1. Introduction

Rice has become one of the key components in global food safety as it is the main ingredient in the daily diets of around three billion people worldwide with Asia being the highest consumer ^[1]. It is considered as the predominant dietary energy source in seventeen countries in Asia and Asia-Pacific region, nine countries in North and South America, and eight countries in Africa ^[2]. According to modern nutrition analysis, rice is a food having higher nutritional value, because it contains protein, starch, fat, vitamins (B1, A, and E), minerals and water ^[3]. However, a detailed analysis of nutrient content of rice suggests that the nutrition value of rice depends on many factors such as nutrient quality of soil used for rice cultivation, whether and how rice is polished or processed, and method of preparation before consumption ^[4]. With improvement in living standards, demand of high-quality rice has been increased. Controlling respiration and transpiration processes by means of waxing has been widely practiced for extending shelf-life of fruits and vegetables with great success ^[5]. This means waxing is already quite mature in technology. But in order to make huge profit and attract consumers, some unscrupulous traders use industrial wax to bring good color and luster in rice. Industrial wax is extracted from oil directly, which contains highly carcinogenic compounds ^[6]. These odorless, colorless substances are deadly when consumed over a long period of time. At present, among the various techniques available for the characterization of waxes, the chromatographic and the spectroscopic techniques are the most efficient ^[7]. Separation methods such as supercritical gas chromatography ^[8], liquid chromatography ^[9], gas chromatography-flame ionization detection (GC-FID) ^[10], and pyrolysis gas chromatography ^[11] can also be used for the detection of waxed rice. However, these techniques are very complicated and involve a considerable amount of manual work. Therefore, a fast, accurate, non-destructive method or device needs to be developed for detection of waxed rice.

Hyperspectral imaging has emerged as a non-destructive, low-cost, and rapid technique, which combines Vis-NIR imaging and spectroscopy to acquire both spatial and spectral information from an object. The resulting three-dimensional dataset containing two spatial dimensions and one spectral dimension is known as data cube. This makes it one of the most powerful approaches to obtain information on different samples that can be used for detecting waxed rice. Similar application of hyperspectral imaging with principal component analysis (PCA) was proposed for the detection of waxed rice ^[12]. The similar study extracted 69 effective wavelengths, which might be complex to set up the multispectral imaging system. A previous study using hyperspectral imaging technique was to test pears’ characteristics^[13]. This study selected more than 50 effective wavelengths for analysis, which also made it hard to set up the simple multispectral imaging system. Hyperspectral imaging has been also successfully applied as a smart and promising analytical tool in food quality control ^[14], such as quality sorting of strawberry ^[15], detection of defects in apples ^{[16, 17]}, vegetable quality assessment ^[18], the fish oil DHA content ^[19] and poultry skin tumor inspection ^[20].

Download as

• PPT
  PowerPoint Slide
• PNG
  Larger image(png format)

Veiw figureFigures index

•  NEW
  View larger figure in new window

View next figure

Figure 1. Scheme of hyperspectral image analysis for waxed rice detection

The study presented an analysis of hyperspectral imaging for detecting waxed rice. Figure 1 shows the proposed scheme of detecting waxed rice. In order to reduce particle size effects, the study used multiplicative scatter correction (MSC) for spectra processing. To minimize collinearity problems ^[21], the study employed successive projections algorithm (SPA) to extract effective wavelengths. Based on the effective wavelengths, the predicted models were then set up by partial least squares regression (PLSR), multiple linear regression (MLR), and linear discriminant analysis (LDA). Additionally, a comparative study of the three methods was carried out. At last, MSC-SPA-MLR and MSC-SPA-LDA model were developed to identify waxed rice. This study is expected to help develop a low-cost and on-line multi-spectral imaging system for waxed rice detection.

