Quality determination of Chinese rice wine based on Fourier transform near infrared spectroscopy

Transcription

1 H. Yu et al., J. Near Infrared Spectrosc. 14, (2006) 37 Quality determination of Chinese rice wine based on Fourier transform near infrared spectroscopy Haiyan Yu, Yibin Ying, * Xiaping Fu and Huishan Lu College of Biosystems Engineering and Food Science, Zhejiang University, Hangzhou , PR China. ybying@zju.edu.cn To euate the applicability of near infrared (NIR) spectroscopy for the determination of five enologi parameters (alcoholic degree, ph ue, total acid, amino acid nitrogen and Brix) in Chinese rice wine, transmission spectra were collected in the spectral range from 800 nm to 2500 nm in a 1 mm path length rectangular quartz cuvette with air as the reference at room temperature. Five ibration equations for the enologi parameters were established between the reference data and NIR spectra by partial least squares (PLS) regression, separately. The best ibration results were achieved for the determination of alcoholic degree and Brix. The coefficients of determination in ibration (R 2 ) for alcoholic degree and Brix were 0.93 and 0.96, respectively. The predictive deviation ratio (RPD) ue of the ibration for alcoholic degree was higher than 3 (3.95), which demonstrated the robustness of the ibration model. The RPD ue for Brix was The performance of the ibration models for ph, total acid and amino acid nitrogen were not as good as that of alcoholic degree and Brix. The RPD ues for the three parameters were 1.41, 1.38 and 1.27, respectively and R 2 ues were 0.97, 0.83 and 0.90, respectively. In the idation step, r 2 ues for alcoholic degree and Brix were all higher than 0.9. The performance of ph and total acid were acceptable [the coefficients of determination in idation (r 2 ) for ph and total acid were 0.82 and 0.77, respectively]. The performance of amino acid nitrogen model was the worst, with an r 2 ue of The results demonstrated that the NIR spectroscopic technique could be used to predict the concentrations of these five enologi parameters in Chinese rice wine. Keywords: NIR spectroscopy, PLS regression, Chinese rice wine, enologi parameters Introduction Determination of food quality is one of the most important issues in food quality control and safety. The quality of Chinese rice wine is regulated by strict guidelines which include sensory euation, chemi analysis and examination of the records kept by the wine producer laid down by the responsible national authorities. Seven enologi parameters are included in chemi analysis: total sugars (addition of glucose), sugar-free extract, alcoholic degree, ph ue, total acid, amino acid nitrogen and cium oxide. The established methods for the determination of these parameters are generally based on wet analysis, gas chromatography (GC) and atomic absorption spectroscopy (AAS). However, the conventional chemi analyses are relatively complex, require sample preparation and are time consuming, laborious and costly. 1 The NIR spectroscopic technique has gained wide acceptance in the field of food chemistry, mainly due to its suitability for recording the spectra of solid and liquid samples at low cost, without any pre-treatment and in a non-destructive way. 2 It is also an easy-to-use, reliable and versatile analyti method for determining different compounds in wine, as well as in many other food products. The NIR region of the electromagnetic spectrum lies between the visible and middle infrared regions, spans the wavelength range between 780 nm and 2526 nm and contains information concerning the relative proportions of C H, N H and O H bonds which are the primary structural components of organic molecules. 3 Quantitative NIR measurements are usually based on the correlation between sample composition, as determined by defined reference methods and the absorption of light at different wavelengths in the NIR region measured by either reflection or transmission mode. 4 The NIR spectroscopic technique has been used to monitor ethanol in alcohol fermentation processes, 5 to predict methanol in wine-fortifying spirit, 6 to euate ethanol content of alcoholic beverages, 7,8 to determine phenolic compounds (malvidin-3-glucoside, pigmented polymers and tannins) in red wine fermentations 4 and to screen up to 15 parameters (alcoholic degree, volumic mass, total acidity, ph, volatile acidity, glycerol, total polyphenol index, reducing sugars, lactic, malic, tartaric and gluconic acids, colour, tonality, total sulphur dioxide and free sulphur dioxide) in different types of wines (red, rose and white wines). 9 NIR Publications 2006, ISSN

