Feature Extraction From Spectroscopy Using LASSO and Net Analyte Signal

Although spectral techniques have been widely used as noninvasive analytical tools in various applications, spectral disturbances impair their performance in quantitative analyses. A new approach for effective signal extraction in spectroscopy is proposed to suppress spectral variations that are not related to the component of interest. First, the collected spectral data were pre-treated to eliminate the noise. The least absolute shrinkage and selection operator (LASSO) was then introduced to select wavelengths that were truly essential to the model, in which the shrinkage parameter was determined using a cross-validation method. The net analyte signal (NAS) was extracted from the selected wavelengths to eliminate the spectrum related to the interfering components using the rank annihilation method. A partial least squares regression (PLSR) model was constructed using the extracted NAS. The protein estimate of corn demonstrated that the extracted effective signal incorporated with PLSR outperforms existing methods with acceptable computational complexity. The results show that the proposed approach achieves the extraction of effective spectroscopic signals.