Food Chemistry 126 (2011) 1856–1861

Contents lists available at ScienceDirect

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

Analytical Methods

Infrared spectroscopy and chemometrics for the starch and protein prediction in irradiated rice Yongni Shao, Yilang Cen, Yong He ⇑, Fei Liu College of Biosystems Engineering and Food Science, Zhejiang University, 268 Kaixuan Road, Hangzhou 310029, China

a r t i c l e

i n f o

Article history: Received 12 December 2008 Received in revised form 2 October 2010 Accepted 30 November 2010 Available online 4 December 2010 Keywords: Infrared spectroscopy Starch Protein Independent component analysis Partial least squares Least squares-support vector machine

a b s t r a c t Infrared spectroscopy was investigated to predict components of starch and protein in rice treated with different irradiation doses based on sensitive wavelengths (SWs). Near infrared and mid-infrared regions were compared to determine which one produces the best prediction of components in rice after irradiation. Partial least-squares (PLS) analysis and least-squares-support vector machine (LS-SVM) were implemented for calibration models. The best PLS models were achieved with NIR region for starch and MIR region for protein. Moreover, independent component analysis (ICA) was executed to select several SWs based on loading weights, and the optimal LS-SVM model was achieved with SWs of 1210–1222, 1315–1330, 1575–1625, 1889–1909 and 2333–2356 nm for starch and SWs of 962–1091, 1232–1298, 1480–1497, 1584–1625 and 2373–2398 cm1 for protein. It indicated that IR spectroscopy combined with LS-SVM could be applied as a high precision way for the determination of starch and protein in rice after irradiation. Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction Rice is a staple food and the concern about rice is much greater than that for other food. However, investigations have revealed that rice would be affected by mould-damage and by insect-damage during the post-harvest storage. This extensive loss of rice has stimulated a great deal of research on minimising these damages, such as spraying insecticides of pyrethrins, malathion or fumigating rice with methyl bromide, ethylene oxide and hydrogen cyanide. However, these processes cause troublesome effects through the intake of hazardous residues of the contaminated insecticides and fumigating chemicals to the human body. Gamma-irradiation treatment has been applied in grain storage for the control of insect infestation, microbial contamination or the prevention of post-harvesting biological activities, e.g. ripening, germination and sprouting. Irradiation-treated grains may lead to the different changes of nutritional contents depending on the different irradiation doses. Generally, whether the grain is irradiated cannot be told by the obvious appearance indication. Some chemical methods have been investigated to inspect the grains under the irradiation treatment. But it always takes long time, high investment, cause some damage and also need specialist to do it. Infrared spectroscopy techniques, such as near infrared (NIRS) and midinfrared spectroscopy (MIRS), offer a quick analysis, non-destruc-

⇑ Corresponding author. Tel./fax: +86 571 86971143. E-mail address: [email protected] (Y. He). 0308-8146/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.foodchem.2010.11.166

tive and low cost. One of the primary advantages is its cause no damage to the environment or the area being viewed. Though it requiring a large amount of samples for the calibration, once the model is build, it only need to collect the reflectance or transmittance of the sample, and then the prediction results such as component content, disease rate and so on will obtained. It does not require any sort of special preparation for samples. Because of its quick analysis, it can finally realise the online using in the agricultural, industrial, environmental science particularly appropriate. There were less study reported in the literature using infrared spectroscopy to analyse different irradiation doses on rice, and the application of mid-infrared spectral region in cereals is rare. Delwiche and Grayboschf (2002) built the identification model of waxy wheat using NIRS. Kim, Rhyu, Kim, and Lee (2003) attempted to use NIRS to authenticate rice of Korean domestic and foreign rice. Baye, Pearson, and Settles (2006) used single kernel infrared spectroscopy to predict the maize seed composition. In this study, we want to find the sensitive wavelengths for starch and protein prediction of irradiated rice, in order to final develop portable instruments and online monitoring for commercial applications of quality detection such as starch and protein content for irradiated rice based on infrared spectroscopy technology combined chemometrics. Recently, a promising chemometrics called support vector machine (SVM) was proposed by Vapnik (1998a). SVM has a good theoretical foundation based on the statistical learning theory. It embodies the structural risk minimisation principle (SRM) instead of the traditional empirical risk minimisation principle (ERM)