2 38 Quality Determination of Chinese Rice Wine by FT-NIR However, no reports were found relating to the use of NIR spectroscopy for determining the quality of Chinese rice wine. The aim of this work was to investigate the applicability of NIR spectroscopy for the determination of five enologi parameters (alcoholic degree, ph ue, total acid, amino acid nitrogen and Brix) in Chinese rice wine. Materials and methods Samples Eighty-eight bottles of Chinese rice wine were collected from supermarkets in Hangzhou, China. Commercially available bottles of wine were sourced from different commercial labels of Shaoxing rice wine, among which 38 were from Tapai, 33 from Kuaijishan, six from Guyuelongshan, four from Nverhong, three from Xianheng, two from Huangzhonghuang and two from Kongyiji. The wine samples ranged in vintage from 2004 to 2005 and were of different wine age (one year, three years and five years). Spectral measurements Samples taken from freshly opened bottles of wine were scanned in transmission mode using a commercial spectro meter, Nexus FT-NIR (Thermo Nicolet Corporation, Madison, WI, USA), which was equipped with an interferometer, an InGaAs detector and a wide band light source (Quartz Tungsten Halogen, 50 W). Samples were scanned in a 1 mm path length rectangular quartz cuvette with air as the reference at room temperature. A reference scan was taken once in every three sample scans. NIR spectra were collected using Omnic software (Thermo Nicolet Corporation, Madison, WI, USA) and stored as absorbance. The spectral range was from 800 nm to 2500 nm, the mirror velocity was cm s 1 and the resolution was 16 cm 1 in this work. The spectrum of each sample was the average of 32 successive scans. Reference analyses Reference analyses for alcoholic degree, ph ue, total acid amino acid nitrogen and Brix were in accordance with the Official Methods of Analysis for Chinese rice wine and Official Methods of Analysis for Shaoxing rice wine [GB/ T and GB (GB means national standard, and were the codes of the two official methods and 2000 was the year in which the methods were revised)]. 10,11 All analyses were done in duplicate. For alcoholic degree measurements, a furnace, a condenser, an alcoholometer (20 C, 0.2 C) and a thermometer (50 C, 0.1 C) were used. The difference between two measurements was less than 0.2% (v v 1 ). For ph measurement, a ph electrode (PHS-3C, Shanghai Scientific Instrument Co., Ltd, Shanghai, China) was used. The difference between the two measurements was less than 0.05 ph. For total acid and amino acid nitrogen automatic measurement, an analyti balance, an electrictitration instrument (accuracy of ± 0.02 ph) (ZD-2, Shanghai Scientific Instrument Co., Ltd, Shanghai, China) and a magnetic stirrer (JB-1A, Shanghai Scientific Instrument Co., Ltd., Shanghai, China) were used. The titer difference between the two measurements for total acid and amino acid nitrogen was less than 0.05 ml and 0.10 ml, respectively. The Brix (%) parameter shows the concentration percentage of the soluble solids content in a sample (water solution). Soluble solids content is the total of all the solids dissolved in the water, including sugar, protein, acids, etc. and the measurement reading is the total sum of these. Basily, Brix (%) is ibrated to the number of grams of sucrose contained in 100 g of sucrose solution. 2 Hence, when a Chinese rice wine sample was measured, Brix (%) represented the sugar equients in the wine sample. In this work, Brix measurement was performed using a digital refractometer (PR-101, Atago Co., Ltd, Tokyo, Japan) with measurement range of Brix %, accuracy of ± 0.2% and an automatic temperature compensation system from 5 to 40 C. Chemometrics and data analysis Chemometrics analysis was performed using the commercial software package, TQ Analyst v6.2.1 (Thermo Nicolet Corporation, Madison, WI, USA). The chemometrics procedure consisted of the following steps. Sample outliers analysis Principal component analysis (PCA) was performed in order to reduce the number of variables showing co-linearity. Thus, the samples were in a new reduction k-dimensional space (k < n). From the k factor score, Mahalanobis distance, which indicated how different a sample spectrum was from the average spectrum of the sample set, was culated. It was defined by the following equation: 1 T i i k i MD ( t t ) S ( t t ) where MD i is the Mahalanobis distance, t i is the score vector of ith sample, t _, is the mean score vector of the sample set, S k is the scores co-variance matrix. Then, the Chauvennet test was applied to the sample with the highest Mahalanobis distance ue to see if it was statistily different from the next highest sample. If the sample failed the test, it was considered as an outlier. Besides the Chauvennet test, leverage diagnostic was applied to the outlier analysis of the same sample set. The leverage diagnostic provided the following information about the samples in a PLS regression method: (1) how much influence each sample had on the method model and (2) how accurately the ibration model described each sample. The information can help to identify samples that might be outliers. In TQ Analyst, the leverage diagnostic showed

3 H. Yu et al., J. Near Infrared Spectrosc. 14, (2006) 39 the relationship between the leverage and the studentised residual ues for each reference parameter. The leverage ue was defined by the following equation: T T 1 h = 1/ n+ t T T ) t i i where h i is the leverage ue, n is the number of the samples and T k is the trimmed scores matrix. The studentised residual ue took leverage ue into account, thus giving a fairer picture of differences among residuals. It was derived from the root mean squared residual for the ibration set: 12 RMSE = 1/( n k) f i k k The studentised residual, r i was then: r = f / RMSE( 1 h ) / i 2 12 / 12 i i i where r i is the Studentised residual ue, RMSE is the root mean squared residual for the ibration set and f i is the residual ue for the ith sample. If a sample had a leverage ue that was noticeably different from the leverage ues for the other samples, it was examined closely in order to know whether it provided any useful information or it must be removed. Selection of the ibration and idation sets Once the outliers without useful information had been removed, the ibration and idation sets were defined. The percentage of samples were 75% and 25% for the ibration and idation sets, respectively. The reference analysis data of the whole sample set for each enologi parameter were sorted into ascending order, then one idation sample was chosen in every four samples. Spectra pre-processing The raw and second derivative spectra were analysed for ibration development. The second derivative was used to reduce baseline variation and enhance the spectral features. 13 Calibration step: cross-idation Calibration models were developed using PLS regression and cross-idation. Cross-idation estimated the prediction error by splitting all samples into two groups. One group was reserved for idation (one or two samples in this case) and the others were used for ibration. The process was repeated until all the standards in the ibration set had been quantified as idation standards. 14 The optimum number of factors used in PLS regression was determined by the lowest ue of the predicted residual error sum of squares (PRESS). Calibration statistics included the standard error of ibration (SEC), the determination coefficient in ibration (R 2 ) and the standard error of cross-idation (SECV). To euate how well the ibration model could predict compositional data, RPD was used in our work. The RPD was defined as the standard deviation (SD) of the population s reference ues divided by SECV for the NIR spectroscopy ibrations. If the error for estimating a constituent (SECV) was large compared to the spread of that compound in all samples (SD), a relatively small RPD was culated, thereby demonstrating that the NIR ibration model was not robust. In contrast, a relatively high RPD ue indicated that the model was able to reliably predict the chemi composition. Generally, if the RPD ue was higher than three, the model could be considered satisfactory for prediction purposes. 4,15 17 Validation step In order to idate the equations, the models were applied to spectra from the idation set, thus statistic parameters as standard error prediction (SEP) and r 2 were obtained. Results and discussion Reference analysis data Table 1 shows information about reference data for the five enologi parameters. Thus, the number, range, mean and SD of the wine samples are summarised in Table 1. Reference data for the five enologi parameters in Chinese rice wine samples. Parameter N Range Mean SD Alcoholic degree (%, v v 1 ) ph (ph units) Total acid Amino acid nitrogen Brix (%)