2. Materials and Methods

Northeast rice samples and “SheYang” rice samples were purchased from local farms in China to ensure no wax on the surfaces of rice. Before processing, all samples were divided into two groups. One group was waxed using industrial wax, which was defined as type 1 and the other group received no treatments defined as type 2. In this study, little horse brush was used to wax rice to cover the surfaces of rice uniformly. The waxing process for each grain of rice was to evenly coat the rice with 3mg of wax. The industrial wax used in this study is mainly made up of C16-C20 N-alkanes, whose composition is the same as those unscrupulous traders used in the international market ^[6]. And wax was acquired from Shanghai national medicine group, China. After waxing, rice was transported to the laboratory and kept at a certain temperature (25°C) for 24 h. All samples were first allowed to equilibrate to room condition (25°C and 30% RH) before Vis-NIR hyperspectral image acquisition. Samples were divided randomly into calibration set and prediction set. The calibration set was used to establish models for identifying waxed rice and non-waxed rice, and the prediction set was used as external test set for validating the actual prediction ability of the models. Calibration set included both waxed rice and non-waxed rice, so did the prediction set. Samples were placed in a glass container (5 cm diameter and 2.5 cm height) to minimize background scattering. All the imaging measurements were completed in the same time period to eliminate the influences of time and temperature.

A Vis–NIR hyperspectral imaging system, schematically shown in Figure 2, was used in the study. A hyperspectral imaging system mainly consists of following parts: an imaging spectrograph (ImSpetorV10E, ZOLIX, CHINA) covering the spectral range of 400–1051nm; a charged couple device (CCD) camera (IGV-B1410M-SC000, Imperx, USA); a focusing lens (Schneider Kreuznach, Germany), attached to the spectrograph; a computer equipped with hyperspectral data acquisition software (Spectral Image-VINR, Wuling Company, Taiwan, China) for controlling exposure time, wavelength range, and generate images; an analysis software (HSI Analyzer, Wuling Company, Taiwan, China) for analyzing data; an assembled light unit consisting of a 150W quartz tungsten halogen lamp(Model 3900, USA) as the light source, a glass optical fiber which divides the light source into two beams and two rectangular focusing lens; a sample table operated by a stepping motor (Model WN232TA300M-F, Weinaguangke Company, Beijing, China); a computer operating sample table software (DyiTV1.1.5, Weinaguangke Company, Beijing, China) for controlling speed and distance. The system is online-scanning configuration, also known as the push broom method.

The camera lens was positioned 330 mm above the surfaces of samples. The lens focus and the light intensity were also adjusted to ensure a clear image. In order to acquire the best image, the spectral response of CCD was kept between 3200 and 3500(80% of the maximum spectral response). The exposure time for data acquisition was fixed as 3.2ms and the corresponding moving speed of the sample table was set at 410µm/s. Throughout the entire process of imaging, the environmental conditions were set at identical values for all the rice samples, i.e. 25°C and 30% RH.

Download as

• PPT
  PowerPoint Slide
• PNG
  Larger image(png format)

Veiw figureFigures index

•  NEW
  View larger figure in new window
• PREV
  View previous figure

View next figure

Figure 2. Hyperspectral imaging system

In the online hyperspectral imaging system, samples were placed on the sample table, and the table was then moved at a speed of 410µm/s to scan line by line. For the purpose of improving detection reliability, waxed rice samples for both calibration and prediction were evenly divided two parts to acquire images, so did the non-waxed rice samples. Eight hyperspectral images were finally obtained. Each point of the hyperspectral image included spectral (λ) as well as spatial (X and Y) information. Figure 3 illustrates the three dimensional hyperspectral data.

Download as

• PPT
  PowerPoint Slide
• PNG
  Larger image(png format)

Veiw figureFigures index

•  NEW
  View larger figure in new window
• PREV
  View previous figure

View next figure

Figure 3. Three dimensional hyperspectral data

CCD in a hyperspectral imaging system collects signal intensity from detector, not spectra-reflectance. Signal intensity not only provides information regarding quality of tested sample, but it also reflects sensitivity of detector and sample illumination source ^[22]. Therefore, raw hyperspectral images should be corrected for reflectance hyperspectral images based on black and white reference images (Figure 4). The white reflectance image (W) was acquired using a white surface board with uniform, stable and high reflectance standard (about 99.9% reflectance). The black one (B) was acquired with the camera lens completely covered with its opaque cap under the same conditions. I_o is the raw image acquired under the same conditions except with a different exposure time. The dark reflectance image of sample (D_s) was acquired at the exposure time same as the raw images. The corrected hyperspectral image (I_C) was then calculated using the following equation:

(1)

The corrected images were used as the basis for subsequent analysis. Finally, the study got eight corrected hyperspectral images, including four images of waxed rice (two for calibration and two for prediction) and four images of non-waxed rice (two for calibration and two for prediction).

