Robust Wavelength Selection Using Filter-Wrapper Method and Input Scaling on Near Infrared Spectral Data.

Area of Science:

• Vibrational Spectroscopy
• Chemometrics
• Machine Learning for Spectroscopy

Background:

• Near Infrared Spectroscopy (NIRS) generates large datasets, posing challenges in extracting relevant wavelengths for accurate chemical analysis.
• Irrelevant wavelengths in multivariate calibration models, like Partial Least Square Regression (PLSR), reduce accuracy and robustness.
• Existing PLSR methods lack standardized procedures for effective wavelength selection.

Purpose of the Study:

• To introduce a novel, robust wavelength selection procedure, modified VIP-MCUVE (mod-VIP-MCUVE), for NIRS data analysis.
• To enhance model interpretability, computational efficiency, and predictive accuracy in chemometric models.
• To address the challenge of irrelevant wavelength inclusion in multivariate calibration.

Main Methods:

• Developed the modified VIP-MCUVE (mod-VIP-MCUVE) using a Filter-Wrapper approach and input scaling.
• Integrated modified Variable Importance in Projection (VIP) and modified Monte Carlo Uninformative Variable Elimination (MCUVE) for scale matrix calculation.
• Utilized orthogonal components of Partial Least Square (PLS) and robust reliability coefficients for variable selection.

Main Results:

• The proposed mod-VIP-MCUVE method demonstrated superior performance compared to classical VIP, MCUVE, and autoscaling in PLSR.
• Evaluated using artificial data via Monte Carlo simulation and real NIRS data from oil palm fruit mesocarp.
• The method successfully improved model interpretability and accuracy while being computationally efficient.

Conclusions:

• The mod-VIP-MCUVE procedure offers significant advantages for wavelength selection in NIRS data analysis.
• This method enhances the accuracy and interpretability of chemometric models derived from spectral data.
• It provides a robust solution for managing complex spectral datasets in vibrational spectroscopy.