Predicting soluble solid content in intact jaboticaba [Myrciaria jaboticaba (Vell.) O. Berg] fruit using near-infrared spectroscopy and chemometrics

Food Chemistry 159 (2014) 458–462

Contents lists available at ScienceDirect

Food Chemistry journal homepage: www.elsevier.com/locate/foodchem

Analytical Methods

Predicting soluble solid content in intact jaboticaba [Myrciaria jaboticaba (Vell.) O. Berg] fruit using near-infrared spectroscopy and chemometrics Nathália Cristina Torres Mariani a, Rosangela Câmara da Costa b, Kássio Michell Gomes de Lima b,⇑, Viviani Nardini a, Luís Carlos Cunha Júnior a, Gustavo Henrique de Almeida Teixeira c a Universidade de São Paulo (USP), Faculdade de Ciências Farmacêuticas de Ribeirão Preto, Departamento de Análises Clínicas, Toxicológicas e Bromatológicas, Av. do Café, s/n, Campus Universitário da USP, Ribeirão Preto, CEP 14040-903 São Paulo, Brazil b Universidade Federal do Rio Grande do Norte, Programa de Pós-Graduação em Química, Grupo de Pesquisa em Química Biológica e Quimiometria, Natal, CEP 59072-970 RN, Brazil c Universidade Estadual Paulista (UNESP), Faculdade de Ciências Agrárias e Veterinárias, Departamento de Produção Vegetal, Via de Acesso Prof. Paulo Donato Castellane, s/n, Jaboticabal, CEP 14884-900 São Paulo, Brazil

a r t i c l e

i n f o

Article history: Received 21 January 2013 Received in revised form 28 November 2013 Accepted 12 March 2014 Available online 20 March 2014 Keywords: NIR spectroscopy PLS BP-ANN LS-SVM Variables selection

a b s t r a c t The aim of this study was to evaluate the potential of near-infrared reflectance spectroscopy (NIR) as a rapid and non-destructive method to determine soluble solid content (SSC) in intact jaboticaba [Myrciaria jaboticaba (Vell.) O. Berg] fruit. Multivariate calibration techniques were compared with pre-processed data and variable selection algorithms, such as partial least squares (PLS), interval partial least squares (iPLS), a genetic algorithm (GA), a successive projections algorithm (SPA) and nonlinear techniques (BP-ANN, back propagation of artificial neural networks; LS-SVM, least squares support vector machine) were applied to building the calibration models. The PLS model produced prediction accuracy (R2 = 0.71, RMSEP = 1.33 °Brix, and RPD = 1.65) while the BP-ANN model (R2 = 0.68, RMSEM = 1.20 °Brix, and RPD = 1.83) and LS-SVM models achieved lower performance metrics (R2 = 0.44, RMSEP = 1.89 °Brix, and RPD = 1.16). This study was the first attempt to use NIR spectroscopy as a non-destructive method to determine SSC jaboticaba fruit. Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction The jaboticaba has origins in Brazil’s south centre. Among the currently known species, the Myrciaria cauliflora (MC) and Myrciaria jaboticaba (Vell.) Berg, produce fruits that are suitable for both fresh consumption and commercial products. The peel of the mature fruit is black, thin and fragile; the flesh is white, sweet and slightly acidic. The jabuticabeira is a fruit tree of great interest to farmers in various regions of Brazil and overseas because of the fruit’s high productivity, hardiness and potential uses (Andersen & Andersen, 1988; Teixeira, Durigan, & Durigan, 2011; Teixeira, Durigan, Santos, Hojo, & Cunha, 2011). The soluble solids content (SSC) is one of the major factors that affects the taste of the fruit and is closely related to the consumer’s perception of maturity in jaboticaba fruits. However, most instrumental techniques used to measure SSC are destructive, time-consuming and costly. Near-infrared (NIR) spectroscopy has been shown to be a rapid and non-destructive method for measuring the internal quality of ⇑ Corresponding author. Tel.: +55 84 3215 3828; fax: +55 83 3211 9224. E-mail address: [email protected] (K.M.G. de Lima). http://dx.doi.org/10.1016/j.foodchem.2014.03.066 0308-8146/Ó 2014 Elsevier Ltd. All rights reserved.

fruit and is ideally suited to the requirements of the agrofood industry for quality control (He, Zhang, Pereira, Gómez, & Wang, 2005). Research has been conducted in the analysis of SSC by NIR spectroscopy in several fruits, such as apples (Nicolai et al., 2008), kiwifruits (Martinsen & Schaare, 1998), pears (Sun, Lin, Xu, & Ying, 2009), mulberry (Huang et al., 2011), oranges (Liu, Sun, & Ouyang, 2010), and mangos (Jha et al., 2012). The nondestructive technique or chemometric approach, however, has not been reportedly used to analyse SSC in jaboticaba fruit. The main challenge of NIR spectroscopy is choosing target wavelengths from the full IR spectrum that result in maximum accuracy, especially when the spectra display unresolved peaks or fail to identify important features. To overcome these difficulties, various chemometric algorithms have been proposed to select optimal variables for multivariate calibration, such as iPLS (interval partial least squares) (Norgaard et al., 2000), GA (genetic algorithm) (Ferrand et al., 2011) and SPA (successive projections algorithm) (Araújo et al., 2001), which improve multivariate models by exploiting relevant variables. Another tool used to improve NIR spectroscopic results is outlier detection, which selects samples that stand out from the bulk data, typically generated by instrumental errors, the presence of

459

N.C.T. Mariani et al. / Food Chemistry 159 (2014) 458–462

another population or human/analytical errors. In this work, the calibration and prediction sets were optimised based on data with extreme leverage, and unmodeled residuals of the overall data and the dependent variables, specifically (Valderrama, Braga, & Poppi, 2007). When spectral data present non-linearity that can arise from the chemical nature of the target analytical parameter, linear methods can lead to inaccurate predictions. These errors can be improved by using non-linear multivariate techniques. Least squares support vector machine (LS-SVM), a new learning algorithm proposed by Suykens, Van Gestel, De Brabanter, De Moor, and Vandewalle (2002) is a promising alternative to existing linear and nonlinear multivariate statistical procedures. Another alternative is back-propagation artificial neural networks (BP-ANN) (Hertz, Krogh, & Palmer, 1991), which can also be effective when non-linear spectral responses are involved. The objectives of this study were (1) to establish the quantitative relationships between the NIR spectra and SSC measurements in intact jaboticaba fruit, and (2) to compare the predictive performances of calibration models established by partial least squares (PLS), back-propagation artificial neural network (BP-ANN), and least squares support vector machine (LS-SVM) on the basis of the selected wavelength variables (iPLS, GA, SPA). 2. Materials and methods 2.1. Fruit samples A total of 100 jaboticaba fruit [M. jaboticaba (Vell) Berg cv. Sabará] were collected in 2011 at the beginning of the harvest season (11/10/2011). The fruits were sorted; 50 were classified as large (10 ± 0.5 g) and 50 as small (5 ± 0.2 g). All fruits harvested were at a stage ready for commercial sale and the skins were completely purple. 2.2. Instrumentation After temperature stabilization (25 °C), spectra were collected by a Fourier transform (FT)-IR spectrophotometer (Spectrum 100N, PerkinElmer) in the diffuse reflectance mode over the range of 10,000–4000 cm 1 (1000–2500 nm). Spectra were randomly collected from the epidermal surface at two different locations on each fruit (64 scans, spectral resolution of 2 cm 1). The mean spectrum was then calculated for each fruit from the measured spectra. 2.3. Reference methods for SSC After the spectra were collected, each piece of fruit was analysed for their soluble solids content (SSC) using reference method 920.151 of the Official Methods of Analysis (AOAC, 1997). The fruit was cut in half and one portion was placed in a hydrophilic bandage and pressed to extract the juice. The juice was then placed on a digital refractometer, ATAGO (model PR101a, Tokyo, Japan) to determine the SSC. This equipment is capable of automatic temperature compensation and a measurement accuracy of ±0.1 °Brix. The descriptive statistics for SSC of the jaboticaba fruit are presented in Table 1.