Download as

• PPT
  PowerPoint Slide
• PNG
  Larger image(png format)

Veiw figureFigures index

•  NEW
  View larger figure in new window
• PREV
  View previous figure

View next figure

Figure 4. The correction of hyperspectral images

In this study, there were 100 grains of waxed rice (50 grains per image) chosen randomly in the two hyperspectral images of waxed rice for calibration, selecting three small regions from each grain of rice to calculate the average relative reflectance. The average relative reflectance can represent the whole grain of rice. With this method, we obtained the information of average relative reflectance of total 300 grains of rice samples. The amount of samples was shown in Table 1. All the data were stored in a computer.

Table 1. The amount of samples

Download as

PowerPoint Slide

Larger image(png format)

• Tables index

Veiw figure

View current table in a new window

View next table

The scattering of light is, however, a rather complicated phenomenon. The degree of scattering varies from sample to sample, and it may create problems in analyses. Martens et al. ^[23] has proposed a method (multiplicative scatter correction) that uses linear regression of spectral variables vs. average spectrum to diminish the influence. Therefore, in the study, multiplicative scatter correction (MSC) was conducted to preprocess the original spectra.

Principal component analysis (PCA) is mainly used in dimensional reduction of the acquired data set while retaining the important characteristic, which contributes most of the variance ^[24]. To determine the feasibility of the study, PCA was conducted on the full-spectrum data.

However, hyperspectral imaging system usually has a high spectral resolution, and the acquired spectral data often contain hundreds of wavelength bands. Indeed, spectral data extracted from hyperspectral images are known to possess variable (wavelength) dimension with redundancy among contiguous variables ^[25]. Variable selection (also called wavelength selection), is a critical step in data analysis for Vis-NIR hyperspectral imaging. Elimination of irrelevant variables can predigest calibration modeling and improve results in terms of accuracy and robustness ^[26]. Successive projections algorithm (SPA) is a novel variable selection algorithm, designed to solve the collinearity problems by selecting variables with minimal redundancy ^[27]. For this purpose, SPA employs a simple projection operation in a vector space to select subsets of variables with minimum collinearity. In this study, SPA was applied to select the effective wavelengths that have the greatest contribution to the detection of waxed rice without retaining redundant information.

In this study, the models were established using three calibration algorithms, partial least squares regression (PLSR) ^{[28, 29]}, multiple linear regression (MLR) ^[30], and linear discriminant analysis(LDA), respectively. PLSR is used for establishing a regression model ^[31]. MLR is a commonly used calibration algorithm with the advantages of being simple and easy to be interpreted. However, it fails when variables are more than samples and gets easily affected by the collinearity between variables ^[32]. In our study, the number of wavelength variables of hyperspectral images was larger than the number of samples. Therefore, it was not possible to perform MLR directly. But we used SPA to select the effective variables before data calibration. Moreover, the selection of variables with less collinearity would be helpful to improve the MLR model. LDA is a well-known scheme for feature extraction and dimension reduction ^[33]. It can form typical discriminant functions to classify samples. Therefore, LDA can work as a model based on reflectance at effective wavelengths for identifying waxed rice and non-waxed rice.

3. Results and Discussion

In order to reduce particle size effects, we employed MSC for processing spectra. After performing MSC on the original spectral matrix Xk (old) it forms a new spectral matrix Xk (new). The preprocessing method was implemented using MATLAB 2011b (The Math Works, Natick, USA).We used spectral matrix Xk (new) to do the following analysis.

