The Characteristic Spectral Selection Method Based on Forward and Backward Interval Partial Least Squares.

Area of Science:

• Analytical Chemistry
• Spectroscopy

Background:

• Near-infrared spectroscopy (NIRS) commonly uses Forward Interval Partial Least Squares (FiPLS) and Backward Interval Partial Least Squares (BiPLS) for wavelength selection.
• These methods, while accurate, suffer from greedy search limitations, potentially selecting suboptimal intervals for analyte information.

Purpose of the Study:

• To propose a novel spectral characteristic intervals selection strategy, FB-iPLS, that combines FiPLS and BiPLS.
• To improve the robustness and predictive performance of NIRS models by addressing the limitations of existing interval selection methods.

Main Methods:

• Developed the FB-iPLS strategy by segmenting spectra and alternatively selecting useful intervals with FiPLS and deleting useless intervals with BiPLS.
• Applied FiPLS, BiPLS, and FB-iPLS to determine ethanol concentration in pure water.
• Evaluated model performance across different interval sizes.

Main Results:

• The FB-iPLS method demonstrated superior prediction performance, particularly when the spectrum was divided into 60 segments.
• FB-iPLS achieved high correlation coefficients (r = 0.9677 for calibration, r = 0.9670 for validation) and low cross-validation root mean square errors (RMSECV = 0.0888 for calibration, RMSECV = 0.0571 for validation).
• The proposed method outperformed both FiPLS and BiPLS in overall predictive performance.

Conclusions:

• The FB-iPLS strategy effectively overcomes the greedy search limitations inherent in FiPLS and BiPLS.
• This two-way selection approach enhances the efficiency and representativeness of characteristic interval selection, leading to improved model predictive performance in NIRS.