Wavelength Optimization Models For NIR Spectroscopy With Application To Analysis Of Zinc Content In Soil

With the chemometrics and computing technology advanced rapidly,near-infrared spectroscopy (NIR), which occupies the advantages of rapidness,non-destructive, and easy-online-analysis, has been successfully applied in manypractical fields. As one of the key NIR analysis techniques, wavelength optimizationis of great significance to improve the prediction effect, reduce computationalcomplexity, and design special spectral instrument. In this paper, NIR analysis forzinc content in soil was taken as an example. Two kinds of wavelength optimalmodels in NIR analysis, which were continuous mode and quasi-continuous mode,were discussed, and provided theoretical basis for soil heavy metal measurement. Thefollowing aspects have acquired corresponding achievements:1) Based on randomness, similarity, stability, a new NIR analysis modelingsystem, which includes calibration, prediction and validation processes, wasbuilt. And a stable, reliable mathematic model was established.2) Based on several divisions of calibration and prediction sets, moving windowpartial least squares regression (MWPLS) was improved and was morestable to establish the continuous wavelength optimal model for soil zinccontent analysis in NIR. The optimal waveband obtained from MW-PLSmodel was812-826nm. The corresponding PLS factor was6, and thevalidation root mean square error and validation correlation coefficients ofprediction were67.980mg kg-1and0.775, respectively.3) Adopted the equidistant combination multiple linear regression (EC-MLR)method that proposed by our team, the quasi-continuous wavelength optimalmodel for soil zinc content NIR analysis was established. The achievedoptimal wavelength combination of EC-MLR model was (1082,1086,1090,1094,1098,1102,1106). And the corresponding V-SEP and V-RPwere75.870mg kg-1and0.712, respectively.4) The equivalent model sets of optimal MW-PLS and EC-MLR models wereproposed, and the number of the corresponding equivalent models were29and11. Among the29equivalent wavebands of MW-PLS equivalent modelsets, waveband812-822nm was also included in other wavebands, whichindicated that it contained sufficient information with lowest computational complexity. The corresponding V-SEP and V-RPwere69.328mg kg-1,0.760. Among the11equivalent wavelength combinations of EC-MLRequivalent model sets, wavelength combination (1094,1098,1102,1106)employed the fewest number of wavelengths (N) which leads to lowestcomputational complexity. The corresponding V-SEP and V-RPwere76.520mg kg-1,0.713.The results indicate that the spectral predicted values and the measured valuesfrom reference method of soil zinc content show high correlation and consistency.The continuous mode and quasi-continuous mode wavelength optimal modelsproposed in this paper provide valuable references for soil zinc content NIR analysisand the design of special instrument.