Authentication of Australian Red Wines Using Fluorescence Spectroscopy and Machine Learning Classification

Abstract

Verification of geographical origin, grape variety, and year of production of wine is essential in validating quality, identifying fraud, and improving the economic value of wine according to those important extrinsic factors. The identity of a wine is influenced mainly by its origin, as reflected in a wine’s composition. Therefore, analytical methods that identify authentication markers to discriminate wine according to the origin (or other variables) are required. Over the years, numerous methods for wine authentication have been identified, from traditional analytical methods to rapid advanced instrumental techniques. However, there is a lack of a robust but simple technique that gives rapid results and is sensitive enough to discriminate wines accurately. This forms the topic of the thesis, which begins with a published book chapter that covers current aspects of wine authenticity and traceability in terms of technological and consumer perspectives (Chapter 1). Different spectroscopic approaches and chemometric methods used in wine authentication in the past decades have been evaluated for their characteristics in the next chapter, published as a review paper (Chapter 2). As a rapid, straightforward, selective, and sensitive method that yields a molecular fingerprint of wine, fluorescence spectroscopy was identified upon reviewing the literature as a promising method to investigate wine authentication. Several original research studies were subsequently performed with the aim of understanding the potential of applying fluorescence spectroscopy in combination with multivariate data analysis for wine authentication (Chapters 3 to 6). Finally, the conclusions and future directions of the study are included in the final chapter (Chapter 7). In the initial research publication using spectrofluorometric analysis (Chapter 3), a method based on absorbance-transmission and fluorescence excitation-emission matrix (known as the A-TEEM technique) was investigated as a tool for regional authentication of commercial Australian Cabernet Sauvignon wines from three different Geographical Indications (GIs) in comparison to wines from Bordeaux, France as an international benchmark. The potential of A-TEEM spectroscopy for wine authentication was assessed in comparison to elemental profiling using inductively coupled plasma-mass spectrometry (ICP-MS) as a reference method for geographical authentication. Among other multivariate algorithms used for classification of the wines, a novel machine learning technique known as extreme gradient boosting discriminant analysis (XGBDA) yielded 100 % correct classification for all tested regions using the fluorescence data, and overall 97.7 % for ICP-MS. This result emphasised the possibility of applying A-TEEM and XGBDA for accurate authentication of wines. With these encouraging GI authentication results, a further study was undertaken to verify the origin of wine according to both geographical and varietal variations. A wide range of commercially-produced but unreleased wines from ten different Australian GIs and three varieties (Shiraz, Cabernet Sauvignon, and Merlot) were studied in the second research publication (Chapter 4). This study identified the effectiveness of combining absorbance and fluorescence data from A- TEEM as a multi-block data set to maximise the model’s robustness. Excellent results were obtained in relation to crossvalidation for each class (100 % for variety and 99.7 % for region of origin), again highlighting the effectiveness of A- TEEM data with XGBDA. In addition, ATEEM data was interrogated using partial least squares regression (PLSR) models to rapidly quantify 24 phenolic compounds of relevance to red wine (i.e., anthocyanins, flavonols, flavan-3-ols, hydroxycinnamates). Principal component analysis of the phenolic compound concentrations revealed differences among the varieties and regions, helping to understand the chemical markers that were important in classification. These findings further strengthen the potential of using the A-TEEM technique for differentiation of wine, not only from GIs at state level but also those from adjacent regions such as Clare, Barossa, and Eden Valleys within a state. Further testing the A-TEEM technique for its ability to discriminate wine at a sub-regional level, research-scale and commercial unreleased Shiraz wines from five different areas within the Barossa Valley GI along with Eden Valley GI were analysed to explore their intra-regional variations. The samples were from three consecutive years, which allowed for authentication testing according to the vintage, as reported in the third original research study submitted for publication (Chapter 5). The sensitivity of the A-TEEM technique allied with XGBDA facilitated 100 % accuracy in classifying Shiraz wines according to the sub-region of origin and year of production. Additionally, A-TEEM data were modelled with PLSR in comparison to reference method data to predict basic chemical parameters of the samples (i.e., pH, alcohol %v/v, titratable acidity), which enhances the utility of the A-TEEM technique as a rapid method for deployment in the wine industry. In wine authentication, it is important to understand the impact of winemaking processes on chemical markers at different stages of production. Hence, variations in molecular fingerprint of wines throughout the process such as after primary fermentation, after malolactic fermentation, and before blending were determined with the A-TEEM technique. XGBDA discriminated wines according to their origin (variety and region) with 100 % accuracy, eliminating the influence of stage of processing on spectral signature. Also, blending different grape varieties or wine from different GIs (as permitted by relevant regulations) is crucial in winemaking. However, it is important to determine whether blending a small proportion (up to 15 % of other varietal or regional wine as per Wine Australia regulations) can be detected for authentication purposes. Unreleased commercially-produced monovarietal wines were prepared with a series of blends containing Shiraz with Cabernet Sauvignon and Shiraz with Grenache and analysed with regression. XGB regression precisely predicted the percentage in the blend, achieving R2 CV of 1.00 and RMSECV of 0.00028 in comparison to PLSR, which did not perform as well. The results of this final study of the thesis were submitted for publication as a short communication (Chapter 6). In summary, this PhD thesis has been devoted to the development of a rapid analytical method to accurately authenticate wine according to geographical origin, variety, and vintage. The use of absorbance and/or fluorescence spectroscopy in conjunction with machine learning classification proved to be highly promising for this purpose. The outcomes of this thesis not only contribute to enriching scientific research but also offer opportunities for potential commercial application in the wine industry as a powerful tool for wine analysis, and in particular, validation of origin and composition.