1857

Y. Shao et al. / Food Chemistry 126 (2011) 1856–1861

employed by conventional neural network to avoid overfitting problems. SVM is used as a binary classification tool but also can be easily extended to regression tasks (Cogdill & Dardenne, 2004). Least-squares-support vector machine (LS-SVM) is the reformulation of the standard SVM simplified by Suykens (Suykens, Van Gestel, De Brabanter, De Moor, & Vandewalle, 2002). It adopts least-squares linear system as the loss function and is applied in the pattern recognition and non-linear evaluation. It is capable of learning in high-dimensional feature space with fewer training data. There have been some reports on the theories and applications of LS-SVM (Cogdill et al., 2004). Being a new chemometric method, only a few studies have been reported on the spectral analysis using LS-SVM (Borin, Ferrao, Mello, Maretto, & Poppi, 2006; Chauchard, Cogdill, Roussel, Roger, & Bellon-Maurel, 2004). The objectives of this study were (1) to investigate the feasibility of using IR spectroscopy to predict the starch and protein in rice after irradiation; (2) to compare the performance of different wavelength bands including NIR (1100–2500 nm) and MIR region (400–4000 cm1) by PLS analysis; (3) to evaluate the feasibility of using sensitive wavelengths (SWs) by SW-LS-SVM models. 2. Materials and methods 2.1. Rice samples Rice samples were unpolished (four rice types: Xie you, Tai you, Feng you and Yue you) and were irradiated with gamma-irradiation in a 60Co irradiator at Zhejiang University at a dosage rate of 2.5 KGy/h., with the following total doses: 0, 0.25, 0.5, 0.75, and 1 KGy. After irradiation, a total of 320 rice samples (80 samples with one rice type) were prepared by milling to rice flour using a flour mill, with 16 samples from each irradiation dose. All samples were divided into calibration sets of 260 samples (four rice types with each dose 52 samples) and prediction sets of 60 samples (four rice types with each dose 12 samples). In order to compare the performance of different calibration models, the samples in the calibration and prediction sets were kept unchanged for all calibration models. The quality parameters of starch and protein were obtained by the standard chemistry methods of GB5006-85 and Kjeldahl method. The statistic values of starch and protein of rice are shown in Table 1. 2.2. Spectra measurement In this study, two spectrometers were used to measure the rice samples: A Foss NIRSystems 6500 (NIRSystems, Silver Spring, MD, USA), with a spectral range of 1100–2500 nm, and a FT/IR-4100 Spectrometer (JASCO, Tokyo, Japan), with a spectral range of 7800–350 cm1. In NIR measurement, about 4 g of rice flour for each sample was scanned in duplicate in a small ring cup (NR7073, internal dia. 35 mm, depth 9 mm). Each spectrum represented the average of 32 scans, and was recorded as log (1/R) at 2 nm increments. For MIR measurement, each sample was mixed with potassium bromide (KBr) at the ratio of 1:49, and then the

mixture was compressed into slices by equipment. After that, samples were loaded on slide holder, and spectral curves were obtained in transmission mode. Duplicates of each sample were scanned twice (rotating the ring cup to a different position). The average spectrum of each sample was used for further analysis. Before the calibration stage, both NIR and MIR spectra data were preprocessed. First to reduce the noise, the Savitzky–Golay (S. Golay) smoothing was used, with a window width of 7 (3-1-3) points (Gorry, 1990; Savitzky & Golay, 1964). The second type of preprocessing was the use of the multiplicative scatter correction (MSC) (Helland, Naes, & Isaksson, 1995). This method was used to correct additive and multiplicative effects in the spectra. The pretreatments were implemented by ‘‘The Unscrambler V 9.600 (CAMO PROCESS AS, OSLO, Norway). 2.3. Partial least square (PLS) In the development of PLS model, calibration models were built between the spectral data and components of starch and protein, full cross-validation was used to evaluate the quality and to prevent overfitting of calibration models. The optimal number of latent variables (LVs) was determined by the lowest value of predicted residual error sum of squares (PRESS). The prediction performance was evaluated by the correlation coefficient (r), the root mean square error of calibration (RMSEC) or prediction (RMSEP) and bias. The ideal model should have higher r value, lower RMSEC and RMSEP values, and lower bias. The models were carried out by ‘‘The Unscrambler V 9.6’’. 2.4. Independent component analysis ICA was originally developed to deal with problems closely to the cocktail-party problem (Amari, Cichocki, & Yang, 1996). As an effective approach to the separation of blind signal, ICA has recently attracted broad attention and has been successfully used in many fields. e.g., medical signal analysis, image processing, dimension reduction, fault detection and near-infrared spectral data analysis (Chen & Wang, 2001; Hyvarinen & Hoyer, 2000; Shao, Wang, Wang, & Su, 2004). ICA is a well-established statistical signal processing technique that aims to decompose a set of multivariate signals into a base of statistically independent components, and with the minimal loss of information content. The independent components (ICs) are latent variables, meaning that they cannot be directly observed, and the independent component must have non-Gaussian distributions. A chief explanation of noise-free ICA model could be written as the following expression:

x ¼ As

ð1Þ

where x denotes the recorded data matrix, s and A represent the independent components and the coefficient matrix, respectively. The goal of ICA is to find a proper linear representation of nonGaussian vectors so that the estimated vectors are as independent as possible, and the mixed signals can be denoted by the linear combinations of these ICs. The ICs were obtained by a high-order

Table 1 The statistic values of starch and protein for rice after irradiation. Parameters

Data set

No. of samples

Range

Mean

Standard deviation

Starch

Calibration set Prediction set All samples Calibration set Prediction set All samples

260 60 320 260 60 320

18.659–28.125 18.892–27.903 18.659–28.125 7.012–8.596 7.031–8.582 7.012–8.596

25.164 24.427 25.016 8.234 8.196 8.226

1.521 1.457 1.519 0.367 0.352 0.361

Protein

1858

Y. Shao et al. / Food Chemistry 126 (2011) 1856–1861

statistic which is much stronger condition than orthogonality. This goal is equivalent to find a separating matrix W that satisfies

^s ¼ Wx

ð2Þ

where ^s is the estimation of s. The separating matrix W can be trained as the weight matrix of a two-layer feed-forward neural network in which x is input and ^s is output. There are lots of algorithms for performing ICA (Hyvarinen, Karhunen, & Oja, 2001; Lee, 1998). Among these algorithms, the fast fixed-point algorithm (FastICA) is a computationally highly efficient method for performing the estimation of ICA, which was developed by Hyvarinen and Oja (2000). The process of FastICA can be described as follows (Chen & Wang, 2001): (1) Choose an initial random weight vector w (0) and let k = 1, where w is an l-dimensional (weight) vector in the weight matrix W, k is an irrelevant constant. (2) Let w(k) = E{xg(w(k  1)T x)}  E{g0 (w(k  1)T x)} w(k  1), where g is the first-derivative of the function G, and G is a practically any nonquadratic function. (3) Let w(k) = w(k)/||w(k)||. (4) If |w(k)T w(k  1)| is not close enough to 1, let k = k + 1 and go back to step 2. Otherwise, output the vector w (k). It was carried out in Matlab 7.0 (The Math Works, Natick, USA). 2.5. Least-squares-support vector machine (LS-SVM) LS-SVM can work with linear or non-linear regression or multivariate function estimation in a relatively fast way (Chen, Zhao, Fang, & Wang, 2007; Suykens & Vanderwalle, 1999). It uses a linear set of equations instead of a quadratic programming (QP) problem to obtain the support vectors (SVs). The details of LS-SVM algorithm could be found in the literature. The LS-SVM model can be expressed as:

yðxÞ ¼

N X

ak Kðx; xk Þ þ b

ð3Þ

k¼1

where K(x, xk) is the kernel function, xi is the input vector, ai is Lagrange multipliers called support value, b is the bias term. In the model development using LS-SVM and radial basis function (RBF) kernel, for each combination of gam(c) and sig2(r2) parameters, the RMSECV was calculated and the optimum parameters were selected when produced smaller RMSECV. In this study,

gam(c) were optimised in the range of 21–210 and 2–215 for sig2(r2) with adequate increments. These ranges were chosen from previous studies where the magnitude of parameters to be optimised was established. The grid search had two steps, and the first was for a crude search with a large step size, while the second step for the specified search with a small step size. The free LSSVMlab toolbox (LS-SVM v 1.5, Suykens, Leuven, Belgium, http:// www.esat.kuleuven.be/sista/lssvmlab/) was applied with MATLAB 7.0 to develop the calibration models. 3. Results and discussion 3.1. Overview of spectra and statistic values of starch and protein Rice contains 90% of starch, and apparent amylase contents were degressive when irradiation dose increased, which was caused by the structure of starch. It is known that amylose is an essentially linear polymer of a-(1–4)-linked-D-glucopyranosyl units with up to 0.1% a-(1–6) linkages, and amylopectin consist of a-(1–4)-linked-D-glucosyl chains which is highly branched with 5–6% a-(1–6)-bonds. Apparently, the content and degree about polymerisation of amylopectin were so higher than that of amylase, and it had higher probability to be broken and cleaved during irradiation (Yu & Wang, 2007). The range of starch was 18.659–28.125, and that of protein was 7.012–8.596 varying with different irradiation doses. The trend of different doses in NIR region is similar, and it could be seen that a high amount of absorption bands is present in NIR spectra, but it is not directly visible since bands are more severely superimposed. So we treated them with 2nd derivative to find some peaks and valleys, shown in Fig. 1. The prominent features are the absorption peaks associated with the first overtone of C–H stretching of the starch around 1131–1155 nm, the first overtone of O–H symmetric and asymmetric stretching of water around 1418 nm, the first overtone of the C–H asymmetric stretching of lipid around 1681 nm, the O–H bending and asymmetric stretching combination band of water around 1852–1906 nm, protein and combination bands around 1950–2250 nm (Barton, Himmelsbach, McClung, & Champagne, 2002). Compared to the NIR region, the curves are smoother, and differences are existed among each dose samples in the original MIR region (Fig. 2), it is because the MIRS provide more information of frequencies and intensities which are richer and stronger than NIRS does (Chung, Ku, & Lee, 1999). This range covered the two large water bands m2 and mL centered around 1640 and 750 cm1, the amide band I around 1558–1705 cm1, the amide bands II around 1480–1613 cm1, the amide bands III around

Fig. 1. The NIR spectral curves of irradiated rice after 2nd derivative preprocessing.

1859

Y. Shao et al. / Food Chemistry 126 (2011) 1856–1861

Fig. 2. The MIR spectral curves of four rice types after five different irradiation doses.

1200–1280 cm1, phosphate groups covalently bind to casein proteins around 1060–1100 cm1. There were also some peaks around 3000, 1480, 1150 and 900 cm1. In both NIR and MIR region, the spectral curves for starch and protein with different irradiation doses exists no certain rule, neither the higher doses with higher transmission value, nor the reverse, and both of them require a chemometrics techniques such as PLS and LS-SVM to separate useful information from irrelevant contributions.

3.2. PLS models PLS model was developed after the spectral data preprocessed by S. Golay smoothing and MSC. With a comprehensive consideration and according to the spectroscopy categories related to the wavelength bands, the NIR region (1100–2500 nm) and MIR region (400–4000 cm1) were separated to establish two models for the measurement of starch and protein, respectively. Different LVs were applied to build the calibration models, and no outliers were detected in the calibration set during the development of PLS models. The results of calibration and prediction sets for starch and protein are shown in Table 2. With a comparison of results for calibration and prediction sets by the aforementioned evaluation standards, the models with NIR region (1100–2500 nm) turned out to be the best for prediction of starch, the rp, RMSEP and bias were 0.913, 0.241 and 6.126e-04, respectively. The models with MIR region (400–4000 cm1) turned out to be the best for the prediction of protein, the rp, RMSEP and bias for prediction set were 0.932, 0.119 and 1.506e-04, respectively. It indicated that MIR spectroscopy in the transmission mode can perform as well or even better than NIR spectroscopy.

3.3. SW-LS-SVM models In order to improve the training speed and reduce the training error, sensitive wavelengths (SWs) obtained from ICA were applied as inputs of LS-SVM models because the training time increased with the square of the number of training samples and linearly with the number of variables (dimension of spectra). From the aforementioned analysis of the performance of PLS models, the SWs from the NIR region was used as new eigenvectors to starch and the SWs from the MIR region was used as new eigenvectors to protein. ICA was applied for the selection of SWs, which could reflect the main features of the raw absorbance spectra. FastICA (one of the algorithms of ICA, introduced above) was used to the preprocessed spectra data, and the main absorbance peaks and valleys were indicated by the spectra of the ICs, and wavelengths with the highest weights of each IC were selected as the SWs. Fig. 3 showed four ICs and the strong peaks and valleys with the highest weights were thought to be the SWs for the prediction of starch, such as 1210– 1222, 1315–1330, 1575–1625, 1889–1909, and 2333–2356 nm, some of them are corresponding to the absorption peaks associated with the C–H stretching of the starch. To the protein, the strong peaks and valleys around 962–1091, 1232–1298, 1480–1497, 1584–1625 and 2373–2398, 1232–1298 cm1 are corresponding to the absorption peaks associated with the amide band III, 1480–1497 cm1 are corresponding to the amide band II, 1584– 1625 cm1 are corresponding to the amide band I. It showed that through the ICA analysis, the main features of the raw absorbance spectra both in NIR and MIR region were remained. In order to evaluate the performance of SWs, they were applied as the input

Table 2 Calibration statistics for rice samples in NIR and MIR region by PLS models. Calibration

Prediction

rc

RMSEC

bias

Slope

Offset

rp

RMSEP

bias

Slope

Offset

NIR Starch Protein

0.926 0.922

0.221 0.127

2.258e-05 9.045e-05

0.857 0.850

0.114 0.103

0.913 0.897

0.241 0.142

6.126e-04 8.091e-04

0.834 0.805

0.117 0.112

MIR Starch Protein

0.907 0.948

0.304 0.101

7.940e-05 3.083e-06

0.823 0.900

0.157 0.092

0.894 0.932

0.392 0.119

1.008e-03 1.506e-04

0.799 0.869

0.169 0.098

1860

Y. Shao et al. / Food Chemistry 126 (2011) 1856–1861

Fig. 3. The first four ICs by ICA to the absorbance spectra in NIR region.

data matrix to develop the SW-LS-SVM models. Rice samples (260) were used for calibration, and the residual 60 samples for prediction. In the model development by using LS-SVM and RBF kernel function, the determination of parameters c and r2 is an important task, which is similar to the process employed to select number of factors for PLS analysis. In this study, these parameters were optimised by grid-search technique using 5-fold cross-validation with values of c in the range of 21–210 and r2 in the range of 2–215 with adequate increments. The optimal range of the parameters was determined in the first step of grid search. Then a comparatively large step width in a 10  10 grid represented as ‘‘’’ was applied, and the values of cost function were calculated at these grids. Through the large step grid searching, a comparatively narrow space with small value of cost function was found. Subsequently, the much smaller step width was used in this narrow space to obtain the optimal combination of these parameters, and the search grid ‘‘’’ represented the small step grid searching. For each combination of c and r2 parameters, the root mean square error of cross-validation (RMSECV) was calculated and the optimum parameters were selected when produced smaller RMSECV. The optimal pair of (c, r2) was found at the value of c = 172.8 and r2 = 27.5 for starch and c = 41.8 and r2 = 11.6 for protein, respectively. The performance of these models was evaluated by 45 samples in prediction set, and the prediction results are shown in Table 3. The rp, RMSEP and bias for prediction set were 0.946, 0.198 and 1.012e-05 for starch, while 0.974, 0.071 and 1.092e-06 for protein, respectively. The prediction results for calibration and prediction sets showed that SW-LS-SVM models outperformed PLS models (seen in Tables 2 and 3). Therefore, the sensitive wavelengths from ICA analysis could represent most of the features and characteristics of the original spectra, and could be applied instead