Table 1 Average concentration of soluble solids content (%) of jaboticaba ‘Sabará’ [Myrciaria jaboticaba (Vell.) Berg] fruit collected across three harvests in 2011–2012. Samples set

N

Average

S.D*

Maximum

Minimum

Large Small

100 50 50

18.9 20.0a 17.7b

2.22 1.62 2.16

23.5 22.5 23.5

12.1 14.8 12.1

Averages followed by the same letter, in the column for each harvest, are significantly different according to Tukey’s test (P < 0.05). Standard deviation.

*

were used to establish the model. The remaining 30 fruit were used for the prediction. To compare the performance of different calibration models, the fruit in the calibration and prediction sets were the same for all calibration models. 2.5. Spectra pre-processing Spectral pre-processing can improve the performance of a model. In this study, the reflectance spectrum (R) for each fruit was first log (1/R) transformed to give an absorbance spectrum. Several spectral pre-processing algorithms, including Savitzky–Golay smoothing (SGS) over 3–11 points, multiplicative scatter correction (MSC), and 1st and 2nd derivatives were investigated over a span of 3–91 points. 2.6. Multivariate analysis The calibration spectra were subject to a partial least squares regression (PLSR) with leave-one-out cross validation. The optimal number of latent variables was determined by minimising the predicted residual error sum of squares (PRESS). The PLSR calibration models were evaluated using the coefficient of correlation (R2) and the root-mean-square error from calibration (RMSEC) and cross validation (RMSECV). In addition, the accuracy of each PLS model was evaluated by comparing their residual predictive deviations (RPDs), which is the ratio of the standard deviation for a specific reference population and RMSEP of the prediction set. According to Nicolaï et al. (2007), a RPD between 1.5 and 2.0 indicates that the model can discriminate low from high values in the response variable. PLS models were calculated using MATLAB version 6.5 (Math-Works, Natick, USA) and the PLS-toolbox (Eigenvector Research, Inc., Wenatchee, WA, USA, version 6.01). Artificial neural network (ANN) models were built from a reduced number of variables based on the scores from a PCA of the original, smoothed and derived spectra plotted over the wavelength range of 1100–2500 nm. The back-propagating network architecture was subject to supervised training on selected samples known as the monitoring set. The ANN models were obtained from the ANN toolbox in MATLAB version 7. The least-squares support vector machines (LS-SVM) approach is an optimised version of the standard supporting vector machines. For detailed in-depth theoretical background on LSSVM, readers are referred to the introduction of Suykens et al. (2002). In this study, the LS-SVM was executed using MATLAB version 7 using the LS-SVM toolbox (LS-SVM v 1.5, Suykens, Leuven, Belgium) to derive the LS-SVM models. 2.7. Selection variables

2.4. Sample distribution The fruit (100) were assigned to one of two groups as follows: a calibration set and a prediction set. A well-known representative sample selection algorithm, Kennard–Stone, was applied to the sample distribution (Kennard & Stone, 1969). In this study, 70 fruit

The predicted results for the calibration models developed by PLS used the spectral regions selected by iPLS, GA, and SPA. These results were also compared to those found by PLS using the whole region. In the iPLS method, the data were divided to non-overlapping sections where each section undergoes independent PLS

460

N.C.T. Mariani et al. / Food Chemistry 159 (2014) 458–462

modeling to determine the most useful variable range. The optimisation of GA can be combined with partial least squares regression (GA-PLS) and is commonly used for analysing NIR spectra. The total number of selected variables was set to 879 wavelengths. The GAPLS models were optimised separately. SPA is a variable selection technique designed to minimise collinearity in multiple linear regression (MLR). The desired number of variables (NSPA) ranges from 4 to 70.

of latent variables used was 5, 6, 7, 8 or 9. Evidently, the PLS model achieved the best prediction precision with R2 = 0.71, RMSECV = 1.64 °Brix and RMSEP = 1.33 °Brix. This model can discriminate low and high values in the response variable as the RPD value was of 1.65. Overall, PLS is a promising method for optimising the efficiency of NIR spectral calibration and has potential for use in practical applications. The relationship between the values of the SSC measurements and predictions from PLS is shown in Fig. 2.

