Discriminant Analysis of Multiple Physicochemical Properties for Thai Rough Rice Varietal Authentication

Abstract

Rough rice variety is a factor influencing milling and eating qualities. To accomplish the prediction of five Thai rough rice varieties with high accuracy, physicochemical properties of those varieties were investigated and used for discriminant analysis. The results indicated that dry weight of a hundred rough rice grains and the content of brown rice protein offered moderate accuracy of 65.38 and 55.38%, respectively, in varietal differentiation in the validation set. Better distinction could be achieved using multiple properties, namely, the hundred-grain weight, brown rice protein content, milled rice protein content, apparent amylose content, and alkali spreading value. The developed model performed the definite prediction with 84.9% explained variances in the first two functions and 100% accuracy.

Keywords:

• Rough rice
• Physicochemical property
• Protein content
• Variety differentiation
• Discrimination analysis

INTRODUCTION

Rice (Oryza sativa L.) is an important energy source of people as it is a staple food in many countries. Thailand is a country producing rice for domestic and exporting purposes. Among rice exporting countries, Thailand has had the highest exported rice volume for at least 10 years. Thai Hom Mali rice, legally comprising only Khao Dawk Mali 105 (KDML105) and Kor Khor 15 (RD 15) varieties, is worldwide famous aromatic rice. It becomes a premium export product as compared to other Thai rice varieties because of its outstanding cooked rice characteristics, such as aroma, softness, and adhesiveness. The unique quality promisingly stimulates an upper price than the others both in domestic and international trades. However, rice quality is very difficult to convey with exact criteria because the preferences usually differ among consumers. Hence, the qualities of rice can be considered from several viewpoints, such as milling quality, grain appearance, cooking characteristics, and eating quality, which vary from variety to variety.[1,2] In rice cultivation and production, there is a possibility to raise a chance in off-variety contamination because of poor agricultural and manufacturing practices. High nutritional and economic value varieties can be substituted with other apparently similar, dietary poor, and lower cost ones. Owing to the similarity of milled rice kernel in shape and size, the cheaper Thai rice varieties, in particular Pathumthani 1 (PTT1), Suphanburi 60 (SPR60), Chainat 1 (CNT1), and Pitsanulok 2 (PSL2), have intentionally replaced costly Hom Mali rice varieties. This adulteration leads to the critical serious problem in trading, especially the PTT1 contamination because of the difficulty to indentify differences between PTT1 rice and KDML105 rice through physical properties and amylose content. Moreover, owing to PTT1 being an off-season rice crop, its annual production volume is infinitely higher than KDML105, which is an in-season rice crop. Newly harvested PTT1 grains can be produced throughout the year, which can be intentionally adulterated to older harvested KDML105 grains to expectantly keep some characteristics that are specific for new rice. However, among premium rice varieties of Thailand, the PTT1 variety was subordinate to the KDML105 variety in domestic and international trade at last.[3] Consumers are then probably exploited from illegitimate intentional making because of poorer cooking and eating qualities. Therefore, the rough rice variety authentication is extremely needed to obtain brown rice or milled rice with the same qualities, in particular for KDML105 and PTT1 varieties.

Physico-chemical properties are generally used as parameters in variety classification. Bett-Garber et al.[4] successfully divided global rice into groups by the differences of amylose, protein, flavor, and texture attributes. Nevertheless, these individual attributes still have a limitation where two varieties have a similarity of that studied property. Srisawas and Jindal[5] reported the potential to discriminate some rice varieties based on the determination of amylose and alkali spreading value because of significant differences, but this approach provided failure for KDML105 and PTT1. Hence, there has been an attempt to use various attributes as multi-parameters with the combination of multivariate analysis to achieve better discrimination. With the technique of discriminant analysis, Suwannaporn et al.[6] reported that the value of paste viscosity in conjunction with trough value could present a potential to discriminate rice varieties across their amylose group. Their results showed the successful classification for Thai milled rice varieties with 100% accuracy of medium- (SPR60) and high-amylose rice groups (CNT1). However, the developed model was also influenced by other factors, such as the effect of protein content and protein-starch complexes on viscosity intensity and rate of starch gelatinization. Thus, more investigation was then suggested due to only 66.7% of accurate classification for this low-amylose rice group (KDML105 and PTT1). Both viscoelastic properties have an involvement with the quantity of amylose and protein in rice in common as well as the structure and type of the protein.[7,8] Therefore, the aims of this study were to investigate physicochemical properties of commercial Thai rough rice varieties consisting of KDML105, PTT1, SPR60, CNT1, and PSL2 based on physical characteristics, constitutions, and dispersion of starch in alkali solution to accomplish their correlations, and to study the application of those multi-properties in discriminant analysis to corroborate their varieties.