Table 3 The prediction results of starch and protein by SW-LS-SVM models in IR region. IR region

Parameters

rp

RMSEP

bias

Slope

Offset

NIR MIR

Starch Protein

0.946 0.974

0.198 0.071

1.012e-05 1.092e-06

0.895 0.949

0.102 0.055

of the whole wavelength region to predict the starch and protein in rice after irradiation. Furthermore, the SWs might be important for the development of portable instruments and online monitoring for commercial applications of quality detection of rice. In order to estimate the SW-LS-SVM models developed, the prediction results for starch and protein contents from the five different dose samples were compared. Table 4 shows the prediction results for each dose samples. It showed that the dose of 0.5 KGy had the highest correlation coefficient for both starch and protein content analysis. In the starch analysis, the correlation coefficient from 0 to 1 KGy is increased first and then decreased, and at the 0.5 KGy it obtained the highest prediction precision. In the protein analysis, the correlation coefficient from 0 to 1 KGy is increased, and the same with starch analysis, the 0.5 KGy had the highest prediction precision. The correlation coefficient for 1 GKy is higher than 0.75 KGy. 3.4. Analysis of the results Compared with the above PLS models, the models with NIR region (1100–2500 nm) turned out to be the best for prediction of starch, and models with MIR region (400–4000 cm1) turned out to be the best for protein prediction, respectively. The NIR range of the electromagnetic spectrum extends from 800 to 2500 nm and is flanked by the MIR region to longer wavelengths. While

Table 4 The prediction results of starch and protein by SW-LS-SVM models in IR region for each dose samples. IR region

Parameters

Doses (KGy)

rp

RMSEP

bias

NIR

Starch

MIR

Protein

0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1

0.923 0.958 0.962 0.941 0.936 0.969 0.978 0.981 0.972 0.975

0.215 0.189 0.182 0.201 0.208 0.078 0.069 0.067 0.074 0.071

1.923e-04 2.391e-05 4.532e-05 1.297e-04 2.093e-04 1.021e-05 2.123e-06 1.245e-06 1.971e-05 3.408e-05

