by Stepwise- five techniques (Biancolillo et al., 2014). In Linear Discriminant Analysis (STEP-LDA). this section we summarized and described In this specific case, NIR spectral data is the researches based on fusion data enough to classify correctly PDO Chianti strategy for food authentication using NIR Classico olive oil. Later, NIR and mid or HIS combined with other analytic infrared (MIR) spectroscopy were techniques. combined using data fusion (low-, mid- and Low-level data fusion showed good high-level) for the quantification of performance for classify black, green, rapeseed oil in olive oil blends (Li et al., white, yellow, dark, and oolong teas 2019). PLS regression models were (Dankowska and Kowalewski, 2019). For constructed using the three conceptions this purpose, UV–Vis, synchronous (low-, mid- and high-level) and using the fluorescence and NIR spectroscopies spectral data of each technique separately. (individually and fused) were combined with The lowest RMSEP (2.86) and highest R2p Linear Discriminant Analysis (LDA), (0.988) was obtained for high-level data Quadratic Discriminant Analysis (QDA), fusion strategy, being the most reliable Regularized Discriminant Analysis (RDA) technique for quantitative analysis. Similar and Support Vector Machine (SVM). Data results (accuracy 99% is test samples) fusion model based on NIR and UV-Vis in were found using high-level data fusion tandem with SVM presented the lowest based on Raman + NIR + Proton Transfer classification error (<1.4%) of tea samples. Reaction – Time of Flight – Mass However, it must be considered that low- Spectrometry (PTR-MS) in tandem with level data fusion requires a greater amount PLS-DA for identify and classify 8 Italian (time) calculation, and may not compensate honey botanical varieties (Ballabio et al., the non-essential variance obtained by 2018). One of the advantages of using high- adding the same blocks (Sun et al., 2017). level data fusion is its flexibility, since On the other hand, data fusion based on FT- classification models do not need to be Raman and NIR spectral (middle- and high- constructed using the same set of samples level) in tandem with Soft Independent (Callao and Ruisánchez, 2018). Modelling of Class Analogy (SIMCA) were Hyperspectral images have the advantage used for classify hazelnut paste adulterated of providing spectral and spatial with almond (Márquez et al., 2016). information, so the fusion of data consists Individually, FT-Raman and NIR showed of obtaining characteristics of the image sensitivity and specificity values between (i.e. texture) and combining it with the 75 – 100%. While, the data fusion in mid- spectral data. Vis-NIR-HIS is also able to level and high-level showed best differentiate between frozen and frozen- performances for sensitivity and specificity: thawed meat samples. Simultaneously, two 96 – 100% and 88 – 100%, respectively. Mid- investigations showed the ability of Vis-NIR- level data fusion based on NIR data HIS to classify between fresh and frozen- combined with high-resolution mass thawed pork Longissimus Dorsi muscles by spectrometry also obtained good combining spectral information and textural performance to classify sulfur-fumigated features (Table 3). Pu et al. (2015) used six Chinese herb using (Dai et al., 2018). features wavelengths (400, 446, 477, 516, Similar results (100% correct classification 592 and 686 nm) and textural features in test samples) were found using mid-level obtained by histogram statistics (HS), gray data fusion based on NIR and MIR spectral level co-occurrence matrix (GLCM) and data for identification of rhubarb (Sun et al., gray level-gradient co-occurrence matrix 2017). The mid-level data fusion is more (GLGCM). The selected wavelengths and efficient than low-level data fusion, since by textural features obtained by GLGCM were previously selecting the relevant variables integrated for in a probabilistic neural of each sensor, the calculation time is network (PNN) model (classification rate = reduced. 90.91 - 93.14%). For its part, Ma et al. (2015) Forina et al. (2015) used data fusion used eight important wavelengths (624, strategy to combine artificial nose, NIR and 673, 460, 588, 583, 448, 552 and 609 nm) UV–visible spectroscopy for authentication and 45 textural features obtained by

-154-

J. Mendez et al. / Scientia Agropecuaria 10(1): 143 – 161 (2019)

GLGCM to obtain the best correct Previous discriminant/regression model classification (97.73%) based on PLS-DA. established, exploratory analysis is com- On the other hand, gray level-gradient co- monly performed. Among the unsupervised occurrence matrix (GLGCM) selected 35 analyzes, principal component analysis textural features, later they were fusion (PCA) has been the most used for analysis with HIS spectral data to differentiate of spectral data NIR or HSI for food successfully (93.3% correct classification) authentication purposes (See Tables 1 – 5). between free-range and broiler chicken PCA is a technique that reduces meats (Xiong et al., 2015). These studies information by creating some new variables suggest that the fusion of spectral called principal components (PC) from a information and textural characteristics of linear combination of the original variables hyperspectral images can improve the (Borràs et al., 2015). PCA allows to observe discrimination capacity of classification if the spectral characteristics allow the models. grouping or separation into groups of samples with specific characteristics 4. Chemometrics in food authentication defined previously (i.e. geographical Chemometrics methods are an analytical origin). If there are no differences between strategy to analyze the spectral information the groups of samples analyzed, it should and generate mathematical models to be considered to reassess the objectives of address the problems related to food fraud the work or verify the information referring (Callao and Ruisánchez, 2018). NIR and HSI to the response variable with which the provide a large amount of complex model is being fed. information, which often cannot be used in After, discriminant models or regression its gross form to generate discriminant / model are established to identify food regression models (Pasquini, 2018). In fraud. There is a great diversity of addition, spectroscopic devices can discriminant and regression methods that generate information with defects such as allow generating mathematical models to noise, edge effect, light scattering and base detect fraud. A classification or prediction line shift (Amigo et al., 2015). First, spectral algorithm is not exclusive to a group of data data should be corrected for improve or to a specific case of fraud. model performance. The most used In several cases, research on food espectral pre-processes are: (1) multipli- authentication are driving to build only cative scatter correction (MSC) and (2) classification models. The main standard normal variate (SNV) for reducing classification supervised techniques are: the spectral variability caused by partial least squares discriminant analysis scattering effects (Rinnan et al., 2009), and (PLS-DA), linear discriminant analysis (3) Norris-Williams (NW) derivates and (LDA) (Vandeginste et al., 1998) and Savitzky-Golay polynomial derivative filters quadratic discriminant analysis (QDA); k for smoothing spectra, removing variations nearest neighbours (KNN) (Callao and in baseline and resolution of overlapping Ruisánchez, 2018), soft independent peaks is used Norris-Williams (NW) modelling of class analogy (SIMCA) derivates and Savitzky-Golay polynomial (Moseholm, 1988), support vector machines derivative filters (Brown et al., 2000). This (SVM) (Sliwinska et al., 2014), random pre-processing spectral tools were forest (Xu et al., 2017) and artificial neural successfully applied for treating NIR and networks (Mu et al., 2016). These models HSI spectral data in order to generate are used to establish specific classes discriminant/regression models for food based on the similarities and differences of authentication. However, care must be the samples analyzed. Then, an external taken with the application of pre-process sample can be classified or not in a certain algorithms, since vital information can be class. The choice of classifier happens not compromised. Therefore, it is important only for having a high precision, but also for that the analyst knows the conditions of the reducing the computational time of sample and can determine if the pre- analysis. Many of the models studied in this process has not compromised the analysis. review have been created using the full In addition, some works show, erroneously, spectrum. However, it is necessary that combinations of pre-process algorithms new studies be directed to evaluate that were designed to correct the same methods of selection of variables in order spectral defect (as SNV and MSC). This to establish more precise models. places multidisciplinary research as an Between regression algorithms, partial essential pillar that allows us to omit basic least square (PLS) is the most popular errors due to lack of knowledge of a supervised techniques used to build specific technique. regression models based on spectral data

