Use of spectral window preprocessing for selecting near-infrared reflectance wavelengths for determination of the degree of enzymatic retting of intact flax stems

A near-infrared reflectance (NIRR) method suitable for air-dried intact flax stems was developed for determination of Fried's test degree-of-retting scores. Performance statistics for preferred wavelength sets from 2 to 12 wavelengths are reported, and 7-, 9-, or 12-wavelength models are recommended. The wavelength sets are taken from the 1432- to 2468-nm spectral range. Calibration samples were prepared by retting flax stems for various lengths of time in an enzyme/chelator solution. Spectral properties of specimens were measured under varying conditions of hydration, stem orientation, and optical geometry, and calibration models were constructed to be insensitive to these effects. Root mean squared model errors were estimated by full cross-validation (RMSECV) that was modified to ensure the independence of each test sample relative to its calibration set. For the 12-wavelength model, the correlation (R-2) between predicted and measured is 0.946 and RMSECV is 0.20 for smosthed Fried's test scores spanning the four visual score levels, 0 to 3, from under-to over-retted. The 12-wavelength NIRR method has lower error than Fried's test scores produced by one observer in a single-blind experiment, where the root mean squared repeatability error is estimated to be 0.25. Furthermore, even the smallest wavelength set has excellent success in classifying enzyme-retted flax stems as either under- or over-retted. Wavelength sets were determined by a novel procedure that first located representative spectral windows by testing the performance of all possible spectral ranges, and then selected the optimum combinations of wavelengths from each window. It is shown that this procedure was able to find multiple linear regression wavelength models that substantially outperform full-spectral partial least squares regression models of similar dimensionality, and furthermore, that the method facilitates observation of the most relevant spectral variations. (C) 2000 Elsevier Science B.V. All rights reserved.