Reflectance Vis/NIR spectroscopy for nondestructive taste characterization of Valencia oranges

Computers and Electronics in Agriculture 85 (2012) 64–69

Contents lists available at SciVerse ScienceDirect

Computers and Electronics in Agriculture journal homepage: www.elsevier.com/locate/compag

Reflectance Vis/NIR spectroscopy for nondestructive taste characterization of Valencia oranges Bahareh Jamshidi a, Saeid Minaei a,⇑, Ezzedin Mohajerani b, Hassan Ghassemian c a

Agricultural Machinery Engineering Department, Faculty of Agriculture, Tarbiat Modares University, Tehran, Iran Laser and Plasma Research Institute, Shahid Beheshti University, Tehran, Iran c Faculty of Electrical and Computer Engineering, Tarbiat Modares University, Tehran, Iran b

a r t i c l e

i n f o

Article history: Received 14 March 2011 Received in revised form 29 December 2011 Accepted 12 March 2012

Keywords: Nondestructive Vis/NIR spectroscopy Valencia orange Taste BrimA

a b s t r a c t The feasibility of reflectance Vis/NIR spectroscopy was investigated for taste characterization of Valencia oranges based on taste attributes including soluble solids content (SSC) and titratable acidity (TA), as well as taste indices including SSC to TA ratio (SSC/TA) and BrimA. The robustness of multivariate analysis in terms of prediction was also assessed. Several combinations of various preprocessing techniques with moving average and Savitzky–Golay smoothing filters, standard normal variate (SNV) and multiplicative scatter correction (MSC) were used before calibration and the models were developed based on both partial least squares (PLS) and principle component regression (PCR) methods. The best models obtained with PLS method had root mean square errors of prediction (RMSEP) of 0.33 °Brix, 0.07%, 1.03 and 0.37, and prediction correlation coefficients (rp) of 0.96, 0.86, 0.87 and 0.92 for SSC, TA, SSC/TA, and BrimA, respectively. It was concluded that Vis/NIR spectroscopy combined with chemometrics could be an accurate and fast method for nondestructive prediction of taste attributes and indices of Valencia oranges. Moreover, the application of this technique was suggested for taste characterization, directly based on BrimA which is the best index related to fruit flavor rather than determination of SSC or TA alone. Ó 2012 Elsevier B.V. All rights reserved.

1. Introduction Taste is one of the most important quality attributes of fruits and vegetables in fresh consumption. Nowadays, number of consumers who attend to the better taste of fruit more than the price is increasing. Fruit taste is composed of sour and sweet taste which is related to titratable acidity (TA) and soluble solids content (SSC), respectively and a distinctive taste which is characteristic for the variety provided by other chemicals (Abu-Khalaf and Bennedsen, 2004). SSC and TA which are the most important attributes of fruit taste are still largely measured destructively and there is a need for development of a reliable and nondestructive method to predict fruit taste. Near infrared (NIR) or visible/near infrared (Vis/NIR) spectroscopy in reflectance or transmittance mode is one of the applicable nondestructive techniques that has been used to measure SSC, TA or other quality attributes of smooth fruits and vegetables such as apple (Abu-Khalaf and Bennedsen, 2004; Zude et al., 2006), apricot (Bureau et al., 2009; Camps and Christen, 2009), bell pepper (Penchaiya et al., 2009), grape (Cao et al., 2010; Jarén et al., 2001), kiwifruit (Lü et al., 2010; McGlone and Kawano, 1998; Slaughter and Crisosto, 1998), mango (Saranwong et al., 2004; ⇑ Corresponding author. Tel.: +98 2148292466; fax: +98 2144180537. E-mail address: [email protected] (S. Minaei). 0168-1699/$ - see front matter Ó 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.compag.2012.03.008