MATERIALS AND METHODS

Rough Rice Samples

Five favorite Thai rough rice varieties, namely, one Hom Mali (aromatic and low amylose variety; KDML105) and four non-Hom Mali (aromatic and low amylose variety; PTT1, non-aromatic and intermediate amylose variety; SPR60, and non-aromatic and high amylose varieties; CNT1 and PSL2), were studied. All 259 rough rice samples were varieties in F3 and F4 generation and were harvested between December 2009 and March 2010. Each sample was collected from different locations in 30 provinces of Thailand (Table 1). To be authentic representatives of those varieties, all samples were derived from rice seeds certified by the Department of Rice, Ministry of Agriculture and Cooperatives, Thailand. Moreover, only samples where farmers assiduously uprooted and cut off-type rice plants in the period of rice development were chosen. The purity testing in the certified seeds production was ordinarily controlled and verified by the official agency, which referred to seed testing methods of International Seed Testing Association (ISTA). All chemical reagents used were of analytical grade.

Table 1  Rough rice samples collected from different areas in Thailand

Variety	Number of samples	Number of locations	Number of provinces
KDML105	66	66	16 ^a
PTT1	50	50	9 ^b
SPR60	39	39	10 ^c
CNT1	51	51	10 ^d
PSL2	53	53	7 ^e
^aKDML105 grains originated from Nakhon Ratchasima, Korn Kaen, Burirum, Mahasarakham, Roi Et, Si Sa Ket, Ubonratchathani, Udon Thani, Surin, Mukdahan, Yasothon, Kamphaeng Phet, Phayao, Lamphun, Chiang Rai, and Chiang Mai.
^bPTT1 grains originated from Pathum Thani, Nakhon Sawan, Nakhon Pathom, Phetchaburi, Ratchaburi, Phra Nakhon Si Ayutthaya, Chai Nat, Suphan Buri, and Kanchanaburi.
^cSPR60 grains originated from Nakhon Pathom, Ratchaburi, Phetchaburi, Phra Nakhon Si Ayutthaya, Suphan Buri, Chai Nat, Kanchanaburi, Sing Buri, Saraburi, and Nakhon Sawan.
^dCNT1 grains originated from Nakhon Pathom, Ratchaburi, Phra Nakhon Si Ayutthaya, Suphan Buri, Chai Nat, Uthaithani, Ang Thong, Kanchanaburi, Nakhon Sawan, and Nakhon Ratchasima.
^ePSL2 grains originated from Nakhon Pathom, Phitsanulok, Ratchaburi, Phra Nakhon Si Ayutthaya, Suphan Buri, Chai Nat, and Kanchanaburi.

Sample Preparation

The selected samples were sieved and blown with air at room temperature, 30 ± 3°C, to eradicate foreign matters. Each assigned sample was kept in a glass bottle that was tightly closed with a plastic cover and stored in a cool room at 6 ± 1°C until testing. To meet an equilibrium prior initial testing, the packed samples were left at room temperature for at least 24 h.

Every cleaned rough rice sample was portioned as approximately 300 g to be shelled with a Satake dehusker THU-35A (Satake Engineering Co., Ltd., Tokyo, Japan) to obtain brown rice. The clearance between two roller rubbers was fixed at 0.8 mm. The wholesome kernels were attained by using a Satake rice grader TR6-05A with a 30 degree sample pot angle (Satake Engineering, Tokyo, Japan). Before dehusking, rough rice samples were dried to reduce moisture content to nearly 15% with wet basis by a hot air drier at certain 45 ± 1°C. Approximately 150 g of sound brown rice samples were milled at about a 10% degree of milling by a grain testing mill machine (Satake Engineering, Tokyo, Japan) to gain milled rice samples.

Hundred-Grain Weight

One hundred cleaned and wholesome kernels of each rough rice sample were counted, and then were weighted by a Shimadzu balance type AX200 with four decimal positions (Shimadzu Corporation, Tokyo, Japan). The works were accomplished in triplicate per sample. The measured values in gram were calculated to change the unit to be dry basis. Afterward, all replicate values were averaged for one sample. The standard deviation was also calculated and reported based on its variety.

Dimensional Characteristics

The length and the width of 20 cleaned and sound rough rice kernels of each sample were measured by a vernier caliper (Mitutoyo Corporation, Japan) in triplicate. The mean of length, width, and ratio of length to width was calculated and reported. The dimensional analysis was also done in brown rice samples (n = 259).

