Advancing grape chemical analysis through machine learning and multi-sensor spectroscopy

Abstract

Chemical testing of grapes is essential for wine-making, allowing informed decisions about grape composition and potential wine characteristics. However, current invasive laboratory methods to identify the key components as total soluble solid contents and acidity, present challenges in terms of time and cost. To address these issues, non-invasive techniques using Visible-Near-Infrared (Vis-NIR) and Raman spectroscopy are explored for rapid and accurate grape chemical analysis. Various machine learning methods are deployed in this study to address the challenges of grape chemical analysis and enhance estimation accuracy. The contributions of this research include pioneering a direct comparison between Vis-NIR and Raman spectroscopy, establishing a regression model benchmark, and providing an open-source dataset for grape composition analysis. We identify Gaussian Process Regression (GPR) and Support Vector Machine Regression (SVMR) as the most effective regression models, with GPR achieving an RMSE of 0.977 °Birx for sugar content estimation in French Colombard grapes and SVMR achieving 0.780 °Brix for Cabernet grapes. For pH estimations, GPR and SVMR also perform exceptionally well.

Keywords

• Non-invasive sensing
• Visible-near-infrared spectroscopy
• Raman spectroscopy
• Regression
• Chemical composition