-155-

J. Mendez et al. / Scientia Agropecuaria 10(1): 143 – 161 (2019) for food identification (see Table 1 – 5). PLS system. Thinking about the final consumer, goal is to analyze or predict a set of the NIR and HSI devices must be dependent variables from a set of implemented in smartphone (Kartakoullis et predictors (independent variable) which al., 2019), with a friendly language and an are rotated to ensure maximum correlation interface that is easy to apply. In addition, efficiency (Alamprese et al., 2013). devices should respond to different food Therefore, PLS generates a mathematical groups. For this, it is important to build model that correlates linearly the spectral these equipment using important wavelen- variables with the variable of interest (i.e. gths for faster and more accurate analysis. adulterant concentration) (Alamprese et al., 2016). These linear combinations are called 6. Conclusions latent variables, and they have a great The application of NIR and HSI for food predictive capacity, especially the first authentication has gained more interest in latent variables (Kumar and Chandrakant- recent years. Both techniques require a Karne, 2017). It is important, especially for minimum or no sample preparation, are HSI applications, that PLS models be cons- non-invasive, do not use reagents ("green tructed using only significant wavelengths. analytical method") and are quite reliable. The spectral information of NIR and HSI 5. Future trends combined with chemometrics are generally This review summarizes and analyzes the sufficient to create mathematical models to latest advances in the use of NIR and HSI identify geographic origin, farming system, for food authentication. In the case of NIR, genetic variety and adulterated samples. although it seems to be a fairly mature The ease of obtaining NIR data allows it to technique (Pasquini, 2018), it can still be be the technique with the greatest explored for new foods and new forms of industrial proximity. Although there are still fraud. For its part, HSI is a technique still in challenges for its implementation in large- exploration for food authentication scale production lines. For its part, HSI can purposes. Although HSI can have the overcome the limitations of the advantage of incorporating spatial heterogeneity of the sample and for certain information, processing that information is cases obtain a greater amount of valuable very tiring, time consuming and impractical information to detect food fraud. Most for industrial applications. Future works papers report models based on HSI based on HSI, but also on NIR, must information using the full spectrum. What generate multispectral models that allow would not be industrially viable, therefore, the technology to be explored at an multispectral models are usually a very industrial level. Multispectral models are useful solution. When both techniques, NIR constructed with the appropriate selection and HSI, do not reach the expected results of variables, therefore, new methods in the individually, the data fusion strategy is selection of variables should be explored. shown as a promising alternative. The On the other hand, although the developed portability of the equipment and the remote models based on NIR and HSI may have access via Wi-Fi or Bluetooth will allow outstanding results, they may not be fraud control to be carried out throughout conclusive. Some limitations in the the production chain. In addition, network applications of both techniques have been connections will allow NIR and HSI devices overcome using data fusion. The fusion of to record information on the network, data, at a certain level, allows to obtain becoming part of the "Internet of things". more robust and precise models. Perhaps Finally, the studies registered in this review an important challenge in the next works is are mostly laboratory-level applications, to establish new ways to merge groups of therefore, it is very likely that with the data to obtain more precise results. continuous development of technology, Regarding the manageability of NIR and more manageable and precise NIR and HSI HSI devices. It is expected that in the not devices allow industrial studies. too distant future, smaller portable devices will be available. These devices must be ORCID J.P Cruz-Tirado i D https://orcid.org/0000-0002-1963-4965 manageable, cheap, robust and easy to R. Quevedo i D http://orcid.org/0000-0001-8132-838X apply (Crocombe, 2018). The loss of R. Siche i D https://orcid.org/0000-0003-3500-4928 precision in portable devices is a topic to be addressed by the scientific community. References Therefore, it is necessary to publish more research using portable devices, Abbas, O.; Zadravec, M.; Baeten, V.; Mikuš, T.; Lešić, T.; Vulić, A.; Prpić, J.; Jemeršić, L.; Pleadin, J. 2018. considering the influence of temperature Analytical methods used for the authentication of and the movement of the sample in on-line food of animal origin. Food Chemistry 246: 6-17.