3. Results and discussion 3.3. Variables selection 3.1. Spectral features Fig. 1 shows the original NIR absorbance spectra (1000– 2500 nm) for the 100 fruit samples. There is an absorption peak that is associated with the second overtone of the hydroxyl group. In the wavelength range 1100–1800 nm, the absorbance curves ascend until 1450 nm and subsequently descend until 1650 nm. Two strong absorption peaks at 1250 nm and 1450 nm are assigned to a combination of the second overtone of the O–H bond from the sugars in the fruit and those in water, and the first overtone of hydroxyl groups in the sugars. The dominant absorption bands (1100–1800 nm) can be attributed to the first overtone from C–H stretching in the sugars. These bands are a composite of different bands because there are many sugars in the jaboticaba fruit. Hence, this NIR fingerprint can be used as an indicator of the sugar content in the analysed samples. A weak band at approximately 1160 nm represents the third overtone from sugars. Absorption peaks were recorded at approximately 1780 nm that correspond with the first overtone signal from the sugars, and the regions from 2110 to 2205 nm and from 2250 to 2260 nm were assigned to the N–H and O–H (water) vibrational modes. Fig. 1 shows baseline shifts in the spectra, which were removed using second-derivative and MSC procedures. The best window for second-derivative fitting was between 3 and 9 points resulting in analytical models with better predictive abilities. 3.2. PLS models Table 2 summarises the results from the calibration and prediction of the optimal PLS models using different spectral pretreatment methods, including smoothing, derivative and MSC. The result of the calibration was determined by RMSEC, RMSEP and R2. A good model should have the lowest RMSECV and RMSEP separated by a small difference and the highest R2. The number

Similar to the PLS analysis, variables selection (iPLS, SPA and GA) was applied to the SSC measurement. The performance of the PLS model was better than the iPLS, GA, and SPA models. The selected variables for the determination of SSC were in the 1300– 1520 nm regions where the C–H stretching of the sugars and two overtone absorption bands for water occur. The best results based on R2, RMSECV and RMSEP are shown in Table 3. The results obtained with the GA model were 0.61, 1.61 and 1.36 °Brix for the prediction set. The correlation coefficients for the set ranged from 0.51 to 0.62 for all the models. The number of latent variables used by iPLS, SPA, and GA was 9 for analysis of the NIR spectra. The benefit of using GA models was that fewer (879) variables are required to build the PLS models. 3.4. BP-ANN models To improve the precision of the predictions of SSC, the first 10 PLS scores were used for each sample. The final BP-ANN model, using PLS N (1-6) scores in a N:4:1 architecture with four nodes in the inner layer, resulted in a RMSEM = 1.20 °Brix, a R2 = 0.68 and slope = 0.35. Slope and bias (0.35 and 0.02 °Brix, respectively) were not improved for the BP-ANN compared to the PLS model (71). The best results generated by BP-ANN are shown in Table 4. The performance of most BP-ANN models was worse than PLS, which suggests that the BP-ANN algorithm could not establish a quantitative relationship between SSC and the spectra with the variables selected. 3.5. LS-SMV model In addition to PLS, variable selection and BP-ANN, we established a SSC determination model based on the LS-SVM (Table 4). This model was run using the radial basis function (RBF) and linear

Fig. 1. Original NIR spectra of 100 samples of jaboticaba.

