Evaluating the generalization ability of deep learning models: An application on sugar content estimation from hyperspectral images of wine grape berries

The assessment of grape ripeness is an extremely important factor in winemaking and has a direct impact on wine quality. This process is usually carried out with a traditional laboratory analysis, a costly procedure that destroys the grapes selected for analysis. Consequently, the research in precision viticulture has shifted focus to the development of digital processes that are fast and non-intrusive. In this context, the use of hyperspectral imaging paired with prediction models for the estimation of oenological parameters has gained wide recognition. The major drawback of these solutions is the extreme variability presented by the data, aligned with a small number of samples for training, derived from the high cost of acquiring new samples infield. Achieving a satisfactory generalization capacity while working on small data sets with such high variability is a serious challenge, and in this work we aim to provide a pipeline on how to properly build validation and test sets that allow for a correct evaluation of performance, avoiding common misconceptions such as using the R2 metric for model selection or creating models based on biased data sets. Additionally, we implement and evaluate different architectures, namely Residual Networks, InceptionTime and MiniRocket, to showcase that deep learning techniques can accurately predict sugar content from different vintages and varieties of wine grape berries, maintaining a strong generalization capacity even in a setting of high variability and small number of samples. Finally, our results also suggest that adding more relevant features to better characterize the data might be enough for the networks to adjust and produce accurate estimates of sugar content, which would eliminate the necessity to capture new samples on a yearly basis.