-156-

J. Mendez et al. / Scientia Agropecuaria 10(1): 143 – 161 (2019)

Alamprese, C.; Amigo, J. M.; Casiraghi, E.; Engelsen, S. characterization of an italian craft beer aimed at its B. 2016. Identification and quantification of turkey authentication. Analytica Chimica Acta 820: 23-31. meat adulteration in fresh, frozen-thawed and Binetti, G.; Del Coco, L.; Ragone, R.; Zelasco, S.; Perri, cooked minced beef by FT-NIR spectroscopy and E.; Montemurro, C.; Valentini, R.; Naso, D.; Fanizzi, chemometrics. Meat Science 121: 175-181. F.P.; Schena, F.P. 2017. Cultivar classification of Alamprese, C.; Casale, M.; Sinelli, N.; Lanteri, S.; Apulian olive oils: Use of artificial neural networks Casiraghi, E. 2013. Detection of minced beef for comparing NMR, NIR and merceological data. adulteration with turkey meat by UV–vis, NIR and Food Chemistry 219: 131-138. MIR spectroscopy. LWT - Food Science and Bona, E.; Marquetti, I.; Link, J.V.; Makimori, G.Y.F.; da Technology 53: 225-232. Costa Arca, V.; Guimarães Lemes, A.L.; Ferreira, J. Amaral, J.; Meira, L.; Oliveira, M. B. P. P.; Mafra, I. 2016. M.G.; dos Santos Scholz, M. B.; Valderrama, P.; 14 - Advances in Authenticity Testing for Meat Poppi, R.J. 2017. Support vector machines in Speciation. In G. Downey (Ed.), Advances in Food tandem with infrared spectroscopy for Authenticity Testing. Woodhead Publishing. Pp. 369 geographical classification of green arabica coffee. - 414. LWT - Food Science and Technology 76: 330-336. Amigo, J.M.; Babamoradi, H.; Elcoroaristizabal, S. 2015. Borràs, E.; Ferré, J.; Boqué, R.; Mestres, M.; Aceña, L.; Hyperspectral image analysis. A tutorial. Analytica Busto, O. 2015. Data fusion methodologies for food Chimica Acta 896: 34-51. and beverage authentication and quality Amigo, J. M.; Martí, I.; Gowen, A. 2013. Hyperspectral assessment – A review. Analytica Chimica Acta 891: imaging and chemometrics: a perfect combination 1-14. for the analysis of food structure, composition and Borràs, E.; Ferré, J.; Boqué, R.; Mestres, M.; Aceña, L.; quality. In data handling in science and technology. Calvo, A.; Busto, O. 2016. Olive oil sensory defects Elsevier. Pp. 343 - 370. classification with data fusion of instrumental Aredo, V.; Velásquez, L.; Siche, R. 2017. Prediction of techniques and multivariate analysis (PLS-DA). beef marbling using Hyperspectral Imaging (HSI) Food Chemistry 203: 314-322. and Partial Least Squares Regression (PLSR). Boyacı, I.H.; Temiz, H.T.; Uysal, R.S.; Velioğlu, H.M.; Scientia Agropecuaria 8: 169-174. Yadegari, R.J.; Rishkan, M.M. 2014. A novel method Azizian, H.; Mossoba, M.M.; Fardin-Kia, A.R.; for discrimination of beef and horsemeat using Karunathilaka, S.R.; Kramer, J.K.G. 2016. Raman spectroscopy. Food chemistry 148: 37-41. Developing FT-NIR and PLS1 Methodology for Branigan, T. 2008. In the Guardian, Guardian News and Predicting Adulteration in Representative Media, London. Varieties/Blends of Extra Virgin Olive Oils. Lipids 51: Brown, C.D.; Vega-Montoto, L.; Wentzell, P.D. 2000. 1309-1321. Derivative Preprocessing and Optimal Corrections Ballabio, D.; Robotti, E.; Grisoni, F.; Quasso, F.; Bobba, for Baseline Drift in Multivariate Calibration. Applied M.; Vercelli, S.; Gosetti, F.; Calabrese, G.; Spectroscopy 54: 1055-1068. Sangiorgi, E.; Orlandi, M.; Marengo, E. 2018. Bázár, G.; Romvári, R.; Szabó, A.; Somogyi, T.; Éles, V.; Chemical profiling and multivariate data fusion Tsenkova, R. 2016. NIR detection of honey methods for the identification of the botanical origin adulteration reveals differences in water spectral of honey. Food Chemistry 266: 79-89. pattern. Food chemistry 194: 873-880. Banerjee, R.; Tudu, B.; Bandyopadhyay, R.; Böhme, K.; Calo-Mata, P.; Barros-Velázquez, J.; Ortea, I. Bhattacharyya, N. 2016. A review on combined odor 2019. Recent applications of omics-based and taste sensor systems. Journal of Food technologies to main topics in food authentication. Engineering 190: 10-21. TrAC Trends in Analytical Chemistry 110: 221-232. Bao, Y.