Non-linear calibration models for near infrared spectroscopy.

Area of Science:

• Spectroscopy
• Chemometrics
• Machine Learning

Background:

• Spectroscopic applications often exhibit nonlinear behavior, necessitating advanced calibration techniques.
• Traditional linear methods like Partial Least Squares (PLS) regression may not capture complex relationships.
• Evaluating nonlinear calibration models is crucial for accurate data analysis.

Purpose of the Study:

• To comprehensively compare various nonlinear calibration techniques for spectroscopic data.
• To assess the performance of Kernel PLS (KPLS), SVM, LS-SVM, RVM, GPR, ANN, and BANN.
• To benchmark against linear methods like PLS and Ridge Regression (RR).

Main Methods:

• Comparative analysis of KPLS, SVM, LS-SVM, RVM, GPR, ANN, and BANN.
• Utilized three real-life near-infrared (NIR) spectroscopic datasets.
• Evaluated methods based on computational time, interpretability, overfitting, and robustness to dataset size and preprocessing.

Main Results:

• Gaussian Process Regression (GPR) and Bayesian ANN (BANN) demonstrated strong performance for both linear and nonlinear systems.
• Least-Squares SVM (LS-SVM) exhibited attractive predictive capabilities across linear and nonlinear calibrations.
• GPR and BANN proved effective even with moderately small datasets.

Conclusions:

• GPR and BANN are powerful and promising methods for spectroscopic calibration, handling both linear and nonlinear data.
• LS-SVM is a valuable alternative due to its good predictive performance.
• The study provides insights into the practical application and comparative strengths of various nonlinear calibration techniques.