Optimal partner wavelength combination method with application to near-infrared spectroscopic analysis

For a complex analyte, appropriate wavelength selection for NIR spectroscopy is necessary and difficult to improve prediction effectiveness, reduce model complexity, and design specialized spectrometer. In this study, a novel approach for wavelengths combination selection, named optimal partner wavelength combination combined with partial least squares (OPWC-PLS), was proposed. NIR spectroscopic analysis of soil organic matter (OM) was taken as an example to evaluate the performance of proposed OPWC-PLS. Additionally, moving window partial least squares (MW-PLS), which is a well-performed and PLS-based wavelength selection method, was also performed for comparison.
A total of 114 soil samples of the same type were collected and then ground after drying. Their OM contents were measured with standard potassium dichromate (K2Cr2O7) oxidation soil analysis method. Spectra were measured using an XDS Rapid ContentTM Grating Spectrometer with a diffuse reflection accessory and 780-2498nm region (2-nm interval). Root-mean-square errors and correlation coefficients of prediction (SECV, RP,CV) for leave-one-out cross validation (LOOCV) were used to evaluate modeling performance. The number of PLS factors (F) was set as 1-20.
For OPWC-PLS, firstly, by finding the optimal partner wavelength of each wavelength based on binary linear regression, a wavelength subset was screened from any wavelength set, which was called partner wavelength subset (PWS). Here, a PWS with 117 wavelengths for the entire region (780-2498nm) was selected by min SECV. Secondly, a new PWS for the first PWS was screened with the same procedure, and the same procedure was performed repeatedly. In this study, the second PWS with 61 wavelengths was screened, and the PWS stopped shrinking after 16 times of the same procedure. The final PWS only included 14 wavelengths (1366, 1370, 2154, 2156, 2192, 2194, 2204, 2218, 2222, 2224, 2232, 2282, 2288, 2320 nm) and was called OPWC, whose PWS was just itself. SECV and RP,CV for OPWC-PLS model were 0.165gkg-1 and 0.967, respectively. While SECV and RP,CV for entire region PLS model (780-2498nm) were 0.250gkg-1 and 0.954, respectively. Obviously, prediction effect and model complexity for OPWC-PLS were both better. In fact, it can be interesting proved that, both experimentally and theoretically, PWS stopped shrinking after limited times of the same procedure, and OPWC formed several loops according to the attribution direction for partner wavelengths. Here, seven loops for the 14 wavelengths were 1366⇆1370, 2154⇆2156, 2192⇆2282, 2194⇆2288, 2204⇆2320, 2218⇆2232 and 2222⇆2224.
For MW-PLS, the parameters included initial wavelength (I), number of wavelengths (N), and number of PLS factors (F), and were set as 780-2498 for I, 1-860 for N, and 1-20 for F. PLS model was established for each combination (I, N, F), and the parameters were selected by min SECV. The selected waveband was 1794-2212nm with 210 wavelengths, and SECV and RP,CV were 0.163gkg-1 and 0.968, respectively. Prediction effects of OPWC-PLS and MW-PLS were almost the same. However, OPWC had obvious lower model complexity.
The results indicate that the proposed OPWC has great performance for wavelength selection. Moreover, OPWC is also feasible to couple with the other multivariate analysis methods.