Food Chemistry 142 (2014) 92100

Contents lists available at SciVerse ScienceDirect

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

Analytical Methods

Optimisation of near-infrared reectance model in measuring protein

and amylose content of rice our
L.H. Xie 1, S.Q. Tang 1, N. Chen, J. Luo, G.A. Jiao, G.N. Shao, X.J. Wei, P.S. Hu
China National Center for Rice Improvement, China National Rice Research Institute, Hangzhou 310006, PR China,
State Key Laboratory of Rice Biology, China National Rice Research Institute, Hangzhou 310006, PR China

a r t i c l e

i n f o

Article history:
Received 3 January 2012
Received in revised form 5 February 2013
Accepted 7 July 2013
Available online 14 July 2013
Keywords:
NIRS
Calibration equation
Protein
Amylose content

a b s t r a c t
Near-infrared reectance spectroscopy (NIRS) has been used to predict the cooking quality parameters of
rice, such as the protein (PC) and amylose content (AC). Using brown and milled ours from 519 rice samples representing a wide range of grain qualities, this study was to compare the calibration models generated by different mathematical, preprocessing treatments, and combinations of different regression
algorithm. A modied partial least squares model (MPLS) with the mathematic treatment 2, 8, 8, 2
(2nd order derivative computed based on 8 data points, and 8 and 2 data points in the 1st and 2nd
smoothing, respectively) and inverse multiplicative scattering correction preprocessing treatment was
identied as the best model for simultaneously measurement of PC and AC in brown ours. MPLS/2,
8, 8, 2/detrend preprocessing was identied as the best model for milled ours. The results indicated
that NIRS could be useful in estimation of PC and AC of breeding lines in early generations of the breeding
programs, and for the purposes of quality control in the food industry.
2013 Elsevier Ltd. All rights reserved.

1. Introduction
Rice is the staple food for over half of the worlds population.
Grain quality is an important concern for rice breeders, producers,
and consumers. Amylose and protein contents are two important
quality indicators that greatly affect the cooking quality of rice
(Champagne et al., 1997, 1998). The amylose content (AC) is closely correlated with the sensory properties of freshly cooked rice
(Champagne, Bett-Garber, McClung, & Bergman, 2004), while the
protein content (PC) dictates the texture of cooked rice by inhibiting water absorption and starch swelling during cooking (Ishima, Taira, & Mikoshiba, 1974; Xie, Chen, Duan, Zhu, & Liao,
2008; Yanase, Ohtsubo, Hashimoto, Sato, & Teranishi,1984). The
AC (more accurately termed as the apparent amylose content,
AAC) is traditionally determined by the iodine-blue complex that
Abbreviations: NIRS, near-infrared reectance spectroscopy; PC, protein content; AC, amylose content; MPLS, modied partial least squares; PLS, partial least
squares regression; PCR, principal component regression; MLR, multiple linear
regressions; INQR, international network for quality rice; GH, Global H distance;
SNV, standard normal variant; MSC, standard multiplicative scattering correction;
RSQ, coefcient of determination; SEC, standard error of calibration; 1 VR, 1
minus variance ratio; SECV, standard error of cross validation; SEP, standard error;
bias, average of the residuals; SEP(C), corrected standard error; SD, standard
deviations; DPS, data processing system.
Corresponding author. Tel.: +86 571 63370221; fax: +86 571 63370482.
E-mail addresses: Qualityh@163.com, xlhxlh18@163.com (P.S. Hu).
1
These authors contributed equally to this work.
0308-8146/$ - see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.foodchem.2013.07.030

is prone to variability. A survey conducted by the international

network for quality rice (INQR) showed that ve different versions of the iodine binding method are currently in use and that
the reproducibility was high within laboratories but low between
laboratories (Fitzgerald et al., 2009). The PC is typically measured
by the conventional Kjeldahl method, which is destructive, tedious, and costly because large amounts of materials are needed.
Thus, more efcient methods to measure rice quality are
stillnecessary.
Near infrared-reectance spectroscopy (NIRS) is a promising
technique with fast, easy-to-use, and nondestructive analytical
potentials (Bart, Himmelsbach, McClung & Champagne, 2007;
Cen & He, 2007). Recently, the technique has gained popularity
as a screening tool for quality traits in cereal breeding programs
and genetic analysis. Using milled rice ours, NIRS calibration
models for the AC value have been developed (Delwiche, Bean,
Miller, Webb & Williams 1995; Shu, Wu, Xia, Gao, & McClung,
1999a; Bao, Cai, & Corke, 2001; Sohn, Barton, McClung, & Champagne, 2004a). NIRS models have also been successfully established for AC in brown rice ours (Shu, Wu, Xia, Gao, and
McClung, 1999b; Wu & Shi, 2007). The literature also shows that
NIRS was able toaccomplish satisfactory analysis for the PC value
in rice (Barton, Himmelsbach, McClung, & Champagne, 2000; Delwiche, McKenzie, & Webb, 1996; Himmelsbach, Barton, McClung,
& Champagne, 2001; Sohn, Himmelsbach, & Champagne, 2004b;
Zhang et al., 2007).

L.H. Xie et al. / Food Chemistry 142 (2014) 92100

The analytical capacity of NIRS depends on the intrinsic correlation between the concentrations of relevant substances and their
associated absorbance spectra. The accuracy of the prediction
models are very much dependent on how well the relationship is
described through the preprocessing spectral data and multivariate
statistical techniques. Spectral data preprocessing removes the
irrelevant information (noise) that cannot be handled properly by
regression techniques, and multiple scatter correction (MSC) is
the most popular normalisation technique used to preprocess the
NIRS spectral data (Ns, Isaksson, Fearn, & Davies, 2004) to compensate for additive (baseline shift) and multiplicative (tilt) effects
(Martens & Stark, 1991). Among different multivariate regression
techniques, principal component regression (PCR), partial least
squares regression (PLS) and modied partial least squares regression (MPLS) are three of the most commonly used techniques to
characterise the correlations between the spectral data and quality
attributes (Martens & Ns, 1998). After comparing the multiple
linear regression (MLR) model, the PCR model and the PLS model,
Geladi and Kowalski (1986) concluded that PLS is the most complete and elegant technique when prediction accuracy is concerned. Furthermore, studies indicated that mathematical
derivations of spectral data affected the predictability of the calibration equations (Wu & Shi, 2004; Bao, Shen, & Jin, 2007a). Sample conditions and sample types also played a signicant role in the
accuracy of the NIRS models (Wu & Shi, 2007).
The popularity of NIR spectroscopy in post-harvest technology
is evidenced by the increasing number of publications on the topic
and by the fact that some online grading service providers have
implemented NIR systems to measure various attributes of rice
quality. However, no research effort was made to evaluate the singular or combined effects of the various types of mathematic and
preprocessing techniques on the accuracy of the MPLS, PLS and
PCR models for measuring the PC and AC in rice. Using an NIRS system equipped with software packages capable of multivariate calibration and data pretreatment, we conducted the current study
with the following objectives: (1) To compare the performance of
different calibration models generated using various mathematical
derivations, preprocessing techniques and regression methods on
measuring the PC and AC values in brown or milled rice ours;
(2) To identify robust NIR models to quickly determine the values
of PC and AC simultaneously; and (3) To identify possible areas to
further improve the model accuracy in PC and AC determination.

93

1978). The PC value (N5.95) was quantied by the Kjeldahl method using a Kjeltec-Foss 2300 Auto-analyser (FOSS Analytical, Denmark) (AOAC, 1999).
2.3. NIRS spectroscopic analysis
Each of the brown and milled rice our samples was scanned on
an NIR Systems III Rapid Content Analyser (Model XDS monochromator type XM-1000, FOSS Analytical AB, Sweden) equipped with
the WinISI III Project Manager Software version 1.50 (FOSS NIRSystems, Silver Springs, MD, USA) to obtain the reectance spectra and
perform calibration and validation. Individual samples weighing
3.0 g each were loaded in a ring cup (with an internal diameter
of 38 mm and a depth of 10 mm) and scanned in duplicate (by
rotating the ring cup to a different position between scans). The
reectance spectra were recorded between 424 and 2498 nm at
8-nm intervals. The NIRS absorption data were expressed as Log
(1/R), where R is the relative reectance. Each sample was subsequently scanned 64 times, and the average spectrum was recorded
and used for analysis.
2.4. Sample selection for NIRS analysis
Sample selection was conducted by the WinISIII Project Manager Software version 1.50 (FOSS NIRSystems, Silver Springs, MD,
USA). To remove outliers, a boundary was set at the Mahalanobis
distance (the Global H distance, GH) of 3.0. This resulted in a total
of 491 brown and 472 milled our samples that qualied for analysis. Optimum NH (GH distance between all pairs of spectra,
termed the Neighbor H distance) cutoff was used to select the calibration sample set. The NH distance calculation is based on a principle component analysis with specic algorithms as different
algorithms could affect the principle component number and further affect NH distance (Wu & Shi, 2007). Since sample sets with
wider range and larger SD tend to generate more reliable calibration equations (Bao, Wang & Shen, 2007b), we set the NH at 0.6
with PCA algorithm to select calibration sample set, which resulted
in 294 brown our samples and 335 milled our samples, respectively. The external validation data sets consisted of the remainder
of the qualied our samples (197 brown our samples and 137
milled our samples).
2.5. Calibration analysis and validation

2. Experimental
2.1. Rice sample preparation
A total of 519 indica rice lines grown in eight major rice-producing provinces of China (Sichuan, Jiangxi, Jiangsu, Hunan, Guangxi,
Guangdong, Anhui and Zhejiang) between 2009 and 2010 were
collected as the original sample population. Grains with a total
weight of 140 g were brought to 10.512.5% moisture (as determined by the Kett Grain Moisture Tester, model PB-1D2, Kiya Seisakusho, Ltd., Japan), dehulled in a Satake Testing Husker (Model
THU-35A, Satake Engineering Co., Ltd., Japan), and debranned with
a Mill (Model McGill No. 2, Seedburo Equipment Co., Chicago, USA).
Brown and milled rice grains were ground to pass through a 0.42mm screen on an Udy Cyclone Mill (Model Cyclotec 1093 Sample
mill, Tecator, Sweden) to obtain brown or milled rice our for
every rice sample.
2.2. Conventional quantication of AC and PC
The AC value was determined by a colorimeter at 620 nm using
an iodine-binding complex of soluble starch (Perez & Juliano,

Calibration and validation were performed by the WinISIII Project Manager Software version 1.50. The calibration models for
measuring the values of AC and PC based on absorbance spectra
were separately established for brown and milled our samples
using MPLS, PLS and PCR regression available in WinISI III Project
Manager. The following combinations of algorithms were applied
to preprocess the spectral data: (1) no spectral data preprocessing
was performed; (2) standard normal variant (SNV) + detrend; (3)
SNV only; (4) detrend only; (5) standard multiplicative scattering
correction (MSC); (6) weighted MSC; and (7) inverse MSC. Mathematical derivations 0, 0, 1, 1, 1, 2, 2, 1, 1, 4, 4, 1 and 2, 8, 8, 2
were also applied. As an example, the notation 0, 0, 1, 1 means
D = 0, G = 0, S1 = 1 and S2 = 1, where D is the derivative order number (0 = no derivative operation, 1 = 1st order derivative, and
2 = 2nd order derivative), G is the gap (the number of data points
computed by the derivation), S1 is the number of data points in
the rst smoothing, and S2 is the number of data points in the second smoothing.
The coefcient of determination (RSQ) and the standard error of
calibration (SEC) were calculated automatically by the equipped
software to evaluate the model. Models with higher RSQ and lower
SEC values are better than models with lower RSQ and higher SEC

94

L.H. Xie et al. / Food Chemistry 142 (2014) 92100

values. Two similar parameters, the 1 minus variance ratio

(1 VR) and the standard error of cross validation (SECV), were
calculated as estimators of the cross-validation coefcient and error, respectively. The models were further independently validated
using external validation set of samples because cross validation
alone is often insufcient for model validation. The prediction ability of each equation in the external validation was evaluated based
on the standard error of performance (SEP), average of the residuals (bias), slope, corrected standard error (SEP(C)), and R2 (coefcient of determination in the external validation) values.

cated that all evaluated models provided generally good calibration for PC in the brown rice.
The 2nd derivative equations were generally superior to the 1st
derivative. In particular, 2, 8, 8, 2 was the best combination and
produced the lowest SEC and SECV values along with the highest
RSQ and 1 VR values. Among the preprocessing treatments,
MSC (5), weighted MSC (6) and inverse MSC (7) produced higher
RSQ and 1 VR values along with lower SEC and SECV values than
others did. MPLS and PLS models were better than PCR models (Table 2). The best model was generated by the MPLS/2, 8, 8, 2/ Inverse MSC (7) combination (RSQ = 0.946, SEC = 0.293).
The statistics of the models and cross validation for milled our
samples are summarised in Table 3. Only preprocessing and
regression methods signicantly affected the RSQ and SEC values
for calibration and the 1 VR and SECV values for cross-validation
(p < 0.01), while the derivation had no signicant effect on these
parameters. Compared to that of brown our, the model quality
of milled our was slightly lower (as indicated by the lower RSQ
and 1 VR values and the higher SEC and SECV values).
Preprocessing algorithms SNV + detrend (combination (2)) and
SNV (combination (3)) produced the lowest RSQ (0.9080 and
0.9063, respectively) and 1 VR (0.9001 and 0.8975, respectively)
values along with the highest SEC (0.3560 and 0.3599, respectively) and SECV (0.3719 and 0.3770, respectively) values (Table 3).
Similar to the performance on brown ours, the MPLS and PLS calibration models were better than PCR models. The averages of RSQ
of MPLS, PLS and PCR calibration models were 0.9204, 0.9209 and
0.9087, respectively.
Tables 2 and 3 also show the combined effects of derivation,
preprocessing methods, and regression techniques on the calibration equations. For milled ours, the regression  preprocessing
combination signicantly affected all model parameters (SEC,
RSQ, SECV and 1 VR) (p < 0.01). However, this interaction was
signicant for only three parameters (SEC, SECV and 1 VR) for
brown ours. The regression  derivation interaction had a significant effect on all model parameters (p < 0.01) in brown ours but
no effect in milled ours. The derivation  preprocessing interaction inuenced the SEC, RSQ and 1 VR values in brown ours signicantly. These evaluations are useful in selecting the best
regression and spectral treatment combinations.

2.6. Statistical analysis

All statistical analyses of the data were conducted using the
data processing system (DPS) v 9.50 (Tang & Zhang, 2013). Oneor two-way ANOVA was performed to evaluate the effect of mathematical derivation, preprocessing and regression techniques on
the statistic parameters of the models and the validations. Means
were compared by the Duncan test at a = 0.05.
3. Results
3.1. Sample representation and homogeneity
The means, standard deviations (SD), standard errors (SE), and
ranges of AC and PC values of all samples measured by conventional methods were summarised in Table 1. The PC values were
distributed between 6.7% and 13.8% with several gaps, while the
AC values were distributed between 1.8% and 28.1% with very
few samples having values <10%. These ranges and distributions
indicated that the selected sample set was a good representation
of the chemical diversity of the rice in the current production system in China. The similarity of the mean ranges and the SDs and
SEs between the calibration and the validation sample sets also
indicated that both subsets represented the potential variations
in the rice samples well.
The raw spectra between the brown and milled ours were similar (Figure not shown). This could indicate that all tested samples
are quite homogenous, with similar spectral shapes but different
intensities (between 800 and 2492 nm). This property enables a
general calibration equation to be developed for the our samples
from different genotypes.

3.3. Comparison of derivation, preprocessing and regression methods

for AC

3.2. Comparison of derivation, preprocessing and regression methods

for PC

Mathematical derivation, preprocessing and regression methods all signicantly affected the RSQ and SEC values of calibration
along with the 1 VR and SECV values of cross-validation for the
AC in brown ours. Different from those of the PC, the AC equations
for brown ours varied greatly (Table 4).
The 2nd order derivative was generally superior to the 1st order
for the ACs. The best treatment was once again 2, 8, 8, 2, which
gave the lowest SEC and SECV values along with the highest RSQ
and 1 VR values. Treatment 0, 0, 1, 1 was the worst, producing

Table 2 shows a summary of the statistics of the calibration and

cross validation of the PC values in brown our samples. Mathematic derivation, preprocessing and regression methods signicantly affected the RSQ and SEC of the calibration models, and
the 1 VR and SECV values of cross validation. However, all RSQ
and 1 VR values were higher than 0.93, and the values of SEC
and SECV were approximately 0.300% (Table 2). These results indi-

Table 1
Protein and amylose content (AC) of brown and milled rice ours measured by conventional methodology.
Parameter

Calibration samples

Mean external validation

No.

Range

Mean

SE

SD

No.

Range

Mean

SE

SD

Brown our
Protein%
AC%

294
294

6.713.8
1.828.1

9.57
19.67

0.073
0.244

1.25
4.19

178
178

7.012.6
12.226.8

9.50
19.40

0.064
0.239

1.0
4.1

Milled our
Protein%
AC%

335
335

6.713.8
1.828.1

9.60
19.40

0.066
0.235

1.20
4.30

137
137

6.913.0
13.326.9

9.60
20.00

0.060
0.218

1.10
4.0

95

L.H. Xie et al. / Food Chemistry 142 (2014) 92100

Table 2
Calibration models and cross validation statistics for protein content in brown our samples.a
Treatments

Calibration

Cross-validation

SEC

RSQ

SECV

Mathematical treatment (A)

0,0,1,1
1,2,2,1
1,4,4,1
2,8,8,2
F, p

VR

0.3034 0.0118bcBC
0.3113 0.0157aA
0.3058 0.0108bB
0.3010 0.0053cC
15.76, <0.001

0.9420 0.0047abAB
0.9389 0.0065cC
0.9412 0.0042bB
0.9432 0.0020aA
15.17, <0.001

0.3093 0.01126bB
0.3175 0.0139aA
0.3120 0.0099bB
0.3079 0.0056bB
10.02, <0.001

0.9401 0.0046abA
0.9365 0.0059cB
0.9393 0.004bA
0.9408 0.0022aA
17.22, <0.001

Preprocessing treatment (B)

1
2
3
4
5
6
7
F, p

0.3082 0.0063aA
0.3085 0.0180aA
0.3109 0.0178aA
0.3075 0.00412aAB
0.3015 0.00897bC
0.2992 0.0084bC
0.3019 0.0101bBC
9.06, <0.001

0.9403 0.0024bC
0.9399 0.0020bC
0.9400 0.0072bC
0.9405 0.0017bBC
0.9430 0.0035aA
0.9439 0.0032aA
0.9428 0.0039aAB
9.24, <0.001

0.3121 0.0060cAB
0.3177 0.0155abA
0.3188 0.0156aA
0.3126 0.0048cAB
0.3062 0.0008 dB
0.3062 0.0075 dB
0.3075 0.0090cdB
8.45, <0.001

0.9387 0.0022cCD
0.9368 0.0066dDE
0.9354 0.0064dE
0.9393 0.0021bcBC
0.9414 0.0032aAB
0.9418 0.0027aA
0.9409 0.0035abABC
15.77, <0.001

Regression method (C)

MPLS
PLS
PCR
F, p

0.2998 0.00697bB
0.3011 0.0773bB
0.3152 0.0134aA
78.15, <0.001

0.9436 0.0027aA
0.9428 0.0031aA
0.9376 0.0056bB
67.60, <0.001

0.3084 0.0072bB
0.3088 0.0077bB
0.3176 0.0145aA
19.64, <0.001

0.9406 0.0028aA
0.9401 0.0035aA
0.9368 0.0060bB
27.33, <0.001

Interaction effect (F value)

A C
B C
A B

3.86**
8.03**
2.39*

3.74**
7.83**
2.12*

2.37*
5.42**
1.49NS

3.41**
9.09**
2.41*

In each category followed by small and big different letters were signicantly different at Duncan test, p < 0.05 and p < 0.01, respectively.
Means SD (n = 21, 14 and 28 for AC, respectively).

p < 0.05.

p < 0.01.
NS
p > 0.05.
a

Table 3
Calibration models and cross validation statistics for protein content in milled our samples.a
Parameters

Calibration

Cross-validation

SEC

RSQ

SECV

VR

Mathematical treatment (A)

0,0,1,1
1,2,2,1
1,4,4,1
2,8,8,2
F, p

0.3392 0.035aA
0.3437 0.0313aA
0.3349 0.015aA
0.3406 0.013aA
1.27, 0.2986

0.9156 0.020aA
0.9141 0.017 a A
0.9193 0.066aA
0.9176 0.0055aA
1.46, 0.2409

0.3538 0.0328aA
0.3594 0.0289aA
0.3494 0.0129aA
0.3542 0.012aA
1.84, 0.1578

0.9083 0.0191abA
0.9063 0.016bA
0.9121 0.006aA
0.9110 0.0053abA
2.19, 0.1066

Preprocessing treatment (B)

1
2
3
4
5
6
7
F, p

0.3351 0.010bB
0.3560 0.0400aA
0.3599 0.0413aA
0.3326 0.0072bB
0.3329 0.010bB
0.3275 0.0090bB
0.3331 0.011bB
8.86, <0.001

0.9188 0.0053aA
0.9080 0.023bB
0.9063 0.023bB
0.9201 0.0030aA
0.9205 0.0039aA
0.9226 0.0039aA
0.9203 0.0045aA
7.11, <0.001

0.3492 0.0063bB
0.3719 0.034aA
0.3770 0.036aA
0.3469 0.0011bB
0.3469 0.006bB
0.3420 0.0101bB
0.3456 0.0096bB
12.67, <0.001

0.9118 0.0035aA
0.9001 0.0203bB
0.8975 0.0209bB
0.9130 0.0005aA
0.9137 0.0021aA
0.9156 0.0044aA
0.9142 0.0040aA
9.85, <0.001

Regression method (C)

MPLS
PLS
PCR
F, p

0.3325 0.0086bB
0.3317 0.009bB
0.3545 0.038aA
21.17, <0.001

0.9204 0.0035aA
0.9209 0.0036aA
0.9087 0.0213bB
18.47, <0.001

0.3479 0.0079bB
0.3509 0.0123bB
0.3638 0.0364aA
10.60, 0.002

0.9129 0.0034aA
0.9115 0.0057aA
0.9039 0.0208bB
9.79, <0.001

Interaction effect (F value)

A C
B C
A B

1.57NS
6.49**
0.93NS

1.81NS
5.66**
0.85NS

1.35NS
6.23**
1.16NS

1.57NS
5.27**
1.07NS

In each category followed by small and big different letters were signicantly different at Duncan test, p < 0.05 and p < 0.01, respectively.
Means SD (n = 21,14 and 28 for AC, respectively).
**
p < 0.01.
NS
p > 0.05.
a

the highest SEC and SECV values along with the lowest RSQ and
1 VR values. The algorithms SNV + detrend (combination (2))
and SNV (combination (3)) reduced the model performance, as

indicated by low RSQ (0.6856 and 0.6783, respectively) and

1 VR (0.6376 and 0.6323, respectively) values along with high
SEC (2.1094 and 1.9779, respectively) and SECV (2.3208 and

96

L.H. Xie et al. / Food Chemistry 142 (2014) 92100

Table 4
Calibration models and cross validation statistics for amylose content in brown our samples.a
Parameters

Calibration

Cross-validation

SEC

RSQ

SECV

Mathematical treatment (A)

0,0,1,1
1,2,2,1
1,4,4,1
2,8,8,2
F, p

VR

2.299 0.7550aA
1.9763 0.9948bB
1.9364 1.0412bB
1.7598 0.1.0427cC
67.33, <0.001

0.6585 0.2269cC
0.7183 0.2646bB
0.7222 0.2706bB
0.7569 0.2544aA
31.93, <0.001

2.4489 0.7512aA
2.2050 0.8974bB
2.0982 0.9313cB
1.9276 0.9799dC
37.14, <0.001

0.6223 0.2333cC
0.6720 0.2628bB
0.6813 0.2734bB
0.7289 0.2546aA
35.50, <0.001

Preprocessing treatment (B)

1
2
3
4
5
6
7
F, p

1.9649 0.9389bBC
2.1094 1.0246aAB
1.9779 0.9945aA
1.8951 1.0151bBC
1.9181 1.0316bC
1.9181 1.0316bC
1.9557 0.9746bC
6.31, <0.001

0.7227 0.2456aAB
0.6856 0.2712bBC
0.6783 0.2727bC
0.7190 0.2586aAB
0.7359 0.2614aA
0.7286 0.2666aA
0.7276 0.2666aA
5.57, <0.001

2.1560 0.8679bcAB
2.3208 0.9822aA
2.2463 0.9166abAB
2.1623 0.9055bcAB
2.0996 0.9625bcB
2.0806 0.0946cB
2.1241 0.9152bcAB
3.26, 0.012

0.6911 0.2374aA
0.6376 0.2804bB
0.6323 0.2821bB
0.6852 0.2503aA
0.6954 0.2678aA
0.7013 0.2610aA
0.6900 0.2585aA
8.70, <0.001

Regression method (C)

MPLS
PLS
PCR
F, p

1.1366 0.3554cC
1.5721 0.3066bB
3.2700 0.1604aA
2257, <0.001

0.9192 0.0513aA
0.8530 0.0612bB
0.3697 0.0551cC
2293.4, <0.001

1.3634 0.3085cC
1.8053 0.3087bB
3.3411 0.2329aA
1119.8, <0.001

0.8887 0.04959aA
0.8083 0.06845bB
0.3315 0.06389cC
2250.1, <0.001

Interaction effect (F value)

A C
B C
A B

13.99**
3.54**
1.07NS

4.09**
2.01NS
0.99NS

4.19**
2.81**
1.08NS

3.20*
3.30**
0.84NS

In each category followed by small and big different letters were signicantly different at Duncan test, p < 0.05 and p < 0.01, respectively.
Means SD (n = 21, 14 and 28 for AC, respectively).
*
p < 0.05.
**
p < 0.01.
NS
p > 0.05.
a

2.2463, respectively) values. MPLS produced the best calibration

model for AC in brown our samples with an average RSQ of
0.9192 (Table 4).
For the AC in milled ours, derivation, preprocessing and
regression methods also signicantly affected the RSQ, SEC,
1 VR and SECV values more than they affected the PC in milled
ours, but the effect was less than that observed on the AC of
brown ours (Table 5). Derivation 0, 0, 1, 1 produced the highest SEC and SECV values along with the lowest RSQ and 1 VR
values. The values produced by the other three derivations were
similar. Preprocessing algorithms SNV + detrend (combination
(2)) and SNV (combination (3)) produced the lowest RSQ and
1 VR values along with the highest SEC and SECV value, indicating that their effects were similar to those on brown our
(Table 5).
Once again, MPLS produced the best calibration model for AC
with milled our samples. The average calibration RSQ was
0.9502, which was higher than that of PLS (0.9049). The average
RSQ of PLS calibration model was 0.9049. The PCR model was not
applicable because the average RSQ only reached 0.5139.
The regression  preprocessing interaction signicantly affected all model parameters (SEC, RSQ, SECV and 1 VR) in both
brown and milled ours (p < 0.01). The regression 
derivation interaction had signicant effects on all 4 model
parameters (p < 0.01) in brown ours and 3 of the model parameters (RSQ, SECV and 1 VR) in milled ours (shown in Tables 4
and 5).

3.4. Comparison of calibration models between sample types (brown

vs. milled)
For both PC and AC, sample type had effect on the statistical
parameters of the models. Even though the models for the PC in
brown our samples were better than those in milled ours, all

RSQ and 1 VR values were high (>0.9100), which suggested that

reliable and accurate results could be achieved regardless of sample types. On the other hand, the calibration models for the AC of
milled ours were superior to those for the AC of brown ours.
Nevertheless, the average RSQ and 1 VR values were generally
low (0.6760.790%) (Data not shown).

3.5. Best models for both PC and AC measurements in brown and

milled ours
The predictive ability of NIRS models was evaluated by lower
SEC and SECV, and higher RSQ and 1 VR values the calibration
set. For the PC and AC of both brown and milled ours, external
validation indicated excellent accuracy evidenced by favorable values of bias, slope, SEP, R2 and SEP(C). For brown ours, the best
model to assess PC and AC simultaneously was MPLS/2, 8, 8, 2/Inverse MSC (7) combination. This model had favorable values of RSQ
(0.946), SEC (0.293%), 1 VR (0.942), SECV (0.311%), and the slope
(0.923), R2 (0.916), SEP (0.309%), bias (0.0203) and SEP(C) (0.309%)
for PC measurements. For AC, the model had RSQ of 0.963, SEC of
0.770%, 1 VR of 0.979, SECV of 0.944%, and the slope of 1.031,
R2 of 0.952, SEP of 0.894%, the bias of 0.021 and SEP(C) of
0.896% for external validation.
For milled ours, the best model to analyse the PC and AC
simultaneously was MPLS/2, 8, 8, 2/SNV (5) combination. For
the PC, the model had favorable values of RSQ (0.924), SEC
(0.324%), 1 VR (0.914), SECV (0.344%), and the slope (1.019), R2
(0.890), SEP (0.346%), bias (0.030%) and SEP(C) (0.346%). For AC,
the model had RSQ of 0.962, SEC of 0.809%, 1 VR of 0.944, SECV
of 0.982%, and the slope of 1.017, R2 of 0.938, SEP of 1.004%, bias
of 0.140% and SEP(C) of 0.998% for external validation. These results demonstrated that estimations of PC and AC simultaneously
are possible with NIRS calibration equations developed based on
a unique calibration sample set.

97

L.H. Xie et al. / Food Chemistry 142 (2014) 92100

Table 5
Calibration models and cross validation statistics for amylose content in milled our samples.a
Parameters

Calibration

Cross-validation

SEC

RSQ

SECV

Mathematical treatment(A)
0,0,1,1
1,2,2,1
1,4,4,1
2,8,8,2
F, p

VR

1.9964 0.9916aA
1.6197 0.8324bB
1.5947 0.9796bB
1.4872 0.8741bB
23.94, <0.001

0.7235 0.2622bB
0.8159 0.1934aA
0.8040 0.2382aA
0.8154 0.2230aA
10.87, <0.001

2.0998 0.9720aA
1.7898 0.7705bB
1.7316 0.9329bcB
1.6596 0.9102cB
23.41, <0.001

0.7006 0.2628bB
0.7889 0.1887aA
0.7843 0.2352aA
0.7857 0.2387aA
8.38, <0.001

Preprocessing treatment (B)

1
2
3
4
5
6
7
F, p

1.5620 0.8408bB
1.9985 1.1308aA
2.0650 1.1580aA
1.5445 0.8574bB
1.5260 0.8171bB
1.5152 0.8179bB
1.5102 0.8196bB
16.70, <0.001

0.8271 0.1833aA
0.7051 0.3064bB
0.6575 0.3380bB
0.8248 0.1896aA
0.8353 0.1695aA
0.8369 0.1637aA
0.8412 0.1653aA
17.88, <0.001

1.6699 0.8011bB
2.1663 1.0618aA
2.3039 1.1533aA
1.6607 0.8109bB
1.6569 0.7799bB
1.6523 0.7654bB
1.6316 0.7831bB
29.19, <0.001

0.8096 0.1830aA
0.6754 0.2979bB
0.6290 0.3274bB
0.8067 0.1891aA
0.8148 0.1721aA
0.8203 0.1654aA
0.7981 0.2065aA
15.99, <0.001

Regression method (C)

MPLS
PLS
PCR
F, p

0.9352 0.1516cC
1.2450 0.3888bB
2.8433 0.5336aA
678.38,< 0.001

0.9502 0.0146aA
0.9049 0.0640bB
0.5139 0.1960cC
421.10, <0.001

1.0929 0.1274cC
1.4327 0.4416bB
2.9350 0.5302aA
798.0, <0.001

0.9312 0.0148aA
0.8757 0.0826bB
0.4875 0.1816cC
354.84, <0.001

Interaction effect (F value)

A C
B C
A B

1.43NS
4.36**
0.96NS

3.28*
7.84**
0.66NS

2.81*
7.18**
0.55NS

3.02*
5.62**
0.76NS

In each category followed by small and big different letters were signicantly different at Duncan test, p < 0.05 and p < 0.01, respectively.
Means SD (n = 21,14 and 28 for AC, respectively).
*
p < 0.05.
**
p < 0.01.
NS
p > 0.05.
a

The regression coefcient plots (load plots) obtained during

computation can be analysed to further evaluate genuine NIR calibration models as spectra variations at different wavelength reect chemical structural information of measured components
(Milica, Jasna, Dragan, & Mladenka, 2010). Fig. 1 shows the analysis
of the regression coefcient plots obtained during computation of
MPLS regression equations for amylose and protein contents in rice
ours. Fig. 1a and b are the regression coefcient analysis for amylose content obtained by using the MPLS/2, 8, 8, 2/ MSC (5) combination for milled rice our and the MPLS/2, 8, 8, 2/inverse MSC
(7) combination for brown rice ours. The plots showed higher
regression coefcients 7479 and 3615 at 896 nm that were related to CH, and 8291 and 1540 at 1540 nm that were related
to OH stretch 1st overtone associated with starch. Higher regression coefcient 6000 and 7465 at 1228 nm were related to CH
2nd overtone and CH combination associated with starch. All
these suggest that the models are authentic and this is why they
yielded high predictive abilities.
Similarly, Fig. 1c and d showed the regression coefcient for
protein content in milled and brown rice ours using the MPLS/
2, 8, 8, 2/MSC (5) combination and MPLS/2, 8, 8, 2/inverse
MSC (7) combination, respectively. The highest regression coefcient 592 and 394 was observed at 1032 nm that associated
with protein. Moreover, high regression coefcient 354 and
228 at 1492 nm were related to NH stretch 1st overtone associated with protein, and 52 and 50 at 2060 nm were related to N
H bend 2nd overtone and NH stretch bend combination associated with protein.
In summary, we identied that MPLS with the mathematic
treatment 2, 8, 8, 2 and inverse MSC preprocessing model as
the best for simultaneously measuring PC and AC in brown ours,
and the MPLS/2, 8, 8, 2/MSC preprocessing model as the best for
milled ours. The data presented here conrmed that the NIRS
method coupled with a single suitable calibration model can be ap-

plied to predict amylose and protein content in brown/milled

ours simultaneously and the predicted values are close to that
measured by conventional chemical analysis.

4. Discussion
For the PC in brown rice ours, the prediction performances of
the different models were very similar, with excellent correlation
coefcients (average RSQ = 0.9413) and low errors (average
SECV = 0.3116). The milled our models showed more variation
(average RSQ = 0.9167, SECV = 0.3542) compared to the models
of brown ours. The variation coefcient and the range of RSQ
were 1.5% and 0.85370.9273, respectively, for milled ours;
these values were much higher than the corresponding values
for brown ours (0.51% and 0.9220.9495, respectively), as calculated (data not shown). Preprocessing treatments SNV and detrend + SNV decreased the RSQ and 1 VR while signicantly
increasing SEC and SECV when the PCR algorithm was employed;
therefore, those combinations should be avoided. Based on the
combined effects of sample type, preprocessing treatment and
regression methods in the current study, we could make the following conclusions: (1) Even though the PC models developed
using brown ours were superior to those developed using milled
ours, both types of models were quite accurate for PC determination; (2) The regression algorithm had effect on the model performance, but the degree of inuence mainly depended on the
sample type. For example, compared to other algorithm, when
the PCR algorithm was employed, the decreases in RSQ and
1 VR values were much greater in milled our than in brown
our. Therefore, further improvement of model performance
may be limited by the sample type (brown our is better); (3)
Preprocessing treatments also inuenced model performance,
and the degree of inuence depended on the regression algo-

98

L.H. Xie et al. / Food Chemistry 142 (2014) 92100

Fig. 1. Regression coefcient plot of modied partial least squares model (MPLS) calibration equations. (1a) MPLS with the mathematic treatment 2, 8, 8, 2 and MSC
preprocessing for amylose content in milled ours. (1b) MPLS /2, 8, 8, 2/ MSC preprocessing for protein content in milled ours. (1c) MPLS/2, 8, 8, 2/ inverse MSC
preprocessing for amylose content in brown ours. (1d) MPLS/2, 8, 8, 2/ inverse MSC preprocessing for protein content in brown ours. Spectra range: 4241092 nm, 1108
2490 nm.

L.H. Xie et al. / Food Chemistry 142 (2014) 92100

rithms used as preprocessing treatment SNV and detrend + SNV

decreased the RSQ and 1 VR values signicantly when PCR
was employed, but had little effect on the same parameters when
the MPLS and PLS algorithms were adopted.
For the AC values, the calibration equations showed low and
varying coefcients (average RSQ = 0.4331 with a range of
0.16480.7161%) for both brown and milled ours with the PCR
algorithm, while the RSQ values of the MPLS and PLS models were
similar. Therefore, the PCR models should not be used regardless of
sample types. The MPLS models had higher RSQ and 1 VR values
with lower errors compared to the PLS and PCR models. The optimal models for the determination of AC were those using MPLS or
PLS for milled ours and using MPLS for brown ours. These results
suggested that (1) PCR regression was not suitable for predicting
the AC in both brown and milled ours; (2) A robust calibration
could be obtained using different combinations of derivations, preprocessing and regression methods regardless of sample types; (3)
The SNV + detrend and SNV preprocessing treatments had a negative effect on the model performance when the PLS and PCR were
used; (4) The selection of regression algorithms can potentially improve the calibration performance; (5) Unlike for PC, the performance of the regression algorithms for AC did not depend on
sample types; and (6) Similar to PC, preprocessing affected model
quality for AC, and the effect was closely associated with the
regression algorithm used.
For both PC and AC, sample type had an effect on the statistical
parameters of the models, though the models for the PC in brown
and milled our samples both performed well (all RSQ and 1 VR
values were high than 0.9100). Since the PC in milled ours are
inuenced by milling degree that tend to introduce more variations, we suggest to analyse the PC using brown our only when
evaluating for selection of early generation of rice in breeding programs or in mutation library construction. For AC, it is better to use
milled our type, because the calibration models for milled ours
were superior to those for brown ours.
The outcome of a research project largely depends on the variation in the materials and methods used. The NIR method tends to
perform best with narrow distribution ranges (Sohn et al., 2004a).
The models for the PC obtained in this work were less precise than
the PLS models reported by Delwiche, Bean, Miller, Webb, and Williams (1995), Barton et al. (2000), and Sohn et al. (2004a), where
the SEP was 0.107% (with a range of 510%, n = 150), the SECV
was 0.14% (with a range of 7.0310%, n = 216), and the SECV was
0.23% (with a range of 4.8912.48%, n = 214), respectively. This is
probably due to the use of a larger sample size and wider range
of PC in the current study(n = 335, a range of 6.713.8%). For the
AC of milled our, the current results were slightly superior to
those obtained by Delwiche et al. (1995), Bao et al. (2007b), and
Bao et al. 2007a generated by PLS, MPLS/SNV + detrend/1,5,5,1
and MPLS/ SNV + detrend/1,4,4,1. However, the prediction models established using PLS calibration by Sohn et al. (2004a) and
Himmelsbach et al. (2001) were slightly better than what was
showed in the current study.
The great advantage of this study is that the same group of samples was used for model combinations in regard of mathematical
derivations, spectrum data prepressing and regression methods,
allowing for direct comparisons among different data treatments
and regression methods. Such direct comparison provides a good
reference for selecting NIRS models.
Acknowledgements
This work was supported by funds from the National 863 Plan
Project of China (2011AA10A101) and the National S&T Major
Project of China (2011ZX08001-006).

99

References
Association of Ofcial Analytical Chemists (AOAC) (1999). Ofcial methods of
analysis (6th ed.). MD, USA: Gaithersburg.
Bao, J. S., Cai, Y. Z., & Corke, H. (2001). Prediction of rice starch quality parameters by
near infrared reectance spectroscopy. Journal of Food Science, 66, 936939.
Bao, J. S., Shen, Y., & Jin, L. (2007a). Determination of thermal and retrogradation
properties of rice starch using near-infrared spectroscopy. Journal of Cereal
Science, 46, 7581.
Bao, J. S., Wang, Y. F., & Shen, Y. (2007b). Determination of apparent amylose
content, pasting properties and gel texture of rice starch by near-infrared
spectroscopy. Journal of the Science of Food and Agriculture, 87, 20402048.
Bart, M. N., Katrien, B., Els, B., Ann, P., Wouter, S., Karen, I. T., et al. (2007).
Nondestructive measurement of fruit and vegetable quality by means of NIR
spectroscopy: A review. Journal of Postharvest Biology and Technology, 46,
99118.
Barton, F. E., Himmelsbach, D. S., McClung, A. M., & Champagne, E. T. (2000). Rice
quality by spectroscopic analysis: Precision of three spectral regions. Journal of
Cereal Chemistry, 77, 669672.
Cen, H. Y., & He, Y. (2007). Theory and application of near infrared reectance
spectroscopy in determination of food quality. Journal of Trends in Food Science
Technology, 18, 7283.
Champagne, E. T., Bett, K. L., Vinyard, B. T., Webb, B. D., McClung, A. M., Barton, F. E.,
et al. (1997). Effects of drying conditions, nal moisture content and degree of
milling on rice avor. Journal of Cereal Chemistry, 74, 566570.
Champagne, E. T., Bett-Garber, K. L., McClung, A. M., & Bergman, C. (2004). Sensory
characteristics of diverse cultivars as inuenced by genetic and environmental
factors. Journal of Cereal Chemistry, 81, 237243.
Champagne, E. T., Lyon, B. G., Min, B. K., Vinyard, B. T., Bett, K. L., Barton, F. E., et al.
(1998). Effects of postharvest processing on rice texture prole analysis. Journal
of Cereal Chemistry, 75, 181186.
Delwiche, S. R., Bean, M. M., Miller, R. E., Webb, B. D., & Williams, P. C. (1995).
Apparent amylose content of milled rice by near-infrared reectance
spectrophotometer. Journal of Cereal Chemistry, 72, 182187.
Delwiche, S. R., McKenzie, K. S., & Webb, B. D. (1996). Quality characteristics in rice
by near-infrared reectance analysis of whole-grain milled samples. Journal of
Cereal Chemistry, 73, 257263.
Fitzgerald, M. A., Bergman, C. J., Resurreccion, A. P., Moller, J., Jimenez, R., Reinke, R.
F., et al. (1986). Partial least squares regression: A tutorial. Journal of Analytical
Chimica Acta, 185, 117.
Himmelsbach, D. S., Barton, F. E., McClung, A. M., & Champagne, E. T. (2001). Protein
and apparent amylose contents of milled rice by NIR-FT/Raman spectroscopy.
Journal of Cereal Chemistry, 78, 488492.
Ishima, T., Taira, H., & Mikoshiba, K. (1974). Effect of nitrogenous fertilizer
application and protein content in milled rice on olganoleptic quality of
cooked rice (in Japanese with English abstract). Report of National Food Research
Institute, Japan, 29, 915.
Martens, H., & Ns, T. (1998). Multivariate calibration. John Wiley & Sons Ltd..
Martens, H., & Stark, E. (1991). Extended multiplicative signal correction and
spectral interference subtractionnew preprocessing methods for nearinfrared spectroscopy. Pharmaceutical Biomedical Analysis, 9, 625635.
Milica, P., Jasna, M., Dragan, P., & Mladenka, P. (2010). The development of nearinfrared spectroscopy (NIRS) calibration for prediction of as content in legumes
on the basis of two different reference methods. Journal of Food Chemistry, 12,
800805.
Ns, T., Isaksson, T., Fearn, T., & Davies, T (Eds). (2004). A user-friendly guide to
multivariate calibration and classication. In Scatter correction of spectroscopic
data (pp. 114118). Charlton, Chichester, UK: NIR publications.
Perez, C. M., & Juliano, B. O. (1978). Modication of the simplied amylose test for
milled rice. Journal of Starch, 30, 424426.
Shu, Q. Y., Wu, D. X., Xia, Y. W., Gao, M. W., & McClung, A. (1999a). Analysis of
grain quality characters in small ground brown rice samples by near
infrared reectance spectroscopy. Journal of Scientica Agronomica Sinica, 32,
9297.
Shu, Q. Y., Wu, D. X., Xia, Y. W., Gao, M. W., & McClung, A. (1999b). Calibration
optimization for rice apparent amylose content by near infrared reectance
spectroscopy (NIRS). Journal of Zhejiang University (Agriculture & Life Science), 25,
343346.
Sohn, M., Barton, F. E., McClung, A. M., & Champagne, E. T. (2004a). Near-Infrared
spectroscopy for determination of protein and amylose in rice our through use
of derivatives. Journal of Cereal Chemistry, 81, 341344.
Sohn, M., Himmelsbach, D. S., & Champagne, E. T. (2004b). A comparative
studies of Fourier transform Raman and NIR spectroscopic methods for
assessment of protein and apparent amylose in rice. Journal of Cereal
Chemistry, 81, 429433.
Tang, Q. Y., & Zhang, C. X. (2013). Data processing system (DPS) software with
experimental design, statistical analysis and data mining developed for use in
entomological research. Journal of Insect Science, 20, 254260.
Wu, J. G., & Shi, C. H. (2004). Prediction of grain weight, brown rice weight and
amylose content in single rice grains using near-infrared reectance
spectroscopy. Journal of Field Crops Research, 87, 1321.
Wu, J. G., & Shi, C. H. (2007). Calibration model optimization for rice cooking
characteristics by near infrared reectance spectroscopy (NIRS). Journal of Food
Chemistry, 103, 10541061.

100

L.H. Xie et al. / Food Chemistry 142 (2014) 92100

Xie, L. H., Chen, N., Duan, B. W., Zhu, Z. W., & Liao, X. Y. (2008). Impact of proteins on
pasting and cooking properties of waxy and non-waxy rice. Journal of Cereal
Science, 47, 372379.
Yanase, H., Ohtsubo, K., Hashimoto, K., Sato, H., & Teranishi, T. (1984). Correlation
between protein contents of brown rice and textural parameters of cooked rice
and cooking quality of rice (in Japanese with English abstract). Report of National
Food Research Institute, Japan, 45, 118122.

Zhang, J., Wu, J. H., Luo, L. J., Li, Y., Yang, H., Yu, X. Q., et al. (2007). Comparison of
near Infrared spectroscopy models for determining protein and amylose
contents between calibration samples of recombinant inbred lines and
conventional varieties of rice. Journal of Agricultural Sciences in China, 6,
941948.