Figure 5 presents the average reflectance curves in the range of 400–1051 nm, collected from the waxed rice and non-waxed rice. The two average relative reflectance curves were obtained using MATLAB2011b and Origin 8.0. As evident from the figure, the average relative reflectance curves of the waxed rice and non-waxed rice show similar trends as there is no difference in variety of rice in both cases. In the visible spectral region (400-780 nm), the intensity of average relative reflectance of the waxed rice is higher than non-waxed rice; the reason for brighter texture of the waxed rice. There are a few effective wavelengths distributed in the region of 400–780 nm. The industrial wax used in this study is mainly made up of C16-C20 N-alkanes, which has many C-H keys. The C-H stretching vibration produces frequency doubling which results in a combination of frequencies when a number of different events occur simultaneously. According to the distribution of C-H group absorption band ^[34], frequency doublings and combination frequency leads to an increase in absorbance values. In detail, level 3 frequency doubling and the third combination frequency of C-H keys in hydrocarbon are appeared at 900–950nm and 1000–1100nm. As shown in Figure 5, initially the average relative reflectance of waxed rice remains higher than that of the non-waxed rice, however, after 778.059nm, the average relative reflectance of the non-waxed rice becomes slightly higher than that of the waxed rice. This is because of the fact that C16-C20 N-alkanes absorb near-infrared light, resulting in C-H stretching vibration and producing absorption spectrum. Therefore, the desired effective wavelength range would be 780–1051 nm.

Download as

• PPT
  PowerPoint Slide
• PNG
  Larger image(png format)

Veiw figureFigures index

•  NEW
  View larger figure in new window
• PREV
  View previous figure

View next figure

Figure 5. Reflectance spectra of waxed rice and non-waxed rice

We performed PCA on the full-spectrum data to verify the separation of waxed and non-waxed rice. Figure 6 illustrates the PCA cluster plots of the first two principal component scores for waxed and non-waxed rice samples. It is noteworthy that the first principal component represents 71.47% of the data variance, while the second principal component represents 23.14%. As depicted in Figure 6, the samples are divided into two parts, but they are observably overlapped. Therefore, new algorithms should be developed to separate waxed rice from non-waxed rice.

Download as

• PPT
  PowerPoint Slide
• PNG
  Larger image(png format)

Veiw figureFigures index

•  NEW
  View larger figure in new window
• PREV
  View previous figure

View next figure

Figure 6. PCA cluster plot

Once spectral data are processed by MSC algorithms, effective wavelengths can be selected using SPA. According to previous studies ^[35], effective wavelengths might be equally or more efficient than full wavelengths, because they carry the most important information relevant to determination. In this study, we used SPA to select effective wavelengths from the full spectra (400–1051nm). Six variables (469.9nm, 666.2nm, 941.5 nm, 1025.8nm, 1034.6nm, and 1050.6nm) were selected as the effective wavelengths, which could later be used to predict whether the rice is waxed or not.

Once the effective wavelengths were selected, the spectral data sets were then reduced to a matrix with a dimension of s × t, where is the number of samples (s = 200) and the number of variables t is reduced from 903 to 6 wavelengths (the number of effective wavelengths). Then MLR, PLSR and LDA were used to establish identification models, respectively, based on the reduced spectral data.

PLSR for the waxed rice detection was performed with Matlab2010b, while MLR modeling was conducted using IBM SPSS Statistics 19.0. In MLR and PLSR models, RMSE (root-mean-square error) and R (coefficient of correlation coefficient) are the two most important parameters. RMSE is more close to 0, whereas R is more close to 1, implying the model’s better prediction ability. As mentioned above, we obtained average relative reflectance for 300 samples, 200 of which were used for calibration, while 100 samples were for prediction. In the analysis of calibration, we used average relative reflectance of the effective wavelengths as independent variables, and the values of type of rice (1 and 2) were set as dependent variables. Thus, we obtained the MSC-SPA-PLSR and MSC-SPA-MLR prediction models, which were established based on the calibration set of rice for identifying whether the rice was waxed.

Table 2. Results of comparison of different calibration models

Download as

PowerPoint Slide

Larger image(png format)

• Tables index

Veiw figure

View current table in a new window

View previous table

View next table

The results of the comparison between the MSC-SPA-PLSR and MSC-SPA-MLR models were shown in Table 2. And from the table we can see MSC and SPA made a significant contribution to the model. R was found to be 0.934 for MSC-SPA-PLSR and 0.938 for MSC-SPA-MLR. RMSE values of MSC-SPA-PLSR and MSC-SPA-MLR were obtained as 0.177 and 0.176, respectively. The obtained MSC-SPA-PLSR and MSC-SPA-MLR equation are shown as follows:

Where X_inm is the relative reflectance value at the wavelength of i nm and TYPE_PLSR and TYPE_MLR are the predicted values.

