Discrimination of stored cachaças using synchronous fluorescence and VIP-PLS-DA: an approach using PCA-oversampling

Storage in wooden barrels is a critical step in determining the sensory profile and commercial value of cachaça. However, monitoring the type of wood used requires analytical methods that are rapid, clean, and highly discriminatory. This study proposes a chemometric strategy based on Synchronous Fluorescence Spectroscopy (SFS), combined with PLS-DA and VIP-PLS-DA modeling, for the discrimination of cachaças aged in five types of wood: Amburana (A, n=38), Balsam (B, n=20), Oak (O, n=35), Freijó (F, n=38), and Jequitibá-rosa (J, n=20). The SFS spectra were acquired in the 300–600 nm range (1 nm intervals), with a scan speed of 2000 nm/min, a response time of 0.1 s, excitation and emission linewidths of 5 nm/5 nm, and low sensitivity. The shift interval between excitation and emission (∆λ) was previously optimized to 80 nm, which allows for maximum resolution of the phenolic fluorophores extracted from the wood samples. To construct the chemometric models, outliers were removed using F-residuals and Hotelling's T² limits. The spectra were evaluated under different preprocessing combinations, including baseline drift correction, linear baseline correction, and the first and second Savitzky-Golay derivatives (2nd-order polynomial and 15-point window). The split between the training set and the independent external test set was performed using the Kennard-Stone algorithm (75%/25%). To address sample imbalance and avoid predictive bias in the classifier, a 1:1 PCA-oversampling strategy was applied to the minority classes (B and J) within the training set. Exploratory PCA (93% total variance, PC1: 58%, PC2: 35%) showed distinct clustering for class O (positive PC1) and class A (positive PC2), whereas class F exhibited high dispersion, partially overlapping with class J. The PLS-DA model built using the full spectrum (300 variables) achieved 95% accuracy on the test set, indicating a limitation in discriminating class J (75% sensitivity, with 2 of the 8 samples classified as class F), reflecting the overlap observed in the PCA. To overcome this limitation and increase the model’s parsimony, variable selection based on importance in the projection (VIP > 0.8) was applied. The optimized VIP-PLS-DA model (17 latent variables) reduced the space to 59 variables (an 80% reduction) and eliminated non-informative noise. As a result, VIP-PLS-DA corrected the previous misclassifications and achieved 100% overall predictive accuracy on the test set, with sensitivity, specificity, and precision all equal to 100% across the five wood classes. The integrated approach based on the VIP-PLS-DA model proved to be an environmentally friendly, fast, highly parsimonious, and robust tool for industrial quality control and for identifying the distillate's aging profile.

Acknowledgments: The authors thank CAPES for the financial support and the research scholarship.