The use of hyperspectral imaging in the visible and near infrared region to discriminate between table grapes harvested at different times

Traditional analytical methods applied to the measurement of grape maturity and quality index in order to assess optimal harvest time have been proved to be slow and destructive. Therefore, non-destructive analytical techniques, including spectroscopy, can be a valid support for the choice of the best time to harvest. This study evaluated the feasibility of using a visible and near infrared spectral scanner (v. 1.4; DV Srl, Padova, Italy) with a detector in the region between 400-1000 nm to discriminate between grapes harvested at different times. Twelve clusters were harvested at 5 different times between October and December 2011. Spectra were acquired with a Spectral scanner on 3 intact berries from each bunch. These were randomly selected from top, medium and bottom zones, for a total of 180 spectra. Classification models were construed comparing 2 methods: soft independent modelling of class analogy (SIMCA) and partial least squares discriminant analysis (PLS-DA). The SIMCA model was developed building individual principal component analysis (PCA) models for the spectra of each harvest time. Different pre-treatment methods were tested in order to enhance the power of the model, thus enhancing the score differences among samples from different harvest times. The transformation that allowed the best statistical separation among scores of grapes from different harvest times was the second derivate of Norris. Therefore, the PCA model obtained from the spectra subjected to this pre-treatment was used for SIMCA classification. The PLS-DA model were developed applying the PLS2 algorithm. In order to construct discriminant models to classify bunch spectra according to the 5 harvest times, spectral variations were correlated with the 5 categories established. No pretreatments were previously applied in this last case since they did not improve the final result. The SIMCA method was unable to correctly classify grapes from harvest time 2 (59% of correct classification) and was less efficient compared to the PLS-DA model. Using the PLS-DA model, all the grapes were correctly classified (100%) with the exception of those from harvest time 5 (94%). The overall results demonstrate that this method has excellent potential for discriminating grape quality.