Towards robust Machine Learning models for grape ripeness assessment

Artificial intelligence methods need to be more transparent for wider acceptance by the industry. In particular deep neural networks (DNN) are not explainable, due to the complex processes the input undergo. The present work addresses model explainability for wine grapes quality assessment through 1D-CNN, using regression activation maps (RAM) to show the contribution score of each wavelength for the prediction of sugar content. This way we identify the relevant regions related to this enological parameter. The results obtained indicate that the proposed approach can successfully highlight important spectral regions related to sugars absorption, improving the current state of the art, and opening way to dimensionality reduction methods and further model interpretation.