The University of Adelaide

Geographic classification of wines using Vis-NIR spectroscopy

Master Thesis

by

Liang Liu

B. Eng. (Shenyang Pharmaceutical Universiy, China)

School of Chemical Engineering

Faculty of Engineering, Computer & Mathematical Sciences

November 2006

I

Declaration

This work contains no material which has been accepted for the award of any other

degree or diploma in any university or other tertiary institution and, to best of my

knowledge and belief, contains no material previously published or written by another

person, except where due reference has been made in the text.

I give consent to this copy of my thesis being made available in the University Library.

The author acknowledges that copyright of published works contained within this

thesis (as listed below) resides with the copyright holders of those works.

Liang Liu

November 2006

II

Summary

The determination of wine authenticity and the detection of adulteration are attracting

an increasing amount of attention for wine producers, researchers and consumers.

Wine authentication and classification based on geographical origin has been widely

studied. Most of these studies have achieved successful classification results.

However, these studies have involved complicated and expensive procedures. Visible

and near infrared spectroscopy (Vis-NIR) is recognized as a rapid and non-destructive

technique. In recent years, several studies have been conducted using Vis-NIR

spectroscopy to analyze wine for both quantitative and qualitative purposes. The aim

of this research was to investigate the geographical classification of wines using Vis-

NIR spectroscopy. The effect of temperature and measurement mode (transmission

and transflectance) on Vis-NIR spectra was investigated to identify optimal conditions

for wine sample analysis. It was found the optimal temperature is between 30 to 35 oC

and the shorter pathlength measurement condition has better prediction ability.

Classification by geographical origin using Vis-NIR spectroscopy was investigated for

sixty-three Tempranillo wines from Spain and Australia, and fifty Riesling wines from

Australia, New Zealand and Europe. Discriminant partial least square regression

(DPLS) and linear discriminant analysis (LDA) based on PCA scores were used to

perform classification. Over 90% of the Tempranillo wines were correctly classified

according to their geographical region using both DPLS and LDA. A classification

rate of 72% was achieved for the Riesling wines. Vis-NIR technique provides a

similar degree of reliability on wine classification comparable to those obtained using

chemical composition. The results of this study demonstrate potential for Vis-NIR

spectroscopy combined with multivariate analysis as a rapid method for classifying

wines by geographical origin.

III

Acknowledgements

I wish to express my sincere gratitude to my research supervisors, Dr. Chris Colby

and Dr. Daniel Cozzolino. I appreciate their expert guidance and never ending

patience. I also thank my co-supervisors, A/Prof. Brian O’Neill, Prof. Derek Abbott

and A/Prof. Graham Jones for their helpful guidance and encouragement.

The major part of the work reported in this thesis was performed at the Australia Wine

Research Institute, Adelaide, SA. I would like to thank the following staff from AWRI,

Dr. Wies Cynkar, Mr. Mark Gishen, Dr. Les Janik, Dr. Robert Dambergs, Mr.

Geoffrey Cowey, Mr Mathew Holdstock, Dr. Paul Smith, Ms. Megan Mecurio, and

their colleagues from the AWRI analytical lab. Their contributions made this work

possible.

Finally, I would like to thank my parents and my girlfriend for their continuing love

and support throughout my academic career.

IV

List of Publications

Liang Liu, Daniel Cozzolino, Wies Cynkar, Mark Gishen, Christopher Colby.

Geographic classification of Spanish and Australian Tempranillo red wines by visible

and near infrared spectroscopy combined with multivariate analysis. Journal of

Agriculture and Food Chemistry. (Published on web 12/08/2006)

Liang Liu, Daniel Cozzolino, Chris Colby, Bob Dambergs, Mark Gishen, Brian

O’Neill, Derek Abbott, (2006) Effect of temperature on visible and near infrared

spectra of wine. The 12th Australian Near Infrared Spectroscopy Conference,

Rockhampton, Queensland, Australia.

V

Table and Contents

SUMMARY............................................................................................................................................II

ACKNOWLEDGEMENTS ................................................................................................................ III

LIST OF PUBLICATIONS................................................................................................................. IV

LIST OF FIGURES ............................................................................................................................VII

LIST OF TABLES ............................................................................................................................... IX

CHAPTER 1 INTRODUCTION...........................................................................................................1

CHAPTER 2 LITERATURE REVIEW ...............................................................................................4

2.1 WINE QUALITY CATEGORY – GEOGRAPHICAL ORIGIN ......................................................................4 2.2 WINE CLASSIFICATION AND AUTHENTICATION.................................................................................5

2.2.1 Sensory evaluation .................................................................................................................5 2.2.2 Instrumental analysis .............................................................................................................6 2.2.3 Spectroscopic methods ...........................................................................................................9 2.2.4 Summary.................................................................................................................................9

2.3 NEAR INFRARED SPECTROSCOPY...................................................................................................10 2.3.1 Introduction..........................................................................................................................10 2.3.2. Effect of Sample presentation on Vis and NIR spectra ........................................................10 2.3.3. Use of NIR to classify food based on geographical origin..................................................12 2.3.4 NIR applications on wine analysis .......................................................................................13

SUMMARY AND RESEARCH GAPS.........................................................................................................15

CHAPTER 3 MATERIAL AND METHODS.....................................................................................16

3.1 WINE SAMPLES .............................................................................................................................16 3.2 WINE REFERENCE ANALYSIS..........................................................................................................17 3.3 SPECTROSCOPIC MEASUREMENTS .................................................................................................17 3.4 SPECTRA DATA ANALYSIS ..............................................................................................................19

3.4.1 Spectra pre-treatment ...........................................................................................................19 3.4.2 Multivariate analysis............................................................................................................20

CHAPTER 4 EFFECT OF SAMPLE PRESENTATION - SAMPLE TEMPERATURE EFFECT

ON THE ANALYSIS OF WINE..........................................................................................................24

4.1 INTRODUCTION .............................................................................................................................24 4.2 RESULTS AND DISCUSSION.............................................................................................................24

4.2.1 Chemical analysis ................................................................................................................24 4.2.2 Spectra interpretation and analysis......................................................................................26

VI

4.2.3 Influence of temperature on the Vis-NIR spectra of wine .....................................................29 Summary .......................................................................................................................................36 4.2.4 Principal component analysis ..............................................................................................36 4.2.5 Comparison of the prediction ability of different temperatures using PLS ..........................40

CHAPTER 5 EFFECT OF SAMPLE PRESENTATION – MEASUREMENT CONDITION

EFFECT ON THE ANALYSIS OF WINE .........................................................................................42

5.1 INTRODUCTION .............................................................................................................................42 5.2 RESULTS AND DISCUSSION.............................................................................................................42

5.2.1 Chemical analysis ................................................................................................................42 5.2.2 Spectra analysis ...................................................................................................................43 5.2.3 Principal component analysis ..............................................................................................45 5.2.4 Comparison using PLS.........................................................................................................47

CHAPTER 6 USE OF VISIBLE AND NIR TO CLASSIFY TEMPRANILLO WINES BASED

ON GEOGRAPHICAL ORIGINS. .....................................................................................................49

6.1 INTRODUCTION .............................................................................................................................49 6.2 RESULTS AND DISCUSSION.............................................................................................................50

6.2.1 Chemical analysis ................................................................................................................50 6.2.2 Spectra interpretation and analysis......................................................................................51 6.2.3 Principal component analysis ..............................................................................................52 6.2.4 Discrimination analysis........................................................................................................54

SUMMARY...........................................................................................................................................58

CHAPTER 7 USE OF VISIBLE AND NIR TO CLASSIFY RIESLING WINES BASED ON

GEOGRAPHICAL ORIGINS.............................................................................................................59

7.1 INTRODUCTION .............................................................................................................................59 7.2 RESULTS AND DISCUSSIONS...........................................................................................................59

7.2.1 Chemical analysis ................................................................................................................59 7.2.2 Spectra interpretation and analysis......................................................................................60 7.2.3 Principal component analysis ..............................................................................................62 7.2.4 Discrimination analysis........................................................................................................65

SUMMARY...........................................................................................................................................67

CONCLUSION .....................................................................................................................................69

REFERENCES .....................................................................................................................................71

VII

List of Figures

FIGURE 3. 1 FOSS NIRSYSTEM6500, SILVER SPRING, MD .....................................................................18 FIGURE 3. 2 SAMPLE PRESENTATION OF TRANSMITTANCE AND TRANSFLECTANCE ...................................19

FIGURE 4. 1 VIS-NIR SPECTRA OF RED WINE SAMPLES AT 5 DIFFERENT TEMPERATURES ..........................27 FIGURE 4. 2 VIS-NIR SPECTRA OF WHITE WINE SAMPLES AT 5 DIFFERENT TEMPERATURES ......................27 FIGURE 4. 3 SECOND DERIVATIVE SPECTRA OF RED AND WHITE WINE SAMPLES .......................................28 FIGURE 4. 4 WATER SPECTRA AT 6 DIFFERENT TEMPERATURES.................................................................28 FIGURE 4. 5 VIS-NIR SPECTRA OF ONE RED WINE SAMPLE (WW1) AT 6 DIFFERENT TEMPERATURES ........29 FIGURE 4. 6 SPECTRA AT 1450NM REGION ...............................................................................................30 FIGURE 4. 7 SPECTRA AT 2270NM TO 2300NM REGION.............................................................................30 FIGURE 4. 8 SECOND DERIVATIVE SPECTRA AT 1450NM REGION...............................................................31 FIGURE 4. 9 SECOND DERIVATIVE SPECTRA AT 2270NM TO 2300NM REGION ............................................31 FIGURE 4. 10 LINEAR RELATIONSHIP OF THE ABSORBANCE OF SECOND DERIVATIVE SPECTRA AT 962NM OF

RED WINES .....................................................................................................................................32 FIGURE 4. 11 LINEAR RELATIONSHIP OF THE ABSORBANCE OF SECOND DERIVATIVE SPECTRA AT 962NM OF

WHITE WINES .................................................................................................................................33 FIGURE 4. 12 LINEAR RELATIONSHIP OF THE ABSORBANCE OF SECOND DERIVATIVE SPECTRA AT 1412NM

OF RED WINES AVERAGE SPECTRA ..................................................................................................34 FIGURE 4. 13 LINEAR RELATIONSHIP OF THE ABSORBANCE OF SECOND DERIVATIVE SPECTRA AT 1462NM

OF RED WINES AVERAGE SPECTRA ..................................................................................................34 FIGURE 4. 14 LINEAR RELATIONSHIP OF THE ABSORBANCE OF RAW SPECTRA OF RED WINES AT 2268 NM.35 FIGURE 4. 15 LINEAR RELATIONSHIP OF THE ABSORBANCE OF RAW SPECTRA OF RED WINES AT 2306 NM.36 FIGURE 4. 16 SCORE PLOT OF PC1 AND PC2 OF THE WHITE WINE SAMPLES. NUMBERS REPRESENT THE

TEMPERATURES ..............................................................................................................................37 FIGURE 4. 17 EIGENVECTORS OF THE FIRST TWO PCS OF THE PCA FOR WHITE WINES .............................37 FIGURE 4.18 SCORE PLOT OF PC1 AND PC2 OF THE RED WINE SAMPLES. NUMBERS REPRESENT THE

TEMPERATURES ..............................................................................................................................38 FIGURE 4. 19 EIGENVECTORS OF THE FIRST TWO PCS OF THE PCA FOR THE RED WINES..........................38 FIGURE 4. 20 LINEAR RELATIONSHIP BETWEEN THE MEAN SCORE VALUES OF TEMPERATURE RELATED PC

OF WHITE RED SAMPLES AND TEMPERATURE VARIATION.................................................................39 FIGURE 4. 21 LINEAR RELATIONSHIP BETWEEN THE MEAN SCORE VALUES OF TEMPERATURE RELATED PC

OF WHITE WINE SAMPLES AND TEMPERATURE VARIATION...............................................................39

FIGURE 5. 1 VIS-NIR SPECTRA OF 6 RED AND 6 WHITE WINE SAMPLES AT 0.2 MM PATH LENGTH

TRANSFLECTANCE MODE................................................................................................................44 FIGURE 5. 2 VIS -NIR SPECTRA OF 6 RED AND 6 WHITE WINE SAMPLES AT 0.4 MM TRANSFLECTANCE MODE

......................................................................................................................................................44 FIGURE 5. 3 VIS-NIR SPECTRA OF THE SAME SAMPLE AT THREE DIFFERENT PATH LENGTHS.....................45

VIII

FIGURE 5. 4 PCA SCORE PLOT OF THE PC1 AGAINST PC2 ........................................................................46 FIGURE 5. 5 PCA SCORE PLOT OF THE PC1 AGAINST PC3 ........................................................................47

FIGURE 6. 1 SECOND DERIVATIVE OF THE VIS-NIR SPECTRA OF AUSTRALIAN AND SPANISH TEMPRANILLO

WINES ............................................................................................................................................52 FIGURE 6. 2 SCORE PLOT OF THE FIRST TWO PRINCIPAL COMPONENTS OF AUSTRALIAN (A) AND SPANISH

(S) TEMPRANILLO WINES USING VIS-NIR AFTER SNV AND SECOND DERIVATIVE PROCESSING ......53 FIGURE 6. 3 EIGENVECTORS OF THE THREE FIRST PRINCIPAL COMPONENTS OF AUSTRALIAN AND SPANISH

TEMPRANILLO WINES USING VIS-NIR AFTER SNV AND SECOND DERIVATIVE PROCESSING ............54 FIGURE 6. 4 PARTIAL LEAST SQUARES SCORE PLOT OF THE FIRST TWO PRINCIPAL COMPONENTS OF

AUSTRALIAN (A) AND SPANISH (S) TEMPRANILLO WINES USING VIS-NIR PRE-PROCESSED SPECTRA

FOR THE CALIBRATION SET .............................................................................................................57

FIGURE 7. 1 VIS -NIR RAW SPECTRA OF RIESLING WINES FROM AUSTRALIA, NEW ZEALAND AND EUROPE

......................................................................................................................................................61 FIGURE 7. 2 SNV AND 2ND DERIVATIVE PROCESSED SPECTRA OF RIESLING WINES FROM AUSTRALIA, NEW

ZEALAND AND EUROPE..................................................................................................................61 FIGURE 7. 3 SCORE PLOT OF THE FIRST 3 PRINCIPAL COMPONENTS OF AUSTRALIAN (AUS), NEW ZEALAND

(NZ) AND EUROPEAN (EUR) RIESLING WINES USING VIS-NIR PRE-PROCESSED SPECTRA...............62 FIGURE 7. 4 EIGENVECTORS OF THE THREE FIRST PRINCIPAL COMPONENTS OF AUSTRALIAN, NEW

ZEALAND AND EUROPEAN RIESLING WINES USING VIS-NIR PRE-PROCESSED SPECTRA..................63 FIGURE 7. 5 SCORE PLOT OF THE FIRST 3 PRINCIPAL COMPONENTS OF AUSTRALIAN (AUS), AND

EUROPEAN (EUR) RIESLING WINES USING VIS-NIR PRE-PROCESSED SPECTRA ...............................64 FIGURE 7. 6 SCORE PLOT OF THE FIRST 3 PRINCIPAL COMPONENTS OF AUSTRALIAN (AUS) AND NEW

ZEALAND (NZ) RIESLING WINES USING VIS-NIR PRE-PROCESSED SPECTRA...................................64 FIGURE 7. 7 SCORE PLOT OF THE FIRST 3 PRINCIPAL COMPONENTS OF NEW ZEALAND (NZ) AND

EUROPEAN (EUR) RIESLING WINES USING VIS-NIR PRE-PROCESSED SPECTRA ...............................65

IX

List of Tables

TABLE 2.1 APPELLATION SYSTEMS OF THE MOST EUROPE WINE PRODUCING COUNTRIES (KOLPAN ET AL.

