|
|
Thanks to the multi co-linearity of the spectra, NIRS spectra can be compressed very readily. It can be easily demonstrated that just a few factors are sufficient to determine the differences between the spectra. It can be shown that often 10 factors are sufficient to illustrate 99% of the variance in the spectra and to explain them. Instead of 700 data items for each sample, only 10 data items are then needed per sample.
Other methods which have been proposed to compress spectra are the Fourier and wavelet transformation. The basis of the Fourier transformation is recognition of the fact that curve traces can be represented mathematically by an infinite series of sine and cosine functions. This Fourier transformation in sine and cosine functions forms the original signal or the original function respectively from what is referred to as the time domain into the frequency domain.
The compression of the spectra is based essentially on the multi co-linearity of the spectra and on the assumption that the slight residual variability in the spectra is to be interpreted as the noise of the spectra. With all the procedures referred to, the transformation is effected by the variability in the spectra (example: Fourier transformation) or between the spectra (example: PCA factors).
Minor nuances in the spectrum are in this case allocated to the noise, and the differentiation between information and noise in the spectra is not materially specified, but to be determined on the basis of statistical tests.
The capability of various different methods to achieve spectral compression has not proved successful in providing loss-free compression for most practical work, due to the absence of generally-valid criteria. | |
|