Journal of Near Infrared Spectroscopy

Two case studies were conducted to verify calibration model transfer methods without standards by multi-step wavelength selection, using 3–7 near infrared spectrometers to predict ingredients in corn and total plant alkaloids (TPA) in tobacco leaves. Based on the characteristic wavelengths of Uc, which are selected using the scale-invariant feature transform (SIFT), this study advances two multistep wavelength selection methods by selecting wavelengths with high independence and a high standard deviation of the sample spectra (SDSS). The first method, SIFT-SDSS-CORX, selects important characteristic wavelengths Uc-i from Uc whose SDSS is greater than a threshold SDSS^a_crit. Subsequently, rx, the correlation coefficient matrix between spectral signals of Uc-i, is calculated, and only one wavelength is retained from those whose correlation coefficients exceed a threshold, rx^a_crit. The wavelength set Uc-i-rx, which is finally screened, is important and independent. In the second method, SIFT-CORX-SDSS, Uc-rx is first selected from Uc by retaining only one wavelength from those whose correlation coefficients between spectral signals of Uc exceed a threshold, rx^b_crit. Subsequently, the wavelengths Uc-rx-i with SDSS exceeding a threshold SDSS^b_crit are selected from Uc-rx. Near infrared spectroscopy calibration models for predicting protein and oil in corn and TPA in tobacco leaves were built using partial least squares regression (PLS) based on different wavelength sets of Uc, Uc-i, Uc-i-rx, Uc-rx, and Uc-rx-i, respectively. The latent variables used in the PLS models were determined by an accumulative contribution ratio over 99.9%. The results indicate that the PLS models built on Uc-i-rx and Uc-rx-i are effective on both primary and secondary units for corn and tobacco samples. This study utilises a three-step wavelength selection method to select highly independent, important, and characteristic spectral variables, thereby enhancing the robustness, simplicity, and interpretability of NIR) calibration models and facilitating their transfer to secondary units without standards.

Partial least squares (PLS) regression is widely used to predict chemical analytes from spectroscopic data, thus reducing the need for expensive and time-consuming wet chemical reference analysis in industrial process monitoring. However, predictions via PLS by definition carry sample-specific errors, and estimation of these errors is essential for correct interpretation of results. To increase trust in PLS regression-based predictions, reliable prediction error estimates must be reported. This can be achieved by determining realistic sample-specific prediction errors using an unbiased mean squared prediction error estimate. This work provides a guide for estimating sample-specific prediction errors, showing the importance of choosing an appropriate error estimator prior to deploying PLS models for industrial applications. We reviewed recent and established methods for estimating the sample-specific prediction error and test them through simulation studies. The methods were subsequently applied for estimating prediction errors in two real-life datasets from the food ingredients industry, where near-infrared spectroscopy was used to quantify i) urea in process water and ii) individual protein concentrations in ultrafiltration retentates from a protein fractionation process. Both the simulations and real data examples showed that the mean squared error of calibration is always a downward biased estimator. Although leave-one-out-cross-validation performed surprisingly well in the data analysed in this work, this paper demonstrated that the appropriate choice of error estimator requires the user to make an informed, data-centered decision.

Near infrared (NIR) spectra contain information regarding the analyte as well as uninformative wavelengths. To build high-performance data-driven models, key wavelengths with a strong correlation to the analyte must be selected. This study proposes a feature selection method called stepwise Bayesian linear regression (SBLR) for eliminating unrelated wavelengths, thereby enhancing the robustness of the constructed model. First, a random wavelength is selected from an optimal variable set, and the other wavelengths are placed in a candidate variable set. A Bayesian linear regression (BLR) is implemented by adding a new variable from the candidate set or removing a variable from the optimal set in each step. Furthermore, the BLR model is utilized to perform the F-test. Comparing with the critical value of the F-test with a significance level of α, the test determines whether the variable is retained in the optimal set. Finally, the extracted variables are used to construct a BLR model. The performance and generalization ability of the proposed method were validated. The physical explanation of extracted wavelengths is consistent with the perspective of chemical analysis based on the experiment, which provides a good understanding of the collected NIR spectral data. In addition, compared with traditional algorithms, such as partial least squares regression, least absolute shrinkage and selection operator, and stepwise regression, the proposed method reserves only a few of the effective wavelengths from the full NIR spectra. The proposed method demonstrates potential for key wavelength selection in NIR spectroscopy.

Near infrared (NIR) spectroscopy has been used to analyze water structures due to the strong absorption of NIR energy by water. The spectral band around 6900 cm⁻¹, corresponding to the first overtone of the OH stretching vibration, is generally studied because the OH in the water molecule with different numbers of hydrogen bonds can be distinguished. In this work, the spectral band around 8600 cm⁻¹, corresponding to the combination of HOH bending and stretching vibration, ν₁+ν₂+ν₃, was studied to extract spectral information about water structures. Continuous wavelet transform was used to enhance the resolution of the spectra. Seven peaks related to the possible molecular structures of water with different numbers of hydrogen bonds were identified based on the spectral changes with temperature. The identification was validated by varying the spectral peaks with molar ratio of H₂O–D₂O in mixtures and the effect of hydration around the cations on the structure of water. NIR spectroscopy is therefore proven to be a powerful technique for identifying water structures with different hydrogen bonds.

The contamination of maize, a major staple food in South Africa, with fumonisin B1 (FB1), has become a major food safety concern. The regulation of this mycotoxin is extremely important and requires efficient detection methods. Near infrared (NIR) spectroscopy has gained widespread interest as a rapid and non-destructive mycotoxin analysis method. The purpose of this study was, therefore, to determine the NIR absorbance bands of FB1. The spectra of 30 FB1 solutions, constituted in methanol, as well as 30 methanol-only samples were recorded in the spectral range of 1000–2500 nm (10,000 – 4000 cm⁻¹). The data was pre-processed with multiplicative scatter correction (MSC) and a partial least squares discriminant analysis (PLS-DA) model was computed. The variable importance in projection (VIP) scores and selectivity ratio (SR) values were used for wavelength selection. A new PLS-DA model was computed with 454 chosen wavelengths and the regression vector of this model was investigated to further identify and remove irrelevant wavelengths. The final model was computed with 150 wavelengths and nine latent variables (LVs) and obtained a classification accuracy of 100% for both the calibration and external validation sets. By investigating the regression vector of the final PLS-DA model, potential FB1 absorbance bands were identified at 1446 nm, 1453 nm, 1891 nm, 2036 nm, 2046 nm, 2148 nm, 2224 nm, 2262 nm and 2273 nm. This study was therefore able to identify the previously unknown NIR absorbance bands of FB1 at 100 ppm.