Application of New Variable Selection Techniques to near Infrared Spectroscopy

Publisher Website
Google Scholar
Abstract
Two non-traditional variable selection techniques, classification and regression trees (CART) and genetic algorithms (GA), were explored for their application to near infrared spectroscopic calibrations. The results were compared to those of multiple linear regression (MLR) and partial least square (PLS) calibrations. Both numerical comparisons and interpretation of the reasons for the wavelength choices of the different techniques were made. A challenging set of mixtures, containing a low level of an alcohol with a spectrum very similar to one of the major components, was used as a test for the various techniques. The genetic algorithm approach succeeded in locating three wavelengths, which together were capable of generating a model which predicted unknown mixtures very well.