Comparison of different approaches for the prediction of sugar content in new vintages of whole Port wine grape berries using hyperspectral imaging

Abstract

Two different approaches, PLS regression and neural networks, were compared for monitoring the quality of grapes using sugar content predictions based on hyperspectral imaging. The present work expands the result analysis and updates the state-of-the-art published in a conference article of the authors which concern the prediction of sugar content for vintages not used in model creation when the measured samples are composed of a small number of whole berries. This is highly innovative. The prediction models were established upon training under each approach and the generalization ability of both methodologies was determined through using n-fold-Cross-Validation and test sets. Sugar content was estimated using a model trained with spectra from samples of 2012. The test sets were composed of samples with six whole berries of 2012 or 2013 vintages.The results for PLS regression and Neural Networks for a test set with 2012 samples, were 0.94 °Brix and 0.96 °Brix for the root mean square error (RMSE), and 0.93 and 0.92 for squared correlation coefficients (R2), respectively, for each approach. When using test data containing 2013 samples, the RMSE values were 1.34 °Brix and 1.35 °Brix, and the R2 values were 0.95 and 0.92. These errors are competitive with those of works from other authors executed under less demanding conditions. The results obtained suggest that when combining hyperspectral imaging with appropriate chemometric techniques or machine learning algorithms, it is possible to have a satisfactory generalization for vintages not employed in model creation.