-d.; Chen, N.; He, Y.; Liu, F.; Zhang, C.; Kong, W.- Cajka, T.; Hajslova, J.; Pudil, F.; Riddellova, K. 2009. w. 2015. Rapid identification of coffee bean variety Traceability of honey origin based on volatiles by near infrared hyperspectral imaging technology. pattern processing by artificial neural networks. Optics and Precision Engineering 23: 349-355. Journal of Chromatography A 1216: 1458-1462. Barbin, D.F.; Felicio, A.L.d.S.M.; Sun, D.-W.; Nixdorf, Callao, M.P.; Ruisánchez, I. 2018. An overview of S.L.; Hirooka, E.Y. 2014. Application of infrared multivariate qualitative methods for food fraud spectral techniques on quality and compositional detection. Food Control 86: 283-293. attributes of coffee: An overview. Food Research Calvini, R.; Amigo, J.M.; Ulrici, A. 2017. Transferring International 61: 23-32. results from NIR-hyperspectral to NIR-multispectral Barbin, D.F.; Kaminishikawahara, C.M.; Soares, A.L.; imaging systems: A filter-based simulation applied Mizubuti, I.Y.; Grespan, M.; Shimokomaki, M.; to the classification of Arabica and Robusta green Hirooka, E.Y. 2015. Prediction of chicken quality coffee. Analytica Chimica Acta 967: 33-41. attributes by near infrared spectroscopy. Food Calvini, R.; Ulrici, A.; Amigo, J.M. 2015. Practical Chemistry 168: 554-560. comparison of sparse methods for classification of Barreto, A.; Cruz-Tirado, J. P.; Siche, R.; Quevedo, R. Arabica and Robusta coffee species using near 2018. Determination of starch content in infrared hyperspectral imaging. Chemometrics and adulterated fresh cheese using hyperspectral Intelligent Laboratory Systems 146: 503-511. imaging. Food Bioscience 21: 14-19. Caporaso, N.; Whitworth, M. B.; Grebby, S.; Fisk, I. D. Basri, K.N.; Hussain, M.N.; Bakar, J.; Sharif, Z.; Khir, 2018. Rapid prediction of single green coffee bean M.F.A.; Zoolfakar, A.S. 2017. Classification and moisture and lipid content by hyperspectral quantification of palm oil adulteration via portable imaging. Journal of Food Engineering 227: 18-29. NIR spectroscopy. Spectrochimica Acta Part A: Chen, H.; Lin, Z.; Tan, C. 2018. Fast quantitative Molecular and Biomolecular Spectroscopy 173: detection of sesame oil adulteration by near- 335-342. infrared spectroscopy and chemometric models. Başar, B.; Özdemir, D. 2018. Determination of honey Vibrational Spectroscopy 99: 178-183. adulteration with beet sugar and corn syrup using Chen, H.; Tan, C.; Lin, Z. 2018. Authenticity Detection of infrared spectroscopy and genetic-algorithm-based Black Rice by Near-Infrared Spectroscopy and multivariate calibration. Journal of the Science of Support Vector Data Description. International Food and Agriculture 98: 5616-5624. journal of analytical chemistry. Bertone, E.; Venturello, A.; Giraudo, A.; Pellegrino, G.; Cheng, J.-H.; Nicolai, B.; Sun, D.-W. 2017. Geobaldo, F. 2016. Simultaneous determination by Hyperspectral imaging with multivariate analysis for NIR spectroscopy of the roasting degree and technological parameters prediction and Arabica/Robusta ratio in roasted and ground coffee. classification of muscle foods: A review. Meat Food Control 59: 683-689. Science 123: 182-191. Biancolillo, A.; Bucci, R.; Magrì, A. L.; Magrì, A. D.; Cheng, J.-H.; Sun, D.-W.; Pu, H.-B.; Chen, X.; Liu, Y.; Marini, F. 2014. Data-fusion for multiplatform Zhang, H.; Li, J.-L. 2015. Integration of classifiers analysis and hyperspectral imaging for rapid

-157-

J. Mendez et al. / Scientia Agropecuaria 10(1): 143 – 161 (2019)

discrimination of fresh from cold-stored and frozen- linear discriminant analysis for authentication of thawed fish fillets. Journal of Food Engineering 161: food products. Food Control 91: 100-112. 33-39. Esteki, M.; Simal-Gandara, J.; Shahsavari, Z.; Zandbaaf, Chiesa, L.; Panseri, S.; Bonacci, S.; Procopio, A.; S.; Dashtaki, E.; Vander Heyden, Y. 2018. A review Zecconi, A.; Arioli, F.; Cuevas, F. J.; Moreno-Rojas, on the application of chromatographic methods, J. M. 2016. Authentication of Italian PDO lard using coupled to chemometrics, for food authentication. NIR spectroscopy, volatile profile and fatty acid Food Control 93: 165-182. composition combined with chemometrics. Food Ferreiro-González, M.; Espada-Bellido, E.; Guillén- Chemistry 212: 296-304. Cueto, L.; Palma, M.; Barroso, C. G.; Barbero, G. F. Correia, R.M.; Tosato, F.; Domingos, E.; Rodrigues, 2018. Rapid quantification of honey adulteration by R.R.T.; Aquino, L.F.M.; Filgueiras, P.R.; Lacerda, V.; visible-near infrared spectroscopy combined with Romão, W. 2018. Portable near infrared chemometrics. Talanta 188: 288-292. spectroscopy applied to quality control of Brazilian Forina, M.; Oliveri, P.; Bagnasco, L.; Simonetti, R.; coffee. Talanta 176: 59-68. Casolino, M.C.; Nizzi Grifi, F.; Casale, M. 2015. Cozzolino, D. 2016. 16 - Authentication of Cereals and Artificial nose, NIR and UV–visible spectroscopy for Cereal Products. In G. Downey (Ed.), Advances in the characterisation of the PDO Chianti Classico Food Authenticity Testing. Woodhead Publishing. olive oil. Talanta 144: 1070-1078. Pp. 441 - 457. Gan, Z.; Yang, Y.; Li, J.; Wen, X.; Zhu, M.; Jiang, Y.; Ni, Y. Crocombe, R.A. 2018. Portable Spectroscopy. Applied 2016. Using sensor and spectral analysis to classify Spectroscopy 72: 1701-1751. botanical origin and determine adulteration of raw Dai, S.; Lin, Z.; Xu, B.; Wang, Y.; Shi, X.; Qiao, Y.; Zhang, honey. Journal of Food Engineering 178: 151-158. J. 2018. Metabolomics data fusion between near Giraudo, A.; Grassi, S.; Savorani, F.; Gavoci, G.; infrared spectroscopy and high-resolution mass Casiraghi, E.; Geobaldo, F. 2019. Determination of spectrometry: A synergetic approach to boost the geographical origin of green coffee beans using performance or induce confusion. Talanta 189: 641- NIR spectroscopy and multivariate data analysis. 648. Food Control 99: 137-145. Danezis, G. P.; Tsagkaris, A. S.; Brusic, V.; Georgiou, C. Gossner, C.M.-E.; Schlundt, J.; Embarek, P.B.; Hird, S.; A. 2016. Food authentication: state of the art and Lo-Fo-Wong, D.; Beltran, J.J.O.; Teoh, K.N.; prospects. Current Opinion in Food Science 10: 22- Tritscher, A. 2009. The melamine incident: 31. implications for international food and feed safety. Dang, T.T.; Vasanthan, T. 2019. Modification of rice bran Environmental health perspectives 117: 1803. dietary fiber concentrates using enzyme and Guelpa, A.; Marini, F.; du Plessis, A.; Slabbert, R.; extrusion cooking. Food Hydrocolloids 89: 773-782. Manley, M. 2017. Verification of authenticity and Dankowska, A.; Kowalewski, W. 2019. Tea types fraud detection in South African honey using NIR classification with data fusion of UV–Vis, spectroscopy. Food Control 73: 1388-1396. synchronous fluorescence and NIR spectroscopies Hong, B.; Jun, L.; Xian-Yun, Z.; Jing, Z. 2003. A fire and chemometric analysis. Spectrochimica Acta detection system based on intelligent data fusion Part A: Molecular and Biomolecular Spectroscopy technology. In Proceedings of the 2003 211: 195-202. International Conference on Machine Learning and da Silva, I.J.S.; Paim, A.P.S.; da Silva, M.J. 2018. Cybernetics (IEEE Cat. No.03EX693) 2: 1096-1101. Composition and estimate of daily mineral intake Jiménez-Carvelo, A.M.; Lozano, V.A.; Olivieri, A.C. 2019. from samples of Brazilian rice. Microchemical Comparative chemometric analysis of fluorescence Journal 137: 131-138. and near infrared spectroscopies for authenticity de la Fuente, E.; Sanz, M.L.; Martínez-Castro, I.; Sanz, J. confirmation and geographical origin of 2006. Development of a robust method for the Argentinean extra virgin olive oils. Food Control 96: quantitative determination of disaccharides in 22-28. honey by gas chromatography. Journal of Kamruzzaman, M.; Makino, Y.; Oshita, S. 2016. Chromatography A 1135: 212-218. Parsimonious model development for real-time De Luca, S.; De Filippis, M.; Bucci, R.; Magrì, A.D.; monitoring of moisture in red meat using Magrì, A.L.; Marini, F. 2016. Characterization of the hyperspectral imaging. Food Chemistry 196: 1084- effects of different roasting conditions on coffee 1091. samples of different geographical origins by HPLC- Karaman, I.; Tanriverdi, H.; Özdemir, D. 2009. Prediction DAD, NIR and chemometrics. Microchemical of Lignin and Extractive Content of Pinus nigra Journal 129: 348-361. Arnold. var. Pallasiana Tree Using Near Infrared Di Rosa, A.R., Leone, F.; Cheli, F.; Chiofalo, V. 2017. Spectroscopy and Multivariate Calibration AU - Fusion of electronic nose, electronic tongue and Uner, B. Journal of Wood Chemistry and computer vision for animal source food Technology 29: 24-42. authentication and quality assessment – A review. Kartakoullis, A.; Comaposada, J.; Cruz-Carrión, A.; Journal of Food Engineering 210: 62-75. Serra, X.; Gou, P. 2019. Feasibility study of Doeswijk, T. G.; Smilde, A. K.; Hageman, J. A.; smartphone-based Near Infrared Spectroscopy Westerhuis, J. A.; van Eeuwijk, F. A. 2011. On the (NIRS) for salted minced meat composition increase of predictive performance with high-level diagnostics at different temperatures. Food data fusion. Analytica Chimica Acta 705: 41-47. Chemistry 278: 314-321. Erich, S.; Schill, S.; Annweiler, E.; Waiblinger, H.-U.; Karunathilaka, S. R.; Kia, A.-R. F.; Srigley, C.; Chung, J. Kuballa, T.; Lachenmeier, D.W.; Monakhova, Y.B. K.; Mossoba, M. M. 2016. Nontargeted, Rapid 2015. Combined chemometric analysis of 1H NMR, Screening of Extra Virgin Olive Oil Products for 13C NMR and stable isotope data to differentiate Authenticity Using Near-Infrared Spectroscopy in organic and conventional milk. Food Chemistry 188: Combination with Conformity Index and Multivariate 1-7. Statistical Analyses. Journal of Food Science 81: Erkinbaev, C.; Henderson, K.; Paliwal, J. 2017. C2390-C2397. Discrimination of gluten-free oats from Kennedy, J.; Delaney, L.; McGloin, A.; Wall, P.G. 2009. contaminants using near infrared hyperspectral Public perceptions of the dioxin crisis in Irish pork. imaging technique. Food Control 80: 197-203. Kumar, Y.; Chandrakant-Karne, S. 2017. Spectral Escuredo, O.; González-Martín, M. I.; Rodríguez-Flores, analysis: A rapid tool for species detection in meat M. S.; Seijo, M. C. 2015. Near infrared spectroscopy products. Trends in Food Science & Technology 62: applied to the rapid prediction of the floral origin 59-67. and mineral content of honeys. Food Chemistry 170: Kumaravelu, C.; Gopal, A. 2015. Detection and 47-54. quantification of adulteration in honey through near Esteki, M.; Shahsavari, Z.; Simal-Gandara, J. 2018. Use infrared spectroscopy. International Journal of of spectroscopic methods in combination with Food Properties 18: 1930-1935.