Subedi et al., 2007), peach (Golic and Walsh, 2006), tomato (Flores et al., 2009; He et al., 2005; Pedro and Ferreira, 2007). Transmittance is not a suitable mode for assessing large produce because the amount of light penetrating the fruit is often very small. This makes it difficult to obtain accurate transmission measurements at grading line speeds. Reflectance mode is the easiest mode to obtain measurements, because they require no contact with the fruit and light levels are relatively high (Cayuela, 2008). Kawano et al. (1993) noted that good results can be obtained in the fruits with a thin peel using reflectance mode. However, it is more difficult to achieve acceptable prediction of internal quality attributes for rough and thick-skinned fruits such as citrus as compared to the smooth and thin ones because of more light scattering and skin thickness effects. In spite of that, good prediction results have been reported on SSC of citrus using reflectance Vis/NIR spectroscopy (Cayuela, 2008; Gómez et al., 2006; Liu et al., 2010) but not on TA because of the relatively low levels of organic acids in citrus (Cayuela, 2008; Guthrie et al., 2005; McGlone et al., 2003). Moreover, the actual taste of each fruit is more closely related to the composite taste indices based on SSC and TA such as SSC to TA ratio (SSC/TA) which is currently used as a sensory acceptability index of fruit taste. One of the drawbacks associated with use of this index is that different levels of SSC and TA may result in the same ratio. Therefore, a new index named BrimA which is expressed as (SSC  k  TA) was proposed by Jordan et al. (2001). The constant

B. Jamshidi et al. / Computers and Electronics in Agriculture 85 (2012) 64–69

(k) reflects the tongue’s higher sensitivity to TA compared to SSC. The index allows smaller amounts of TA than SSC to make the same numerical change to BrimA and in the opposite direction (Jordan et al., 2001). Therefore, it can be more useful to assess the correlation of these indices with Vis/NIR spectrum than solely SSC or TA to characterize fruit taste. Moreover, there is no report on NIR or Vis/ NIR spectroscopy predictions of taste indices. This research aims to use reflectance Vis/NIR spectroscopy for determination of taste attributes (SSC, TA) and to assess the feasibility of using the technique for prediction of taste indices (SSC/TA, BrimA) of Valencia oranges directly and nondestructively. In addition, different multivariate analysis models including PLS and PCR with various combinations of preprocessing methods were also investigated. 2. Material and methods 2.1. Fruit samples A total of 120 Valencia oranges used for the experiments were purchased from different local markets in different days (approximately 10 samples each time) to obtain enough variability in the range of the samples. The fruits were stored at 25 °C for 1 day before Vis/NIR measurements to reach equilibrium temperature with laboratory environment. The weight and maximum equatorial diameter of each sample were measured before spectroscopy. All reference measurements were carried out on the next day. 2.2. Vis/NIR spectroscopy set up and measurements The spectroscopy equipment consisted of a USB2000 spectrometer with charge coupled device (CCD) detector (Oceanoptics Inc., USA), tungsten halogen lamp (12 V–50 W) as a light source, an optical fiber (SMA905) with diameter of 0.4 mm, and a sample holder. The wavelength range of the spectrometer was 350–1050 nm. The fiber optic was placed vertically at a height of 100 mm from the holder. The light source was set up about 200 mm above the holder center at an angle of 45°. Before spectra acquisition, a white reference was used to obtain the relative reflectance. For each sample, reflectance spectra at three positions around equatorial locations (approximately 120°) and six scans at each position were collected by OOIBase32 software (Oceanoptics Inc., USA). Next, the mean spectrum was calculated from a total of 18 scans for each fruit. The acquired spectra were stored for later data analysis. 2.3. Taste reference analysis After spectroscopy measurements, each sample was peeled and the juice was extracted. SSC was measured using a digital refractometer DR-A1 (Atago Co. Ltd., Japan). Moreover, TA was determined by titration of 10 ml sample juice with 0.1 M NaOH using phenolphthalein as an indicator and expressed as percentage of citric acid. Finally, the ratio of SSC/TA and BrimA (SSC  k  TA) indices were calculated. The constant (k) of four suggested by Obenland et al. (2009), which provided the highest correlation between the calculated BrimA and the hedonic flavor scores given by the panelists for navel oranges, was used to determine the BrimA index. 2.4. Chemometrics and data analysis 2.4.1. Preprocessing The overall mean reflectance spectra of fruits were converted to absorbance values (log 1/R) to obtain linear correlation between the spectrum and the sample molecular concentration. Also, the

