Characterization of neural network generalization in the determination of pH and anthocyanin content of wine grape in new vintages and varieties

Abstract

The generalization ability of hyperspectral imaging combined with neural networks (NN) in estimating pH and anthocyanin content during ripening was evaluated for vintages and varieties not employed in the NN creation. A NN, from a previously published work, trained with grape samples of Touriga Franca (TF) variety harvested in 2012 was tested with TF from 2013 and two new varieties, Touriga Nacional (TN) and Tinta Barroca (TB) from 2013. Each sample contained a small number of whole berries. The present work results suggest that, under certain conditions, it might be possible for the NN to provide for new vintages and varieties results comparable to those of the vintages and varieties employed in the NN training. For pH, the results are state-of-the-art for the new vintage and varieties tested. For anthocyanin, generalization is bad for TB from 2013 but presents state-of-the-art absolute percentage error for TF and TN from 2013.

Keywords: Grape berries; Hyperspectral imaging; Neural networks; Prediction; Wine quality.