-158-

J. Mendez et al. / Scientia Agropecuaria 10(1): 143 – 161 (2019)

Laroussi-Mezghani, S.; Vanloot, P.; Molinet, J.; Dupuy, coffee: A preliminary chemometric study. Food N.; Hammami, M.; Grati-Kamoun, N.; Artaud, J. Control 91: 276-283. 2015. Authentication of Tunisian virgin olive oils by Moran, L.; Andres, S.; Allen, P.; Moloney, A. P. 2018. chemometric analysis of fatty acid compositions Visible and near infrared spectroscopy as an and NIR spectra. Comparison with Maghrebian and authentication tool: Preliminary investigation of the French virgin olive oils. Food Chemistry 173: 122- prediction of the ageing time of beef steaks. Meat 132. Science 142: 52-58. Li, H.; Liang, Y.; Xu, Q.; Cao, D. 2009. Key wavelengths Moseholm, L. 1988. Analysis of air pollution plant screening using competitive adaptive reweighted exposure data: the soft independent modelling of sampling method for multivariate calibration. class analogy (SIMCA) and partial least squares Analytica Chimica Acta 648: 77-84. modelling with latent variable (PLS) approaches. Li, S.; Zhang, X.; Shan, Y.; Su, D.; Ma, Q.; Wen, R.; Li, J. Environmental Pollution 53: 313-331. 2017. Qualitative and quantitative detection of Mossoba, M.M.; Azizian, H.; Fardin-Kia, A.R.; honey adulterated with high-fructose corn syrup Karunathilaka, S.R.; Kramer, J.K.G. 2017. First and maltose syrup by using near-infrared Application of Newly Developed FT-NIR spectroscopy. Food Chemistry 218: 231-236. Spectroscopic Methodology to Predict Authenticity Li, Y.; Xiong, Y.; Min, S. 2019. Data fusion strategy in of Extra Virgin Olive Oil Retail Products in the USA. quantitative analysis of spectroscopy relevant to Lipids 52: 443-455. olive oil adulteration. Vibrational Spectroscopy 101: Mouazen, A. M.; Al-Walaan, N. 2014. Glucose 20-27. adulteration in Saudi honey with visible and near Liu, Y.; Ma, D.-h.; Wang, X.-c.; Liu, L.-p.; Fan, Y.-x.; Cao, infrared spectroscopy. International journal of food J.-x. 2015. Prediction of chemical composition and properties 17: 2263-2274. geographical origin traceability of Chinese export Mu, T.; Chen, S.; Zhang, Y.; Chen, H.; Guo, P.; Meng, F. tilapia fillets products by near infrared reflectance 2016. Portable Detection and Quantification of Olive spectroscopy. LWT - Food Science and Technology Oil Adulteration by 473-nm Laser-Induced 60: 1214-1218. Fluorescence. Food Analytical Methods 9: 275-279. Luna, A.S.; da Silva, A.P.; Alves, E.A.; Rocha, R.B.; Lima, Murniece, I.; Straumite, E. 2014. The information I.C.A.; de Gois, J.S. 2017. Evaluation of presented on labels for bread produced in Latvia. chemometric methodologies for the classification of Food Chemistry 162: 117-121. Coffea canephora cultivars via FT-NIR Márquez, C.; López, M.I.; Ruisánchez, I.; Callao, M.P. spectroscopy and direct sample analysis. Analytical 2016. FT-Raman and NIR spectroscopy data fusion Methods 9: 4255-4260. strategy for multivariate qualitative analysis of food López-Maestresalas, A.; Insausti, K.; Jarén, C.; Pérez- fraud. Talanta 161: 80-86. Roncal, C.; Urrutia, O.; Beriain, M.J.; Arazuri, S. Naila, A.; Flint, S.H.; Sulaiman, A.Z.; Ajit, A.; Weeds, Z. 2019. Detection of minced lamb and beef fraud 2018. Classical and novel approaches to the using NIR spectroscopy. Food Control 98: 465-473. analysis of honey and detection of adulterants. Food Ma, J.; Pu, H.; Sun, D.-W.; Gao, W.; Qu, J.-H.; Ma, K.-Y. Control 90: 152-165. 2015. Application of Vis–NIR hyperspectral imaging Nakyinsige, K.; Man, Y.B.C.; Sazili, A.Q. 2012. Halal in classification between fresh and frozen-thawed authenticity issues in meat and meat products. Meat pork Longissimus Dorsi muscles. International science 91: 207-214. Journal of Refrigeration 50: 10-18. Nenadis, N.; Tsimidou, M.Z. 2017. Perspective of Maione, C.; Barbosa, R.M. 2018. Recent applications of vibrational spectroscopy analytical methods in on‐ multivariate data analysis methods in the field/official control of olives and virgin olive oil. authentication of rice and the most analyzed European Journal of Lipid Science and Technology parameters: A review. Critical reviews in food 119: 1600148. science and nutrition 1-12. Nolasco, I.M.; Badaró, A.T.; Barbon, S.; Barbon, Marquetti, I.; Link, J.V.; Lemes, A.L.G.; Scholz, M.B.d.S.; A.P.A.C.; Pollonio, M.A.R.; Barbin, D.F. 2018. Valderrama, P.; Bona, E. 2016. Partial least square Classification of Chicken Parts Using a Portable with discriminant analysis and near infrared Near-Infrared (NIR) Spectrophotometer and spectroscopy for evaluation of geographic and Machine Learning. Applied Spectroscopy 72: 1774- genotypic origin of arabica coffee. Computers and 1780. Electronics in Agriculture 121: 313-319. Ortea, I.; O'Connor, G.; Maquet, A. 2016. Review on Mendes, T.O.; da Rocha, R.A.; Porto, B.L.S.; de Oliveira, proteomics for food authentication. Journal of M.A.L.; dos Anjos, V.d.C.; Bell, M.J.V. 2015. Proteomics 147: 212-225. Quantification of Extra-virgin Olive Oil Adulteration Ouchemoukh, S.; Louaileche, H.; Schweitzer, P. 2007. with Soybean Oil: a Comparative Study of NIR, MIR, Physicochemical characteristics and pollen and Raman Spectroscopy Associated with spectrum of some Algerian honeys. Food Control Chemometric Approaches. Food Analytical Methods 18: 52-58. 8: 2339-2346. Ozen, B. F.; Mauer, L. J. 2002. Detection of hazelnut oil Minaei, S.; Shafiee, S.; Polder, G.; Moghadam-Charkari, adulteration using FT-IR spectroscopy. Journal of N.; van Ruth, S.; Barzegar, M.; Zahiri, J.; Alewijn, M.; agricultural and food chemistry 50: 3898-3901. Kuś, P. M. 2017. VIS/NIR imaging application for Paradkar, M.M.; Irudayaraj, J. 2002. Discrimination and honey floral origin determination. Infrared Physics classification of beet and cane inverts in honey by & Technology 86: 218-225. FT-Raman spectroscopy. Food Chemistry 76: 231- Mo, C.; Lim, J.; Kwon, S.W.; Lim, D.K.; Kim, M. S.; Kim, 239. G.; Kang, J.; Kwon, K.-D.; Cho, B.-K. 2017. Pasquini, C. 2018. Near infrared spectroscopy: A Hyperspectral imaging and partial least square mature analytical technique with new perspectives discriminant analysis for geographical origin – A review. Analytica Chimica Acta 1026: 8-36. discrimination of white rice. Journal of Biosystems Peršurić, Ž.; Saftić, L.; Mašek, T.; Kraljević Pavelić, S. Engineering 42: 293-300. 2018. Comparison of triacylglycerol analysis by Monahan, F.J.; Schmidt, O.; Moloney, A.P. 2018. Meat MALDI-TOF/MS, fatty acid analysis by GC-MS and provenance: Authentication of geographical origin non-selective analysis by NIRS in combination with and dietary background of meat. Meat Science 144: chemometrics for determination of extra virgin olive 2-14. oil geographical origin. A case study. LWT - Food Monteiro, P.I.; Santos, J.S.; Alvarenga Brizola, V.R.; Science and Technology 95: 326-332. Pasini Deolindo, C.T.; Koot, A.; Boerrigter-Eenling, Picouet, P.A.; Gou, P.; Hyypiö, R.; Castellari, M. 2018. R.; van Ruth, S.; Georgouli, K.; Koidis, A.; Granato, Implementation of NIR technology for at-line rapid D. 2018. Comparison between proton transfer detection of sunflower oil adulterated with mineral reaction mass spectrometry and near infrared oil. Journal of Food Engineering 230: 18-27. spectroscopy for the authentication of Brazilian