65

first 50 and the last 50 wavelength values were ignored because of the existence of considerable noise in these regions. Thus, all the spectra from 400 to 1000 nm were considered for preprocessing. The moving average filter with segment size of five for averaging and a Savitzky–Golay filter with five smoothing points and polynomial order of two were used to smooth and remove random noise from Vis/NIR spectra. First derivative of the spectra based on Savitzky–Golay algorithm was also used to increase the spectral resolution and to interpret the spectra. To correct both multiplicative and additive effects of the spectra due to scattering, particle size, and the change of light distance (Cen and He, 2007), the spectra were preprocessed using multiplicative scatter correction (MSC) and standard normal variate correction (SNV). MSC attempts to remove these effects by linearizing each spectrum to the average spectrum. Moreover, in SNV, each individual spectrum is normalized to zero mean and unit variance (Nicolaï et al., 2007). All preprocessing methods were conducted by the Unscrambler software v9.7 (CAMO Software AS, Norway). 2.4.2. Multivariate regression models Before developing determination models, the sample data was randomly divided to calibration set (75% of the whole samples) and prediction set (25% of the whole samples). Then, principal component regression (PCR) and partial least squares (PLS) with cross validation based on 20 segments were used for obtaining the calibration models. Finally, prediction set was used to evaluate the robustness of the models. PCR decomposes the spectral matrix X (X-variables) by a principal component analysis (PCA) and then fits a multiple linear regression (MLR) model, using a small number of principal components (PCs) or latent variables (LVs) instead of the original variables as predictors. However, in PLS regression an orthogonal basis of latent variables is constructed one by one in such a way that they are oriented along directions of maximal covariance between the spectra matrix X and the response value y (Y-variable such as SSC and TA) (Nicolaï et al., 2007). The number of LVs investigated for both PLS and PCR models were 10 to avoid over-fitting and noise modeling. The outliers due to chemical measurement errors were determined as large values on the influence plot. After removing the outliers (three samples), the sets of calibration and prediction were reconstructed and the models were developed again. The Unscrambler software v9.7 (CAMO Software AS, Norway) was also used for multivariate analysis. 2.4.3. Evaluation of the models The calibration models were mainly assessed in terms of correlation coefficient of calibration and cross validation (rc and rcv), as well as root mean squared error of calibration and cross validation (RMSEC and RMSECV). In assessing the soundness of calibration performance, the main considerations were correlation coefficient of prediction (rp) and root mean squared error of prediction (RMSEP). RMSEC, RMSECV or RMSEP are defined as follows (Nicolaï et al., 2007):

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Pnc 2 ^ i¼1 ðyi  yi Þ RMSEC ¼ nc sP ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi np 2 ^ i¼1 ðyi  yi Þ RMSECV or RMSEP ¼ np where nc is the number of observations in calibration set and np is ˆi and yi are the predicted and the number of validated objects. y measured values of the ith observation, respectively.

66

B. Jamshidi et al. / Computers and Electronics in Agriculture 85 (2012) 64–69

Table 1 Statistics of both calibration and prediction sets for morphological properties, taste attributes and taste indices. Calibration set (88 samples)

Diameter (mm) Weight (g) SSC (°Brix) TA (%) SSC/TA BrimA

Prediction set (29 samples)