Y. Shao et al. / Food Chemistry 126 (2011) 1856–1861

the MIR region relates essentially to transitions between vibrational states of molecules, the NIR spectrum records the so-called overtone and combination bands of molecular vibrations (Chalmers, 2000). This fundamental difference makes interpretation of specific functionality easier for MIR spectra (since NIR bands are broader and less defined) and can be seen as an advantage for the replacement of NIR by MIR spectroscopy. Previous works have also shown that MIR spectroscopy in the transmission mode can perform as well or even better than NIR spectroscopy for the determination of properties of wheat samples (Reeves & Delwiche, 1997), forages and by-products (Reeves, 1994). The SW-LS-SVM models had a better performance than the PLS models, and the reason might be that the LS-SVM models took the non-linear information of the spectral data into consideration and the non-linear information had improved the prediction precision. The ICs from ICA were obtained by a high-order statistic which is much stronger condition than orthogonality, so the SWs selected from ICs were more effective, and it could be very helpful for the development of portable instrument or real-time monitoring of the rice quality. Many factors affect the precision and reliability of calibration and prediction, such as sample preparation, accuracy of the reference data, etc. The differences in rice sample preparation may influence the results, where the rice sample was milled to flour (Bao, Shen, & Jin, 2007). Delwiche, McKenzie, and Webb (1996) used milled whole grain samples, and showed that flour spectra were superior to those whole ones. Shu, Wu, Xia, Gao, and McClung (1999) also reported that the milled rice flour was superior to brown rice flour in calibration for alkali spreading value. In addition, the grain sizes of the milled rice still vary with different rice genotypes, so all flours passed through a fixed mesh sieve and in homogenous grain size may result in better calibration performance. 4. Conclusions The determination of starch and protein of rice after irradiation could be successfully performed through IR spectroscopy combined with chemometric methods of PLS and SW-LS-SVM models. In PLS models, the models with NIR region (1100–2500 nm) turned out to be the best for prediction of starch, the rp, RMSEP and bias were 0.913, 0.241 and 6.126e-04. The models with MIR region (400– 4000 cm1) turned out to be the best for the prediction of protein, the rp, RMSEP and bias for prediction set were 0.932, 0.119 and 1.506e-04. SWs selected from ICs were applied as the input data matrix of SW-LS-SVM models, and a two-step grid-search technique was used for the optimal RBF kernel parameters of (c, r2). The SW-LS-SVM models were developed and the best prediction performance was achieved. The rp, RMSEP and bias for prediction set were 0.946, 0.198 and 1.012e-05 for starch, while 0.974, 0.071 and 1.092e-06 for protein, respectively, which were better than PLS models. The overall results indicted that IR spectroscopy had the capability to determine the starch and protein of rice after irradiation. The ICA was a powerful way for the selection of sensitive wavelengths, and infrared spectroscopy combined with LSSVM models had powerful capability to predict the components in irradiation rice. Further interpretation of the input data selection, parameter optimisation and results explanation would be needed in order to improve the calibration generalisation and stability. Acknowledgements This study was supported by Natural Science Foundation of China (Project Nos.: 31071332, 60802038), Zhejiang Provincial Natural Science Foundation of China (Project No.: Z3090295), National Agricultural Science and Technology Achievements