1996; MACNEIL 2001) .....................................................................................................................5 TABLE 2. 2 GEOGRAPHICAL CLASSIFICATION OF WINES USING MULTI-ELEMENT AND STABLE ISOTOPE

RATIO. ..............................................................................................................................................7

TABLE 4. 1 WHITE WINE SAMPLES’ CODE, VARIETY, CHEMICAL COMPOSITIONS AND THE CORRESPONDING

STATISTICS OF SAMPLE ANALYZED..................................................................................................25 TABLE 4. 2 RED WINE SAMPLES’ CODE, VARIETY, CHEMICAL COMPOSITIONS AND THE CORRESPONDING

STATISTICS OF SAMPLE ANALYZED..................................................................................................26 TABLE 4. 3 STANDARD ERROR IN CROSS VALIDATION (SECV) OF PLS PREDICTION FOR CHEMICAL

ANALYSIS PARAMTERS....................................................................................................................41

TABLE 5. 1 SAMPLES’ CODE, CHEMICAL COMPOSITIONS AND THE CORRESPONDING STATISTICS OF SAMPLE

ANALYSED......................................................................................................................................43 TABLE 5. 2 THE STANDARD ERROR IN CROSS VALIDATION (SECV) OF THE PREDICTION MODELS FOR

EACH PARAMETER AT DIFFERENT PATH LENGTH..............................................................................48

TABLE 6. 1 VINTAGE AND ORIGIN OF COMMERCIAL TEMPRANILLO WINE SAMPLES ANALYSED ................49 TABLE 6. 2 RANGE OF CHEMICAL COMPOSITION FOR THE AUSTRALIAN AND SPANISH WINE ANALYSED...51 TABLE 6. 3 LDA CLASSIFICATION RESULTS OF AUSTRALIAN AND SPANISH TEMPRANILLO WINES USING

VIS-NIR RAW SPECTRA BASED ON THE FIRST 3 PCS (98% OF THE TOTAL VARIATION) ....................55 TABLE 6. 4 LDA CLASSIFICATION RESULTS OF AUSTRALIAN AND SPANISH TEMPRANILLO WINES USING

VIS-NIR PRE-PROCESSED SPECTRA BASED ON THE FIRST 3 PCS (77% OF THE TOTAL VARIATION) ..55 TABLE 6. 5 LDA CLASSIFICATION RESULTS OF AUSTRALIAN AND SPANISH TEMPRANILLO WINES USING

VIS-NIR PRE-PROCESSED SPECTRA BASED ON THE FIRST 9 PCS (95% OF THE TOTAL VARIATION) ..56 TABLE 6. 6 DISCRIMINANT PARTIAL LEAST SQUARES (DPLS) CLASSIFICATION RESULTS OF AUSTRALIAN

AND SPANISH TEMPRANILLO WINES USING VIS-NIR PRE-PROCESSED SPECTRA..............................58

TABLE 7. 1 VINTAGE AND ORIGIN OF RIESLING WINE SAMPLES ANALYZED ..............................................59 TABLE 7. 2 STATISTICS OF CHEMICAL COMPOSITION FOR RIESLING WINES FROM DIFFERENT

GEOGRAPHICAL REGION .................................................................................................................60 TABLE 7. 3 LDA CLASSIFICATION RESULTS OF AUSTRALIAN, NEW ZEALAND AND EUROPEAN RIESLING

WINES USING VIS-NIR PRE-PROCESSED SPECTRA...........................................................................66 TABLE 7. 4 LDA CLASSIFICATION RESULTS OF EACH TWO REGIONS OF AUSTRALIAN, NEW ZEALAND AND

EUROPEAN RIESLING WINES USING VIS-NIR PRE-PROCESSED SPECTRA .........................................66 TABLE 7. 5 DPLS CLASSIFICATION RESULTS OF AUSTRALIAN, NEW ZEALAND AND EUROPEAN RIESLING

WINES USING VIS-NIR PRE-PROCESSED SPECTRA ...........................................................................67

1

Chapter 1 Introduction

Authenticity is an important food quality criterion. Wine as one of the most important

beverages around the world requires meticulous and continuous control to maintain its

quality. Geographical origin of the wine is often used as an indicator of quality,

especially for fine wines where that higher quality production from particular regions

has long been recognized. Moreover, there is growing enthusiasm among consumers

for high quality food with a clear regional identity (Kelly et al. 2005). To classify and

authenticate wine samples based on their geographical origin is of obvious meaning

for both industry and consumers. Therefore, this problem has attracted great interest

from researchers.

Historically, sensory evaluation is the most direct and widely applied method to assess

and authenticate wine products. However, it is subjective and interferences may easily

lead to incorrect conclusions.

As a consequence, more objective methods, including routine chemical, instrumental

methods, based on the chemical composition of wines have been introduced as on

alternative. In these methods, multivariate analysis techniques (chemometrics) are

widely employed to enhance the classification ability. For example, instrumental

analyses in conjunction with pattern recognition techniques have been able to classify

wines from different geographical and varietals origin (Reid et al. 2006). These

studies employ advanced chromatographic (high performance liquid chromatography

(HPLC), gas chromatography (GC)) and/or spectroscopic (nuclear magnetic

resonance (NMR), mid infrared spectroscopy (MIR)) techniques (Arvanitoyannis et al.

1999; Reid et al. 2006). DNA-based and immunological techniques have also been

applied (Lockley and Bardsley 2000). Most studies have achieved

satisfactory/successful outcomes, with wines from different geographical origins were

correctly classified. However, most techniques require high cost instruments and

complicated procedures, and are not likely to gain wide application in the wine

industry unless the instrument and running costs are lowered.

2

The near infrared (NIR) spectroscopy technique has been recognized as a rapid and

non-destructive technique and has been widely applied in the agriculture and food

fields in recent decades. By recording the overtone and combination vibrations of the

molecular bonds, the resulting NIR spectrum produces a fingerprint of the sample.

However, the characteristics of broad, superimposed and weak absorption bands in the

NIR spectrum has limited its direct use. Consequently, the technique has been

neglected by spectroscopists for a long time. Fortunately, with the information

extraction tool – chemometrics and advances in computing power, NIR spectroscopy

now provides a practical alternative to classical chemistry methods.

A large number of studies have been conducted using NIR spectroscopy to perform

wine analysis for quantitative and for qualitative purposes. Different wine chemical

parameters have been quantified and different wine varieties have been classified

based on NIR spectra (Dambergs et al. 2002; Cozzolino et al. 2005; Urbano Cuadrado

et al. 2005). However, no workers have attempted to classify wines by NIR

spectroscopy based on their geographical origin. Several successes have been

achieved using NIR to geographically classify food products. With the success of

these previous studies with food products, classification of red (Tempranillo wines)

and white wines (Riesling wines) from different geographical origins using NIR

spectroscopy combined with chemometrics was investigated in this thesis.

To date, basic protocols for wine sample presentation employed in NIR studies were

mainly based on the experience with other food products. No systematic investigation

has been conducted of optimal experiment conditions for taking NIR measurements.

Therefore, in this study, the effects of sample temperature and measurement mode

(transmission and transflectance) were firstly examined before conducting the

classification study.

Therefore, the objectives of the research undertake was to:

• Examine the effect of experimental protocols for wine analysis using visible

and near infrared (Vis-NIR) spectroscopy, including sample temperature effect

and measurement mode (transmission and transflectance).

3

• Analyze and classify red (Tempranillo wine) and white (Riesling wine) wine

samples using Vis-NIR spectroscopy according to the geographical origin.

Chapter 2 of the thesis presents a literature review and identifies the knowledge gaps

that this work is filling. This is followed by the investigation of sample presentation

effects (temperature and scanning modes) for the Vis-NIR spectra for wine samples

(Chapters 4 and 5, respectively). Finally the study of wine classification based on the

geographical origin using Vis-NIR spectroscopy (Chapters 6 and 7) is presented.

4

Chapter 2 Literature review

2.1 Wine quality category – geographical origin

“Great wines do not come from just any where” (MacNeil 2001). Wine is one of the

agricultural products whose aroma and taste are influenced by where it grows (such as

tea, coffee, honey, and olive oil) (Pigott 2004). Geographical origin plays a role in

wine just like a mother does in the birth of her child. It is reflected in the aromas and

flavors of the vineyard from where the wine comes and the specific environment

where the vine grew. The harmonic convergence of every facet of nature produces the

finest wines. “Terroir” a word originally from France, presents the idea that the site

determines the quality of wine, and is now a buzz word all over the world (Kolpan et

al. 1996).

In the 1930’s, France was the first to develop a system of regulations based on the

geographical origin known as the ‘Appellation d’Origine Controlee’ (AOC) (Kolpan

et al. 1996; MacNeil 2001). Soon the system became a model for wine producing

countries around the world. Table 2.1 shows the category systems of several European

wine producing countries. All of these systems were designed to define and protect

wines from specific geographic areas, and also imply an indicator of wine quality

(Arvanitoyannis et al. 1999).

5

Table 2.1 Appellation systems of the most Europe wine producing countries (Kolpan et

al. 1996; MacNeil 2001)

France Appellation d’Origine Controlee(AOC)

Italy Denominazione di Origine Controllata (DOC)

Spain Denomination of origin (DO)

Portugal Denominations of controlled origin (DOC)

Germany Qualitätswein mit Prädikat (QmP)

Greece Appellation of Origin of Superior Quality and

Controlled Appellation of Origin (OPAP)

2.2 Wine classification and authentication

As a consequence, classification and authentication of wine based on their

geographical origin has attracted interest from researchers and the wine industry and

has been subject to extensive research using both sensory analysis and instrumental

methods (Kallithraks et al. 2001).

2.2.1 Sensory evaluation

Sensory evaluation by experienced tasters remains the widely used method to inspect

and authenticate wine by the industry and consumers. However, it does not always

lead to the correct conclusions. Frank and Kowalski (1984) have shown that sensory

data did not provide sufficient information to separate wines from different regions of

France and USA. Sensory descriptive scores also have been applied in conjunction

with pattern recognition analysis. Sivertsen et al. (1999) used 17 sensory attributes to

classify wines in conjunction with multivariate analysis methods from different

French wine regions. Only 63.3% of the samples were correctly classified, compared

with an 81.8% correct rate achieved by using 12 chemical parameters. Although

sensory experts are well trained and have an outstanding ability to “identify” coded

6

wine samples, incorrect conclusions commonly occur because of changes in

vinification, differences between vintage, and even the mental or physical fatigue of

the tasters. Therefore, objective methods based on wine chemical composition are

considered superior.

2.2.2 Instrumental analysis

Wine is a complicated mixture of chemical components, including various organic

and inorganic constituents. The chemical matrix reflects the character of the wine

sample. It has been used to explore wine classification based on geographical origins.

A diverse range of chemical parameters have been measured in wine to classify

samples according to their geographic origins.

2.2.2.1 Trace element and isotope analysis

Several authors have performed trace element and stable isotope analysis to identify

wine geographical origin. Isotope ratios are dependent on climate and variety.

Measurement of strontium (Sr) isotope ratios (87Sr/86Sr) by thermal ionization-mass

spectrometry (TIMS) was one of the earliest trials to discriminate wines growing in

different regions within a given country (e.g. France, Italy) (Horn et al. 1993). The

variation of δ18O value, associated with water in wine can indicate production region

(Breas et al. 1994). Deuterium content of water and ratios of the methyl group of

ethanol analyzed by a comprehensive NMR technique known as Site-specific Isotope

Fractionation measured by Nuclear Magnetic Resonance (SNIF-NMR) were also used

to identify geographical origin of wines (Day et al. 1995).

Metal elements are considered good indicators of wine origin since generally they are

not metabolized or modified during the vinification process (Kelly et al. 2005). The

most frequently quantified elements are Na, Fe, Zn, Rb, Ca, Mg, Mn, Cu, Cr, Co, Sb,

Cs, Br, Al, Ba, As, Li, Ag (Arvanitoyannis et al. 1999; Kelly et al. 2006). The

potential of multi-element analysis for determining wine geographical origin was

demonstrated by McCurdy et al. (1992). The separation of 112 Spanish and English

7

wines according to geographical origin was achieved by analysis of 48 elemental

concentrations using inductively coupled plasma mass spectrometry (ICP-MS)

(Baxter et al. 1997) Canonical discriminant analysis was applied to extract

geographical information from elemental composition. Table 2.2 provides examples

of the studies that have used elements or isotope ratios to authenticate wines based on

their provenance.

Table 2. 2 Geographical classification of wines using multi-element and stable isotope

ratio.

Chemical

content Instrument

Parameters

analysed

Geographical

origins

Multivariate

analysis Reference

Multi-

elements AAS

Ba, Ca, Mg,

Rb, Sr, Ba,

V

Slovakian and

European wines PCA, PCF

Korenovska

& Suhaj

2005

ICP-MS 40 elements Three areas of

South Africa DA

Coetzee et

al. 2005

AAS, AES 11 elements

Three areas of

Canary Islands

(Spain)

PCA, LDA,

SIMCA

Frias et al.

2003

ICP-MS 13 elements Four German

regions QDA

Gomez et

al. 2004

Isotope

analysis

IRMS &

SNIF-NMR

D/H, δC-13,

δO-18

Three regions of

Slovenia PCA, LDA

Orginc et al.

2001

ICP-MS B-11/B-10 South Africa,

France, and Italy

Coetzee et

al. 2005

2.2.2.2 Amino acids

Amino acids have been used to characterize wine by geographical origin. Forty-two

Greek white wines were analyzed for primary amino acids by HPLC and using

discriminant analysis, the amino acid profiles were demonstrated to be useful in

classification of wine provenance (Soufleros et al. 2003).

8

2.2.2.3 Phenolic compounds

Phenolic compounds, one of the most important constituents in wine, have also been

successfully applied to differentiate wines based on geographical origin. The phenolic

composition of red and white wines from four Spanish Appellations of Origin was

investigated using HPLC combined with statistical analysis methods (Pena-Neira et al.