Figure 7 presents the results for multivariate statistical analysis. MSC-SPA-PLSR and MSC-SPA-MLR models developed using the effective wavelengths for the detection of waxed rice are shown in Figure 7 (a) and Figure 7 (b). In contrast, the results of calibration in MSC-SPA-MLR are comparatively better but slightly. Hence, the two models were evaluated by comparing the prediction results. In this study, we used 100 samples from prediction set for evaluating the prediction ability of the two models, MSC-SPA-PLSR and MSC-SPA-MLR. The waxed and non-waxed samples from prediction set were divided into two sets based on the value of TYPEP_PLSR and TYPE_MLR. And the threshold value of the two sets is 1.5. Figure 7 (c) shows that eight samples that are miscalculated in the prediction, giving 92% detection rate for MSC-SPA-PLSR model. On the other hand, Figure 7(d) shows four samples are miscalculated in the prediction, giving 96% detection rate for MLR model. Therefore, we can conclude that the prediction ability of MSC-SPA-MLR is better than that of MSC-SPA-PLSR.

Download as

• PPT
  PowerPoint Slide
• PNG
  Larger image(png format)

Veiw figureFigures index

•  NEW
  View larger figure in new window
• PREV
  View previous figure

Figure 7. The results of multivariate statistical analyses

In the analysis of LDA, we imported the average relative reflectance of the effective wavelengths and types (1 and 2) of calibration set into the software of IBM SPSS Statistics 19.0. Next a typical discriminant function was formed, and the independent variables of typical discriminant function were the average relative reflectance of the effective wavelengths. The obtained MSC-SPA- LDA function is shown as follows:

Where X_inm is the average relative reflectance spectral value at wavelength i nm and TYPE_LDA is the predicted values in the model. Similarly, we used 100 samples from the prediction set to demonstrate the reliability of the MSC-SPA-LDA model. We imported the average relative reflectance of the effective wavelengths into the IBM SPSS Statistics 19.0. The software could predict the type of rice based on the typical discriminant function. The sample is waxed rice when TYPE_LDA ≥ 0, but the sample is non-waxed rice if TYPE_LDA < 0. There were 4 grains of rice samples out of 100 samples that were miscalculated in the prediction set. Therefore, MSC-SPA- LDA provides 96% detection rate.

MSC-SPA-PLSR, MSC-SPA-MLR, and MSC-SPA-LDA algorithms were employed to detect whether the rice was waxed. Table 3 presents the prediction results based on different algorithms. The data presented in Table 3 indicate that the results based on MSC-SPA-MLR and MSC-SPA- LDA are better. All the results together indicate that MSC-SPA-MLR and MSC-SPA-LDA could be considered as the better methods for identifying whether a rice sample is waxed or not.

Table 3. Prediction results based on three models

Download as

PowerPoint Slide

Larger image(png format)

• Tables index

Veiw figure

View current table in a new window

View previous table

4. Conclusions

The Vis-NIR hyperspectral imaging technique exhibited good performance as a fast and non-destructive method for detecting waxed rice. The spectra measurement technique is also simple and convenient. In order to reduce particle size effects, MSC was conducted to preprocess the original spectra. Then out of 903 wavelengths, only six wavelengths were selected by SPA as the effective wavelengths that were found to be suitable for waxed rice prediction. Compared with MSC-SPA-PLSR, MSC-SPA-MLR and MSC-SPA-LDA models produced more acceptable precision and accuracy in predicting waxed rice. The identification accuracy for these methods on prediction set was 96%. In further research, a simple multispectral imaging system based on six effective wavelengths could replace the hyperspectral imaging system for identifying waxed rice and non-waxed rice. This system can obtain reflectance values corresponding to six effective wavelengths used as independent variables for the MSC-SPA-MLR and MSC-SPA-LDA models. Furthermore, other variable selection algorithms should be considered to select the best wavelength variables with higher accuracy and fewer numbers for detecting waxed rice.

Acknowledgements

This work was partially supported by National Basic Research Program of China (973 Program) (2015CB352001), National Science Instrument Important Project (2011YQ14014704), Shanghai Municipal Science Instrument Important Project (14142200902), National Natural Science Foundation of China (61378060, 61205156).

