Machine Learning Model Stability for Sub-Regional Classification of Barossa Valley Shiraz Wine Using A-TEEM...

Area of Science:

• Analytical Chemistry
• Food Science
• Machine Learning

Background:

• Wine fraud poses economic risks and damages regional reputation.
• Molecular fingerprinting using A-TEEM offers potential for wine authentication.
• Previous studies lacked data on A-TEEM model stability and wine blend classification.

Purpose of the Study:

• To assess the stability and application of A-TEEM models for wine classification over time.
• To develop and validate machine learning models for classifying Shiraz wines by vintage and Barossa Valley sub-region.
• To investigate the classification accuracy of wine blends using spectral fingerprinting.

Main Methods:

• Utilized absorbance-transmission and fluorescence excitation-emission matrix (A-TEEM) spectroscopy for molecular fingerprinting.
• Employed extreme gradient boosting discriminant analysis (XGBDA) for building classification models.
• Applied cross-validation and training/test set splitting for model evaluation on Shiraz wines from five sub-regions across four vintages.

Main Results:

• Achieved 100% cross-validation accuracy for vintage year and 98.8% for unknown sample prediction.
• Reached 99.5% cross-validation accuracy for sub-region and 93.8% for unknown sample prediction.
• Demonstrated 100% vintage prediction for recent vintages and high accuracy for sub-region prediction with new data, including successful classification of wine blends.

Conclusions:

• A-TEEM coupled with XGBDA provides a robust method for wine authentication and origin verification.
• The developed models exhibit stability and effectiveness in classifying wines by vintage, sub-region, and blend composition.
• This spectral fingerprinting approach supports data-driven terroir classification and combats wine fraud.