Reproduce “Data Augmentation of Spectral Data for Convolutional Neural Network (CNN) Based Deep Chemometrics” on OpenResearch

About this paper

Esben Jannik Bjerrum, Mads Glahder, Thomas Skov

Convolutional Neural Networks (CNNs) were applied to Near-Infrared (NIR) spectra for quantitative drug content prediction, integrating a novel spectral data augmentation technique. The approach achieved superior prediction accuracy, particularly in extrapolation scenarios, compared to Partial Least Squares models, while also revealing how CNN kernels extract relevant spectral features.

Key insights

• CNNs, particularly when enhanced with spectral data augmentation, can achieve predictive performance superior to or on par with traditional PLS models for quantitative chemometric analysis, especially in robust extrapolation scenarios.
• The combination of spectral data augmentation and Extended Multiplicative Scatter Correction (EMSC) provides the best CNN performance, suggesting that the CNN learns robustness to subtle EMSC residuals while benefiting from its overall corrections.
• Qualitative analysis of CNN kernel activations revealed that these networks can intrinsically learn signal processing operations akin to smoothing, differentiation, thresholding, and spectral region selection, potentially reducing the need for manual, expert-driven preprocessing.

Summary by alphaXiv.