Moisture Content

Initial moisture content of rough rice samples was analyzed by applying the approved method 44-15A of the American Association of Cereal Chemists,[9] which is suggested for corn and beans. The rough rice samples, modified as approximately 3 g, were weighed with the exact four decimals, and then were dried at 103 ± 1°C for 72 h in triplicate. A whole-grains test was applied to avoid moisture loss during grinding. Two-stage drying was employed for the high moisture rough rice sample. It used air temperature at 45 ± 1°C in the first step for water removing till the moisture content remained at approximately 13%. The moist grains were weighed and calculated the moisture content of the first stage drying. Afterward, the samples were secondarily dried at 103 ± 1°C until obtaining the steady weight. The moisture content in rough rice based on 100 g wet basis was calculated and reported.

Crude Protein Content

Crude protein analysis was accomplished in brown rice samples (n = 259) and milled rice samples (n = 259) in triplicate. An approved micro-Kjeldahl method of the American Association of Cereal Chemists was applied.[7] Five-hundred milligram grounded samples were digested with 8 ml concentrated sulfuric acid in the presence of a catalyst (potassium sulfate:copper sulfate:Selenium = 50:10:1). Steam distillation with 40% sodium hydroxide solution was then completed to liberate ammonia into 4% saturated boric acid. Back titration was applied to achieve total crude nitrogen content. Exact concentration of hydrochloric acid solution (1N HCl) was used as the standard acid solution. Methyl red (0.1 gram methyl red and 0.1 gram bromcresol green in 100 ethyl alcohol solution) was added as an indicator. The nitrogen content was converted to protein content through using a conversion factor of 5.95.[10]

Apparent Amylose Content

Apparent amylose content in each milled rice sample was determined by a simplified colorimetric assay suggested by Juliano.[11] A milled rice sample was ground to be a flour sample. Distilled water was used to disperse the 100 mg of flour, and then 95% ethanol and 2N sodium hydroxide solution were added to treat the dispersion. The sample solution was subsequently heated in nearly boiling water for an hour and was left at room temperature overnight. The pH was then adjusted by 1N acetic acid solution before adding the iodine solution. As the amylose formed a complex with the iodine, the blue color would be developed. The absorbance of the developed color solution was measured by a Jenway-6405 spectrophotometer (Jenway Ltd., Essex, UK) at 620 nm. All tests were carried out in triplicate. Pure amylose extracted from potatoes was used for creating a standard curve.

Alkali Spreading Value Test

The alkali spreading value of milled rice samples was investigated corresponding to the procedures of Little et al.[12] and Delwiche et al.[13] Six milled rice grains of each sample were immersed in 1.7% potassium hydroxide solution at room temperature for 23 h. Those kernels were separately rested in a Petri dish that was placed on a black board. The level of intactness of an individual grain was examined and given a score. The score ranged from 2 to 7 referring to the disintegration of a kernel from lowest to highest, respectively.

Statistical Analysis

A multivariate analysis through SPSS® software version 14.0 (SPSS Inc., Chicago, IL, USA) was applied for statistical analysis. Relationships between physical attributes and chemical properties of all samples (n = 259) were achieved via the Pearson correlation coefficients at 0.05 and 0.01 of significant level (P ≤ 0.05 and P ≤ 0.01). One-way analysis of variance (ANOVA) was utilized to assess differences between the average values of all variables among all five observed varieties at 0.05 of significant level (P ≤ 0.05). The analyzed significances that displayed in homogeneity of the variance test were used to choose the methods of post hoc tests. The least significant difference (LSD) technique was applied for parameters in which higher analyzed significant values than the set significant level were achieved. The parameters that had a lower analyzed significant value were compared through the Dunnett's T3 technique. Discriminant analysis by using the stepwise Wilks’ Lambda technique was employed to acquire predictive models for rough rice varietal differentiation. Samples were divided to calibration set and validation set. The samples in the calibration set (n = 130) consisted of 33, 25, 20, 25, and 27 samples of KDML105, PTT1, SPR60, CNT1, and PSL2 variety, respectively. The other samples (n = 129) were in the validation set. These validation samples were used to prove the performance of the obtained models. The discriminant analysis is generally accomplished on the basis of observed characteristics in each group and assumes a sample group-membership by using a prior knowledge of those groups.[14] The discrimination of rough rice varieties was explored by using both individual and multiple attributes.

RESULTS AND DISCUSSION

All samples (n = 259) were determined by their physicochemical properties, namely, the hundred-grains weight, the length and the width of rough rice and brown rice kernels, the ratios of the length to the width, the initial moisture content of rough rice, the crude protein content in brown rice and milled rice, the apparent amylose content, and the alkali spreading value. Although the association among rice quality attributes has been performed by many researchers,[15–18] non-consistency of the correlation coefficients were occasionally perceived because there was diversity among rice genomes and the attributes involving rice qualities were very complicated.[19] In this study, the relationship among these properties is summarized in Table 2. In rough rice samples, the negative correlation was significantly determined between the initial moisture content and the dry weight of hundred kernels with the correlation coefficient (R) value of -0.747 (P ≤ 0.01). There were significantly positive relations between the grain length and the ratio of length to width in rough rice kernels (R = 0.729, P ≤ 0.01) and in brown rice kernels (R = 0.482, P ≤ 0.01). Conversely, the correlation of the grain width and the ratio length to width showed significantly negative for rough rice and brown rice with the R values of -0.628 and -0.504, respectively (P ≤ 0.01). These results were in line with the study of Thongbam et al.,[20] who worked in milled rice samples, and reported the significant positive and negative correlations between the ratio of length to width and the length (R = 0.72, P < 0.01) or the width (R = -0.77, P < 0.01), respectively. Likewise, the ratios of length to width in rough rice and brown rice samples had a low correlation with brown rice protein content and milled rice protein content, respectively. The R values were revealed at -0.191 (P ≤ 0.01) and -0.158 (P ≤ 0.05), respectively. In addition, these four dimensional characteristics expressed negative relations with the alkali spreading value. Among the physicochemical properties, low relations in negative were observed between the apparent amylose content and the crude protein content in brown rice and milled rice with R values of -0.228 and -0.213, respectively, at 99% significance level. The apparent amylose content had a higher negative relation with the alkali spreading value with R values of -0.524 at 99% significance level, which was similar to the results in many previous studies.[15,21] The positive correlation was detected between the crude protein content and the alkali spreading value (R = 0.29, P ≤ 0.01). However, the negative correlation of these two properties was reported by Thongbam et al. through R value of -0.47 at 95% significance level.[20]

Table 2  Pearson correlation coefficients for association of physicochemical properties of Thai rough rice samples

	A	B	C	D	E	F	G	H	I	J	K	L
A	1
B	0.042	1
C	0.059	0.072	1
D	−0.004	0.729**	−0.628**	1
E	−0.075	0.296**	−0.164**	0.346**	1
F	−0.174**	−0.356**	0.304**	−0.485**	0.513**	1
G	0.106	0.658**	−0.479**	0.844**	0.482**	−0.504**	1
H	−0.747**	−0.001	−0.013	0.006	0.056	0.106	−0.053	1
I	0.207**	−0.065	0.207**	−0.191**	−0.174**	−0.106	−0.065	−0.219	1
J	−0.400**	−0.060	0.088	−0.108	−0.074	0.080	−0.158*	0.211	0.692**	1
K	−0.023	0.072	−0.009	0.066	0.472**	0.151*	0.313**	−0.039	−0.228**	−0.213**	1
L	−0.221**	−0.195**	0.062	−0.200**	−0.271**	−0.105	−0.160**	0.368**	0.346**	0.290**	−0.524**	1
A: hundred-rough rice grain weight (g dry basis); B: rough rice length (mm); C: rough rice width (mm); D: rough rice length to width ratio; E: brown rice length (mm); F: brown rice width (mm); G: brown rice length to width ratio; H: rough rice moisture content (g/100 g); I: brown rice crude protein content (g/100 g dry solids); J: milled rice crude protein content (g/100 g dry solids); K: apparent amylose content (g/100 g dry solids); L: alkali spreading value.
*Significant at P ≤ 0.05 (2-tailed); **P ≤ 0.01 (2-tailed).

In addition, based on the variety, some physicochemical properties significantly differed. The mean and standard deviation of these variables are shown in Table 3. The dry solids of wholesome KDML105, PTT1, SPR60, CNT1, and PSL2 rough rice samples containing various initial moisture content varied from 2.44 to 2.52, 2.22 to 2.40, 2.05 to 2.49, 2.31 to 2.50, and 2.37 to 2.46 g/100 kernels, respectively. Mean values of this hundred-grain weight revealed the significant difference among all five varieties. The overlapping in ranges of the length, the width, and the ratios of length to width of rough rice and brown rice were observed over five varieties. However, the mean values of the CNT1 rough rice and brown rice lengths expressed significant differences from that of the KDML105 variety. The significant difference in rough rice length to width ratio was achieved between the CNT1 and the PSL2 varieties. The SPR60, CNT1, and PSL2 brown rice varieties had significantly bigger width than the KDML105 variety. Overall physical characteristics of KDML105 and PTT1 varieties were very similar because of no significant differences in the average values of the length, width, and ratio of length to width of rough rice, and the length and ratio of length to width of brown rice.

Table 3  Physicochemical properties of five Thai rough rice varieties

	Mean and standard deviation
Properties	KDML105	PTT1	SPR60	CNT1	PSL2
Hundred grains weight (g dry basis)	2.50 ± 0.02^a	2.36 ± 0.04^d	2.31 ± 0.07^e	2.48 ± 0.04^b	2.43 ± 0.02^c
Rough rice length (mm)	10.64 ± 0.38^bc	10.79 ± 0.49^ab	10.60 ± 0.36^bc	10.90 ± 0.50^a	10.68 ± 0.41^ac
Rough rice width (mm)	2.67 ± 0.09^a	2.69 ± 0.06^a	2.63 ± 0.08^b	2.67 ± 0.11^a	2.70 ± 0.10^a
Rough rice length to width ratio	3.99 ± 0.20^ad	4.02 ± 0.21^ac	4.04 ± 0.17^ab	4.10 ± 0.26^a	3.96 ± 0.20^bcd
Brown rice length (mm)	7.67 ± 0.24^b	7.72 ± 0.29^b	7.98 ± 0.27^a	8.01 ± 0.23^a	7.99 ± 0.30^a
Brown rice width (mm)	2.18 ± 0.10^c	2.21 ± 0.09^b	2.26 ± 0.08^a	2.23 ± 0.07^ab	2.21 ± 0.08^b
Brown rice length to width ratio	3.53 ± 0.13^b	3.49 ± 0.14^b	3.54 ± 0.13^b	3.60 ± 0.11^a	3.61 ± 0.14^a
Initial moisture content (g/100 g)	16.16 ± 1.72^c	19.85 ± 2.41^a	21.18 ± 3.08^a	15.72 ± 0.82^c	18.35 ± 1.93^b
Brown rice protein content (g/100 g dry solids)	9.23 ± 0.23^b	9.60 ± 0.58^a	8.03 ± 0.51^d	8.52 ± 0.24^c	9.81 ± 0.24^a
Milled rice protein content (g/100 g dry solids)	6.98 ± 0.20^d	8.14 ± 0.49^a	7.29 ± 0.47^c	6.81 ± 0.18^e	7.60 ± 0.35^b
Apparent amylose content (g/100 g dry solids)	15.58 ± 1.12^d	16.30 ± 1.07^c	24.07 ± 1.93^b	28.43 ± 1.14^a	28.58 ± 1.18^a
Alkali spreading value	6.96 ± 0.08^a	6.91 ± 0.14^ab	6.62 ± 0.53^c	5.21 ± 0.20^d	6.87 ± 0.11^b
In a row, means and standard deviation tagged with the same alphabet are not significant difference (P < 0.05).

The results displayed that crude protein content in brown rice samples ranged from 8.77 to 9.84, 8.29 to 10.72, 7.05 to 8.98, 7.79 to 8.97, and 9.21 to 10.31 g/100 g dry solids for KDML105, PTT1, SPR60, CNT1, and PSL2 varieties, respectively. Their milled rice, with 10% degree of milling, contained crude protein content from 6.60 to 7.54, 7.02 to 9.13, 6.38 to 8.16, 6.27 to 7.18, and 6.73 to 8.26 g/100 g dry solids, respectively. Likewise, the apparent amylose content also varied subsequently covering from 13.01 to 17.76, 15.00 to 19.45, 20.21 to 26.00, 25.76 to 30.14, and 26.65 to 31.35 g/100 g dry solids, respectively. The average quantities of apparent amylose showed a bit lower content in KDML105 rice than PTT1 rice (15.58 and 16.30 g/100 g dry solids, respectively) and the quantity was correspondingly raised in SPR60, CNT1, and PSL2 samples. These results affirmed that the KDML105 and the PTT1 rice varieties were in the same low-amylose group, but the SPR60 rice variety was in the intermediate-amylose group. Moreover, the CNT1 and PSL2 rice varieties were in the same high-amylose group. The results of alkali spreading value testing corroborated the specific trait of KDML105 and PTT1 variety presenting the score from 6 to 7 only. Moreover, it asserted that, although the alkali spreading value was trendily decreasing in low-, medium-, and high-amylose groups in order, some high-amylose varieties, in particular SPR60 and PSL2, could have the score in the same range of KDML105 and PTT1 varieties.

In addition, the mean values of physicochemical properties based on varieties revealed that there were significant differences in crude protein content of brown rice and milled rice and apparent amylose content between KDML105 variety and PTT1 or SPR60 or CNT1 or PSL2 variety. Champagne et al.[3] have also indicated that KDML105 milled rice had lower significant differences in apparent amylose and protein content than PTT1 milled rice. However, no significant difference between the averaged brown rice crude protein content of PTT1 and PSL2 varieties were observed. Likewise, the averaged apparent amylose content of CNT1 and PSL2 varieties were not significantly different. The alkali spreading value could be used to distinguish the CNT1 variety from KDML105 and PTT1 varieties because the range and mean values of those varieties were notably disparate.[5] Nonetheless, the alkali spreading value of KDML105, PTT1, SPR60, and PSL2 rice varieties were too close to use for distinguishing their varieties.

The varietal discrimination of all rough rice varieties based on one of these parameters could not be accomplished although there were some properties significantly different among studied rough rice varieties. The wide variations in those properties resulted in the failure. In this study, the discriminant analysis with Wilks’ Lambda technique was used for proving the variety of pure rough rice samples based on an individual physicochemical attribute. The coefficient and constant of each attribute for varietal identification are shown in Table 4. The validation results revealed that the highest total correction of 65.38% could be achieved as the hundred-grain weight in dry basis was used for modeling for identifying rough rice varieties (Table 5). The models could correctly identify 23, 19, 8, 13, and 22 samples from the total of 33, 25, 20, 25, and 27 samples of KDML105, PTT1, SPR60, CNT1, and PSL2 samples, respectively. Among chemical attributes, crude protein in brown rice offered the highest total correction at 55.38%. Chandi and Sogi[22] suggested that specific proteins could be associated with the clear distinction of traditional and developed Basmati rice besides the dimensional changes. However, the performance of the models that were developed using the individual parameter of brown rice crude protein was too low for future prediction to authenticate those varieties. The analysis proffered the accuracy of 64.7% for KDML105, 35.3% for PTT1, 66.7% for SPR60, 51.4% for CNT1, and 56.5% for PSL2 variety. The apparent amylose content, alkali spreading value, and milled rice crude protein could correctly identify 54.62, 52.31, and 47.69% of all samples. The authentication of variety on the basis of the apparent amylose content offered the correction of 57.1% for KDML105, 43.5% for PTT1, 90.5% for SPR60, 42.3% for CNT1, and 45.8% for PSL2 variety. Entire false negative identification generally occurred among varieties within the same amylose type, in particular between KDML105 and PTT1 varieties. Literally, both KDML105 and PTT1 varieties have had the similarities in amylose content that led to the inability to detect the adulteration dyed grain method with a Thymol blue indicator.[23] The high-amylose rice, CNT1, could be misidentified to be another high-amylose rice, PSL2, and the intermediate-amylose rice, SPR60. For the attributes of dimensional characteristics, they were unsuitable to use for differentiating these five varieties because the total accuracy values were lower than 30% (data not shown). The length, the width, and the ratio of length to width of rough rice were not modeled in this regard because, at first step, the entering F values were lower than the minimum numbers set as 3.84. An interesting discriminant result was disclosed that there was an impact of rough rice moisture content variation on varietal differentiation. The 30.77% accuracy of varietal authentication based on the rough rice moisture content could confirm the effect. In other words, the varietal distinguishing analysis through using multi-attributes simultaneously then can offer a prejudice if the rough rice moisture content is included for model construction. Thus, this attribute was left out from modeling to avoid the error due to moisture difference.

Table 4  Discriminant coefficients of variety group based on individual physicochemical attribute

	Identification function coefficients*
Attribute	KDML105	PTT1	SPR60	CNT1	PSL2
Hundred-grain weight (g dry basis)	1250.309	1181.725	1149.807	1241.338	1213.609
	(-1563.783)	(-1397.101)	(-1322.737)	(-1541.446)	(-1473.419)
Brown rice length (mm)	125.303	127.544	132.269	132.472	131.730
	(-480.818)	(-498.114)	(-535.579)	(-537.220)	(-531.238)
Brown rice width (mm)	337.495	348.501	352.823	348.251	345.702
	(-366.360)	(-390.536)	(-400.356)	(-389.979)	(-384.314)
Brown rice length to width ratio	216.514	213.389	218.425	221.750	222.006
	(-385.134)	(-374.144)	(-391.935)	(-403.908)	(-404.838)
Rough rice moisture content (g/100 g)	3.252	4.017	4.339	3.187	3.773
	(-27.409)	(-40.992)	(-47.548)	(-26.395)	(-36.3538)
Brown rice crude protein (g/100 g dry solids)	67.068	69.155	58.114	61.528	71.144
	(-314.070)	(-333.821)	(-236.205)	(-264.584)	(-353.208)
Milled rice crude protein (g/100 g dry solids)	61.059	70.526	63.381	59.170	66.847
	(-217.255)	(-289.309)	(-233.972)	(-204.122)	(-260.076)
Apparent amylose content (g/100 g dry solids)	8.371	8.952	12.818	15.247	15.380
	(-66.917)	(-76.289)	(-154.729)	(-218.251)	(-222.043)
Alkali spreading value	106.825	106.202	101.160	79.267	104.997
	(-372.744)	(-368.432)	(-334.424)	(-205.960)	(-360.156)
*Based on Fisher's linear discrimination function.
Values in parentheses are constant (β0). Significances of the discriminate function by Wilks’ Lambda method equal to 0.000.

Table 5  Varietal discrimination results of samples in validation set based on individual physicochemical attribute

Attribute and varieties	n	TP	FP	Sensitivity	Specificity	% Total accuracy
Hundred-grain weight (g dry basis)						65.38
KDML105	33	23	11	0.6765	0.8866
PTT1	25	19	13	0.5938	0.8762
SPR60	20	8	3	0.7273	0.9727
CNT1	25	13	14	0.4815	0.8667
PSL2	27	22	3	0.8800	0.9709
Rough rice moisture content (g/100 g)						30.77
KDML105	33	10	26	0.2778	0.7320
PTT1	25	3	16	0.1579	0.8476
SPR60	20	11	14	0.4400	0.8727
CNT1	25	8	13	0.3810	0.8762
PSL2	27	8	20	0.2857	0.8058
Brown rice crude protein (g/100 g dry solids)						55.38
KDML105	33	22	12	0.6471	0.8763
PTT1	25	6	11	0.3529	0.8952
SPR60	20	12	6	0.6667	0.9455
CNT1	25	19	18	0.5135	0.8286
PSL2	27	13	10	0.5652	0.9029
Milled rice crude protein (g/100 g dry solids)						47.69
KDML105	33	11	13	0.4583	0.8660
PTT1	25	16	3	0.8421	0.9714
SPR60	20	5	13	0.2778	0.8818
CNT1	25	22	26	0.4583	0.7524
PSL2	27	8	12	0.4000	0.8835
Apparent amylose content (g/100 g dry solids)						54.62
KDML105	33	20	15	0.5714	0.8454
PTT1	25	10	13	0.4348	0.8762
SPR60	20	19	2	0.9048	0.9818
CNT1	25	11	15	0.4231	0.8571
PSL2	27	11	13	0.4583	0.8738
Alkali spreading value						52.31
KDML105	33	27	34	0.4426	0.6495
PTT1	25	3	12	0.2000	0.8857
SPR60	20	8	7	0.5333	0.9364
CNT1	25	26	1	0.9630	0.9905
PSL2	27	4	7	0.3636	0.9320
TP: true positive; FP: false positive; Sensitivity: the proportion of correctly classified actual positives that was calculated from the division of the number of true positives by all testing samples number; Specificity: the proportion of correctly classified actual negatives that was calculated from the ratio of the number of true negatives to all testing samples number.

Based on the varietal authentication by using multiple attributes, the developed predictive models could completely distinguish all five rough rice varieties by referring to only five physicochemical attributes, which were the hundred-grain weight of rough rice, the crude protein content of brown rice and milled rice, the apparent amylose content, and the alkali spreading value. The results indicated that only four functions were used for analysis. Those five discriminators supported high performance of identification because an Eigen value of the best discriminators at the first function equaled to 54.22. In general, the coefficients for discriminant function interpret the correlation of the variables. Weight of each variable and constant are indicated for a model of each group. Hence, the assessment of single variable importance using the coefficients is unsuitable. However, the assessment is possible when the occurred unequal means and standard deviations are adjusted by standardizing the coefficients.[24] Consequently, the standardized canonical discriminant function coefficients (Table 6) implied that the rough rice variety group produced larger function values in terms of brown rice crude protein, milled rice crude protein, apparent amylose content, and alkali spreading value. The term of hundred-grain weight generated the highest function value of -0.420 for function 3, but this value is still lower than the highest value of other attributes. The relative order and function values of all variables suggested that the crude protein both in brown rice and milled rice are the best discriminant predictors for the study relating to varietal differentiation because very high function values of -4.666 and 4.959 were generated for function 1. Similar to other researches, significant differences could be pointed out among rice varieties where the protein content was compared.[25] Based on protein distribution, the purity of Japanese Koshihikari milled rice bulk at 5% adulteration of Akitakomachi milled rice.[26] The different levels of albumin and glutelin could cause the variety differences.[27,28]

Table 6  Standardized canonical discriminant function coefficients based on calibration samples