461

N.C.T. Mariani et al. / Food Chemistry 159 (2014) 458–462

Table 2 Averaged performance of PLS models developed for raw spectra and various transformed spectra based on experimental data set. Root mean square error of cross validation (RMSECV) and prediction (RMSEP) and coefficient of correlation (R2). The number of PLS factors are represented in parentheses. Modelsa

PLS PLS PLS PLS PLS PLS a A

Calibration

(7), original (8), smoothing 3 pts, second derivate 3 pts, MSC (5), smoothing 9 pts, derivate 3 pts, MSC (9), original, MSC (6), smoothing 9 pts, second-derivate 3 pts, MSC (7)A, Original

Prediction

RPD

R2

RMSECV (%)

R2

RMSEP (%)

0.55 0.99 0.99 0.59 0.99 0.58

1.93 0.12 0.17 1.21 1.01 1.64

0.62 0.50 0.55 0.62 0.49 0.71

1.55 1.85 2.46 1.27 1.40 1.33

1.41 1.18 0.89 1.73 1.57 1.65

Pts, points. One application of outliers detection.

Fig. 2. Concentration predicted versus measured concentration by reference of calibration and validation samples for SSC in jaboticaba fruit using PLS model after outliers detection. ( ) calibration set; ( ) validation set.

Table 3 Averaged performance of iPLS, GA, and SPA models developed for SSC measurement in jaboticaba. Root mean square error of cross validation (RMSECV) and prediction (RMSEP) and coefficient of correlation (R2). The number of factors in iPLS, PLS-SPA and PLS-GA models are represented in parentheses. Modelsa

Calibration 2

iPLS (9), original PLS-SPA (9), original PLS-SPA (9), original PLS-GA (9), original a

Prediction 2

R

RMSECV (%)

R

0.67 0.61 0.60 0.60

1.71 1.84 1.27 1.84

0.51 0.60 0.62 0.61

RPD

RMSEP(%) 1.84 1.62 1.22 1.61

1.19 1.35 1.80 1.36

Pts, points.

kernels. Grid search and leave-one-out cross validation, generated the following values: 51.278 and 3.757 for c and 1.80e+5 and 7.01e+4 for r2 that were input into the LS-SVM RBF model. In LSSVM linear models, an acceptable pair of values was found at 3.05e 4 and 6.22e 5 for c. For the combination of c and r2 parameters, a RMSECV was calculated and the optimal parameters, which produced the smallest RMSECV, were selected. Alternatively, the LS-SVM model performed best when R2 was 0.44, RMSEP was 1.89 °Brix and RPD was 1.16.

3.6. Comparison analysis with conventional methods and other papers To compare conventional (destructive) methods and the PLS model (original spectra), the paired t-test method was applied. The calculated t value is 1.22, while the value of t (0.05, 30) is 2.04. The result of the paired t-test (t = 1.22 < t(0.05,30) = 2.04)

proves that there was no significant difference between the two methods at the 95% confidence interval. In this study, the PLS model results for SSC in intact jaboticaba fruit were R2 = 0.71, RMSECV = 1.64 °Brix and RMSEP = 1.33 °Brix. The majority of the studies involving SSC prediction with apples presented a variation of 0.5 °Brix. However, few were externally validated against data from fruits collected from different orchards and/or seasons, which could lead to higher RMSEP values 1.0 and 1.5 °Brix (Nicolaï et al., 2007). Reasonable models for SSC (root mean squared error of prediction, RMSEP = 0.44 °Brix and coefficient of determination, R2pre = 0.60) were obtained in the spectral range 780–1700 nm and the number of latent variables was 14. Martinsen and Schaare (1998) achieved an error in prediction of 1.2 °Brix over a range of 4.7–14.1 °Brix of kiwifruit with reflectance spectra from 650 nm to 1100 nm. Furthermore, Sun et al. (2009) focused on evaluating the use of Vis/NIR spectroscopy for measuring SSC of intact ‘Cuigan’ pears on-line. The best model for SSC of this fruit species was partial least square (PLS) regression of the original spectra, where the calculated R2 and RMSEP were 0.91% and 0.53%, respectively. In addition, Huang et al. (2011) has applied near-infrared spectroscopy in the determination of SSC in mulberry. PLS, least squares supporting vector machines (LS-SVM) and multiple linear regression (MLR) were used to calibrate the model and a successive projections algorithm (SPA) was used for the informed selection of variables. The R2pre for PLS, LS-SVM and MLR values were 0.77, 0.74 and 0.93, respectively. Liu et al. (2010) measured the SSC of navel orange fruit using a Vis/NIR spectrometric technique in conjunction with PLS and a back-propagating neural network (BPNN) based on principal component analysis (PCA). The best predicted results for correlation coefficient, RMSEP, and average difference between predicted and measured biases

