A comparative analysis of deep learning and chemometric approaches for spectral data modeling

Abstract

Background:

This study presents a comprehensive comparison of five different modeling approaches for spectroscopic data analysis. The first approach uses PLS combined with classical chemometric pre-processing (9 models). The second and third approaches use interval PLS (iPLS) with either classical pre-processing or wavelet transforms (28 models). The fourth uses LASSO with wavelet transforms (5 models). Finally, the fifth approach combines CNN with spectral pre-processing (9 models).

Results:

In this paper we consider two low dimensional case studies: a regression problem for a beer dataset (40 training samples) and a classification problem for a waste lubricant oil dataset (273 training samples). The results show that, after exhaustive pre-processing selection, iPLS variants show better performance for the first case study and remain competitive in the second case study. Wavelet transforms proved to be a viable alternative to classical pre-processing, improving performance for both linear and CNN models while maintaining interpretability. For the second case study, with more data, CNNs present good performance when applied on raw spectra and could potentially be used to avoid exhaustive pre-processing selection. However, it was found that CNNs can benefit from some types of pre-processing, leading to improved performance for the first case study and overall better performance for the second case study.

Significance and Novelty:

This study provides a critical and exhaustive comparison of combinations of pre-processing methods and models for spectroscopic data analysis. It was found that no single combination of pre-processing and model that can be identified as optimal beforehand in low data settings.