	Function
Properties	1	2	3	4
Hundred-grain weight (g dry basis)	−0.300	−0.298	−0.420	−0.017
Brown rice crude protein (g/100 dry solids)	−4.666	0.768	1.010	0.238
Milled rice crude protein (g/100 dry solids)	4.959	−0.020	0.101	0.420
Apparent amylose content (g/100 dry solids)	0.016	−0.600	0.883	−0.111
Alkali spreading value (g/100 dry solids)	0.362	1.104	0.726	−0.386

Table 7  Discriminant coefficients of variety group multiple attributes excluding rough rice moisture content and identification results in validation set

	Identification function coefficients*
Attribute	KDML105	PTT1	SPR60	CNT1	PSL2
Hundred-rough rice weight (g dry basis)	830.8727	708.1769	709.7672	890.0582	741.6108
Brown rice crude protein (g/100 g dry solids)	259.2553	104.5739	−2.06733	180.2293	239.0522
Milled rice crude protein (g/100 g dry solids)	−103.357	92.06089	194.8669	−40.1259	−53.6603
Apparent amylose content (g/100 g dry solids)	14.14344	15.74797	19.15129	21.4284	23.01027
Alkali spreading value	228.0604	256.3647	237.979	171.7532	241.9112
Constant	−2785.2	−2732.65	−2534.9	−2485.92	−3030.8
Correct identification (%)	100	100	100	100	100
Sensitivity	1.00	1.00	1.00	1.00	1.00
Specificity	1.00	1.00	1.00	1.00	1.00
*Based on Fisher's linear discrimination function.
Significance of the discriminate function by Wilks’ Lambda method equal to 0.000.

The Wilks’ Lambda values of this study also expressed high ability in discrimination through those variables because of very low values at 0.000, 0.001, 0.042, and 0.670 for the first four canonical discriminant functions. The discrimination function coefficients of those attributes are shown in Table 7. The constants for each group are also presented. Moreover, in the first two functions, there were big differences among the group centroid of each variety, which certainly supported an excellent varietal discrimination of samples in the calibration set. The discrimination score of calibration samples is illustrated in Fig. 1. The KDML105, PTT1, and SPR60 varieties were distinguished with the first function. A somewhat overlapping area between CNT1 and PSL2 varieties was observed under the first function as well as that between PSL2 and KDML105 varieties. However, the differentiations of those varieties were accomplished in function 2. With these models, although rough rice moisture content was not considered in varietal discriminant analysis, an excellent varietal identification, 100% of accuracy, was still obtained for all 129 independent rough rice samples in the validation set. The sensitivity and specificity displayed in the highest level of 1.00. The discrimination plot (Fig. 2) illustrates that samples distribute around their variety group centroid correctly. The first function provided very clear differentiation among KDML105, PTT1, and SPR60 group members. Likewise, members of KDML105, CNT1, and PSL2 were projected in the plot without an overlapping. However, it was suggested that the application of this technique may have a limitation because the conventional methods to obtain those attributes can cause the delay measurement.

Figure 1 Discrimination plot of first two functions of rough rice varieties (KDML105, PTT1, SPR60, CNT1 and PSL2) in calibration set based on hundred rough rice weight, brown rice crude protein, milled rice crude protein, apparent amylose content, and alkali spreading value.

Figure 2 Discrimination plot of first two functions of rough rice varieties (KDML105, PTT1, SPR60, CNT1 and PSL2) in validation set based on hundred rough rice weight, brown rice crude protein, milled rice crude protein, apparent amylose content, and alkali spreading value.

CONCLUSION

Our findings strongly support that proteins in brown rice and milled rice are key factors for variety discrimination, while the dimensional characteristics of rough rice and brown rice provide less ability to differentiate varieties of KDML105, PTT1, SPR60, CNT1, and PSL2 from each other. The discriminant analysis for authenticating these varieties implies that efficient models can be developed by using several attributes, such as hundred-rough rice weight, brown rice protein, milled rice protein, apparent amylose content, and alkali spreading value. This approach can support an efficient quality control system at drying and milling facilities and help to protect a deception, which leads to the misunderstanding that it is a desirable rough rice variety, in particular a premium rice variety, such as KDML105. Therefore, an overall result ascertains the useful approach, accounting for the use of multiple physiochemical attributes in combination with multivariate analysis for rice varietal identification in routine works, as rapid indirect methods for determining these chemical attributes are used together.

ACKNOWLEDGMENTS

The authors are grateful to the Asian Institute of Technology for funds for this research. Additional special thanks go to the Ministry of Agriculture and Cooperatives, Thailand for financial support to pursue a doctoral study for Namaporn Attaviroj.