Min

Max

Mean

SD

Min

Max

Mean

SD

69.000 186.090 7.900 0.538 8.681 4.583

87.900 283.290 11.300 1.088 17.113 8.535

76.175 229.014 9.489 0.807 12.142 6.260

0.309 23.220 0.815 0.151 2.352 0.940

72.400 198.120 7.600 0.538 9.153 4.928

87.00 282.000 11.500 1.082 17.241 8.454

77.662 241.126 9.348 0.737 12.988 6.400

0.336 22.980 0.914 0.135 2.111 0.856

The optimal number of LVs was determined by selection of the minimum on RMSECV versus LVs plot. 3. Results and discussion 3.1. Statistics of the samples Table 1 shows the statistics of morphological properties, taste attributes and indices of Valencia oranges for both calibration and prediction sets. The weight and the maximum equatorial diameter of the whole samples were in the range of 186.09– 283.29 g and 69–87.9 mm, respectively. Therefore, the samples were quite varied in terms of morphology which was the main reason to use normalizing methods (MSC and SNV) and correct the multiplicative and additive effects on the spectra. The range of SSC and TA were from 7.6 to 11.5 °Brix and 0.538% to 1.088%, respectively. The levels and standard deviation of TA were low which is normal in citrus varieties. Moreover, the range of SSC/ TA and BrimA were 8.681–17.241 and 4.583–8.535, respectively. 3.2. Vis/NIR spectra interpretation Fig. 1(a) and (b) show the typical absorbance spectrum of Valencia orange and its first derivative in the wavelength range of 400–1000 nm, respectively. The spectrum had some absorbance peaks in specific frequencies due to stretching vibration of the overtones of O–H, C–H or N–H functional groups relative to the concentration of some inner compositions with these bands such as sugars and acids. The absorbance in the range of 400–500 nm was due to the pigments. In the visible region (after 500 nm), the curve had decreasing trend and there was a perceptible peak around 675 nm because of the absorption of chlorophyll a similar to that described for mandarin by Gómez et al. (2006) and for Valencia orange by Cayuela (2008), respectively. Moreover, a broad peak around 750 nm was found which could be due to the third overtone of O–H and the forth overtone of C–H. In NIR region, the curve had increasing trend and a broad peak was found around 850 nm which could be due to the third overtone of C–H. There was also a perceptible peak around 970 nm because of the second overtone of O–H similar to that described by Cayuela (2008) and Gómez et al. (2006), and the second overtone of N–H. First derivative of the spectrum confirmed all above absorbance peaks. More details were also found in visible region regarding the absorbance around 650 and 550 nm due to chlorophyll b and other pigments, respectively. In spite of that, the noises were become sharp in the first and the end wavelength values using first derivative.

Fig. 1. Typical absorbance Vis/NIR spectrum of Valencia orange (a), and its first derivative (b).

3.3. Preprocessing and multivariate analysis

taste attributes but also taste indices. Although PCR models could predict SSC and BrimA as well as the PLS models but they did not give an accurate prediction of TA and SSC/TA. Moreover, all PLS models obtained better prediction results of taste attributes and indices than the PCR models. This result was similar to that reported for SSC determination of Satsuma mandarin with PLS and PCR models and other different preprocessing techniques by Gómez et al. (2006). This could be due to the drawback of PCR which uses the first irrelevant PCs for the regression model (Wold et al., 2001). Therefore, PLS was selected as the best model for characterization of all taste parameters studied.

The results of calibration and prediction of both PLS and PCR models with several combinations of preprocessing methods for SSC, TA, SSC/TA, and BrimA are also summarized in Table 2. Results indicated that PLS models had the potential to estimate not only

3.3.1. Soluble solids content (SSC) PLS models could predict SSC as well and better than the other taste characteristics. Withal, in each smoothing method, the PLS model with different normalizing techniques was evaluated and

67

B. Jamshidi et al. / Computers and Electronics in Agriculture 85 (2012) 64–69 Table 2 The results of calibration and prediction of PLS and PCR models with several combinations of preprocessing methods for taste attributes and indices. Preprocessing

SSC (°Brix)

Smoothing

Normalizing

Savitzky–Golay

MSC SNV

Moving average

MSC SNV

TA (% citric acid)

Savitzky–Golay

MSC SNV

Moving average

MSC SNV

SSC/TA

Savitzky–Golay

MSC SNV

Moving average

MSC SNV

BrimA

Savitzky–Golay

MSC SNV

Moving average

MSC SNV

Model

LVs

Calibration set

Prediction set

RMSEC

rc

RMSECV

rcv

RMSEP

rp

PLS PCR PLS PCR PLS PCR PLS PCR PLS PCR PLS PCR PLS PCR PLS PCR PLS PCR PLS PCR PLS PCR PLS PCR PLS PCR PLS PCR PLS PCR PLS PCR

5 5 5 7 4 6 4 6 9 9 9 9 8 9 8 9 9 9 9 9 8 9 9 9 9 9 9 9 8 9 9 9

0.328 0.348 0.329 0.344 0.349 0.351 0.351 0.352 0.093 0.119 0.094 0.119 0.095 0.119 0.095 0.119 1.413 1.790 1.414 1.802 1.422 1.787 1.361 1.792 0.432 0.529 0.425 0.530 0.423 0.529 0.421 0.529