References

[1]	Smith BD, Nesbitt M. The emergence of agriculture: Scientific American Library New York. 1995.
	In article
[2]	Kwak HS, Kim M, Jeong Y. Physicochemical Properties and Determination of Key Instrumental Quality Measurement Parameters of Frozen-Cooked Rice by Correlating Consumer Acceptance. Journal of Food Quality , 38(3). 192-200. 2015.
	In article
[3]	Lourenço LFH, Tavares TS, Araujo EAF, Pena RS, Joele MRSP, Carvalho AV. Optimization of extrusion process to obtain shrimp snacks with rice grits and polished rice grains. CyTA - Journal of Food, 1-9. 2015.
	In article
[4]	Juliano BO. Rice in human nutrition: Int. Rice Res. Inst. 1993.
	In article
[5]	Praditdoung S, Vananuvat N, Jutamongkone R. Utilization of rice bran waxes. 1991.
	In article
[6]	Bhuvenesh, Rajamony, Nabil, Khalil, Mohd, Faizal, et al. Development of Palm Oil Based Bio-product Industrial Wax and Its Application Performance Study. Life Science（ISSN1934-7391）23-27. 2010.
	In article
[7]	Soomro RK, Sherazi STH. Extraction and characterization of seed oil waxes by using chromatographic techniques. International Journal of Industrial Chemistry, 4(1). 2013.
	In article
[8]	Mondello L, Tranchida PQ, Dugo P, Dugo G. Comprehensive two-dimensional gas chromatography-mass spectrometry: A review. Mass Spectrometry Reviews, 27(2). 101-124. 2008.
	In article
[9]	Asperger A, Engewald W, Fabian G. Thermally assisted hydrolysis and methylation – a simple and rapid online derivatization method for the gas chromatographic analysis of natural waxes. Journal of Analytical & Applied Pyrolysis , 61(01). 91-109. 2001.
	In article
[10]	Gysels K, Delalieux F, Deutsch F, Grieken RV, Camuffo D, Bernardi A, et al. Identification of binding media in works of art by gas chromatography–mass spectrometry. Journal of Cultural Heritage, 5(2). 231-240. 2004.
	In article
[11]	Regert M, Langlois J, Colinart S. Characterisation of wax works of art by gas chromatographic procedures. Journal of Chromatography A , 1091(1-2). 124-136. 2005.
	In article
[12]	Zhao M. Hyperspectral visible-near infrared imaging for the detection of waxed rice. Hyperspectral visible-near infrared imaging for the detection of waxed rice. Optical Metrology & Inspection for Industrial Applications III 2014, 9276.
	In article
[13]	Li B, Hou B, Zhang D, Zhou Y, Zhao M, Hong R, et al. Pears characteristics (soluble solids content and firmness prediction, varieties) testing methods based on visible-near infrared hyperspectral imaging. Optik - International Journal for Light and Electron Optics, 127(5). 2624-2630. 2015.
	In article
[14]	Yavari AA, Hamedi M, Haghbin S. Retraction Note: Investigation of Frying Oil Quality Using VIS/NIR Hyperspectral Analysis. Journal of the American Oil Chemists Society, 90(1). 163-163. 2013.
	In article
[15]	ElMasry G, Wang N, ElSayed A, Ngadi M. Hyperspectral imaging for nondestructive determination of some quality attributes for strawberry. Journal of Food Engineering, 81(1). 98–107. 2007.
	In article
[16]	Vol. N. Multispectral detection of fecal contamination on apples based on hyperspectral imagery: Part I. Application of visible and near-infrared reflectance imaging. Transactions of the ASAE, 45(6). 2027-2038. 2002.
	In article
[17]	Mehl PM. Detection of Defects on Selected Apple Cultivars Using Hyperspectral and Multispectral Image Analysis. Applied engineering in agriculture, 18(2). 219-226. 2002.
	In article
[18]	D. Lorente , N. Aleixos , J. Gómez-Sanchis , S. Cubero , O. L. García-Navarrete , Blasco J. Recent Advances and Applications of Hyperspectral Imaging for Fruit and Vegetable Quality Assessment. Neuroimage,5(4). 1121-1142. 2012.
	In article
[19]	Jianhua C. Near-infrared Spectrum Detection of Fish Oil DHA Content Based on Empirical Mode Decomposition and Independent Component Analysis. Journal of Food and Nutrition Research, 2(2). 62-68. 2014.