-159-

J. Mendez et al. / Scientia Agropecuaria 10(1): 143 – 161 (2019)

Pieszczek, L.; Czarnik-Matusewicz, H.; Daszykowski, M. Soares, S.; Amaral, J.S.; Oliveira, M.B.P.P.; Mafra, I. 2018. Identification of ground meat species using 2017. A Comprehensive Review on the Main Honey near-infrared spectroscopy and class modeling Authentication Issues: Production and Origin. techniques – Aspects of optimization and validation Comprehensive Reviews in Food Science and Food using a one-class classification model. Meat Safety 16: 1072-1100. Science 139: 15-24. Su, W.-H.; Sun, D.-W. 2016. Facilitated wavelength Pita-Calvo, C.; Guerra-Rodríguez, M.E.; Vázquez, M. selection and model development for rapid 2017. Analytical Methods Used in the Quality determination of the purity of organic spelt Control of Honey. Journal of Agricultural and Food (Triticum spelta L.) flour using spectral imaging. Chemistry 65: 690-703. Talanta 155: 347-357. Pu, H.; Sun, D.-W.; Ma, J.; Cheng, J.-H. 2015. Su, W.-H.; Sun, D.-W. 2017. Evaluation of spectral Classification of fresh and frozen-thawed pork imaging for inspection of adulterants in terms of muscles using visible and near infrared common wheat flour, cassava flour and corn flour in hyperspectral imaging and textural analysis. Meat organic Avatar wheat (Triticum spp.) flour. Journal Science 99: 81-88. of Food Engineering 200: 59-69. Rady, A.; Adedeji, A. 2018. Assessing different Sun, J.; Lu, X.; Mao, H.; Jin, X.; Wu, X. 2017. A Method for processed meats for adulterants using visible-near- Rapid Identification of Rice Origin by Hyperspectral infrared spectroscopy. Meat Science 136: 59-67. Imaging Technology. Journal of Food Process Ravikanth, L.; Singh, C.B.; Jayas, D.S.; White, N.D.G. Engineering 40: e12297. 2015. Classification of contaminants from wheat Sun, W.; Zhang, X.; Zhang, Z.; Zhu, R. 2017. Data fusion using near-infrared hyperspectral imaging. of near-infrared and mid-infrared spectra for Biosystems Engineering 135: 73-86. identification of rhubarb. Spectrochimica Acta Part Revilla, I.; Lastras, C.; González-Martín, M.I.; Vivar- A: Molecular and Biomolecular Spectroscopy 171: Quintana, A.M.; Morales-Corts, R.; Gómez-Sánchez, 72-79. M.A.; Pérez-Sánchez, R. 2019. Predicting the Thorburn-Burns, D.; Tweed, L.; Walker, M.J. 2017. physicochemical properties and geographical Ground Roast Coffee: Review of Analytical ORIGIN of lentils using near infrared spectroscopy. Strategies to Estimate Geographic Origin, Species Journal of Food Composition and Analysis 77: 84- Authenticity and Adulteration by Dilution. Food 90. Analytical Methods 10: 2302-2310. Rinnan, Å.; Berg, F.v.d.; Engelsen, S.B. 2009. Review of Trifković, J.; Andrić, F.; Ristivojević, P.; Guzelmeric, E.; the most common pre-processing techniques for Yesilada, E. 2017. Analytical methods in tracing near-infrared spectra. TrAC Trends in Analytical honey authenticity. Journal of AOAC International Chemistry 28: 1201-1222. 100: 827-839. Sanaeifar, A.; Jafari, A.; Golmakani, M.-T. 2018. Fusion Tähkäpää, S.; Maijala, R.; Korkeala, H.; Nevas, M. 2015. of dielectric spectroscopy and computer vision for Patterns of food frauds and adulterations reported quality characterization of olive oil during storage. in the EU rapid alert system for food and feed and in Computers and Electronics in Agriculture 145: 142- Finland. Food Control 47: 175-184. 152. van Ruth, S. M.; Luning, P. A.; Silvis, I. C. J.; Yang, Y.; Sanz, J.A.; Fernandes, A.M.; Barrenechea, E.; Silva, S.; Huisman, W. 2018. Differences in fraud vulnerability Santos, V.; Gonçalves, N.; Paternain, D.; Jurio, A.; in various food supply chains and their tiers. Food Melo-Pinto, P. 2016. Lamb muscle discrimination Control 84: 375-381. using hyperspectral imaging: Comparison of Vandeginste, B.G.M.; Massart, D.L.; De Jong, S.; various machine learning algorithms. Journal of Massaart, D.L.; Buydens, L.M.C. 1998. Handbook of Food Engineering 174: 92-100. chemometrics and qualimetrics: Part B: Elsevier. Schmutzler, M.; Beganovic, A.; Böhler, G.; Huck, C. W. Verdú, S.; Vásquez, F.; Grau, R.; Ivorra, E.; Sánchez, A. 2015. Methods for detection of pork adulteration in J.; Barat, J. M. 2016. Detection of adulterations with veal product based on FT-NIR spectroscopy for different grains in wheat products based on the laboratory, industrial and on-site analysis. Food hyperspectral image technique: The specific cases Control 57: 258-267. of flour and bread. Food Control 62: 373-380. Sentandreu, M.Á.; Sentandreu, E. 2014. Authenticity of Vermeulen, P.; Suman, M.; Fernández Pierna, J. A.; meat products: Tools against fraud. Food Research Baeten, V. 2018. Discrimination between durum and International 60: 19-29. common wheat kernels using near infrared Shafiee, S.; Polder, G.; Minaei, S.; Moghadam-Charkari, hyperspectral imaging. Journal of Cereal Science N.; van Ruth, S.; Kuś, P. M. 2016. Detection of Honey 84: 74-82. Adulteration using Hyperspectral Imaging. IFAC- Wang, P.; Sun, J.; Zhang, T.; Liu, W. 2016. Vibrational PapersOnLine 49: 311-314. spectroscopic approaches for the quality evaluation Shao, Y.; Xuan, G.; Hu, Z.; Gao, X. 2018. Identification of and authentication of virgin olive oil. Applied adulterated cooked millet flour with Hyperspectral Spectroscopy Reviews 51: 763-790. Imaging Analysis. IFAC-PapersOnLine 51: 96-101. Wu, J.; Dong, J.; Dong, W.; Chen, Y.; Liu, C. 2016. Rapid Shi, J.; Tang, Y.; Wei, H.; Zhang, L.; Zhang, D.; Shi, J.; authentication of adulteration of olive oil by near- Gong, W.; He, X.; Yang, K.; Liu, D. 2012. infrared spectroscopy using support vector Temperature dependence of threshold and gain machines. In International Symposium on coefficient of stimulated Brillouin scattering in Optoelectronic Technology and Application 2016, water. Applied Physics B 108: 717-720. vol. 10157. Pp. 12: SPIE. Shi, J.; Yuan, D.; Hao, S.; Wang, H.; Luo, N.; Liu, J.; Wu, L.; Du, B.; Vander Heyden, Y.; Chen, L.; Zhao, L.; Zhang, Y.; Zhang, W.; He, X.; Chen, Z. 2019. Wang, M.; Xue, X. 2017. Recent advancements in Stimulated Brillouin scattering in combination with detecting sugar-based adulterants in honey – A visible absorption spectroscopy for authentication challenge. TrAC Trends in Analytical Chemistry 86: of vegetable oils and detection of olive oil 25-38. adulteration. Spectrochimica Acta Part A: Wójcicki, K.; Khmelinskii, I.; Sikorski, M.; Caponio, F.; Molecular and Biomolecular Spectroscopy 206: Paradiso, V. M.; Summo, C.; Pasqualone, A.; 320-327. Sikorska, E. 2015. Spectroscopic techniques and Sivakesava, S.; Irudayaraj, J. 2001. Detection of chemometrics in analysis of blends of extra virgin inverted beet sugar adulteration of honey by FTIR with refined and mild deodorized olive oils. spectroscopy. Journal of the Science of Food and European Journal of Lipid Science and Technology Agriculture 81: 683-690. 117: 92-102. Sliwinska, M.; Wisniewska, P.; Dymerski, T.; Namiesnik, Xiao, R.; Liu, L.; Zhang, D.; Ma, Y.; Ngadi, M. O. 2018. J.; Wardencki, W. 2014. Food analysis using Discrimination of organic and conventional rice by artificial senses. Journal of agricultural and food chemometric analysis of NIR spectra: a pilot study. chemistry 62: 1423-1448.

