Detection of smoke contamination in grapevine canopies and berries using infrared thermography, near infrared spectroscopy, and machine learning modelling

Abstract

Bush fires are becoming more frequent, severe and extensive due to changing climate and intentional fires. When bush fires happen to be close to vineyards, especially around veraison, smoke contamination of grapevines and grapes can occur, which is passed to the winemaking process producing smoke-taint in wines. At the moment, detection of contaminated fruit can be achieved with tedious and expensive laboratory analysis. This research proposed a non-invasive detection system for smoke contaminants in berries from seven grapevine cultivars using Near Infrared spectroscopy (NIR) and machine learning models. For this purpose, grapes were collected from non-smoked and smoked grapevines within an experiment performed at The University of Adelaide. All grapes collected were measured as whole and halves using a portable near infrared spectrometer (Field Spec Pro). Models for smoke contaminant detection were constructed using artificial neural networks (ANN) for pattern recognition of changes in the transformed (second-derivative) spectra from berries within the 700 – 1100 nm range. The ANN models obtained for whole berries from seven cultivars had 89.3% accuracy in the detection of smoked berries compared to only 61.6% accuracy for the model found for half berries. These results are consistent to levels of smoke-taint compounds found in different berry organs with higher concentration in the skin compared to pulp and seeds. The smoke detection method proposed, and single model developed for seven cultivars, can offer grape growers a quick, affordable, accurate, non-destructive and in-field screening tool that can be used for differential harvest of fruit and respective winemaking.