Answers are grounded in this paper and cite their sources.
Deep learning models have emerged as a promising solution for complex, non-linear chemometric problems. However, training deep models typically requires large datasets, which are often difficult to obtain in chemometrics. In deep learning, a popular solution to artificially augment the size of the training dataset and encourage generalization capabilities is data augmentation. Data augmentation consists in generating modified copies of existing samples during training. Although data augmentation methods for spectroscopic data exist, they are based on disparate principles and the lack of a unifying theoretical framework limits their practical applicability. Motivated by these limitations, this work proposes a new generic data augmentation framework based on the well-established formalism of Vicinal Risk Minimization (VRM). It exploits classic chemometric preprocessing transformations to quantify and analyze spectral distortions, then, generates vicinal samples by sampling within these distortions, ensuring a match with original reference values. Implementations of this framework using additive transformations are evaluated for the prediction of mango dry matter content from Vis-NIR spectra using a convolutional neural network. The results show that these implementations can improve the generalization capabilities of the model and that data augmentation seems to be a better paradigm than preprocessing for training deep learning models. Comparing these implementations to another existing data augmentation method shows the benefits of the VRM formalism. While promising, the obtained results do not yet reach state-of-the-art performance on this dataset, leaving room for further research on spectral data augmentation, advanced deep learning architectures and foundation models tailored to spectral data.
No takes yet. Share an insight, caveat, or question.
Tumoine et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: