(25) Optimized Noise Reduction in Extended Derivative Spectroscopy for NIR
Calibration
¡¡¡¡¡¡of Rice Constituents ¡Ý a Preliminary Study Proceedings of Japan-Korea
Joint Symposium
¡¡¡¡¡¡on Near Infrared Spectroscopy,pp.195-200,2006-6
¡¡Differentiation processing is sometimes used as effective pretreatment
of NIR spectra. However, there are several drawbacks in ordinary derivative
spectra; spectral peaks are shifted in first derivatives and inversed in
second derivatives, not only weak absorption peaks but feeble noises are
also enhanced, undesired side lobes are generated as by-produces, and the
type of derivatives is practically limited only to those of first and second
orders. To overcome these drawbacks and limitations, the conventional derivative
has been modified on the basis of Fourier transform approach of numerical
differentiation to define two types of extended derivatives with fractional
order: fractional derivatives (FD) and fractional absolute derivatives (FAD).
Although the extended derivatives provide us with a solution to the above
problems, suitable reduction of noise is still necessary for the new derivatives
to work in practice. In the present paper, the reduction of noise in
extended derivatives are discussed in the context of the calibration of rice
constituents by means of PLS regressions. The reduction of noise is performed
by a Gaussian filter in the Fourier domain of original spectra. It is shown that
the width of the Gaussian filter should be optimized to yield best calibration
result, depending upon target constituents and the fractional order of derivatives.