4 40 Quality Determination of Chinese Rice Wine by FT-NIR the table. The ranges for ph, total acid and amino acid nitrogen ( , g L 1 and g L 1, respectively), were narrower than those for alcoholic degree and Brix ( %, v v 1 and %, respectively). According to Official Methods of Analysis for Chinese rice wine and Official Methods of Analysis for Shaoxing rice wine (GB/T and GB ) the ranges of ph and total acid for semi-dry wine are limited to and g L 1, respectively and the content of amino acid nitrogen must be higher than 0.6 g L 1 for the first-rate Chinese rice wine. All the samples in this work met these requirements. Spectral analysis Figure 1 shows the NIR spectra of the Chinese rice wine samples. The main features of the spectra are absorption bands at 1455 nm and nm which are related to the first overtone of the O H stretch of H 2 O and a combination of stretch and deformation of the O H group in H 2 O and ethanol, respectively. 18,19 From the spectrum of the pure water (Figure 2), it can be seen that the strongest absorption of water is also at nm and 1455 nm. Spectra of natural products tend to be dominated by any water that is present in the samples. For this reason, quantitative analysis often relies on minor changes in the spectra. 20 When the spectrum of absolute ethanol (99.7% v v 1 ) is presented in Figure 3, specific bands between 1695 and 1735 nm, 2270 and 2305 nm can be observed. However, as the concentration of the ethanol decreases, the absorption intensity of bands at 1930 nm and 1455 nm increases, while that of bands between 1695 and 1735 nm, 2270 and 2305 nm decreases. The spectrum of aqueous solutions containing 10% v v 1 ethanol is very close to the water spectrum except that it shows two small characteristic absorption bands between 2270 and 2305 nm. Then, the absorption band of the wine sample (Figure 1) around 2300 nm is likely associated with the CH 2 group of ethanol. Absorption bands at 1450, 1790 and 2266 nm were also reported to be associated with sucrose, fructose and glucose in fruit juices. 21,22 The spectra of the aqueous solution of each individual component, as well as the mixture of sucrose, fructose and glucose, were collected in our work. As was expected, the similarity between the spectra was confirmed. Sucrose, fructose and glucose seem to be quite difficult to differentiate at any wavelength and e n c a rb A bso Figure 1. NIR spectra of Chinese rice wine samples Figure 3. NIR spectra of aqueous solutions of ethanol (10%, 20%, 25%, 50%, 75%, 99.7% v v 1 ) n ce a rb A bso Figure 2. NIR spectra of pure water. 0.0 (nm) Figure 4. NIR spectra of the aqueous solution of each individual component as well as the mixture of sucrose (0.5 g L 1 ), fructose (0.5 g L 1 ) and glucose (10 g L 1 ).

5 H. Yu et al., J. Near Infrared Spectrosc. 14, (2006) 41 absorption bands at 1450 and 1790 nm can be observed in Figure 4. Transforming the spectra by the second derivative inverted the spectra so that peaks became narrow leys that corresponded to each major absorption band in the raw spectra. 13 The highest variations in the second derivative of the spectra are also around 1902, 1958, 2266, 2305 and 2348 nm, which can be seen in Figure 5. Sample outliers The histogram in Figure 6 shows the Mahalanobis distance ues of the samples ranked from smallest (closely clustered) to largest (most unlike the others) for the alcoholic degree analysis. A broken line is used in the histogram to separate samples (no sample identifier) that pass the Chauvennet test (left side of the separator) from samples that fail (right side of the separator). For alcoholic degree analysis, four samples from 88 are in the right part, which means the spectra of the four samples are most unlike the spectra of the other standards. The possible reason for two of the four outliers was that they were of a different variety to the other samples. The two bottles of wine were dry wine and semi-sweet wine, respectively, while the other samples were semi-dry wine. These two samples were also considered as outliers for the other four reference parameters. Figure 7 shows the leverage diagnostic result for alcoholic degree analysis. It can be seen that there are six samples Arbitrary units u n its y rbitra r A Figure 5. Second derivative NIR spectra of Chinese rice wine samples. 