462

N.C.T. Mariani et al. / Food Chemistry 159 (2014) 458–462

Table 4 Averaged performance of BP-ANN and LS-SVM models developed for SSC measurement in jaboticaba. Root mean square error of cross validation (RMSECV) and prediction (RMSEP) and coefficient of correlation (R2). The number of factors in models are represented in parentheses. Modelsa

BP-ANN (5), original BP-ANN-iPLS (6), original BP-ANN-iPLS (4), original BP-ANN-iPLS (3), original LS-SVM-RBF, smoothing 3 LS-SVM-linear, smoothing LS-SVM-RBF, smoothing 3 LS-SVM-linear, smoothing

Calibration

pts, second-derivate 3 3 pts, second-derivate pts, second-derivate 9 3 pts, second-derivate

pts, MSC 3 pts, MSC pts, MSC 9 pts, MSC

Prediction

RPD

R2

RMSECV (%)

R2

RMSEP (%)

0.60 0.48 0.45 0.53 0.94 0.94 0.82 0.83

1.86 1.30 1.30 1.99 1.29 1.25 1.68 1.64

0.62 0.65 0.68 0.63 0.44 0.44 0.45 0.45

1.57 1.20 1.20 1.48 1.89 1.89 2.74 2.73

were 0.90, 0.68 °Brix and 0.16 °Brix, respectively. Jha et al. (2012) applied NIR spectroscopy and multivariate calibration to find the SSC and pH of whole mangos. They developed models based on MLR and PLS, applying multiple correlation coefficients (0.782 and 0.762 for total soluble solids; 0.715 and 0.703 for pH) for calibration and validation. 4. Conclusion NIR is a promising technique for the non-destructive quantification of SSC content in jaboticaba fruit. The results tended to be overfit, depending on the initial parameters, and repeated network training was required to generate non-unique solutions. In contrast, LS-SVM produced a global model capable of efficiently addressing non-linear relationships. The prediction performance, however, could not be improved using the LS-SVM model. These results suggest that the combination of NIR spectroscopy and a chemometric approach offers the best method to predict SSC in jaboticaba fruits. Acknowledgements The authors would like to thank the Graduate Program in Chemistry (PPGQ) of UFRN and FAPERN (proc. 005/2012) for supporting this research. The authors also acknowledge the PróReitoria de Pesquisa da Universidade de São Paulo for partially sponsorship of this work (Novos Docentes proc. 10.1.25403.1.1 and 2011.1.6858.1.8), FAPESP (proc. 2008/51408-1) and CNPq (477386/2011-3). References Andersen, O., & Andersen, V. U. (1988). As Fruteiras Silvestres Brasileiras (2nd ed.). Rio de Janeiro: Editora Globo. AOAC (1997). Official methods of analysis of AOAC international (16th ed). Gaithersburg: Ed. Patricia Cunniff. Araújo, M. C. U., Saldanha, T. C. B., Galvão, R. K. H., Yoneyama, T., Chame, H. C., & Visani, V. (2001). The successive projections algorithm for variable selection in spectroscopic multicomponent analysis. Chemometrics and Intelligent Laboratory Systems, 57, 65–73. Ferrand, M., Huquet, B., Barbey, S., Barillet, F., Faucon, F., Larroque, H., et al. (2011). Determination of fatty acid profile in cow’s milk using mid-infrared spectrometry: Interest of applying a variable selection by genetic algorithms