	In article
[20]	Kong SG, Yud-Ren C, Intaek K, Kim MS. Analysis of hyperspectral fluorescence images for poultry skin tumor inspection. Applied Optics , 43(4). 824-833. 2004.
	In article
[21]	Grünhut M, Centurión ME, Fragoso WD, Almeida LF, de Araújo MC, Band BSF. Flow-batch technique for the simultaneous enzymatic determination of levodopa and carbidopa in pharmaceuticals using PLS and successive projections algorithm. Talanta, 75(4). 950-958. 2008.
	In article
[22]	Wu D, Sun D W, He Y. Application of long-wave near infrared hyperspectral imaging for measurement of color distribution in salmon fillet. Innovative Food Science & Emerging Technologies, 16(39). 361-372. 2012.
	In article
[23]	Martens H. S. Jensen, and P. Geladi, “Multivariate Linearity Transformation for Near-Infrared Reflectance Spectrometry,”. Proc. Nordic Syrup. on Applied Statistics , 205-234. 1983.
	In article
[24]	Klarić A, Klarić D, Mornar I, Ana. Polyphenol content and antioxidant activity of commercial blackberry wines from Croatia : The application of multivariate analysis for the geographical origin differentiation. Journal of Food & Nutrition Research, 2011.
	In article
[25]	Kamruzzaman M, Elmasry G, Sun DW, Allen P. Application of NIR hyperspectral imaging for discrimination of lamb muscles. Journal of Food Engineering, 104(3). 332-340. 2011.
	In article
[26]	Qing Z, Ji B, Zude M. WAVELENGTH SELECTION FOR PREDICTING PHYSICOCHEMICAL PROPERTIES OF APPLE FRUIT BASED ON NEAR-INFRARED SPECTROSCOPY. Journal of Food Quality, 30(4). 511-526. 2007.
	In article
[27]	Araújo MCU, Saldanha TCB, Galvão RKH, Yoneyama T, Chame HC, Visani V. The successive projections algorithm for variable selection in spectroscopic multicomponent analysis. Chemometrics and Intelligent Laboratory Systems , 57(2). 65-73. 2001.
	In article
[28]	Jiewen Z, Qin O, Quansheng C, Hao L. Simultaneous determination of amino acid nitrogen and total acid in soy sauce using near infrared spectroscopy combined with characteristic variables selection. Food science and technology international = Ciencia y tecnologia de los alimentos internacional, 19(4). 305-314. 2013.
	In article
[29]	Piñabarrera A, Mezamárquez OG, Osoriorevilla G, Gallardovelázquez T. Identification and quantification of corncob as adulterant in corn dough and tortilla by MIR–FTIR spectroscopy and multivariate analysis. CyTA - Journal of Food, 12(1). 65-72. 2013.
	In article
[30]	Wu D, Shi H, Wang S, He Y, Bao Y, Liu K. Rapid prediction of moisture content of dehydrated prawns using online hyperspectral imaging system. Analytica Chimica Acta, 726(9). 57-66. 2012.
	In article
[31]	Rohman A, Man YBC, Ismail PH, A. FTIR spectroscopy combined with chemometrics for analysis of lard adulteration in some vegetable oils Espectroscopia FTIR combinada con quimiometría para el análisis de adulteración con grasa de cerdo de aceites vegetales. CyTA - Journal of Food, 9(2). 96-101. 2011.
	In article
[32]	Næs T, Mevik BH. Understanding the collinearity problem in regression and discriminant analysis. Journal of Chemometrics, 15(4). 413-426. 2001.
	In article
[33]	Perriere G, Thioulouse J. Use of correspondence discriminant analysis to predict the subcellular location of bacterial proteins. Computer Methods and Programs in Biomedicine, 70(2). 99-105. 2003.
	In article
[34]	Bruno TJ, Svoronos PDN. CRC handbook of fundamental spectroscopic correlation charts: CRC Taylor & Francis, 2006.
	In article
[35]	D W, X C, P S, S W, F F, Y. H. Determination of alpha-linolenic acid and linoleic acid in edible oils using near-infrared spectroscopy improved by wavelet transform and uninformative variable elimination. Anal Chim Acta, 634(2). 166-171. 2009.
	In article