Modelling Cabernet-Sauvignon wine sensory traits from spectrofluorometric data

Abstract

Understanding how wine compositional traits can be related to sensory profiles is an important and ongoing challenge. Enhancing knowledge in this area could assist producers to select practices that deliver wines of the desired style and sensory specifications. This work reports the use of spectrofluorometry in conjunction with chemometrics for prediction, correlation, and classification based on sensory descriptors obtained using a rate-all-that-apply sensory assessment of Cabernet-Sauvignon wines (n = 26). Sensory results were first subjected to agglomerative hierarchical cluster analysis, which separated the wines into five clusters represented by different sensory profiles. The clusters were modelled in conjunction with excitation-emission matrix (EEM) data from fluorescence measurements using extreme gradient boosting discriminant analysis. This machine learning technique was able to classify the wines into the pre-defined sensory clusters with 100 % accuracy. Parallel factor analysis of the EEMs identified four main fluorophore components that were tentatively assigned as catechins, phenolic aldehydes, anthocyanins, and resveratrol (C1, C2, C3, and C4, respectively). Association of these four components with different sensory descriptors was possible through multiple factor analysis, with C1 relating to ‘dark fruits’ and ‘savoury’, C2 with ‘barnyard’, C3 with ‘cooked vegetables’ and ‘vanilla/chocolate’, and C4 with ‘barnyard’ and a lack of C1 descriptors. Partial least squares regression modelling was undertaken with EEM data and sensory results, with a model for perceived astringency being able to predict the panel scores with 68.1 % accuracy. These encouraging outcomes pave the way for further studies that relate sensory traits to fluorescence data and move research closer to the ultimate goal of predicting wine sensory expression from a small number of compositional factors.

Predicting Bordeaux red wine origins and vintages from raw gas chromatograms

“…Given our strong performance with estate recognition, artificial and GC-based systems might be able to complement human tasters on wine recognition. We note that other groups have obtained promising results using chemical analysis to predict Rate-All-That-Apply sensory attributes, using techniques that would complement the GC approach presented here 26,27 .…”

Use of Machine Learning with Fused Spectral Data for Prediction of Product Sensory Characteristics: The Case of Grape to Wine

“…The second dataset involved feature-level data fusion of A-TEEM data and CIELAB colour coordinates using principal component analysis (PCA). A-TEEM and CIELAB datasets were chosen because fluorescent fingerprints and colour matrices have been shown to contain valuable information for the prediction of sensory characteristics in studies specifically investigating wine sensory predictions [13,26,27] as well as other food products [8]. Furthermore, fluorescence and colour data can be readily obtained from an Aqualog spectrophotometer via a simple and quick procedure for spectral data acquisition.…”

“…Fluorescence spectroscopy, which yields a molecular fingerprint of a sample, is particularly appealing because of its high sensitivity compared to other spectroscopic methods [9]. Using a spectrophotometer originally employed for water quality analysis, the absorbance-transmission and fluorescence excitation-emission matrix (A-TEEM) method [10] offers a powerful option, being used recently in conjunction with chemometrics and machine learning for wine authentication, modelling the sensory and chemical characteristics of wine, and chemotyping of cannabis [11][12][13][14][15][16]. A-TEEM spectral fingerprints are based on compounds containing a fluorophore or chromophore (i.e., UV and visible wavelengths), which in the case of foods and beverages includes aromatic amino acids, phenolic compounds, vitamins, and porphyrins (e.g., chlorophyll) [17,18].…”

“…Seeking to improve on the previous research with an innovative approach, a gradientboosted decision tree algorithm known as XGBoost [29] was deemed to be a suitable machine learning method that could be used with spectral data for the prediction of wine sensory properties. XGBoost is a non-linear algorithm and has previously been used to good effect for classification and regression modelling [11][12][13][14][15][16]. Thus, the aim of the present work was to conduct a proof-of-concept study involving the novel application of A-TEEM analysis coupled with XGBoost for the prediction of ratings (scores) for a diverse range of wine sensory attributes from five sensory stimuli-colour, aroma, flavour, taste, and mouthfeel-using Cabernet Sauvignon grape extract spectra collected from different regions over three vintages.…”