1.40 1.83 1.83 1.48 1.16 1.16 0.80 0.80

before a PLS regression. Chemometrics and Intelligent Laboratory Systems, 106, 183–189. He, Y., Zhang, Y., Pereira, A. G., Gómez, A. H., & Wang, J. (2005). Nondestructive determination of tomato fruit quality characteristics using VIS/NIR spectroscopy technique. International Journal of Information Technology, 11, 97–108. Hertz, J., Krogh, A., & Palmer, R. (1991). Introduction to the theory of neural computation. USA: Addison Wesley. Huang, L., Wu, D., Jin, H., Zhang, J., He, Y., & Lou, C. (2011). Internal quality determination of fruit with bumpy surface using visible and near infrared spectroscopy and chemometrics: A case study with mulberry fruit. Biosystems Engineering, 109, 377–384. Jha, S. N., Jaiswal, P., Narsaiah, K., Gupta, M., Bhardwaj, R., & Singh, A. K. (2012). Nondestructive prediction of sweetness of intact mango using near infrared spectroscopy. Scientia Horticulturae, 138, 171–175. Kennard, R. W., & Stone, L. A. (1969). Computer aided design of experiments. Technometrics, 11, 137–148. Liu, Y., Sun, X., & Ouyang, A. (2010). Nondestructive measurement of soluble solid content of navel orange fruit by visible-NIR spectrometric technique with PLSR and PCA-BPNN. LWT – Food Science and Technology, 43, 602–607. Martinsen, P., & Schaare, P. (1998). Measuring soluble solids distribution in kiwifruit using near-infrared imaging spectroscopy. Postharvest Biology and Technology, 14, 271–281. Nicolaï, B. M., Beullens, K., Bobelyn, E., Peirs, A., Saeys, W., Theron, K. I., et al. (2007). Nondestructive measurement of fruit and vegetable quality by means of NIR spectroscopy: A review. Postharvest Biology and Technology, 46, 99–118. Nicolai, B. M., Verlinden, B. E., Desmet, M., Saevels, S., Theron, K., Cubeddu, R., et al. (2008). Time-resolved and continuous wave NIR reflectance spectroscopy to predict firmness and soluble solids content of Conference pears. Postharvest Biology and Technology, 47, 68–74. Norgaard, L., Saudland, A., Wagner, J., Nielsen, J. P., Munck, L., & Engelsen, S. B. (2000). Interval partial least-squares regression (iPLS): A comparative chemometric study with an example from near-infrared spectroscopy. Applied Spectroscopy, 54, 413–419. Sun, T., Lin, H., Xu, H., & Ying, Y. (2009). Effect of fruit moving speed on predicting soluble solids content of ‘Cuiguan’ pears (Pomacea pyrifolia Nakai cv. Cuigan) using PLS and LS-SVM regression. Postharvest Biology and Technology, 51, 86–90. Suykens, J. A. K., Van Gestel, T., De Brabanter, J., De Moor, B., & Vandewalle, J. (2002). Least squares support vector machines. Singapore: World Scientific Publishing Company. Teixeira, G. H. A., Durigan, M. F. B., & Durigan, J. F. (2011a). Jaboticaba (Myrciaria cauliflora (Mart.) O.Berg. [Myrtaceae]. In E. M. Yahia & J. K. Brecht (Eds.), Postharvest biology & technology of tropical and sub-tropical fruits (Vol. 3, pp. 246–274). Cambridge: Woodhead Publishing Limited. Teixeira, G. H. A., Durigan, J. F., Santos, L. O., Hojo, E. T. D., & Cunha, L. C. J. (2011b). Changes in the quality of jaboticaba (Myriciaria jabuticaba (Vell) Berg. cv. Sabará) stored under controlled atmosphere. Journal of the Science of Food and Agriculture, 91(15), 2844–2849. Valderrama, P., Braga, J. W. B., & Poppi, R. J. (2007). Variable selection, outlier detection, and figures of merit estimation in a partial least-squares regression multivariate calibration model. A case study for the determination of quality parameters in the alcohol industry by near-infrared spectroscopy. Journal of Agricultural and Food Chemistry, 55, 8331–8338.