A graphical method to evaluate spectral preprocessing in multivariate regression calibrations: example with...

Area of Science:

• Chemometrics
• Spectroscopy
• Data Analysis

Background:

• Spectral preprocessing is crucial in multivariate regression for spectroscopy data, aiming to reduce noise and enhance features.
• Common preprocessing techniques include smoothing and derivatives, which can improve calibration models but also introduce risks.
• Over-reliance on preprocessing can lead to inaccurate models, particularly when dealing with limited sample sizes or weak spectral responses.

Purpose of the Study:

• To develop and demonstrate a graphical method for evaluating the impact of spectral preprocessing on partial least squares (PLS) regression.
• To highlight the potential pitfalls of preprocessing in multivariate calibration using near-infrared (NIR) spectroscopy of wheat.

Main Methods:

• Applied PLS regression to NIR reflection spectra of ground wheat meal, analyzing protein content and SDS volume.
• Utilized Savitzky-Golay smoothing and first/second derivative preprocessing techniques.
• Incorporated an artificial component to demonstrate the risks of preprocessing.

Main Results:

• Preprocessing can be misleading when sample size is low (<50), analyte spectral response is weak, or model evaluation relies heavily on R(2) instead of residual error.
• The study identified specific conditions where preprocessing introduces significant calibration errors.
• A graphical method was developed to visually assess the influence of preprocessing on PLS models.

Conclusions:

• Excessive or inappropriate spectral preprocessing poses a significant risk in multivariate calibration, leading to unreliable models.
• The developed graphical method provides a valuable tool for assessing preprocessing effects and ensuring model robustness.
• This approach is applicable to various spectroscopy data and preprocessing functions, enhancing the reliability of chemometric analyses.