0.0 Distance Rank 88 Figure 6. Ranking of the Chinese rice wine samples for spectrum outlier analysis by Mahalanobis distance. Samples to the right of the dashed line failed the Chauvennet test. Studentized Residual Leverage Figure 7. Leverage diagnostic results for alcoholic degree analysis. with leverage ues that are noticeably different from the leverage ues for the other samples. After re-examination, the six samples were removed as outliers. Between the Chauvennet test and the leverage diagnostic result, there were two common samples. Accordingly, eight samples were removed from the sample set for alcoholic degree analysis. For the outlier analysis of each enologi parameter, different samples were removed as outliers. However, among the sample outliers for each parameter, four were the same samples. After the outlier remo, there were 80, 81, 82, 80 and 82 samples for alcoholic degree, ph, total acid, amino acid nitrogen and Brix analysis, respectively. Calibration and idation results Infl uence of spectral pre-processing The PLS regression models developed on the raw spectra showed the better ibration statistics compared with the ones on the second derivative spectra (data not presented). Thus, raw spectra were used for ibration and idation analysis. Calibration results The use of narrow spectral ranges for ibration models led to poor results, thus the full spectral range was applied to the ibration analysis of the five enologi parameters. Calibrations for the five enologi parameters were developed using PLS regression and cross-idation. The ues of R 2 and SECV indicated the precision achieved in ibration. The choice of model was based on the RPD ue. Table 2 shows the results obtained in ibration models. The best results were achieved for the determination of alcoholic degree and Brix. The RPD ue of the ibration for alcoholic degree was higher than three (3.95), which demonstrates the robustness and power of the ibration model. And the RPD ue for Brix was The R 2 ues for the correlation between the reference and NIR meth- ods were 0.93 and 0.96 for the determination of alcoholic degree and Brix, respectively. The performance of the two parameters related to acidity (ph and total acid) was not as good as that of alcoholic degree and Brix. The RPD ues of the ibrations for ph and total acid were 1.41 and 1.38,

6 42 Quality Determination of Chinese Rice Wine by FT-NIR Table 2. Descriptive statistics for the ibration set. Parameter N PLS factors Mean Min. Max. SECV r 2 RPD Alcoholic degree (%, v v 1 ) ph (ph units) Total acid Amino acid nitrogen Brix (%) Table 3. Descriptive statistics for the idation set. Parameter N Mean Min. Max. r 2 SEP Alcoholic degree (%, v v -1 ) ph (ph units) Total acid Amino acid nitrogen Brix (%) respectively, which indicates that the ibration models for the two parameters are not very robust. The R 2 ues for ph and total acid were 0.97 and 0.83, respectively. The performance of amino acid nitrogen was the worst among the five enologi parameters due to the low concentration range of the reference data (0.90, 0.06 g L 1 and 1.27 for R 2, SECV and RPD ue, respectively). Validation equations The ibration models were tested with the idation set. Table 3 shows the number of idation samples, mean, minimum, maximum, r 2 and SEP. The ues of SEP and r 2 were used to euate the analyti quality of the models. From Table 3, it can be seen that the best results were achieved for alcoholic degree and Brix, with r 2 ue higher than 0.9. The performance of ph and total acid were acceptable (r 2 ues for ph and total acid of 0.82 and 0.77, respectively), while the performance of amino acid nitrogen was the worst of the five parameters, with r 2 ue Figures 8 11 show the correlations between the ues determined by the reference analysis and the ues predicted by the NIR spectroscopy technique on the whole sample set of Chinese rice wine after outlier remo. In summary, Figures 8 11 demonstrate that the NIR spectroscopy technique could be used to predict the Predicted ue/%, v v ibration idation Reference ue/%, v v -1 Figure 8. Comparison of the alcohol degree determined by reference analysis with those predicted by NIR.

