Determination of sugar content in whole Port Wine grape berries combining hyperspectral imaging with neural networks methodologies

The potential of hyperspectral imaging combined with machine learning algorithms to measure sugar content of whole grape berries is presented, as a starting point for developing generalized and flexible frameworks to estimate enological parameters in wine grape berries. In this context, to evaluate the generalization ability of the used machine learning procedure, two neural networks were trained with different training data to compare the performance of each one when tested with the same data set. Six whole grape berries were used for each sample to draw the hyperspectral spectrum in reflectance mode between 308 and 1028 nm. The sugar content was estimated from the spectra using feedforward multiplayer perceptrons in two different neural networks trained each one with a data set from a different year (2012 & 2013); the validation for both neural networks was done by n-fold cross-validation, and the test set used was from 2013. The test set revealed R
²
values of 0.906 and RMSE of 1.165 °Brix for the neural network trained with 2012 data and R
²
of 0.959 and RMSE of 1.026 °Brix for the 2013 training data neural network. The results obtained indicate that both neural networks present good results and that the 2012 training data neural network exhibits a good performance when compared with the other NN, suggesting that the approach is robust since a generalization (without further training) over years may be obtainable.