2000). Several phenolic compounds were identified and quantified. The multivariate

analysis result indicated that the different geographical origins strongly influence the

phenolic composition of the final wine.

Thirty nine Galician certified brand of origin (CBO) red wines from Ribeira Sacra and

non-Ribeira Sacra area of Galicia were authenticated based on phenolic composition

(Rebolo et al. 2000). Nineteen major polyphenolic phytochemicals have been

determined by HPLC in forty experimental red wines. Multivariate chemometric

classification procedures were employed. The results indicated good performance in

terms of classification and differentiation of CBO Ribeira Sacra wines from wine

produced in other geographical areas.

2.2.2.4 Routine chemical analyses

Alcohol content, pH, colour parameters (color density, hue), etc. are routinely

analyzed chemical parameters for quality control. Normally, a single one of these

chemical parameters cannot explain the variation between wine provenance; however,

when combined with multivariate analysis techniques, wine patterns can be

recognized. Sivertsen et al. (1999) used a set of chemical parameters, including

alcohols, esters, pH and color, to identify twenty-two wines from four main wine

regions in France. An 81% correct classification rate was achieved. Although the

limitation in wine sample numbers decreased the reliability of this study, the result

demonstrated the potential of using general chemical data for geographical

classification.

9

2.2.3 Spectroscopic methods

In recent years, spectroscopic methods have been applied for wine authentication and

classification. Spectroscopic methods do not require complicated analytical

preparation procedures.

NMR spectroscopy has permitted the discrimination of red wines from three areas of

Italy’s Apulia region (Brescia et al. 2002).The results were comparable with those

obtained from chromatographic and ICP-AES analyses. However, the major

disadvantage of NMR is that it is one of the most expensive analytical techniques to

employ, both in initial capital outlay and running costs (Reid et al. 2006).

Mid infrared (MIR) spectroscopy reveals information about the fundamental

vibrations of molecular bonds. Palma and Barroso (2002) investigated MIR to

characterize and classify wines, brandies and other distilled drinks. Brandy samples

from four countries have been characterized according to their provenance.

2.2.4 Summary

Identification and quantification of trace elements, isotope ratios, phenolic

compounds, amino acids, and general chemical parameters are useful for the

authentication and classification of wines according to geographical origin. However,

to measure those chemical components, expensive instrument or complicated

analytical procedure were required, such as HPLC, NMR and ICP-MS. Routine

chemical analysis parameters, eg. alcohol, pH etc. cost less to achieve the

classification, but several parameters are needed simultaneously. Spectroscopic

methods (such as MIR, Ultra-violet (UV), NIR) provide more convenient procedures.

10

2.3 Near infrared spectroscopy

2.3.1 Introduction

In recent years, near infrared (NIR) spectroscopy has become an effective and

economical analytical technique for measuring food quality parameters. NIR

spectroscopy can analyse the entire sample in 30 seconds and can determine

multivariate parameters simultaneously. It is non-destructive and no sample pre-

treatment is required (Burns and Ciurczak 2001). Furthermore, spectroscopic

instruments are significantly cheaper than other instrumental methods (eg.

chromatography).

The wavelengths of the NIR region range between 750 and 2500 nm. This wavelength

region contains the overtones and combination vibration information of O-H, C-H,

and N-H bonds (Osborne et al. 1993), which are the principal structural components

of organic molecules. The NIR spectrum provides an overall fingerprint of the sample.

The spectra from NIR are very complicated. Hense, it is impossible to realize

meaningful information based on molecular structure (Williams and Norris 2001).

However, analytical information can be extracted from the NIR spectrum through

application of mathematical multivariate analysis techniques. This approach has been

demonstrated in a number of studies for NIR spectroscopy of food and wines. These

are presented and discussed in following sections.

2.3.2. Effect of Sample presentation on Vis and NIR spectra

2.3.2.1. Effect of sample temperature

When building a spectral database, the data is often sourced under varying conditions.

These variations may cause spectra variations. Sample temperature is an important

factor, especially for the liquid samples.

Previous study of the NIR spectrum of pure water has showed that an increase in the

11

temperature results in an intensity increase and peak shifting towards lower

wavelengths (Swierenga et al. 2000). While the hydroxyl group gives rise to two

bands for its stretching mode: a sharper band for the "free" OH groups and a broader

one for the stretch mode of hydrogen-bonded OH groups, as mentioned by Wulfert et

al. (1998), following a temperature increase, the absorption band for the free OH

groups increases in intensity.

Many researchers have studied the effect of temperature on spectra of water and

model wine solutions. Bianco et al. (2000) studied the influence of temperature on the

water spectra in the 1.2 µm region. The amplitudes and widths of the peaks varied

linearly with the temperature, and therefore, the authors believed that it was possible

to mathematically model the water spectrum with a high degree of precision. Spectra

variation has also been studied for ternary mixtures of ethanol, water and iso-propanol

(Swierenga et al. 2000). Absorption band variations occurred around 970 nm

wavelength and affected calibration models.

Multivariate calibration techniques have been used to deal with temperature

influences by including the temperature change in the calibration set (Wulfert et al.

2000; Hageman et al. 2005). Both linear regression and non-linear regression methods

were employed. For linear regression, principal component regression (PCR) and PLS

are the main techniques (Wulfert et al. 2000) and for non-linear regression, locally

weighted regression (LWR) and neural networks are frequently applied (Hageman et

al. 2005)

However, wine is a complex mixture of chemical components, not as simple as the

model solution containing only ethanol and water. Therefore, the temperature effect

for a real wine is expected to be more complicated than that for ethanol-water mixture

or model wine solutions.

2.3.2.2 Measurement mode effect

Measurement mode is another factor which may affect the application of NIR

spectroscopy. Reflectance (R) and transmission (T) are the two scanning modes that

12

have been applied for NIR spectroscopy. Transmission spectroscopy, where light

passes through a liquid sample and measured by the detector placed behind the

specimen, is well understood. Transflectance (TR) originally developed by Technicon

(now Bran+Luebbe, Germany) for the InfraAlyzer, was designed to study liquids in

an instrument using the reflectance measurement mode (Kawano 2002). In

transflectance, light passes through a sample, is reflected back from a diffuse mirror

placed behind the sample, then passes back through the sample and is measured by a

detector (Murray & Cowe 2004). Transflectance is not as popular as the transmission

mode. However, it can be successfully used for a liquid stream, frequently in

conjunction with optical bundle probes. It is suitable for in-line analysis and may be

more appropriate for industrial applications (Pasquini 2003).

Transmission mode has been adopted in most studies using NIR spectroscopy for

wine and other alcoholic beverage samples (Dambergs et al. 2002; Sauvage et al.

2002; Cozzolino et al. 2003). Some authors have used the tranflectance mode to

determine wine chemical parameters (Urbano Cuadrado et al. 2004). However, no

literature was found to compare the effect on the spectra and the analysis results for

different measurement modes for wine samples.

2.3.3. Use of NIR to classify food based on geographical origin

Several authors utilized NIR to classify food products based on their geographical

region. NIR was exploited to classify Japanese soy sauce based on their geographic

regions. Thirty-eight soy sauce samples were collected from three regions in Japan

(Iizuka and Aishima 1997). Approximately eighty percent of samples were correctly

classified using three pattern recognition methods, including linear discriminant

analysis (LDA), PLS and artificial neural network (ANN).

Bertran et al. (2000) has successfully applied NIR with pattern recognition methods to

authenticate virgin olive oils from very close geographical origins. Two chemometric

techniques, ANN and logistic regression (LR), were employed as classifying tools for

NIR spectra.

13

Samples of Emmental cheese sourced from six regions have been analyzed by NIR.

Using a combination of PCA and LDA, classification by region of origin of the

cheeses was achieved (Pillonel et al. 2002).

Honey samples produced in the European Community from different geographical

and botanical sources were characterized by NIR spectroscopy (Davies et al. 2002).

With chemometric methods, the botanical origins of the honey samples were

identified. However, separation based on the geographical origin could not be

achieved.

Arana et al. (2005) have authenticated the origins of white grapes from two different

wine production zones in Spain. The authors initially used the weight of the berries

and the soluble solids content, as important parameters to evaluate the grape. Only

59% of grapes were correctly classified. By contrast, classification using NIR spectra

achieved a 79.2% accuracy. The results from this study demonstrated the ability of

NIR spectra to identify the origin of white grapes and also indicated that NIR spectra

may achieve superior discrimination. Furthermore, if geographical differences can be

observed in grape spectra, similar information may also appear in the wine’s spectra.

2.3.4 NIR applications on wine analysis

The first application for NIR in wine analysis was the determination of some of the

main components of wine, such as ethanol, fructose, and tartaric acid, by Kaffka and

Norris. This preliminary study was performed on small number of test samples

prepared by standard addition of the components of interest to red or white wines.

Through their work, the critical wavelengths that could be used for multi-linear

regression analysis were identified (Dambergs et al. 2004).

Dambergs et al. (2002) accurately predicted the methanol concentration in wine-

fortifying spirit samples using NIR spectroscopy. Calibration models were developed

by combining sample NIR spectra and with the concentration measured by gas

chromatography (GC) tby several regression techniques, including partial least square

regression (PLS) and multiple linear regression (MLR). Comparisons of the standard

14

error of prediction demonstrated that the most useful NIR calibration model was built

by using continuous spectra, rather than a smaller number of fixed wavelengths.

Trace metals in twenty-four white wines were studied by atomic absorption

spectrometry (AAS) and NIR spectroscopy. Both MLR and PLS were applied to

construct calibration models by coupling NIR spectra with trace metal concentrations

analyzed by AAS. The regression correlation coefficients (R) and the standard error

of cross validation (SECV) for the calibration models indicated that the models were

acceptable for K, Na, Mg and Ca, but not for Cu and Fe (Sauvage et al. 2002).

Cozzolino et al. (2004) examined the potential of NIR spectroscopy to predict the

concentration of phenolic compounds of Australian red wine. The calibration

equations were built using PLS regression of the reference method (HPLC) and NIR

data. The calibration equations proved robust for the prediction of unknown samples.

This experiment demonstrated the ability of NIR for the quantitative analysis of wine

samples. The relationship between sensory analysis and NIR spectroscopy in

Australian Riesling and Chardonnay wines was also investigated (Cozzolino et al.

2005). The research suggested a correlation between sensory and NIR data but results

were considered unreliable due to the small sample size.

The feasibility of utilizing fifteen common wine parameters was studied using NIR

reflectance spectroscopy in conjunction with partial least square regression (Urbano

Cuadrado et al. 2005). Major components, such as alcohol, total acidity, pH, glycerol,

color and total polyphenol index were accurately determined. The SECV values

achieved from NIR spectra were close to those from reference methods. However, for

some organic acids including malic acid, tartaric acid and lactic acid, the accuracy of

prediction results were not as good as the above components.

The studies listed above demonstrated that NIR spectra contain the chemical

information for the analyzed samples and also can reveal characteristic information

about wine quality. This information can be extracted and applied for quality control

purposes. As well, this information is also predictive of geographical origin.

15

Summary and research gaps

As outlined in this review of the literature, wine classification based on geographical

origin may be successfully achieved using different instrumental analyses of the

chemical composition of wines. Unfortunately, most methods require expensive

instruments and/or complicated analysis procedures, and are problematic for industry

application. Visible and near infrared spectroscopy (Vis-NIR) is a relatively rapid and

low cost analytical technique. It has been employed to analyze wine samples and to

predict the value of several chemical parameters. However, previous work has not

focused on the geographical classification of wine using Vis-NIR spectroscopy.

Furthermore, no research has been performed to study the effect of sample

temperature and measurement mode for wine samples. Therefore, to fill these gaps in

knowledge, the aims of this research were to

1. to study the Vis-NIR spectra variation of wine samples and the corresponding

effect for the model calibration caused by sample temperature changes;

2. to examine the effect of measurement mode, including transmission and

transflectance, for wine Vis-NIR spectra and their influence on model

calibration;

3. to classify wine samples of same variety using Vis-NIR spectroscopy based on

their geographical origins, for red (Tempranillo) and white (Riesling) wines

respectively.

16

Chapter 3 Material and methods

The experimental investigation in this thesis involved measurement of Vis-NIR

spectra of wine samples and mathematical analysis of these spectra. This chapter

describes the wine samples analysed, the instruments, their method of use, and the

mathematical techniques which were employed.

3.1 Wine samples

Different wine samples were analyzed:

1. to study the effect of temperature on Vis-NIR spectra (Chapter 4). Ten red and

ten white wine samples (see Table 4.1) were collected randomly from the

Analytical Lab of the Australian Wine Research Institute (AWRI). All samples

were commercially available Australian wine. The red wines included three

wine varieties (Cabernet Sauvignon, Shiraz, and Pinot Noir) and one blend

(blend of Cabernet Sauvignon and Shiraz). There was one Rośe wine. The

white wines included six varieties (Chardonnay, Pinot Gris, Riesling,

Sauvignon Blanc, Semillon, Verdelho). Each sample was given a unique code,

initialled with R for red wine and W for white wine;

2. to study the measurement mode effect on Vis-NIR spectra (Chapter 5). Six red

and six white wines were randomly sourced from Analytical Lab of the

Australian Wine Research Institute (AWRI) (see Table 5.1). Each wine was

allocated a unique code;

3. to classify red wines based on geographical origin using Vis-NIR spectra

(Chapter 6). A total of sixty three bottles (8 labels x 4 replicates; 14 labels x 2

replicates and three bottles for three different labels) comprised of 25

Australian (n =15) and Spanish (n =10) Tempranillo wines were collected. All

samples were commercially available. The vintage of these wines ranged from

1999 to 2004 for the Spanish wines, and 2001 to 2004 for the Australian wines,

17

wines (Table 6.1).

4. to classify white wines based on geographical origin using Vis-NIR spectra

(Chapter 7). A total of fifty bottles (4 labels x 3 replicates and 19 labels x 2

replicates) comprised of 23 commercially available Australian (n =10), New

Zealand (n =5) and European (France and Germany) (n =8) Riesling wines

were collected. The vintage of these wines ranged from 2001 to 2005 (see

Table 7.1).

3.2 Wine reference analysis

Prior to spectra scanning, wine samples were analyzed to determine the key chemical

characteristics at the AWRI Analytical Service (http://www.awri.com.au/analytical_

service/analyses/). The chemical parameters included alcohol content, pH, titratable

acidity (TA), and glucose plus fructose (G+F) and (for geographical classification

samples only) total phenolics, color density and hue. The value of alcohol, pH, TA and

G+F were obtained from Wine Scan analysis (FOSS WineScan FT 120, Silver Spring,

MD, USA). The WineScan is a simple-to-use instrument for rapid analysis of wine. It

delivers results for the major quality parameters in a single analysis, including ethanol,

total acid, volatile acid, pH, glucose, fructose and reducing sugar. Total phenolics

were calculated by the absorbance at 280 nm measured using a UV/Vis

spectrophotometer (Cary 300, Varian, Inc., Palo Alto, CA, USA). For the Tempranillo

wines, wine color density was calculated by measuring the optical density (OD) of the

wine sample at two wavelengths at the actual wine pH (OD 520 nm plus OD420 nm).