7 H. Yu et al., J. Near Infrared Spectrosc. 14, (2006) ibration idation ibration idation Predicted ue/ph units Predicted ue/g L Reference ue/ph units Figure 9. Comparison of ph ues determined by reference analysis with those predicted by NIR Reference ue/g L -1 Figure 10. Comparison of total acid concentrations determined by reference analysis with those predicted by NIR. Predicted ue/g L ibration idation Predicted ue/ph units ibration idation Reference ue/g L -1 Figure 11. Comparison of amino acid nitrogen concentrations determined by reference analysis with those predicted by NIR Reference ue/ph units Figure 12. Comparison of Brix determined by reference analysis with those predicted by NIR concentration of the five enologi parameters in Chinese rice wine. Comparison with the results obtained by other authors The determination of alcoholic degree, ph and total acid yielded R 2 ues which were close to the ues in the literature. The RPD ues of ph and total acid were lower than those in the literature, because the range of the two parameters in this work were narrower and the concentrations were lower than those in the literature. 9 The determination of amino acid nitrogen and Brix in wine has not been reported previously. five enologi parameters using the NIR spectroscopic technique together with the spectroscopic technique. The models for alcoholic degree (R 2 =0.93, r2 =0.91, RPD=3.95), Brix (R 2 =0.96, r2 =0.92, RPD=2.34) showed good prediction performance. The performance of ph and total acid was acceptable (R 2 for ph and total acid was 0.97 and 0.83, respectively, r 2 was 0.82 and 0.77, respectively and RPD ues were 1.41 and 1.38, respectively.). The prediction performance of amino acid nitrogen was the worst of the five enologi parameters. Further studies are needed in order to improve accuracy and robustness of the ibration and to extend to other Chinese rice wine varieties. Conclusion The applicability of NIR spectroscopy in determining alcoholic degree, ph, total acid, amino acid nitrogen and Brix in Chinese rice wine has been studied in this work. The results obtained showed that it was possible to determine the Acknowledgements The authors gratefully acknowledge the financial support provided by the Program for New Century Excellent Talents in University, No. NCET and the National Natural Science Foundation of China, No

8 44 Quality Determination of Chinese Rice Wine by FT-NIR References 1. A. Urtubia, J.-R. Pérez-Correa, M. Meurens and E. Agosin, Talanta 64, 778 (2004). 2. M.-J. Sáiz-Abajo, J.-M. González-Sáiz and C. Pizarro, J. Agric. Food Chem. 52, 7711 (2004). 3. B.G. Osborne, T. Fearn and P.H. Hindle, Practi NIR Spectroscopy with Applications in Food and Beverage Analysis. Longman Scientific and Techni, Harlow, UK (1993). 4. D. Cozzolino, M.J. Kwiatkowski, M. Parker, W.U. Cynkar, R.G. Dambergs, M. Gishen and M.J. Herderich, Anal. Chim. Acta 513, 73 (2004). 5. A.G. Cavinato, D.M. Mayes, Z. Ge and J.B. Callis, Anal. Chem. 62, 1977 (1990). 6. R.G. Dambergs, A. Kambouris, I.L. Francis and M. Gishen, J. Agric. Food Chem. 50, 3079 (2002). 7. M. Gishen and R.G. Dambergs, Aust. N Z Grapegrower Winemaker 414(a), 43 (1998). 8. L.S. Mendes, F.C.C. Oliveira, P.A.Z. Suarez and J.C. Rubim, Anal. Chim. Acta 493, 219 (2003). 9. M. Urbano-Cuadrado, M.D. Luque de Castro, P.M. Pérez-Juan, J. García-Olmo and M.A. Gómez-Nieto, Anal. Chim. Acta 527, 81 (2004). 10. National Standards of People s Republic of China. GB/ T : Chinese rice wine. (2000). 11. National Standards of People s Republic of China. GB : Shaoxing rice wine. (2000). 12. R.D. Cook and S. Weisberg. Residuals and Infl uence in Regression. Chapman and Hall, New York, USA (1982). 13. D. Cozzolino, H.E. Smyth and M. Gishen, J. Agric. Food Chem. 51, 7703 (2003). 14. Y. Liu and Y. Ying, J. Zhejiang Univ. Sci. 5, 651 (2004). 15. P.C. Williams, Cereal Chem. 52, 561 (1975). 16. L. Sauvage, D. Frank, J. Stearne and M.B. Millikan, Anal. Chim. Acta 458, 223 (2002). 17. T. Fearn, NIR news 13, 13 (2003). 18. H. Maeda, Y. Wang, Y. Ozaki, M. Suzuki, M.A. Czarnecki and M. Iwahashi, Chemometr. Intell. Lab. Syst. 45, 121 (1999). 19. H. Büning-Pfaue, Food Chem. 82, 107 (2003). 20. C.M. Garcia-Jares and B. Médina. Fresenius J. Anal. Chem. 357, 86 (1997). 21. F.J. Rambla, S. Garrigues and M. de la Guardia, Anal. Chim. Acta 344, 41 (1997). 22. Y. Roggo, L. Duponchel and J.-P. Huvenne, J. Agric. Food Chem. 52, 1055 (2004). Received:14 June 2005 Revised: 22 November 2005 Accepted: 5 January 2006 Web Publication: 4 April 2006