Machine learning for classifying and predicting grape maturity indices using absorbance and fluorescence spectra

Claire E J Armstrong; Adam M Gilmore; Paul K Boss; Vinay Pagay; David W Jeffery

doi:10.1016/j.foodchem.2022.134321

Machine learning for classifying and predicting grape maturity indices using absorbance and fluorescence spectra

Food Chem. 2023 Mar 1:403:134321. doi: 10.1016/j.foodchem.2022.134321. Epub 2022 Sep 20.

Authors

Claire E J Armstrong¹, Adam M Gilmore², Paul K Boss³, Vinay Pagay¹, David W Jeffery⁴

Affiliations

¹ Australian Research Council Training Centre for Innovative Wine Production, The University of Adelaide, PMB 1, Glen Osmond, South Australia 5064, Australia; School of Agriculture, Food and Wine, and Waite Research Institute, The University of Adelaide, PMB 1, Glen Osmond, South Australia 5064, Australia.
² HORIBA Instruments Inc., 20 Knightsbridge Road, Piscataway, NJ 08854, United States.
³ Australian Research Council Training Centre for Innovative Wine Production, The University of Adelaide, PMB 1, Glen Osmond, South Australia 5064, Australia; CSIRO Agriculture and Food, Locked Bag 2, Glen Osmond, SA 5064, Australia.
⁴ Australian Research Council Training Centre for Innovative Wine Production, The University of Adelaide, PMB 1, Glen Osmond, South Australia 5064, Australia; School of Agriculture, Food and Wine, and Waite Research Institute, The University of Adelaide, PMB 1, Glen Osmond, South Australia 5064, Australia. Electronic address: david.jeffery@adelaide.edu.au.

PMID: 36191418
DOI: 10.1016/j.foodchem.2022.134321

Abstract

Absorbance-transmission and fluorescence excitation-emission matrix (A-TEEM) spectroscopy was investigated as a rapid method for predicting maturity indices using Cabernet Sauvignon grapes produced under four viticulture treatments during two growing seasons. Machine learning models were developed with fused spectral data to predict 3-isobutyl-2-methoxypyrazine (IBMP), pH, total tannins (Tannin), total soluble solids (TSS), and malic and tartaric acids based on the results from traditional analysis methods. Extreme gradient boosting (XGB) regression yielded R² values of 0.92-0.96 for IBMP, malic acid, pH, and TSS for externally validated (Test) models, with partial least squares regression being superior for TSS prediction (R² = 0.97). R² values of 0.64-0.81 were achieved with either approach for tartaric acid and Tannin predictions. Classification of grape maturity, defined by quantile ranges for red colour, IBMP, malic acid, and TSS, was investigated using XGB discriminant analysis, providing an average of 78 % correctly classified samples for the Test model.

Keywords: A-TEEM; Chemometrics; Data fusion; Discriminant analysis; Regression; XGBoost.

MeSH terms

Machine Learning
Tannins / analysis
Vitis* / chemistry
Wine* / analysis

Substances

malic acid
Tannins