Old advice on multivariate calibration: still in force, but not always followed. A tutorial

Area of Science:

• Chemometrics
• Spectroscopy
• Machine Learning

Background:

• Deep learning is increasingly used in multivariate calibration, often overlooking simpler methods like partial least-squares (PLS) regression.
• There's a lack of analysis on spectral preprocessing necessity, statistical comparison of prediction errors, and reference technique uncertainty impact.

Purpose of the Study:

• To compare deep learning and PLS regression for spectral property estimation.
• To analyze the necessity of spectral preprocessing and proper error comparison methods.
• To evaluate the impact of reference technique uncertainty on model performance.

Main Methods:

• Analysis of a publicly available dataset for dry matter content estimation in fruits using UV-Vis-NIR spectra.
• Comparison of global and local PLS models against literature shallow and deep learning methods.
• Evaluation of prediction errors relative to reference technique uncertainty.

Main Results:

• Local PLS models demonstrated superior performance compared to deep learning methods in terms of simplicity and accuracy.
• Prediction errors from PLS models were found to be close to the uncertainty of the reference technique.
• The study highlights the potential preference for simpler models in specific multivariate calibration tasks.

Conclusions:

• Adopting parsimonious concepts in multivariate calibration enhances robustness and reliability.
• Prediction errors should align with the uncertainty of reference values for credible calibration.
• Simpler models like PLS can be highly effective and should be considered alongside complex methods.