Wine color hue was calculated as the ratio OD420/OD520 (Jackson 2000).

3.3 Spectroscopic measurements

Wine samples from freshly opened bottles were scanned in transmission or

transflectance mode using a FOSS NIRSystems6500 spectrophotometer (see Figure

3.1) (FOSS NIRSystems, Silver Spring, MD, USA). A 1mm path length cuvette was

used to contain the wine sample for transmission measurement mode. 0.2 mm and 0.4

18

mm depth transflectance sample cells were used for transflectance mode. Figure 3.2

depicts the operation of these two different measurement modes. Samples were pre-

equilibrated at the measurement temperature for 3 minutes before scanning. In the

investigation of temperature effect on Vis-NIR spectra, six temperatures were applied,

including ambient, 30 ºC, 35 ºC, 40 ºC, 45 ºC and 50 ºC. For other analyses, samples

were pre-equilibrated at 33 ºC before measuring spectra.

Figure 3. 1 FOSS NIRSystem6500 spectrophotometer, Silver Spring, MD

19

Figure 3. 2 Sample presentation of transmittance and transflectance

Spectral data were recorded using Vision software (version 1.0, FOSS NIRSystems,

Silver Spring, USA). The full wavelength range (400- 2500 nm) including the visible

region was analyzed. Spectral data were stored as the logarithm of the reciprocal of

transmittance [log (1/T)] or reflectance [log (1/R)] at 2 nm intervals. Instrument

performance was checked following the diagnostic protocols provided by the

manufacturer.

3.4 Spectra data analysis

Spectra were exported from the Vision software in NSAS format into The

Unscrambler software (version 9.2, CAMO ASA, Oslo, Norway) for chemometric

analysis.

3.4.1 Spectra pre-treatment

Standard Normal Variate (SNV) and second derivative transformation were used to

pre-process the spectra. SNV was performed for scatter correction. SNV was invented

to reduce spectral noise and eliminate background effects of NIR data (Barnes et al.

1989). It is a row-oriented transformation which centres and scales individual spectra

such that:

20

SNVi =(Yi -Ymean) / Stdev (Y) [3.1]

Where: SNV = Standard normal variate for the value of log (1/T) at the ith wavelength

Yi = value of log (1/T) at the ith wavelength

Stdev = standard deviation of the log (1/T) at all wavelengths

Second derivative transformation was performed using Savitzky-Golay derivative and

smoothing (left 5 and right 5 points and 2nd order filtering operation) to reduce

baseline variation and enhance the spectral features (Hruschka 1992). In NIR spectra,

peak overlapping is commonly observed (Williams and Norris 2001). Second

derivative transformation is a simple and frequently used approach to improve

spectral resolution, by which peak width is decreased and more peaks appear (Naes et

al. 2002). As a result, the spectral features are enhanced. However, a disadvantage of

derivatives is that they can amplify noise. This problem was overcome by using

Savitsky-Golay smoothing (Brereton 2003).

3.4.2 Multivariate analysis

3.4.2.1 Principal component analysis

Principal component analysis (PCA) was used in this investigation to reduce the

dimensionality of the data to a smaller number of components, to examine any

possible grouping of samples, and to visualize the presence of outliers.

PCA is a mathematical procedure widely used to transform sets of possibly correlated

data into a new set of orthogonal components, which are called principal components

(PCs) (Naes et al. 2002). The PCs reduce the data in a way that maximises between-

sample spectral variation. A set of n spectra can be expressed as a n x p data matrix X

containing n values of transmission absorbance at each of the p wavelength. The

general equation for PC calculation (Otto 1999) is:

X = TPT + E [3.2]

21

Where T is the score matrix,

P is the transposed eigenvector matrix,

E is the residual matrix.

The scores are the new values of spectra in the coordinate system defined by PCs and

the eigenvectors are the link between the wavelengths of the X matrix and the

principal component space (Otto 1999).

Fifteen PCs were derived from the spectral data in the studies of this thesis. PCA

models were developed using both raw and pre-processed data.

3.4.2.2 Partial least square regression

Calibration models between chemical composition and NIR spectra were developed

using partial least square regression (PLS). PLS is a multivariate data analysis

technique which can be used to relate several response (Y) variables to several

explanatory (X) variables (Otto 1999). The method aims to create new explanatory

variables by linear combination of the original X variables that maximize the

covariance between the response variables and the explanatory variables (Massart et

al. 1988; Otto 1999; Naes et al. 2002). This process is similar to PCA in the fact that

PLS also creates scores and loadings. However, the difference is that PCA creates its

scores by finding the linear combination of the explanatory variable that has

maximum variance among those X variables, but without considering the response

variables.

Calibration statistics, including the coefficient of determination in calibration (R2) and

the standard errors of cross validation (SECV), were calculated to evaluate the

accuracy of PLS calibration models.

To compare the calibration models achieved at different experiment conditions, the

SECVs were compared using an F test (Naes et al. 2002):

F = SECV22 / SECV1

2, where SECV1 < SECV2 [3.3]

22

The calculated F value was compared with the confidence limit F limit (1-α, n1-1, n2-2),

obtained from the distribution F table, where α is the test significance level (α=0.05 in

this experiment), n1 is the sample number measured at the first condition and n2 is the

sample number measured at the second condition. The differences between the SECV

are significant when F > F limit.

3.4.2.3 Discriminant analysis

Discrimination models were developed using linear discriminant analysis based on

PCA scores and the discriminant PLS techniques, respectively (Massart et al. 1988;

Naes et al. 2002).

Linear discriminant analysis (LDA) is a supervised classification technique where the

number of categories and the samples that belong to each category are previously

defined (Otto 1999; Naes et al. 2002). The criterion of LDA for selection of latent

variables is maximum differentiation between the categories that minimizes the

variance within categories (Naes et al. 2002). The method produces a number of

orthogonal linear discriminant functions, equal to the number of categories minus one,

that allow the samples to be classified in one or another category (Naes et al. 2002).

LDA was carried out on the PCA sample scores using JMP software (version 5.01,

SAS Institute Inc., Cary, NC, USA). This procedure has previously been used to

authenticate instant coffee and differentiate apple juice samples and meats (Charlton

et al. 2002; Reid et al. 2005; Cozzolino et al. 2005). The first several components

which yield the highest level of separation and explain most variation of the spectra

matrix in the PCA models, were input to the LDA analysis.

Discriminant partial least square regression (DPLS) is a variant of partial least square

regression (PLS). In this technique, each sample in the calibration set is assigned a

dummy variable as a reference value (eg. for Tempranillo wines, set to 1 = Australian

wines and 2 = Spanish wines) (Naes et al. 2002; Brereton 2003). The classification of

the wine samples accordingly to geographic origins was based on a 0.5 cut-off value.

The coefficient of determination in calibration (R2) and SECV were calculated to

evaluate the DPLS calibration models. The sample numbers correctly classified in

23

prediction were counted to calculate the classification rate.

The PCA, PLS, LDA and DPLS models were developed using full cross validation

(CV) (leave one out method). Cross-validation estimates the prediction error by

splitting the calibration samples into groups, where in the case of full cross validation,

each sample can be seen as one group (Otto 1999; Naes et al. 2002; Brereton 2003).

24

Chapter 4 Effect of sample presentation - sample

temperature effect on the analysis of wine

In order to study the classification ability of Vis-NIR spectroscopy, the initial step in

this research was to determine the optimal experimental conditions for obtaining

“fingerprints” of wine samples. The dominant parameters influencing the Vis-NIR

spectrum include the scanning temperature and measurement mode.

4.1 Introduction

NIR records the overtones and combination vibrational information of the molecular

bonds. Temperature changes can affect the vibration intensity of molecular bonds,

hence, the vibrational spectrum will change according to the temperature variation.

Consequently different temperatures may affect the result of a classification or

calibration model. There has been no study of the impact of temperature on the Vis-

NIR spectra of real wines. This chapter will focus on this question.

4.2 Results and discussion

4.2.1 Chemical analysis

Tables 4.1 and 4.2 summarized the profiles of the red and white wine samples

analysed.

25

Table 4. 1 White wine sample codes, variety, chemical composition and the

corresponding statistics of samples analyzed

Sample

code Variety Alcohol

(% v/v)pH TA

(g L-1)

G+F

(g L-1) Ww1 Chardonnay 13.75 3.36 6.55 1.80 Ww2 Chardonnay 13.37 3.41 6.73 4.10

Ww3 Chardonnay 13.43 3.26 6.54 0.70

Ww4 Pinot Gris 13.76 3.13 7.48 1.60

Ww5 Riesling 13.21 3.18 6.52 8.0

Ww6 Sauvignon Blanc 12.99 3.27 6.26 2.10

Ww7 Semillon 11.70 3.27 6.3 0.70

Ww8 Semillon 11.09 3.18 7.05 2.60

Ww9 Unwooded Chardonnay 13.55 3.36 6.32 9.30

Ww10 Verdelho 12.90 3.41 6.91 4.10

Mean 12.98 3.28 6.67 3.50 S.D. 0.89 0.10 0.39 2.97

Min 11.09 3.13 6.26 0.70

Max 13.76 3.41 7.48 9.30 a TA, titratable acidity; G+F, glucose + fructose; S.D., standard deviation; Min,

minimum value; Max, maximum value.

26

Table 4. 2 Red wine sample codes, variety, chemical composition and the corresponding

statistics of samples analyzed

Sample

code Variety

Alcohol

(% v/v) pH

TA

(g L-1)

G+F

(g L-1)

Rw1 Cabernet Sauvignon 13.61 3.53 6.63 0.30 Rw2 Cabernet Sauvignon 13.08 3.57 6.04 1.80

Rw3 Cabernet Sauvignon 12.49 3.43 7.35 0.40

Rw4 Cabernet Sauvignon 13.21 3.49 7.78 3.90

Rw5 Cabernet Sauvignon 12.92 3.36 7.22 0.20

Rw6 Pinot Noir 13.51 3.62 7.31 0.50

Rw7 Rośe 13.08 3.36 6.06 4.60

Rw8 Shiraz 13.65 3.54 6.44 0.20

Rw9 Shiraz 14.08 3.63 6.73 0.60

Rw10 Blend of Cabernet Sauvignon

and Shiraz 14.15 3.43 6.58 0.30

Mean 13.29 3.50 6.84 1.34

S.D. 0.52 0.10 0.58 1.64

Min 12.49 3.43 6.04 0.2

Max 14.15 3.63 7.78 4.6

4.2.2 Spectra interpretation and analysis

Figures 4.1 and 4.2 present the Vis-NIR spectra for the red and white wines at six

temperatures. Both varieties have high absorption at 1450 nm and 1950 nm.

Absorption at 1450 nm is the first overtone of O-H stretching vibration and absorption

at 1950 nm is a combination band of OH stretch and deformation (Osborne et al.

1993). Minor peaks appear around 976 nm, 1690 nm, 2268 nm, and 2306 nm. The

976 nm area is associated with the O-H stretch second overtone of water and ROH

(Osborne et al. 1993). The absorption at 1690 nm is related to C-H stretch first

overtones (Osborne et al. 1993). The absorption band at 2268 nm is related to C-H

combination and O-H stretch overtones and absorption at 2306 nm is related to C-H

combinations (Burns and Ciurczak 2001). Red wines have a peak in the visible region

around 540 nm which is related to the pigments (Somers 1998).

27

500 1000 1500 2000 25000

1

2

3

Red wines

2268 & 2306 nm

1950 nm

1450 nm

540 nm

log

1/T

wavelength (nm)

1690 nm

Figure 4. 1 Vis-NIR spectra of red wine samples at six different temperatures

500 1000 1500 2000 25000.0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

log

1/T

Wavelength (nm)

1450 nm

1690 nm

1950 nm

2268 & 2306 nm

White wines

Figure 4. 2 Vis-NIR spectra of white wine samples at six different temperatures

Figure 4.3 shows the second derivative of spectra of red and white wine samples. The

second derivative transformation inverts the spectra, so the peaks of the original

spectra become troughs. The peaks become sharper, and some of the overlapping

peaks are separated (Hruschka 1992). From Figure 4.3, it can be noticed that peaks at

1450 nm and 1950 nm were separated into two peaks, and an additional peak has

appeared at 2306 nm.

28

500 1000 1500 2000 2500-0.025

-0.020

-0.015

-0.010

-0.005

0.000

0.005

0.010

0.015

0.020

Seco

nd d

eriv

ativ

e

Wavelength (nm)

Figure 4. 3 Second derivative spectra of red and white wine samples

Figure 4.4 presents the spectra of water. It is similar to that for wine, also exhibiting

strong absorptions at 1450 nm and 1950 nm. However, peaks do not occur at 1690 nm

and 2200 to 2300 nm, and the absorption of the 1450 nm peak is higher.

500 1000 1500 2000 25000.0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

Wavelenth (nm)

Log

1/T

ambient 30oC 35oC 40oC 45oC 50oC

water

Figure 4. 4 Water spectra at six different temperatures

29

4.2.3 Influence of temperature on the Vis-NIR spectra of wine

Figure 4.5 shows the spectra of one red sample Rw1 at six different temperatures.

These spectra are typical of all wine samples. Figures 4.6 and 4.7 present

enlargements of the spectra at 1450 nm and 2270 to 2300 nm. Peak shifting and

increased peak height were observed with temperature. However, peak shifting at

1450 nm is different to that occurring at 2270 to 2300 nm. These displacements can

also be observed in the second derivative of the spectra. Figures 4.8 and 4.9 show the

second derivative spectra at 1450 nm and 2270 to 2300 nm. To establish the

relationship between the temperature and spectral variation, four wavelengths regions:

960~1000 nm, 1410~1470 nm, 1660~1706 nm and 2250~2360 nm were analysed.

500 1000 1500 2000 25000.0

0.5

1.0

1.5

2.0

2.5

3.0 ambient 30oC 35oC 40oC 45oC 50oC

Log

1/T

Wavelength (nm)

2268 & 2306nm

1450nm

540nm

1950nm

Figure 4. 5 Vis-NIR spectra for a red wine sample (Ww1) at six different temperatures

30

1410 1420 1430 1440 1450 1460 1470 1480 14900.75

0.80

0.85

0.90

0.95

1.00

1.05

1.10

1.15

1.20

Wavelenth (nm)

Log

1/T

ambient 30oC 35oC 40oC 45oC 50oC

Figure 4. 6 Wine spectra at 1450 nm region

2250 2260 2270 2280 2290 2300 2310 2320 23300.90

0.95

1.00

1.05

1.10

1.15

1.20

1.25

1.30

Wavelength (nm)

Log

1/T

ambient 30oC 35oC 40oC 45oC 50oC

Figure 4. 7 Wine spectra at 2270 nm to 2300 nm region

31

1410 1420 1430 1440 1450 1460 1470 1480

-3.0x10-3

-2.8x10-3

-2.6x10-3

-2.4x10-3

-2.2x10-3

-2.0x10-3

-1.8x10-3

-1.6x10-3

-1.4x10-3

-1.2x10-3

-1.0x10-3

-8.0x10-4

-6.0x10-4

-4.0x10-4

-2.0x10-4

ambient 30oC 35oC 40oC 45oC 50oC

Seco

nd d

eriv

ativ

e

Wavelength (nm)

Figure 4. 8 Second derivative wine spectra at 1450 nm region

2260 2280 2300 2320 2340-0.003

-0.002

-0.001

0.000

0.001

0.002

ambient 30oC 35oC 40oC 45oC 50oC

Sec

ond

Der

ivat

ive

Wavelength (nm)

Figure 4. 9 Second derivative wine spectra at 2270 nm to 2300 nm region

a. 960~1000 nm

The peak in this region occurs at approximately 976 nm, and is related to the O-H

second overtone (Osborne et al. 1993). Wine samples produced a peak at 976 nm at

30 ºC, which shifts to 972 nm at 50 ºC. The peak for water shifts from 974 nm at 30

32

ºC to 970 nm at 50 ºC. Plotting the peak height against temperature change displayed

no apparent linear or other kind of relationship (data not shown).

In the second derivative spectra, a peak in all wine samples and water at six

temperatures occurs at 962 nm. Figures 4.10 and 4.11 present the plots of the second

derivative peak heights at 962 nm for the average spectra of red and white wines

versus temperature. A linear relationship was observed.

30 35 40 45 50

-1.30x10-4

-1.25x10-4

-1.20x10-4

-1.15x10-4

-1.10x10-4

-1.05x10-4

-1.00x10-4

Sec

ond

deriv

ativ

e

Temperature oC

y = -1.394E-06x - 5.835E-05

R2 = 9.990E-01P < 0.0001

Figure 4. 10 Linear relationship of the absorbance of second derivative spectra at 962nm

of red wines

33

30 35 40 45 50

-1.35x10-4

-1.30x10-4

-1.25x10-4

-1.20x10-4

-1.15x10-4

-1.10x10-4

-1.05x10-4

-1.00x10-4

Sec

ond

Der

ivat

ive

Temperature oC

y = -1.358E-06x - 6.223E-05

R2 = 9.975E-01P < 0.0001

Figure 4. 11 Linear relationship of the absorbance of second derivative spectra at 962nm

of white wines

b. 1410~1470 nm

The peaks in this region experienced a similar shifting trend as that observed for the

960~1000 nm region. However, the extent of shifting was larger. The peak points

moved from 1454 nm at 30 ºC to 1444 nm at 50 ºC. No obvious linear relationship

existed between absorbance and temperature of the peaks in the raw spectra (data not

shown).

After transforming the spectra to second derivative, the major peak located at

approximately 1450 nm separated into two peaks, around 1414 nm and 1462 nm.

Both peaks displayed a linear relationship between height and temperature. However,

the peak height at 1414 nm decreased with temperature increase, whilst the peak

height at 1460 nm increased with temperature. Figures 4.12 and 4.13 show the linear

relationships of second derivative peaks against temperature of red wines. White

wines have similar relationships (data not presented). This behavior might be

explained by the increase in free hydroxyl bonds as temperature increases (Wulfert et

al. 1998). The 1414 nm peak may be attributed to the free hydroxyl bond, whilst the

1462 nm peak might be associated with stretch mode of hydrogen-bonded OH groups.

34

30 35 40 45 50-0.0036

-0.0034

-0.0032

-0.0030

-0.0028

-0.0026

-0.0024

-0.0022

-0.0020

Sec

ond

deriv

ativ

e

Temperature oC

y = -6.380E-05x - 2.680E-04R2 = 9.992E-01P < 0.0001

Figure 4. 12 Linear relationship of the absorbance of second derivative spectra at 1412

nm of red wines average spectra

30 35 40 45 50-2.06x10-3

-2.04x10-3

-2.02x10-3

-2.00x10-3

-1.98x10-3

-1.96x10-3

-1.94x10-3

-1.92x10-3

-1.90x10-3

-1.88x10-3

-1.86x10-3

-1.84x10-3

Sec

ond

deriv

ativ

e

Temperature oC

y = 9.000E-06x - 2.308E-03

R2 = 9.985E-01P < 0.0001

Figure 4. 13 Linear relationship of the absorbance of second derivative spectra at 1462

nm of red wines average spectra

35

c. 1660~1710 nm

There was no wavelength shifting of raw and second derivative spectra in the

1660~1710 nm region. The peak of the raw spectra appeared at 1694 nm and at 1688

nm in the second derivative. In both cases, the relationship between peak height and

temperature appears non-linear (data not presented).

d. 2250~2360nm

In this region, spectral absorptions increased with temperature. No noticeable peak

shifts were observed in this area. All peaks occurred at identical wavelengths for the

raw and second derivative spectra: 2268 nm and 2306 nm.

A linear relationship was found in the raw spectra between absorbance and

temperature. Figures 4.14 and 4.15 show the linear relationships of raw spectra peaks

against temperature of red wines. Similar relationships were observed for white wines

(data not presented). However, no linear relation existed in the second derivative

spectra (data not presented).

30 35 40 45 501.095

1.100

1.105

1.110

1.115

1.120

1.125

1.130

1.135

Log

1/T

Temperature oC

y = 1.653E-03x + 1.051

R2 = 9.802E-01P= 0.0012

Figure 4. 14 Linear relationship of the absorbance of raw spectra of red wines at 2268

nm

36

30 35 40 45 50

1.23

1.24

1.25

1.26

1.27

1.28

Log

1/T

Temperature oC

y = 1.838E-03x + 1.179

R2 = 9.669E-01P = 0.0025

Figure 4. 15 Linear relationship of the absorbance of raw spectra of red wines at 2306

nm

Summary

From the analysis of these four wavelength regions, it was noticed that the second

derivative of the spectra minimizes the peak shifting effect caused by temperature

variation. The peaks related to O-H overtones exhibit linear relationships between

peak absorbance and temperature variation in the second derivative spectra. For the

peaks related to C-H bonds (2250 – 2360 nm), a linear relationship was observed in

the raw spectra between peak absorbance and temperature increase.

4.2.4 Principal component analysis

PCA was performed to analyze the wine raw spectra. Figures 4.16 and 4.18 illustrate

the score plots of PC1 and PC2 for the white wines and red wines, respectively.

Samples scanned at the same temperature were clustered. The groups were dispersed

from left to right as the temperature increased. However, the behaviour of red and

white wines was different. Comparing the first two PCA eigenvectors of red and white

wine samples (see Figures 4.17 and 4.19), it was found that the second PC of the red

37

wine is similar to the first PC of the white samples. Eigenvectors of PC1 of red wines

are primarily associated with absorbance at 540 nm which is related to the pigments

of the wine. PC2 of the red wines and the PC1 of the white samples were associated

with changing temperature.

-0.8 -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6 0.8

-0.4

-0.3

-0.2

-0.1

0.0

0.1

0.2

0.3

0.4

0.5

0.6 ambient 30oC 35oC 40oC 45oC 50oC

PC

2

PC1

Figure 4. 16 Score plot of PC1 and PC2 of the white wine samples.

500 1000 1500 2000 2500-0.15

-0.10

-0.05

0.00

0.05

0.10

0.15

0.20

PC1 (76%) PC2 (15%)

PC

A E

igen

vect

ors

Wavelength (nm)

Figure 4. 17 Eigenvectors of the first two PCs of the PCA for white wines

38

-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

-0.8

-0.6

-0.4

-0.2

0.0

0.2

0.4

0.6

0.8

1.0

ambient 30oC 35oC 40oC 45oC 50oC

PC2

PC1

Figure 4.18 Score plot of PC1 and PC2 of the red wine samples.

500 1000 1500 2000 2500

-0.12

-0.10

-0.08

-0.06

-0.04

-0.020.00

0.02

0.04

0.06

0.08

0.100.12

0.14

0.16

0.18

PC

A E

igen

vect

ors

Wavelength

PC1 (55%) PC2 (32%)

Figure 4. 19 Eigenvectors of the first two PCS of the PCA for the red wines

A linear relationship was observed between the mean score value of the wine samples

at the same temperature and temperature variation in the first PC for white and second

PC for red wines. Figures 4.20 and 4.21 show these linear relationships. The linear

relationships of red and white wine samples possessed similar slopes and intercepts.

39

30 35 40 45 50

-0.6

-0.4

-0.2

0.0

0.2

0.4

0.6

0.8

PC

A m

ean

scor

e va

lue

Temperature oC

y = 0.0681x - 2.6259R2 = 0.9929P = 0.0002

Figure 4. 20 Linear relationship between the mean score values of temperature related

PC of white red samples and temperature variation

30 35 40 45 50

-0.6

-0.4

-0.2

0.0

0.2

0.4

0.6

0.8

1.0

30 35 40 45 50

-0.6

-0.4

-0.2

0.0

0.2

0.4

0.6

0.8

1.0

PC

A m

ean

scor

e va

lue

Temperature oC

y = 0.0676x - 2.5913R2 = 0.9981

PC

A m

ean

scor

e va

lue

Temperature oC

y = 0.0676x - 2.5913R2 = 0.9981P < 0.0001

Figure 4. 21 Linear relationship between the mean score values of temperature related

PC of white wine samples and temperature variation

40

4.2.5 Comparison of calibration performance at different temperatures using

PLS

To compare the influence of temperature on wine spectra, calibration models of wine

chemical data and Vis-NIR spectra were performed using PLS. The SECV obtained

from different temperatures were compared. A smaller SECV presents a better

prediction result. Table 4.3 lists the standard error in cross validation (SECV) of the

prediction models for each parameter at different temperatures. Temperature affects

the SECV of the red and white wines differently. For red wines, no significant

difference was observed between 30 ºC and 35 ºC in the four chemical parameters.

However, the SECVs of one or several parameters were significantly different at

ambient temperature, 40 ºC, 45 ºC and 50 ºC. For white wines, the SECVs of the four

parameters did not differ significantly with temperature (from ambient to 45 ºC).

However, at 50 ºC, the SECVs for alcohol and G+F were significantly different from

the valures at other temperatures. Moreover, for both red and white wines, the SECV

at 50 oC were the maximal. Clearly, the model at 50 oC has the worst prediction

capability. Generally, for both red and white wines, the SECV of the chemical

parameter at 30 ºC and 35 ºC, were smaller than those ones obtained at other

temperatures. This implies that the optimal temperature for wine analysis using Vis-

NIR spectroscopy lies between 30 ºC to 35 ºC.

41

Table 4. 3 Standard error in cross validation (SECV) of PLS prediction using Vis-NIR

raw spectra for chemical analysis paramters

SECV

Red wine Alcohol pH TA GF

ambient 0.084 ab 0.038 b 0.18 b 0.54 b

30 ºC 0.059 a 0.013 a 0.12 a 0.18 a

35 ºC 0.062 a 0.017 a 0.071 a 0.27 a

40 ºC 0.14 b 0.029 b 0.18 b 0.51 b

45 ºC 0.30 c 0.059 c 0.11 a 0.43 b

50 ºC 0.097 b 0.027 b 0.17 b 0.59 b

White wine

ambient 0.077 a 0.056 a 0.19 a 0.64 a

30 ºC 0.070 a 0.058 a 0.17 a 0.66 a

35 ºC 0.074 a 0.059 a 0.22 a 0.80 a

40 ºC 0.12 a 0.065 a 0.23 a 1.04 a

45 ºC 0.069 a 0.040 a 0.17 a 0.58 a

50 ºC 0.23 b 0.08 a 0.24 a 2.58 b

a,b: Levels in column not connected by the same letter are significantly different, α <

0.05

42

Chapter 5 Effect of sample presentation –

measurement condition effect on the analysis of wine

5.1 Introduction

Following recent advances in NIR spectroscopy, new NIR instruments have been

developed. Some of these instruments offer convenience and easier scanning

procedures. Different scanning modes are available. For wine or grape juice samples,

some instruments can analyse the sample in a sample cup in transflectance mode and

others use transmittance. However, no studies have compared differences between the

transmittance and transflectance scanning mode for wine analysis. Does the mode

affect the analysis result when using NIR, and which one produces better prediction

results? To answer these questions, different scanning modes and different path

lengths for transflectance mode were applied to wine samples and the prediction

errors were compared.

5.2 Results and discussion

5.2.1 Chemical analysis

Table 5.1 lists the chemical profiles of the red and white wine samples analysed.

43

Table 5. 1 Sample codes, chemical composition and the corresponding statistics of

samples analysed

Sample code Alcohol

(% v/v) pH TA (g/L) G+F (g/L)

Red wines

3012 14.25 3.46 6.81 0.7

0443 13.68 3.48 6.79 0.7

2138 13.41 3.61 7.29 0.3

2139 12.81 3.82 6.84 0.4

R590 13.08 3.57 6.04 1.8

R598 13.13 3.48 7.71 4

Mean 13.39 3.57 6.91 1.32

S.D. 0.51 0.14 0.56 1.42

Min 12.81 3.46 6.04 0.3

Max 14.25 3.82 7.71 4

White wines

1930 11.05 3.26 7.38 77.7

0107 13.27 3.48 6.77 9.7

1162 13.21 3.4 6.68 1.8

1724 11.95 3.37 6.58 0.5

2705 12.47 3.27 6.83 1.8

2166 13.03 3.22 6.61 2.2

Mean 12.50 3.33 6.81 15.62

S.D. 0.87 0.10 0.30 30.59

Min 11.05 3.22 6.58 0.5

Max 13.27 3.48 7.38 77.7

5.2.2 Spectra analysis

Figures 5.1 and 5.2 show the spectra of the wine samples scanned in transflectance

44

mode for 0.2 mm and 0.4 mm depth. Figure 5.3 presents a comparison of the spectra

of the identical sample scanned under different measurement conditions. The 0.2 mm

transflectance spectrum exhibits the lowest absorbance among the spectra. The 0.4

mm spectrum has higher peaks at 540 nm, 1450 nm and 2300 nm, and has a flat peak

around 1950 nm, lower than that for the 1 mm transmission spectrum. Although the

spectra absorptions differ for the spectra obtained from different scanning conditions,

the absorbance peaks occurred at the identical wavelengths.

400 600 800 1000 1200 1400 1600 1800 2000 2200 24000.0

0.5

1.0

1.5

2.0

2.5

log

1/T

Wavelength (nm)

0.2 mm Transflectance

Figure 5. 1 Vis-NIR spectra of six red and six white wine samples at 0.2 mm

transflectance mode

400 600 800 1000 1200 1400 1600 1800 2000 2200 24000.0

0.5

1.0

1.5

2.0

2.5

log

1/T

Wavelength (nm)

0.4mm Transflectance

Figure 5. 2 Vis -NIR spectra of six red and six white wine samples at 0.4 mm

transflectance mode

45

400 600 800 1000 1200 1400 1600 1800 2000 2200 24000.0

0.5

1.0

1.5

2.0

2.5

3.0

log

1/T

Wavelength (nm)

0.2mm (TR)pathlength 0.4mm (TR) 1mm (T)

Figure 5. 3 Vis-NIR spectra of the same sample at three different path lengths

5.2.3 Principal component analysis

The spectra data were analyzed by PCA. The first three PCs together explained more

than 99% of the variation of the spectral data set: PC1 73%, PC2 26% and PC3 1%.

The first two PCs account for most of the variation. Figure 5.4 is the score plot of the

first two PCs. It can be observed that spectra of the same path length were clustered

together. The spectra obtained using transflectance mode were negative on the PC2

axis, while spectra obtained at transmittance mode were positive. This suggests that

PC2 is associated with the measurement mode.

46

-10 -8 -6 -4 -2 0 2 4 6-4

-3

-2

-1

0

1

2

3

4

5

0.2mm TR 0.4mm TR 1mm T

PC

2

PC1

Figure 5. 4 PCA score plot of the PC1 against PC2 using Vis-NIR raw spectra

It was found in the score plot of PC1 and PC3 (Figure 5.5) that the samples were

placed in a same sequence along the PC3, no matter which scanning condition was

used. This suggests that PC3 explains wine information. It also demonstrates that

although the spectra look visibly differently, the information they contain was similar.

The sample presentation does not affect or change the sample information recorded by

Vis-NIR.

47

-10 -8 -6 -4 -2 0 2 4 6 8 10-1.2

-1.0

-0.8

-0.6

-0.4

-0.2

0.0

0.2

0.4

0.6

0.8

1.0

1.2

R598

3012

0443

R590

2139

213821382139

R590

0.2mm TR 0.4mm TR 1mm T

PC3

PC1

R598

3012

0443

R5902139

2138

R598

3012

0443

Figure 5. 5 PCA score plot of the PC1 against PC3 using Vis-NIR raw spectra

5.2.4 Comparison using PLS

Calibration models were constructed using PLS regression between the spectra and

the wine chemical parameters, including alcohol content, pH, titratable acidity and

glucose plus fructose. Table 5.1 describes the chemical parameters of the red and

white wine samples. The predicted values of each chemical parameter of the wine

samples were compared between scanning arrangements by ANOVA (Table 5.2). The

differences between the predicted value and the reference value of each sample using

the scanning conditions were also compared. No significant difference was observed

between these mean values, which mean the spectra acquired from different scanning

conditions produced a similar prediction result for the same sample.

However, Table 5.2 indicates that the standard error using cross validation (SECV)

increased with the path length whilst the coefficient of correlation decreased, which

indicates that the calibration accuracy was reduced. The SECV of the measurement

conditions were compared by F test. No significant difference was observed between

0.4 mm transflectance mode and 1 mm transmittance mode of all the chemical

48

parameters for both red and white wines. However, the SECV of all the chemical

parameters for red wine and alcohol and pH for white wine were significant between

0.2 mm transflectance mode and 0.4 mm transflectance or 1mm transmittance mode.

While the effective pathlength of 0.4 mm transflectance mode is approximately 0.8

mm, the light pathlength is comparable with the 1 mm transmittance mode. Since no

significant difference was observed between these two modes, it can be concluded

that under an analogous effective pathlength, the different measurement modes can

produce similar prediction results. Both the SECV and the coefficient of correlation

indicated that the shorter path length measurement mode provides a more precise

prediction ability. However, the sample loading procedure for transflectance

measurement mode was more complicated than the transmission mode. Therefore it

was decided to use transmission mode for further study.

Table 5. 2 The Standard Error in Cross Validation (SECV) of the prediction models for

each parameter at different scanning modes

Alcohol pH TA G+F

R SECV R SECV R SECV R SECV

1 mm T 0.992 0.0615a 0.982 0.024a 0.983 0.0948a 0.988 0.2a

0.4 mm TF 0.995 0.048a 0.994 0.0132a 0.987 0.0865a 0.996 0.108a Red

wines 0.2 mm TF 0.999 0.0107b 0.999 0.0047b 0.999 0.0243b 0.998 0.0784b

Alcohol pH TA G+F

R SECV R SECV R SECV R SECV

1 mm T 0.996 0.0696a 0.97 0.024a 0.989 0.0413a 0.998 1.88a

0.4 mm TF 0.995 0.078a 0.989 0.0137ab 0.997 0.0199a 0.996 2.45a White

wines 0.2 mm TF 0.999 0.0154b 0.997 0.0065b 0.993 0.0334a 0.99 3.86a

T: transmission mode; TF: transflectance mode;

R: Correlation; a,b: Levels not connected by the same letter in column are

significantly different, α < 0.05

49

Chapter 6 Use of Visible and NIR to classify

Tempranillo wines based on geographical origins.

6.1 Introduction

Tempranillo is the most abundant indigenous grape variety in Spain, especially in

Rioja region (MacNeil 2001). “Tempranillo” means early, which is why the grape was

given that name, because it ripens earlier than most red varietals. Its wine was

characterized by fruity mouth feel and subtle aromas (Kolpan et al. 1996). In recent

years, it has been planted in many Australia vineyards and the wines have become

popular with consumers. Since the different regions may bring different wine

characters, the objective of this experiment was to explore the use of visible and near

infrared spectroscopy to analyze Tempranillo wines from Australia and Spain and

classify them accordingly to their geographical origin (refer Table 6.1).

Table 6. 1 Vintage and origin of commercial Tempranillo wine samples analysed

1999 2000 2001 2002 2003 2004 Total

Australia 2 6 10 18 36

South Australia 2 2 13

New South Wales 2

Victoria 4 4 5

Western Australia 4

Spain (D.O.) 2 2 6 4 13 27

Rioja 4 2 5

La Mancha 2

Ribera del Duero 2 2 2 4

Toro 4

D.O. = denomination of origin

50

6.2 Results and discussion

6.2.1 Chemical analysis

Table 6.2 shows the alcohol content, pH, titratable acidity (TA), glucose plus fructose

(G+F), total phenolics, wine colour density and wine colour (Hue) of the wine

samples analysed. It was noticed that the range in chemical composition for the

Australian wines varied the most, compared with the Spanish wines. No statistically

significant differences were observed for alcohol content, G+F, total phenolics, colour

density, and hue values in the set of wine analysed. Statistically significant differences

were observed in pH and TA, suggesting that the Australian wines contain more

acidity than the Spanish wines. It was noticed that some Australian wines

corresponding to 2004 vintage have high alcohol content (higher than 15% Alc). This

tendency of high ethanol content was also observed for Australian Cabernet

Sauvignon, Shiraz and Merlot wines and reported elsewhere (Godden and Gishen

2005).

51

Table 6. 2 Range of chemical composition for the Australian and Spanish wine analysed

Alcohol

(%) pH

TA

pH8.2

(g L-1)

G+F

(g L-1)

Total

Phenolic

(A.U.)

Colour

Density

(A.U.)

Hue

Mean 13.8 3.6 6.2 1.2 53.2 7.6 8.8

S.D. 0.8 0.1 0.5 1.6 7.2 1.1 0.6

Min 12.6 3.4 5.4 0.0 40.5 5.8 7.9 Australia (n=36)

Max 15.2 3.8 7.1 5.7 63.2 9.7 10.3

Mean 14.0 3.7 5.2 0.3 57.2 7.1 8.6

S.D. 0.2 0.1 0.6 0.2 4.7 1.1 0.5

Min 13.6 3.5 4.4 0.2 50.1 5.2 7.9 Spain (n=27)

Max 14.2 3.9 5.9 0.9 62.8 9.2 9.7

Significance of

difference NS * * NS NS NS NS

TA, titratable acidity; G+F, glucose + fructose; S.D., standard deviation; Min, the minimum

value; Max, the maximum value; A.U., absorbance unit; NS, not significant; *, significant

difference between the mean value (p < 0.05).

6.2.2 Spectra interpretation and analysis

Figure 6.1 show the Vis and NIR spectra of wines after SNV and second derivative

transformation. The second derivative inverts the spectra, so the peaks of the original

spectra become troughs (Hruschka 1992). Two wavelength regions were not used for

calibration, namely between 1000-1100 nm (changes of detector in the instrument);

and between 1880-2000 nm (off-scale, absorbance higher than 2.6 absorbance units).

Absorptions at 1450 nm, 1950 nm (not included for chemometric analysis) are related

to first overtone of O-H stretching vibration and combination band of OH stretch and

deformation (Osborne et al. 1993). Additionally, absorptions were observed around

976 nm related to O-H stretch third overtone associated with water and ethanol, at

1690 nm related to C-H stretch first overtones, and at 2268 nm, and 2306 nm, due to

52

C-H combination and overtones (Burns and Ciurczak 2001). Absorption in the visible

region occurred at around 540 nm related to wine pigments (anthocyanins and

pigmented tannins) (Somers 1998). No obvious differences between wine samples

from different geographical origins were observed; therefore the spectra were

processed by means of multivariate analysis.

500 1000 1500 2000 2500-0.03

-0.02

-0.01

0.00

0.01

0.02

Seco

nd d

eriv

ativ

e an

d SN

V

Wavelength (nm)

Tempranillo wines

Figure 6. 1 Second derivative of the Vis-NIR spectra of Australian and Spanish

Tempranillo wines

6.2.3 Principal component analysis

PCA was applied to both the raw and the pre-processed spectra (SNV and 2nd

derivative). It was noticed that the separation in the score plot of the first two

principal components (PCs) using raw spectra was less clear than that showed using

SNV and second derivative pre-processed spectra (data not presented). Figure 6.2

shows the score plot of the first two PCs from the Vis and NIR pre-processed spectra.

The first three PCs explained 68% of the total variance of the spectra in the set of

wines analysed. A separation between wines according to the geographical origin was

observed. However, it was noticed that some Australian wines overlapped with some

Spanish wine samples.

53

-0.003-0.002

-0.0010.000

0.0010.002

0.003

PC1

-0.002

-0.001

0.000

0.001

0.002

PC

2

AAAA

AAAA

AAAA

AAAA

SSS

S

SS

SS

SSSS

SSSS

SSSS

SS

AAA

A

AA

AA

AA

AA

AA

AA

SS

SS

Figure 6. 2 Score plot of the first two principal components of Australian (A) and

Spanish (S) Tempranillo wines using Vis-NIR after SNV and second derivative

processing

Figure 6.3 shows the eigenvectors corresponding to PC1 (51%), PC2 (17%) and PC3

(9%). The highest eigenvectors in PC1 were observed in the NIR region around 2180

to 2300 nm wavelengths. This wavelength region is related to C-H and O-H

combinations, which indicated that the difference is caused by organic components in

the wine such as alcohol content, sugars, phenolic compounds, organic acids that

contribute to variations among the wines produced in different geographical origins.

The highset eigenvectors in PC2 were observed in the Vis region around 400 to 700

nm, related to wine pigments (colour). The highest eigenvectors in PC3 were

observed around Vis region (450 - 700 nm) and 2200 to 2300 nm, are related to wine

pigments and ethanol and phenolic compounds, respectively. These four absorption

regions explained most of the variance between the spectra of the samples, which may

also relate to differences arising from different geographical regions.

54

500 1000 1500 2000 2500

-0.2

-0.1

0.0

0.1

0.2

0.3

0.4

Eige

nvec

tors

Wavelenth (nm)

PC1 (51%) PC2 (17%) PC3 (9%)

Figure 6. 3 Eigenvectors of the three first principal components of Australian and

Spanish Tempranillo wines using Vis-NIR after SNV and second derivative processing

6.2.4 Discrimination analysis

6.2.4.1 Linear discrimination based on PCA scores

LDA was carried out using the PCA sample scores on PCs. Table 6.3 shows the LDA

classification according to geographical origin based on the first three PC scores of

PCA using raw spectra, which account for more than 98% of the variance of the

spectra data. A total of 45 (71%) samples were correctly classified. Table 6.4 lists the

classification result using the first three PCs of pre-processed spectra. A 76% of

correct classification was achieved. In this case, Australian wines were 70% correctly

classified, while 85% of the Spanish wines were correctly classified.

55

Table 6. 3 LDA classification results of Australian and Spanish Tempranillo wines using

Vis-NIR raw spectra based on the first 3 PCs (98% of the total variation)

Prediction

Australia Spain

Overall correct

classification

Australia (n=36) 26 (72%) 10

Spain (n=27) 8 19 (70%) 45 (71%)

Table 6. 4 LDA classification results of Australian and Spanish Tempranillo wines using

Vis-NIR pre-processed spectra based on the first 3 PCs (77% of the total variation)

Prediction

Australia Spain

Overall correct

classification

Australia (n=27) 25 (70%) 11

Spain (n=27) 4 23 (85%) 48 (76%)

The classification obtained from the pre-processed spectra was slightly better than

obtained from the raw spectra. However, when using the first three PCs of the pre-

processed spectra, only 77% of the total variation was explained. This implies that

20% less information was involved in the classification. Consequently, the first nine

PCs of the pre-processed data were included for LDA, which explained 95% variation

of the spectra (Table 6.5). These nine PCs improved the overall correct classification

rate to 90%, where 89% of Australian wine and 93% of Spanish wines were correctly

classified.

56

Table 6. 5 LDA classification results of Australian and Spanish Tempranillo wines using

Vis-NIR pre-processed spectra based on the first 9 PCs (95% of the total variation)

Prediction

Australia Spain Overall

Australia (n=27) 32 (89%) 4

Spain (n=27) 2 25 (93%) 57 (90%)

It was concluded that the pre-processed spectra achieved a better classification result

compared to that derived from the raw spectra. In another words, these spectra contain

more information than the raw spectra for geographical classification. The reasons for

this improvement may include: second derivative spectra resolve more peaks than the

raw spectra; and SNV and second derivative remove influence of baseline shifts and

improve the signal/noise ratio of the spectra.

6.2.4.2 DPLS classification

The DPLS model was developed using the Vis and NIR pre-processed spectra. Wine

samples were split randomly into calibration (n = 32) and validation sets (n= 31). The

validation set was used to evaluate the accuracy of the models to classify samples

according to the geographical origin. Figure 6.4 presents the score plot of the first two

PCs of the DPLS model using the calibration set. It is similar to the PCA score plot;

however the separation of wines according to the geographical origin was more

obvious than in the PCA. This is probably be explained by the fact that the DPLS

algorithm may maximise the variance between-groups rather than in the group

(Kemsley, 1996). The DPLS loadings for the calibration models were similar to those

observed in the PCA analysis (eigenvectors) (data not presented).

57

-0.002 -0.001 0.000 0.001 0.002-0.002

-0.001

0.000

0.001

AAAA

AAAAAAAA

AAAA

SSS S

SS

SS

S SSS

SSSS

SS

SS

SS

AA A A

AA

A A

AA

AA

AA

AA

AA

SSSS

AA S

PC

2

PC1

Figure 6. 4 Partial least squares score plot of the first two principal components of

Australian (A) and Spanish (S) Tempranillo wines using Vis-NIR pre-processed spectra

for the calibration set

The R2 and RMSECV in calibration were 0.95 and 0.16 (6 PLS latent variables),

respectively. The calibration statistics indicated that the model developed will be

acceptable to classify new samples. Table 6.6 shows the DPLS classification rate

(percent of classification) for the validation set according to geographical origin. The

DPLS yield an overall rate of correct classification of 93.5%. Wine samples

belonging to Australia were 100% correctly classified, while Spanish wines were

85.7% correctly classified.

58

Table 6. 6 Discriminant partial least squares (DPLS) classification results of Australian

and Spanish Tempranillo wines using Vis-NIR pre-processed spectra

Prediction

Australia Spain

Overall correct

classification

Australia 18 (100%) 0

Spain 2 (15.3%) 11 (84.7%) 29 (93.5%)

Summary

Both methods, DPLS and LDA, achieved overall correct classification rates exceeding

90%. The DPLS models achieved the highest rate of classification. These

discrimination results verified that differences existed between the wines from

different geographical origins and confirmed that the Vis-NIR spectra contain

information sufficient to discriminate between samples using these mathematical

techniques.

59

Chapter 7 Use of Visible and NIR to classify Riesling

wines based on geographical origins

7.1 Introduction

Riesling is the leading white grape variety of Germany’s noble wine (MacNeil 2001).

Wines made from Riesling are usually characterised of a light to medium body, floral

and fruity, and an implicit sweetness (Kolpan et al. 1996). Excellent dry Riesling

wines are also made in Alsace of France, Australia, and New Zealand. The objective

of this experiment was to explore the use of visible and near infrared spectroscopy to

analyze Riesling wines from Australia, New Zealand and Europe and to classify them

accordingly to their geographical origin (refer Table 7.1).

Table 7. 1 Vintage and origin of Riesling wine samples analyzed

2001 2002 2003 2004 2005 Total

Australia 4 9 8 21

New Zealand 2 10 12

Europe 2 3 8 4 17

7.2 Results and discussions

7.2.1 Chemical analysis

Table 7.2 shows the statistics of the chemical compositions of the Riesling wine

samples from different regions. No statistically significant differences were observed

for the mean values of alcohol content, total phenolics, and volatile acid in the set of

wine analysed grouped by geographical areas. Statistically significant differences

were observed in pH, TA and glucose plus fructose contents.

60

Table 7. 2 Statistics of chemical composition for Riesling wines from different

geographical region

Alcohol

(% v/v) pH

VA

(g/L)

TA

(g/L)

G+F

(g/L)

Total

Phenolics

(A.U.)

Mean 12.02A 3.01A 0.44A 7.32B 1.74B 8.51A

Min 11.06 2.84 0.38 6.14 0.00 6.83

Max 12.97 3.20 0.59 8.64 5.38 11.64

Australia

(n=21)

S.D. 0.54 0.13 0.06 0.65 1.94 1.18

Mean 11.52A 2.95A 0.49A 8.35A 17.10A 8.12A

Min 8.87 2.57 0.40 7.14 4.45 6.58

Max 13.19 3.25 0.67 10.81 56.49 12.18

New

Zealand

(n=12)

S.D. 1.40 0.21 0.11 1.43 20.00 1.96

Mean 11.86A 3.26B 0.46A 6.97B 12.40A 8.90A

Min 10.07 3.07 0.34 5.87 2.15 7.68

Max 13.19 3.76 0.81 7.96 41.44 11.91

Europe

(n=17)

S.D. 1.02 0.17 0.13 0.73 12.70 1.11

* A, B, Levels in columns not connected by the same letter are significantly different,

p<0.05

7.2.2 Spectra interpretation and analysis

Figures 7.1 and 7.2 present the raw and SNV and second derivative processed spectra

of the Riesling wines. No obvious differences were detected from visual observation

of the spectra between the wine samples from different geographical origin. All wines

possessed absorption bands at 1450 nm, related to O-H first overtone of both water

and ethanol (Osborne et al. 1993). Absorption regions were observed at 1690 nm

61

related to C-H stretch first overtones, and at 2268 nm, and 2306 nm, associated with

C-H combination and overtones (Burns and Ciurczak 2001). The absorption bands at

1790 and 2268 nm are believed associated with sucrose, fructose, and glucose in fruit

juices (Dambergs et al. 2002; Cozzolino et al. 2003).

500 1000 1500 2000 25000.0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

1790 nm

2306 nm

2268 nm

1950 nm

1690 nm

1450 nm

Riesling wine

log

(1/T

)

Wavelength (nm)

Figure 7. 1 Vis -NIR raw spectra of Riesling wines from Australia, New Zealand and

Europe

500 1000 1500 2000 2500-0.025

-0.020

-0.015

-0.010

-0.005

0.000

0.005

0.010

0.015

0.020

Sec

ond

deriv

ativ

e

Wavelength (nm)

Figure 7. 2 SNV and 2nd Derivative processed spectra of Riesling wines from Australia,

New Zealand and Europe

62

7.2.3 Principal component analysis

The pre-processed spectra of all wine samples were firstly analysed by PCA. Figure

7.3 is the score plot of the first three PCs. These PCs account for 97% of the variation

in the spectra. A general grouping was observed among the samples from different

regions; however, the separation was not very clear and an intense overlap occurred

around the centre of the 3D space. The overlapped samples were predominantly for

wines from New Zealand and Europe. Eigenvectors for the first three PCs were

investigated (Figure 7.4). PC1 explains 93% of the total variance in the samples’

spectra, with the highest eigenvectors occurring around 2250-2350 nm which is

related to C-H combinations and O-H stretch overtones. Eigenvectors also occurred

around 1400-1460 nm and around 1660-1760 nm region related to O-H first overtones

and C-H first overtones, respectively. The highest eigenvectors in PC2 (3%) appeared

around 1400-1460 nm and 2250-2350 nm. The highest eigenvectors in PC3 (2%) and

some eigenvectors in PC2 occurred in the visible region, around 410 – 540nm.

NZ

NZ

NZ

NZ

Aus

Aus

Aus

Aus

Eur

Eur

AusAus

Eur

Eur

EurEur

AusAus

AusAus

AusAus

AusAus

AusAus

AusAus

NZNZ

NZNZ

NZNZ

EurEur

AusAus

EurEur

Aus

NZ

NZ

EurEurEur

Eur

Eur

Eur

Eur

PC1

PC2

PC3

Figure 7. 3 Score plot of the first 3 principal components of Australian (Aus), New

Zealand (NZ) and European (Eur) Riesling wines using Vis-NIR pre-processed spectra

63

When analyzing all the samples from three regions together, overlapping was

observed. This may indicate the similarity among some samples, and also may be

caused by the sample matrix, with insufficient samples to present the pattern of their

region. To enhance the comparison, samples from geographical regions were analyzed

as pairs by PCA (e.g. samples from Australia versus Europe, Australia versus New

Zealand and New Zealand versus Europe).

Figures 7.5, 7.6 and 7.7 show the score plots of the first three PCs of the PCA of

samples for each pair. Clearer separations were observed between the wines from

Australia and Europe or New Zealand. However, poor separation was observed

between the samples from Europe and New Zealand.

500 1000 1500 2000 2500

-0.3

-0.2

-0.1

0.0

0.1

0.2

0.3

Eig

enve

ctor

s

Wavelength(nm)

PC1 (92%) PC2 (3%) PC3 (2%)

Figure 7. 4 Eigenvectors of the three first principal components of Australian, New

Zealand and European Riesling wines using Vis-NIR pre-processed spectra

64

AusAus

AusAus

Aus

Aus

EurEur

EurEur

AusAus

Aus

Aus

AusAus

AusAus

AusAus

AusAus

EurEur

Aus

Aus

EurEur

Aus

Eur

Eur

EurEur

Eur

EurEur

PC1

PC2

PC3

Figure 7. 5 Score plot of the first three principal components of Australian (Aus), and

European (Eur) Riesling wines using Vis-NIR pre-processed spectra

NZ

NZ

Aus

Aus

Aus

Aus

Aus

Aus

AusAus

Aus

Aus

AusAus

AusAus

AusAus

AusAus

NZ

NZ

NZNZNZ

NZ

Aus

Aus

Aus

NZ

NZ

PC1

PC2

PC3

Figure 7. 6 Score plot of the first three principal components of Australian (Aus) and

New Zealand (NZ) Riesling wines using Vis-NIR pre-processed spectra

65

NZ

NZ

NZ

NZEur

Eur

Eur

Eur

EurEur

NZNZ

NZNZ

NZ

NZ

EurEurEur

Eur

NZ NZ

Eur

EurEur

Eur

EurEur

Eur

PC1

PC2

PC3

Figure 7. 7 Score plot of the first three principal components of New Zealand (NZ) and

European (Eur) Riesling wines using Vis-NIR pre-processed spectra

7.2.4 Discrimination analysis

7.2.4.1 Linear discriminant analysis

LDA was performed on the first several PCs from the PCA result of the total samples.

The best classification rate was achieved by using the first three PCs, which account

for 97% variation of the spectral data. Table 7.3 shows the classification result

according to the wine provenance. Up to 72% of the total samples were correctly

classified. It was noticed that a similar classification rate, around 75%, was achieved

for the samples from Australia and New Zealand; however, a lower classification rate

(65%) was obtained for the samples from Europe, where several samples were

misclassified.

66

Table 7. 3 LDA classification results of Australian, New Zealand and European Riesling

wines using Vis-NIR pre-processed spectra

3 PCs Australia Europe New Zealand Overall correct

classification

Australia (n=21 ) 16 (76%) 2 (10%) 3 (14%)

Europe (n=17 ) 2 (11%) 11 (65%) 4 (24%)

New Zealand (n=12 ) 1 (8%) 2 (17%) 9 (75%)

36 (72%)

As displayed in the PCA score plots of wine samples from side-by-side comparisons

of geographical regions, Figures 7.5, 7.6 and 7.7, clearer groupings were observed.

LDA was performed based on PC scores of PCA. The number of PCs involved were

selected based on the best classification result obtained. The classification results

were summarized in Table 7.4. Most of the wines from Australia were correctly

classified as distinct from wines from Europe or New Zealand, 95% (Australia versus

Europe) and 86% (Australia versus New Zealand) respectively. It was more difficult

to discriminate between Riesling samples when comparing with wines from Europe

and New Zealand. An overall correct classification rate of 72% was achieved between

these samples. This comparatively low classification rate might indicate style

similarity of Riesling wines from New Zealand and Europe whilst wines from

Australia were different.

Table 7. 4 LDA classification results of each two regions of Australian, New Zealand and

European Riesling wines using vis-NIR pre-processed spectra

Number

PCs

Variance

Explained Sample geographical origins

Overall correct

classification

Australia Europe 3 94%

20 (95%) 13(76%) 33 (87%)

Australia New Zealand 3 96%

18 (86%) 9 (75%) 27 (82%)

Europe New Zealand 4 98%

11(65%) 10(83%) 21 (72%)

67

7.2.4.2 DPLS analysis

DPLS was employed to build a calibration models for classification of the samples

However, a disappointing result was obtained. Half of the samples were misclassified

and classification by DPLS was unsuccessful.

DPLS was applied to discriminate the samples based on side-by-side comparisons

between geographical regions. Table 7.5 lists classification rate of Rieslings from

these comparisons. Similar classification rates were obtained to those achieved from

LDA based on the PCA scores. Most of the wines from Australia were correctly

classified. However, wines from Europe and New Zealand remained difficult to

discriminate.

Table 7. 5 DPLS classification results of Australian, New Zealand and European Riesling

wines using vis-NIR pre-processed spectra

Sample geographical origins Overall correct

classification

Australia Europe

20 (95%) 12(71%) 32 (84%)

Australia New Zealand

21 (100%) 9 (75%) 30 (91%)

Europe New Zealand

11(65%) 7(58%) 18 (62%)

Summary

White wine classification was achieved between the samples from three geographical

origins. Wines from Australian were most easily classified. Greater than 86% of

samples from Australia were correctly classified using different multivariate analysis

methods when compared with samples from New Zealand or Europe. However, lower

classification rates were achieved between samples from New Zealand and Europe.

68

The style similarity of wines from New Zealand and Europe might explain the poor

classification result. Furthermore, the small sample number may affect the data matrix,

which may shape the result. To further test the ability of Vis-NIR to classify white

wine samples, a larger sample set is required.

69

Conclusion

Vis-NIR spectra of wine samples are affected by different factors of sample

presentation, such as sample temperature, optical path length and measurement mode.

For temperature, changes in the spectra were observed, with peak shifting and

absorbance increasing with temperature. The use of second derivative transformation

minimizes the effect of peak shifting in the NIR spectra due to temperature variation.

When PCA was performed on the wine spectra, temperature related changes on the

scores and eigenvectors were observed for both red and white wines.

In relation to the effect of measurement mode, transmittance versus transflectance,

variations in the spectra were observed. However, the absorption peaks of the spectra

appeared at the same wavelength regardless of scanning mode. When PCA was

performed on the wine spectra, measurement related changes on the scores and

eigenvectors were observed. The prediction of chemical composition using PLS

calibration models showed that the spectra acquired using transmission and

transflectance modes with similar pathlength produced equivalent prediction results.

However, longer pathlengths appeared to increase the standard error of cross

validation (SECV) and the coefficient of correlation, implying a lower prediction

accuracy.

It has been demonstrated that Vis-NIR spectroscopy combined with multivariate

analysis can be used as a classification tool to differentiate geographical origin of both

red and white wine samples.

Tempranillo wines from Australia and Spain, were classified using discriminant

partial least squares and linear discriminant analysis based on the PCA scores. The

models developed achieved an overall correct classification rates over 90%. Riesling

wine samples from three geographical origins were also correctly classified with

acceptable rates, over 75%.

The discrimination results demonstrated the differences between the wines from

70

different geographical origins and suggested that the Vis and NIR spectra (fingerprint)

store information able to discriminate among the wine samples.

Vis-NIR spectroscopy is a secondary method relying strongly on reference methods

during the modeling step. Therefore the calibration sample matrix should present as

much as possible the variability of the aimed feature. To employ this technique for

industry application with the objective of geographical classification of wines, further

research is recommended. The future work should expand the set of wine samples

analysed to build a “fingerprint data bank” which includes as many as representative

wines from specific or different regions to collect as much information as possible.

This will maximise the predictive reliability for geographical classification of the

method.

