[Isomap-PLS nonlinear modeling method for near infrared spectroscopy]

Area of Science:

• Chemometrics
• Machine Learning
• Spectroscopy

Context:

• Near-infrared (NIR) spectroscopy generates high-dimensional data.
• Modeling the nonlinear relationship between NIR spectra and sample properties is challenging.
• Existing methods like Partial Least Squares (PLS) primarily address linear correlations.

Purpose:

• To introduce a novel nonlinear modeling method, Isomap-PLS, for NIR spectra analysis.
• To leverage Isomap's manifold learning capabilities for nonlinear dimension reduction.
• To combine Isomap with PLS for enhanced calibration model development.

Summary:

• Isomap extracts intrinsic nonlinear structures from high-dimensional NIR spectra.
• PLS is then applied to reduce linear redundancy and build a calibration model.
• Parameter optimization for Isomap was performed using a grid search approach.

Impact:

• Isomap-PLS demonstrated superior performance over PLS on two benchmark NIR datasets.
• Rooted mean square error of cross-validation (RMSECV) was significantly reduced (2-5 fold for some properties).
• The method effectively models nonlinear correlations, enhancing calibration and prediction power.