A review of different dimensionality reduction methods for the prediction of sugar content from hyperspectral images of wine grape berries

Abstract

Several dimensionality reduction techniques were applied to hyperspectral reflectance images of wine grape berries, leading a study of the machine learning models’ efficiency in the prediction of sugar content for training, validation and independent test sets, and for generalization sets with vintages not previously seen in the training phase. The results obtained across all settings were up to par, either matching or improving state-of-the-art results, and showcasing that a model capable of generalizing predictions from one vintage year to another without further training is achievable in a very accurate way. For the dimensionality reduction techniques studied, the results show that the PCA outperforms the nonlinear techniques for the case of real-world hyperspectral data while also suggesting that, for the case of hyperspectral images of wine grape berries, local nonlinear techniques more frequently have a better performance than their global nonlinear counterparts. This review highlights that more complex methods for dimensionality reduction may not be necessary for the case of hyperspectral images, since the PCA still yields the best results when using the transformed dataset for the prediction of oenological parameters.