-160-

J. Mendez et al. / Scientia Agropecuaria 10(1): 143 – 161 (2019)

Journal of Food Measurement and Characterization Zhao, H.; Guo, B.; Wei, Y.; Zhang, B. 2014. Effects of 1-12. grown origin, genotype, harvest year, and their Xiong, Z.; Sun, D.-W.; Pu, H.; Zhu, Z.; Luo, M. 2015. interactions of wheat kernels on near infrared Combination of spectra and texture data of spectral fingerprints for geographical traceability. hyperspectral imaging for differentiating between Food Chemistry 152: 316-322. free-range and broiler chicken meats. LWT - Food Zheng, X.; Li, Y.; Wei, W.; Peng, Y. 2019. Detection of Science and Technology 60: 649-655. adulteration with duck meat in minced lamb meat by Xu, J.-L.; Riccioli, C.; Sun, D.-W. 2017. Comparison of using visible near-infrared hyperspectral imaging. hyperspectral imaging and computer vision for Meat Science 149: 55-62. automatic differentiation of organically and Zhou, X.; Yang, Z.; Haughey, S. A.; Galvin-King, P.; Han, conventionally farmed salmon. Journal of Food L.; Elliott, C. T. 2015. Classification the Engineering 196: 170-182. geographical origin of corn distillers dried grains Zhang, C.; Liu, F.; He, Y. 2018. Identification of coffee with solubles by near infrared reflectance bean varieties using hyperspectral imaging: spectroscopy combined with chemometrics: A influence of preprocessing methods and pixel-wise feasibility study. Food Chemistry 189: 13-18. spectra analysis. Scientific Reports 8: 2166. Zhu, H.; Basir, O. 2006. A novel fuzzy evidential Zhang, X.; Li, W.; Yin, B.; Chen, W.; Kelly, D. P.; Wang, X.; reasoning paradigm for data fusion with Zheng, K.; Du, Y. 2013. Improvement of near applications in image processing. Soft Computing infrared spectroscopic (NIRS) analysis of caffeine in 10: 1169-1180. roasted Arabica coffee by variable selection Ziegler, J.U.; Leitenberger, M.; Longin, C.F.H.; method of stability competitive adaptive reweighted Würschum, T.; Carle, R.; Schweiggert, R.M. 2016. sampling (SCARS). Spectrochimica Acta Part A: Near-infrared reflectance spectroscopy for the Molecular and Biomolecular Spectroscopy 114: rapid discrimination of kernels and flours of 350-356. different wheat species. Journal of Food Zhao, H.; Guo, B.; Wei, Y.; Zhang, B. 2013. Near infrared Composition and Analysis 51: 30-36. reflectance spectroscopy for determination of the geographical origin of wheat. Food Chemistry 138: 1902-1907.

-161-