1861

Transformation Fund Programs (2009GB23600517), Natural Science Foundation of Ningbo (Project No: 2009A610173) and the Fundamental Research Funds for the Central Universities. References Amari, S., Cichocki, A., & Yang, H. H. (1996). A new learning algorithm for blind signal separation. Advances in Neural Information Processing Systems, 8, 757–763. Bao, J. S., Shen, Y., & Jin, L. (2007). Determination of thermal and retrogradation properties of rice starch using near-infrared spectroscopy. Journal of Cereal Science, 46, 75–81. Barton, F. E., Himmelsbach, D. S., McClung, A. M., & Champagne, E. T. (2002). Twodimensional vibration spectroscopy of rice quality and cooking. Cereal Chemistry, 79, 143–147. Baye, T. M., Pearson, T. C., & Settles, A. M. (2006). Development of a calibration to predict maize seed composition using single kernel near infrared spectroscopy. Journal of Cereal Science, 43, 236–243. Borin, A., Ferrao, M. F., Mello, C., Maretto, D. A., & Poppi, R. J. (2006). Least-squares support vector machines and near infrared spectroscopy for quantification of common adulterants in powdered milk. Analytica Chimica Acta, 579, 25–32. Chalmers, J. M. (2000). Spectroscopy in process analysis. UK: Sheffield Academic Press. Chauchard, F., Cogdill, R., Roussel, S., Roger, J. M., & Bellon-Maurel, V. (2004). Application of LS-SVM to non-linear phenomena in NIR spectroscopy: Development of a robust and portable sensor for acidity prediction in grapes. Chemometrics and Intelligent Laboratory Systems, 71(2), 141–150. Chen, J., & Wang, X. Z. (2001). A new approach to near-infrared spectral data analysis using independent component analysis. Journal of Chemical Information and Computer Science, 41, 992–1001. Chen, Q. S., Zhao, J. W., Fang, C. H., & Wang, D. M. (2007). Feasibility study on identification of green, black and Oolong teas using near-infrared reflectance spectroscopy based on support vector machine (SVM). Spectrochimica Acta Part A-Molecular and Biomolecular Spectroscopy, 66, 568–574. Chung, H., Ku, M. S., & Lee, J. S. (1999). Comparison of near-infrared and midinfrared spectroscopy for the determination of distillation property of kerosene. Vibrational Spectroscopy, 20, 155–163. Cogdill, R. P., & Dardenne, P. (2004). Least-squares support vector machines for chemometrics: An introduction and evaluation. Journal of Near Infrared Spectroscopy, 2, 93–100. Cogdill, R. P., Schimleck, L. R., Jones, P. D., Peter, G. F., Daniels, R. F., & Clark, A. (2004). Estimation of the physical wood properties of Pinus taeda L. radial strips using least squares support vector machines. Journal of Near Infrared Spectroscopy, 12, 263–269. Delwiche, S. R., & Grayboschf, R. A. (2002). Identification of waxy wheat by nearinfrared reflectance spectroscopy. Journal of Cereal Science, 35, 29–38. Delwiche, S. R., McKenzie, K. S., & Webb, B. D. (1996). Quality characteristics in rice by near-infrared reflectance analysis of whole-grain milled samples. Cereal Chemistry, 73, 257–263. Gorry, P. A. (1990). General least-squares smoothing and differentiation by the convolution (Savitzky–Golay) method. Analytical Chemistry, 62, 570–573. Helland, I. S., Naes, T., & Isaksson, T. (1995). Related versions of multiple scatter correction. Chemometrics and Intelligent Laboratory Systems, 29, 233–241. Hyvarinen, A., & Hoyer, P. O. (2000). Emergence of phase and shift invariant features by decomposition of natural images into independent feature subspaces. Neural Computation, 12, 1705–1720. Hyvarinen, A., Karhunen, J., & Oja, E. (2001). Independent component analysis. New York: Wiley. Hyvarinen, A., & Oja, E. (2000). Independent component analysis: Algorithms and applications. Neural Networks, 13, 411–430. Kim, S. S., Rhyu, M. R., Kim, J. M., & Lee, S. H. (2003). Authentication of rice using near-infrared reflectance spectroscopy. Cereal Chemistry, 80, 346–349. Lee, T. W. (1998). Independent component analysis: Theory and application. Boston, MA: Kluwer. Reeves, J. B. III, (1994). Near- versus mid-infrared diffuse reflectance spectroscopy for the quantitative determination of the composition of forages and byproducts. Journal of Near Infrared Spectroscopy, 2, 49–57. Reeves, J. B., III, & Delwiche, S. R. (1997). Determination of protein in ground wheat samples by mid-infrared diffuse reflectance spectroscopy. Applied Spectroscopy, 51, 1200–1204. Savitzky, A., & Golay, M. J. E. (1964). Smoothing and differentiation of data by simplified least squares procedures. Analytical Chemistry, 36, 1627–1639. Shao, X. G., Wang, G. Q., Wang, S. F., & Su, Q. D. (2004). Extraction of mass spectra and chromatographic profiles from overlapping GC/MS signal with background. Analytical Chemistry, 76, 5143–5148. Shu, Q. Y., Wu, D. X., Xia, Y. W., Gao, M. W., & McClung, A. (1999). Analysis of grain quality characters in small ground brown rice samples by near infrared reflectance spectroscopy. Scientia Agricultura Sinica, 19, 92–97. 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. Suykens, J. A. K., & Vanderwalle, J. (1999). Least squares support vector machine classifiers. Neural Processing Letters, 9, 293–300. Vapnik, V. (1998a). Statistical learning theory. John Wiley and Sons: New York. Yu, Y., & Wang, J. (2007). Effect of c-ray irradiation on starch granule structure and physicochemical properties of rice. Food Research International, 40, 297–303.