0.914 0.902 0.913 0.905 0.902 0.901 0.901 0.900 0.786 0.611 0.783 0.610 0.777 0.614 0.774 0.612 0.797 0.641 0.797 0.638 0.794 0.645 0.813 0.642 0.887 0.824 0.890 0.823 0.892 0.824 0.893 0.825

0.361 0.379 0.367 0.378 0.377 0.381 0.382 0.385 0.105 0.130 0.106 0.132 0.106 0.130 0.108 0.131 1.620 1.998 1.638 2.005 1.635 2.011 1.576 1.973 0.546 0.587 0.534 0.590 0.532 0.602 0.529 0.600

0.895 0.884 0.892 0.885 0.885 0.882 0.882 0.881 0.723 0.509 0.715 0.499 0.720 0.521 0.708 0.510 0.727 0.531 0.720 0.527 0.726 0.526 0.743 0.544 0.816 0.779 0.823 0.777 0.824 0.768 0.827 0.769

0.331 0.344 0.331 0.332 0.356 0.341 0.370 0.343 0.070 0.102 0.070 0.103 0.070 0.104 0.071 0.104 1.039 1.673 1.049 1.686 1.052 1.690 1.026 1.697 0.407 0.509 0.370 0.512 0.379 0.507 0.370 0.512

0.958 0.954 0.958 0.955 0.956 0.955 0.953 0.954 0.858 0.652 0.858 0.652 0.857 0.639 0.852 0.642 0.875 0.608 0.871 0.605 0.870 0.597 0.871 0.598 0.899 0.851 0.917 0.852 0.911 0.857 0.918 0.858

Bold values indicate the best developed model for each taste attribute or index.

found that MSC gave slightly better results than SNV method. This is in agreement with the results reported by Liu et al. (2010) for SSC prediction of navel orange fruit with PLS model preprocessed using MSC and SNV without any smoothing. Not only the calibration models and normalizing methods but also the different smoothing techniques influenced the results for SSC and the obtained result using Savitzky–Golay smoothing was better against moving average filter. Therefore, the PLS model with combination of Savitzky–Golay smoothing filter and MSC normalization for preprocessing can yield better prediction of SSC with RMSEC = 0.328, RMSECV = 0.361, RMSEP = 0.331, rc = 0.914, rcv = 0.895, rp = 0.958, and number of LVs = 5. The scatter plot of correlation between the measured and predicted values of SSC for the best model is shown in Fig. 2a. The prediction result of SSC in this research was in agreement with those reported by Guthrie et al. (2005) for Imperial mandarin in the 720–950 nm region and interactance mode using modified PLS (MPLS) with derivate and scatter correction preprocessing methods (RMSEC of 0.75). Moreover, these results were better than those achieved by Gómez et al. (2006) for Satsuma mandarin in 400–2350 nm region and reflectance mode using PLS and PCR models with moving average and MSC pretreatment (RMSEP = 0.33, rp = 0.94), and by Liu et al. (2010) for navel orange in 350–1800 nm region and reflectance mode using principle component analysis-back propagation (PCA-BPNN) model preprocessed using MSC and SNV without any smoothing (RMSEP = 0.68, rp = 0.90). RMSEP value obtained in this research for SSC was better than that reported by Cayuela (2008) for Valencia orange in 580–1850 nm and reflectance mode using PLS model preprocessed by SNV method (RMSEP = 0.51, r2c ¼ 0:91), and was similar to those reported by Kawano et al. (1993) for Satsuma mandarin in 680– 1235 nm and transmittance mode (SEP = 0.32, r = 0.989), and by

McGlone et al. (2003) for Satsuma mandarin in 700–930 nm and transmittance mode using PLS model (RMSEP = 0.32, r2c ¼ 0:93) although they achieved better values for r or r 2c . However, a comparison is not recommended due to the variety of spectrometers and spectral ranges used in these researches. As described by Cayuela (2008), oranges are larger and thicker than mandarin fruit and have a different structure. Therefore, histological differences such as skin thickness, pulp texture and composition, segment number and seediness between citrus cultivars and varieties may influence spectroscopy measurements.