71

References

Arana, I., Jaran, C. and Arazuri, S., (2005). Maturity, Variety and Origin

Determination in White Grapes (Vitis Vinifera L.) Using Near Infrared Reflectance

Technology. Journal of Near Infrared Spectroscopy 13 349-357.

Arvanitoyannis, I. S., Katsota, M. N., Psarra, E. P., Soufleros, E. H. and Kallithraka,

S., (1999). Application of Quality Control Methods for Assessing Wine Authenticity:

Use of Multivariate Analysis. Trends in Food Science & Technology 10 321-336.

Barnes, R. J., Dhanoa, M. S. and Lister, S. J., (1989). Standard Normal Variate

Transformation and De-Trending of Near-Infrared Diffuse Reflectance Spectra.

Applied Spectroscopy 43(5), 772-777.

Baxter, M. J., Crews, H. M., Dennis, M. J., Goodall, I. and Anderson, D., (1997). The

Determination of the Authenticity of Wine from Its Trace Element Composition. Food

Chemistry 60(3), 443-450.

Bertran, E., Blanco, M., Coello, J., Iturriaga, H., Maspoch, S. and Montoliu, I., (2000).

Near Infrared Spectrometry and Pattern Recognition as Screening Methods for the

Authentication of Virgin Olive Oils of Very Close Geographical Origins. Journal of

Near Infrared Spectroscopy 8(1), 45-52.

Bianco, A. D., Benes, R., Trinkel, M., Reininger, F. and Leitner, M. (2000).

Temperature Influence on the 1.2um Band of Water. Near Infrared Spectroscopy: 9th

International Conference, Verona, Italy, NIR Publications, 81-87.

Bréas, O., Reniero, F., Serrini, G., Martin, G. J. and Rossmann, A., (1994). Isotope

Ratio Mass Spectrometry: Analysis of Wines from Different European Countries.

Rapid Communications in Mass Spectrometry 8(12), 967-970.

72

Brereton, R. G. (2003). Chemometrics. Chichester, John Wiley & Sons Ltd. 498 pp.

Brescia, M. A., Caldarola, V., De Giglio, A., Benedetti, D., Fanizzi, F. P. and Sacco, A.,

(2002). Characterization of the Geographical Origin of Italian Red Wines Based on

Traditional and Nuclear Magnetic Resonance Spectrometric Determinations.

Analytica Chimica Acta 458(1), 177-186.

Burns, D. A. and Ciurczak, E. W. (2001). Handbook of Near-Infrared Analysis. New

York, Marcel Dekker, Inc. 814 pp.

Charlton, A. J. and Farrington, W. H. B., P., (2002). Application of H-1 Nmr and

Multivariate Statistics for Screening Complex Mixtures: Quality Control and

Authenticity of Instant Coffee. Journal of Agriculture and Food Chemistry 50 3098-

3103.

Coetzee, P. and Vanhaecke, F., (2005). Classifying Wine According to Geographical

Origin Via Quadrupole-Based ICP-Mass Spectrometry Measurements of Boron

Isotope Ratios. Analytical and Bioanalytical Chemistry 383(6), 977-984.

Cozzolino, D., Kwiatkowski, M. J., Parker, M., Cynkar, W. U., Dambergs, R. G.,

Gishen, M. and Herderich, M. J., (2004). Prediction of Phenolic Compounds in Red

Wine Fermentations by Visible and Near Infrared Spectroscopy. Analytica Chimica

Acta 513(1), 73-80.

Cozzolino, D., Smyth, H. E., Lattey, K. A., Cynkar, W., Janik, L., Dambergs, R. G.,

Francis, I. L. and Gishen, M., (2005). Relationship between Sensory Analysis and

Near Infrared Spectroscopy in Australian Riesling and Chardonnay Wines. Analytica

Chimica Acta 539(1-2), 341-348.

Cozzolino, D., Vaz Martins, D. and Murray, I., (2002). Visible and Near Infrared

Spectroscopy of Beef Longissimus Dorsi Muscle as a Means of Discriminating

between Pasture and Corn Feeding Regimes. Journal of Near Infrared Spectroscopy

10 187-193.

73

Dambergs, R. G., Esler, M. and Gishen, M. (2004). Analysis of beverages and brewing

products. Near-Infrared Spectroscopy in Agriculture. Roberts, C. A., Workman, J. and

Reeves, J. B. Madison, American Society of Agronomy, Inc., Crop Science Society of

America, Inc., Soil Science Society of America, Inc.

Dambergs, R. G., Kambouris, A., Francis, I. L. and Gishen, M., (2002). Rapid

Analysis of Methanol in Grape-Derived Distillation Products Using Near-Infrared

Transmission Spectroscopy. Journal of Agriculture and Food Chemistry 50 3079-

3084.

Davies, A. M. C., Radovic, B., Tom, F. and Anklamb, E., (2002). A Preliminary Study

on the Characterisation of Honey by Near Infrared Spectroscopy. J. Near Infrared

Spectroscopy 10 121-135.

Day, M. P., Zhang, B. L. and Martin, G. J., (1995). Determination of the Geographical

Origin of Wine Using Joint Analysis of Elemental and Isotopic Composition. Journal

of the Science of Food and Agriculture 67(1), 113-123.

Frank, I. E. and Kowalski, B. R., (1984). Prediction of Wine Quality and Geographic

Origin from Chemical Measurements by Partial Least-Squares Regression Modelling.

Analytica Chimica Acta 162 241-251.

Frias, S., Conde, J. E., Rodriguez-Bencomo, J. J., Garcia-Montelongo, F. and Perez-

Trujillo, J. P., (2003). Classification of Commercial Wines from the Canary Islands

(Spain) by Chemometric Techniques Using Metallic Contents. Talanta 59(2), 335-344.

Godden, P. and Gishen, M., (2005). Trends in the Composition of Australian Wine.

Australian New Zealand Wine Industry Journal 20 21-46.

Gómez, M. d. M. C., Feldmann, I., Jakubowski, N. and Andersson, J. T., (2004).

Classification of German White Wines with Certified Brand of Origin by

Multielement Quantitation and Pattern Recognition. Journal of Agriculture and Food

Chemistry 52(10), 2962-2974.

74

Hageman, J. A., Westerhuis, J. A. and Smilde, A. K., (2005). Temperature Robust

Multivariate Calibration: An Overview of Methods for Dealing with Temperature

Influences on Near Infrared Spectra. Journal of Near Infrared Spectroscopy 13 53-62.

Horn, P., Schaaf, P., Holbach, B., HÃzl, S. and Eschnauer, H., (1993). 87Sr/86Sr from

Rock and Soil into Vine and Wine. European Food Research and Technology 196(5),

407-409.

Hruschka, W. R. (1992). Spectral Reconstruction. Handbook of Near-Infrared

Analysis. Burns, D. A. and Ciurczak, E. W. New York, M. Dekker, 401-418.

Iizuka, K. and Aishima, T., (1997). Soy Sauce Classification by Geographical Region

Based on Nir Spectra and Chemometrics Pattern Recognition. Journal of Food

Science 62(1), 101-104.

Jackson, R. S. (2000). Wine Science: Principles, Practice, Perception San Diego,

Academic. 648 pp.

Kallithraka, S., Arvanitoyannis, I. S., Kefalas, P., El-Zajouli, A., Soufleros, E. and

Psarra, E., (2001). Instrumental and Sensory Analysis of Greek Wines;

Implementation of Principal Component Analysis (PCA) for Classification According

to Geographical Origin. Food Chemistry 73(4), 501-514.

Kawano, S. (2002). Sample Presentations for near Infrared Analysis of Intact Fruits,

Single Grains, Vegetable Juice, Milk and Other Agricultural Products. Near Infrared

Spectroscopy: 10th International Conference, Kyongju, Korea, NIR Publications.

Kelly, S., Heaton, K. and Hoogewerff, J., (2005). Tracing the Geographical Origin of

Food: The Application of Multi-Element and Multi-Isotope Analysis. Trends in Food

Science & Technology 16(12), 555-567.

Kemsley, E. K., (1996). Discriminant Analysis of High-Dimensional Data: A

Comparison of Principal Components Analysis and Partial Least Squares Data

Reduction Methods. Chemometrics and Intelligent Laboratory Systems 33 47-61.

75

Kolpan, S., Smith, B. H. and Weiss, M. A. (1996). Exploring Wine: The Culinary

Institute of America's Complete Guide to Wines of the World. New York, Van Nostrand

Reihold. 699 pp.

Korenovska, M. and Suhaj, M., (2005). Identification of Some Slovakian and

European Wines Origin by the Use of Factor Analysis of Elemental Data. European

Food Research and Technology 221(3), 550-558.

Lockley, A. K. and Bardsley, R. G., (2000). DNA-Based Methods for Food

Authentication. Trends in Food Science & Technology 11(2), 67-77.

MacNeil, K. (2001). The Wine Bible. New York, Workman Publishing. 910 pp.

Massart, D. L., Vandenginste, B. G. M., Deming, S. N., Michotte, Y. and Kaufmann, L.

(1988). Chemometrics: A Textbook. Amsterdam, Elsevier, 488 pp.

McCurdy, E., Potter, D. and Medina, M., (1992). Trace Elements in Wines.

Laboratory News 9, 10-11.

Medrano, R., Yan, S. H., Maudoux, M., Baeten, V. and Meurens, M. (1995). Wine

Analysis by NIR. Leaping Ahead with Near Infrared Spectroscopy, NIR spectroscopy

group, Royal Australian Chemical Institute.

Murray, I. and Cowe (2004). Sample Preparation. Near-Infrared Spectroscopy in

Agriculture. Roberts, C. A., Workman, J. and Reeves, J. B. Madison, American

Society of Agronomy, Inc., Crop Science Society of America, Inc., Soil Science

Society of America, Inc.

Naes, T., Isaksson, T., Fearn, T. and T., D. (2002). A User-Friendly Guide to

Multivariate Calibration and Classification. Chichester, UK, NIR Publications, 344

pp.

76

Ogrinc, N., Kosir, I. J., Kocjancic, M. and Kidric, J., (2001). Determination of

Authenticy, Regional Origin, and Vintage of Slovenian Wines Using a Combination of

IRMS and SNIF-NMR Analyses. Journal of Agriculture and Food Chemistry 49(3),

1432-1440.

Osborne, B. G., Fearn, T. and Hindle, P. H. (1993). Practical NIR Spectroscopy with

Applications in Food and Beverage Analysis, 2nd Ed. Harlow, Essex, U.K., Longman

Scientific and Technical, 227 pp.

Otto, M. (1999). Chemometrics. Hemsbach, Germany, Wiley-VCH, 314 pp.

Palma, M. and Barroso, C. G., (2002). Application of FT-IR Spectroscopy to the

Characterisation and Classification of Wines, Brandies and Other Distilled Drinks.

Talanta 58(2), 265-271.

Pasquini, C., (2003). Near Infrared Spectroscopy: Fundamentals, Practical Aspects

and Analytical Applications. Journal of Brazil Chemistry Society 14(2), 198-219.

Pena-Neira, Hernandez, T., Garcia-Vallejo, C., Estrella, i. and Suarez, J. A., (2000). A

Survey of Phenolic Compounds in Spanish Wines of Different Geographical Origin.

European Food Research and Technology 210 445-448.

Pigott, S. (2004). Planet Wine. London, Mitchell Beazley, 155 pp.

Pillonel, L., Picque, S. and Tabacchi, B., (2003). Analytical Methods for the

Determination of the Geographic Origin of Emmental Cheese: Mid- and Near-

Infrared Spectroscopy. European Food Research and Technology 216(2), 174-178.

Rebolo, S., Pena, R. M., Latorre, M. J., Garcia, S., Botana, A. M. and Herrero, C.,

(2000). Characterisation of Galician (NW Spain) Ribeira Sacra Wines Using Pattern

Recognition Analysis. Analytica Chimica Acta 417(2), 211-220.

77

Reid, L. M., O'Donnell, C. P. and Downey, G., (2006). Recent Technological

Advances for the Determination of Food Authenticity. Trends in Food Science &

Technology 17(7), 344-353.

Reid, L. M., Woodcock, A., O'Donnell, C. P., Kelly, J. D. and Downey, G., (2005).

Differentiation of Apple Juice Samples on the Basis of Heat-Treatment and Variety

Using Chemometric Analysis of MIR and NIR Data. Food Research International 38

1109-1115.

Sauvage, L., Frank, D., Stearne, J. and Millikan, M. B., (2002). Trace Metal Studies of

Selected White Wines: An Alternative Approach. Analytica Chimica Acta 458 223-

230.

Sivertsen, H. K., Holen, B., Nicolaysen, F. and Risvik, E., (1999). Classification of

French Red Wines According to Their Geographical Origin by the Use of Multivariate

Analyses. Journal of the Science of Food and Agriculture 79(1), 107-115.

Somers, C., (1998). The Wine Spectrum: An Approach Towards Objective Definition

of Wine Quality, 135 pp.

Soufleros, E. H., Pissa, I., Petridis, D., Lygerakis, M., Mermelas, K., Boukouvalas, G.

and Tsimitakis, E., (2001). Instrumental Analysis of Volatile and Other Compounds of

Greek Kiwi Wine; Sensory Evaluation and Optimisation of Its Composition. Food

Chemistry 75(4), 487-500.

Swierenga, H., Wulfert, F., de Noord, O. E., de Weijer, A. P., Smilde, A. K. and

Buydens, L. M. C., (2000). Development of Robust Calibration Models in Near Infra-

Red Spectrometric Applications. Analytica Chimica Acta 411(1-2), 121-135.

Urbano-Cuadrado, M., de Castro, M. D. L., Perez-Juan, P. M., Garcia-Olmo, J. and

Gomez-Nieto, M. A., (2004). Near Infrared Reflectance, Spectroscopy and

Multivariate Analysis in Enology - Determination or Screening of Fifteen Parameters

in Different Types of Wines. Analytica Chimica Acta 527(1), 81-88.

78

Urbano-Cuadrado, M., Luque De Castro, M. D., Perez Juan, P. M. and Gomez-Nieto,

M. A., (2005). Comparison and Joint Use of Near Infrared Spectroscopy and Fourier

Transform Mid Infrared Spectroscopy for the Determination of Wine Parameters.

Talanta 66(1), 218-224.

Williams, P. and Norris, K. (2001). Near-Infrared Technology in the Agricultural and

Food Industries. St. Paul, American Association of Cereal Chemists, Inc.

Wulfert, F., Kok, W. T., de Noord, O. E. and Smide, A. K., (2000). Linear Techniques

to Correct for Temperature-Induced Spectral Variation in Multivariate Calibration.

Chemometrics and Intelligent Laboratory Systems 51 189-200.

Wulfert, F., Kok, W. T. and Smide, A. K., (1998). Influence of Temperature on

Vibrational Spectra and Consequences for the Predictive Ability of Multivariate

Models. Analytical Chemistry 70(9), 1761-1767.