Discriminant analysis of volatile compounds in wines obtained from different managements of vineyards obtained by e-nose

•A prototype device used to obtain volatile compounds to evaluate different beverage sample fingerprints.•Construction of a dataset with 28 different wine samples, obtained from different vineyard managements.•Development of a data processing methodology to assess the separability of these fingerprints.•Discrimination analysis by treatment, harvest and cultivar. Volatile organic compounds (VOCs) in wines are indicators commonly employed to evaluate the quality of the beverage. They are used with visual analyses (clarity, brightness, and color shade) and taste analyses (balance, acidity, and taste clarity according to its fruity or woody notes, among others). However, to perform critical analyses and identify the characteristics of wines, sommeliers and oenologists must have well-honed sensory skills. Different studies analyze olfactory parameters using technical analyses, such as sensory and gas chromatography. Some of those analyses are subjective since there is a high variation of compounds in cultivars, crops, and cultural management; others are expensive and not affordable to small winemakers. Thus, electronic noses (e-noses) are an alternative to evaluate wines; they are relatively easy to implement and adapt to any experiment and also provide efficient results. This study analyzed the performance of a prototype of an e-nose, composed of 13 sensors, to classify 28 samples of wine. The samples were from two consecutive seasons (harvests of 2017/2018 and 2018/2019) of two cultivars (Cabernet-Sauvignon and Merlot). The berries used in this study received seven treatments to increase the quality of the grapes and then evaluate the chemical and phenolic quality of the wines. After acquiring the signals and generating the database, pre-processing steps were applied to adjust the data and extract the signal features. The entire set was divided into training and test data so that, in the processing stage, different classifiers were applied again to evaluate the separability of the samples in terms of treatments, cultivars and harvests. Different classification methods were evaluated (including variations of KNN (k-Nearest Neighbors); SVM (Support Vector Machine), and RF (Random Forest)) in addition to dimensionality reduction techniques (Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA)). The results showed accuracy rates greater than 94 % in the separability of the 28 classes of wine samples using the RF classifier imputed with LDA, w