3.3.2. Titratable acidity (TA) The smoothing and normalizing methods positively influenced the results of PLS models for TA prediction as MSC was slightly better than SNV when combined with each smoothing technique. However, the PLS model preprocessed with the combination of Savitzky–Golay filter and MSC method with RMSEC = 0.093, RMSECV = 0.105, RMSEP = 0.07, rc = 0.786, rcv = 0.723, rp = 0.858, and number of LVs = 9 was preferred. The scatter plot of correlation between the measured and predicted values of TA for the best model is shown in Fig. 2b. The prediction result of TA was not as accurate as the results of SSC prediction because of low levels and standard deviation of TA (SD = 0.135%) in the samples similar to the results reported by Cayuela (2008) for Valencia orange (RMSEP = 0.331, r2c ¼ 0:637), and by McGlone et al. (2003) for Satsuma mandarin (RMSEP = 0.15, r 2c ¼ 0:65). However, the RMSEP obtained for TA prediction in this research was better (RMSEP = 0.07). Moreover, the developed model for predicting TA was superior compared to that developed by Guthrie et al. (2005) for Imperial mandarin (RMSEP = 0.2, r2c ¼ 0:3). As mentioned before, it is not recommended to compare the results

68

B. Jamshidi et al. / Computers and Electronics in Agriculture 85 (2012) 64–69

(a)

Elements: 29 Correlation: 0.957808 RMSEP: 0.331672 SEP: 0.334525

Predicted TA (%)

Predicted SSC (°Brix)

12

(b)

Slope: 1.134188 Offset: -1.298669 Bias: -0.044246

10

8

Elements: 29 Correlation: 0.857799 RMSEP: 0.070605 SEP: 0.071166

Slope: 0.837750 Offset: 0.129349 Bias: 0.009754

1 0.8 0.6 0.4

6 6

7

8

9

10

11

0.5

12

0.6

Measured SSC (°Brix)

(c)

Elements: 29 Correlation: 0.871156 RMSEP: 1.026404 SEP: 1.039832

Slope: 0.725577 Offset: 3.466382 Bias: -0.097668

0.7

0.8

0.9

1

1.1

Measured TA (%)

(d)

Elements: 29 Correlation: 0.918436 RMSEP: 0.370143 SEP: 0.362980

Slope: 0.983650 Offset: 0.005668 Bias: -0.097668

Predicted BrimA

Predicted SSC/TA

9 15

12

9

8 7 6 5 4

9

10 11 12 13 14 15 16 17 18

4

5

6

7

8

9

Calculated BrimA

Calculated SSC/TA

Fig. 2. Vis/NIR prediction results for SSC (a) and TA (b) with PLS model preprocessed using Savitzky–Golay smoothing filter + MSC normalization, and for SSC/TA (c) and BrimA (d) with PLS model preprocessed using moving average smoothing filter + SNV normalization.

achieved from different wavelength ranges, instruments and citrus varieties. 3.3.3. SSC to TA ratio (SSC/TA) For SSC/TA index, when Savitzky–Golay filter was used for smoothing, MSC was preferable against SNV normalizing method to develop the PLS model. In contrast, the moving average filter combined with SNV obtained better results of PLS prediction than using MSC technique. However, between these two preprocessing combinations, the PLS model preprocessed with moving average filter combined with SNV method had slightly better prediction results as RMSEC = 1.361, RMSECV = 1.576, RMSEP = 1.026, rc = 0.813, rcv = 0.743, rp = 0.871, and number of LVs = 9. The scatter plot of the correlation between calculated (from measured SSC and TA) and predicted SSC/TA index using the best selected model is shown in Fig. 2c. Results indicated that Vis/NIR spectroscopy had the potential to predict SSC/TA directly as accurate as it predicts TA. The reason is that SSC/TA depends on SSC and TA, as well as their chemical bands such as C-H and O-H groups. On the other hand, when TA (acidity) is low, SSC/TA becomes high and because of its relation to SSC, has a higher standard deviation than TA which is good for modeling. 3.3.4. BrimA index Results indicated that PLS models can predict BrimA well. In terms of preprocessing methods, SNV was better against MSC to predict BrimA for each smoothing technique. Nevertheless, better results were achieved using a combination of moving average filter

and SNV for predicting BrimA by PLS model. Fig. 2d shows the scatter plot of calculated (from measured SSC and TA) and predicted BrimA when the best model with RMSEC = 0.421, RMSECV = 0.529, RMSEP = 0.370, rc = 0.893, rcv = 0.827, rp = 0.918, and number of LVs = 9 was used. It was noted that BrimA as a taste index based on SSC and TA is related to chemical bands (just a SSC/TA) which makes it possible to predict taste using NIR spectroscopy, directly. Moreover, the accuracy of BrimA prediction was much better than that of TA and SSC/TA. BrimA is a better taste index than SSC/TA because it is more closely related to flavor especially at low levels of TA as reported by Jordan et al. (2001) and Obenland et al. (2009), and the good prediction results obtained in our research, it is concluded that Vis/NIR spectroscopy can be used for taste characterization directly and nondestructively. 4. Conclusions The feasibility of utilizing Vis/NIR spectroscopy using CCD spectrometer combined with different preprocessing methods to nondestructively characterize the taste of Valencia oranges was investigated. Taste attributes included SSC and TA as well as SSC/ TA and BrimA indices. The prediction models were developed using both PCR and PLS with four different combinations of smoothing (moving average and Savitzky–Golay filters) and normalizing (MSC and SNV) techniques. Results indicated that all models, especially PLS, had the potential to estimate taste characteristics and that preprocessing methods influenced the results. The best obtained models for SSC (RMSEP = 0.331, rp = 0.958) and TA

B. Jamshidi et al. / Computers and Electronics in Agriculture 85 (2012) 64–69

(RMSEP = 0.07, rp = 0.858) with Savitzky–Golay + MSC, and for SSC/ TA (RMSEP = 1.026, rp = 0.871) and BrimA (RMSEP = 0.37, rp = 0.918) with moving average + SNV showed the good prediction ability of Vis/NIR spectroscopy. More accurate predictions were achieved for SSC and BrimA compared to TA and SSC/TA. A noticeable result was the capability of this technique to predict orange fruit taste based on BrimA index directly and nondestructively which could be used for taste characterization before or after harvesting.

Acknowledgments This research was supported by Tarbiat Modares University and Laser & Plasma Institute of Shahid Beheshti University. The authors are grateful to Mrs. Afkhami for her help in spectral measurements.

References Abu-Khalaf, N., Bennedsen, B.S., 2004. Near infrared (NIR) technology and multivariate data analysis for sensing taste attributes of apples. Int. Agrophys. 18, 203–211. Bureau, S., Ruiz, D., Reich, M., Gouble, B., Bertrand, D., Audergon, J.M., Renard, C.M.G.C., 2009. Rapid and nondestructive analysis of apricot fruit quality using FT-near-infrared spectroscopy. Food Chem. 113, 1323–1328. Camps, C., Christen, D., 2009. Nondestructive assessment of apricot fruit quality by portable visible-near infrared spectroscopy. Food Sci. Technol. 42, 1125–1131. Cao, F., Wu, D., He, Y., 2010. Soluble solids content and pH prediction and varieties discrimination of grapes based on visible-near infrared spectroscopy. Comput. Electron. Agric. 71s, s15–s18. Cayuela, J.A., 2008. VIS/NIR soluble solids prediction in intact oranges (Citrus sinensis L.) cv. Valencia Late by reflectance. Postharvest Biol. Technol. 47, 75–80. Cen, H., He, Y., 2007. Theory and application of near infrared reflectance spectroscopy in determination of food quality. Trends Food Sci. Technol. 18, 72–83. Flores, K., Sánchez, M., Pérez-Marín, D., Guerrero, J., Garrido-Varo, A., 2009. Feasibility in NIRS instruments for predicting internal quality in intact tomato. J. Food Eng. 91, 311–318. Golic, M., Walsh, K., 2006. Robustness of calibration models based on near infrared spectroscopy for the in-line grading of stone fruit for total soluble solids content. Anal. Chim. Acta 555, 286–291. Gómez, A.H., He, Y., Pereira, A.G., 2006. Nondestructive measurement of acidity, soluble solids and firmness of Satsuma mandarin using VIS/NIR-spectroscopy techniques. J. Food Eng. 77, 313–319.

69

Guthrie, J.A., Walsh, K., Reid, D.J., Liebensberg, C.J., 2005. Assessment of internal quality attributes of mandarin fruit. 1. NIR calibration model development. Aust. J. Agric. Res. 56, 405–416. He, Y., Zhang, Y., Pereira, A.G., Go9 mez, A.H., Wang, J., 2005. Nondestructive determination of tomato fruit quality characteristics using VIS/NIR spectroscopy technique. Int. J. Info. Technol. 11 (11), 97–108. Jarén, C., Ortuño, J.C., Arazuri, S., Arana, J.I., Salvadores, M.C., 2001. Sugar determination in grapes using NIR technology. Int. J. Infrared Millimet. Waves 22 (10), 1521–1530. Jordan, R., Seelye, R., McGlone, A., 2001. A sensory-based alternative to brix/acid ratio. Food Technol. 55 (6), 36–44. Kawano, S., Fujiwara, T., Iwamoto, M., 1993. Determination of sugar content in satsuma mandarin using near infrared (NIR) transmittance. J. Jpn. Soc. Hort. Sci. 62 (2), 465–470. Liu, Y., Sun, X., Ouyang, A., 2010. Nondestructive measurement of soluble solids content of navel orange fruit by visible-NIR spectrometric technique with PLSR and PCA-BPNN. LWT-Food Sci. Technol. 43 (4), 602–607. Lü, Q., Tang, M., Cai, J., Lu, H., 2010. Long-term prediction of Zhonghua kiwifruit dry matter by near infrared spectroscopy. Sci. Asia 36, 210–215. McGlone, V.A., Kawano, S., 1998. Firmness, dry-matter and soluble solids assessment of postharvest kiwifruit by NIR spectroscopy. Postharvest Biol. Technol. 13, 131–141. McGlone, V.A., Fraser, D.G., Jordan, R.B., Künnemeyer, R., 2003. Internal quality assessment of mandarin fruit by Vis/NIR spectroscopy. J. Near Infrared Spectrosc. 11, 323–332. Nicolaï, B.M., Beullens, K., Bobelyn, E., Peirs, A., Saeys, W., Theron, K.I., Lammertyn, J., 2007. Nondestructive measurement of fruit and vegetable quality by means of NIR spectroscopy: a review. Postharvest Biol. Technol. 46, 99–118. Obenland, D., Collin, S., Mackey, B., Sievert, J., Fjeld, K., Arpaia, M.L., 2009. Determinants of flavor acceptability during the maturation of navel oranges. Postharvest Biol. Technol. 52, 156–163. Pedro, A.M.K., Ferreira, M.M.C., 2007. Simultaneously calibrating solids, sugars and acidity of tomato products using PLS2 and NIR spectroscopy. Anal. Chim. Acta. 595, 221–227. Penchaiya, P., Bobelyn, E., Verlinden, B.E., Nicolaï, B.M., Saeys, W., 2009. Nondestructive measurement of firmness and soluble solids content in bell pepper using NIR spectroscopy. J. Food Eng. 94, 267–273. Saranwong, S., Sornsrivichai, J., Kawano, S., 2004. Prediction of ripe-stage eating quality of mango fruit from its harvest quality measured nondestructively by near infrared spectroscopy. Postharvest Biol. Technol. 31, 137–145. Slaughter, D.C., Crisosto, C.H., 1998. Nondestructive internal quality assessment of kiwifruit using near-infrared spectroscopy. Semin. Food Anal. 3, 131–140. Subedi, P.P., Walsh, K.B., Owens, G., 2007. Prediction of mango eating quality at harvest using short-wave near infrared spectroscopy. Postharvest Biol. Technol. 43, 326–334. Wold, S., Sjöström, M., Eriksson, L., 2001. PLS-regression: a basic tool of chemometrics. Chemom. Intell. Lab. Syst. 58, 109–130. Zude, M., Herold, B., Roger, J.M., Bellon-Maurel, V., Landahl, S., 2006. Nondestructive tests on the prediction of apple fruit flesh firmness and soluble solids content on tree and in shelf life. J. Food Eng. 77, 254–260.