134
K. M. Sørensen et al.
single-seed sample cassette, and transmittance spectra were recorded in the range
850–1050 nm in 2 nm steps. A tungsten lamp (50 W) and a diffraction grating were
used to create monochromatic light. The light passed through the kernel, reaching
the silicon detector. The time required for scanning (single scans) 23 single kernels
in the cassette was about 90s. The dataset is available for download at http://food.
ku.dk/foodomics.
The dataset is assembled from 7 specimens of 5 kernel varieties. Each of these 35
seeds has been measured in two alternative orientations in the carrousel—with the
germ facing either up or down—each with four alternative positions (front, left, back
and right); the different measurements are made by four times stopping the carrousel
and manually changing the same 35 kernels. The dataset composition can thus be
written as (in total, 280 NIR samples):
varieties (5) × individuals (7) × positions (4) × orientations (2)
(7.3)
Unfortunately, 6 kernel samples were measured using a faulty carousel well and
we thus have only 264 samples. After recording the spectra of the intact wheat kernels,
each one was crushed, and the single kernel protein content was determined directly
by a modified Kjeldahl method [9].
7.2 Spectral Inspection and Pre-processing
Before any spectral exploration, regression or classification, it is of
fundamental importance to visualize and understand the quantitative
aspects of the recorded spectra
NIR spectra consist of a complex superimposition of linear concentration effects
and several different nonlinear contributions such as intermolecular interactions, light
scattering by particles, surfaces and phase transitions. In theory, these are highly
complex phenomena, but in practice they can be removed by simple pre-processing
techniques.
A necessary and implicit pre-processing step for spectral data, automatically
applied in most spectrometer software, is the linearization of chemical concentration
in the measured spectra to absorbance through application of the Lambert–Beer law
on the amount of light transmitted through the sample:
A λ = −log 10 (T λ ) = ε λ · l · c
(7.4)
where A λ is the absorbance at wavelength λ, T λ is ratio of light transmitted through the
sample (a.k.a. transmittance) at wavelength λ, ε λ is the wavelength-dependent molar
absorptivity for the chemical constituent of interest, l is the effective path length of
K. M. Sørensen et al.
single-seed sample cassette, and transmittance spectra were recorded in the range
850–1050 nm in 2 nm steps. A tungsten lamp (50 W) and a diffraction grating were
used to create monochromatic light. The light passed through the kernel, reaching
the silicon detector. The time required for scanning (single scans) 23 single kernels
in the cassette was about 90s. The dataset is available for download at http://food.
ku.dk/foodomics.
The dataset is assembled from 7 specimens of 5 kernel varieties. Each of these 35
seeds has been measured in two alternative orientations in the carrousel—with the
germ facing either up or down—each with four alternative positions (front, left, back
and right); the different measurements are made by four times stopping the carrousel
and manually changing the same 35 kernels. The dataset composition can thus be
written as (in total, 280 NIR samples):
varieties (5) × individuals (7) × positions (4) × orientations (2)
(7.3)
Unfortunately, 6 kernel samples were measured using a faulty carousel well and
we thus have only 264 samples. After recording the spectra of the intact wheat kernels,
each one was crushed, and the single kernel protein content was determined directly
by a modified Kjeldahl method [9].
7.2 Spectral Inspection and Pre-processing
Before any spectral exploration, regression or classification, it is of
fundamental importance to visualize and understand the quantitative
aspects of the recorded spectra
NIR spectra consist of a complex superimposition of linear concentration effects
and several different nonlinear contributions such as intermolecular interactions, light
scattering by particles, surfaces and phase transitions. In theory, these are highly
complex phenomena, but in practice they can be removed by simple pre-processing
techniques.
A necessary and implicit pre-processing step for spectral data, automatically
applied in most spectrometer software, is the linearization of chemical concentration
in the measured spectra to absorbance through application of the Lambert–Beer law
on the amount of light transmitted through the sample:
A λ = −log 10 (T λ ) = ε λ · l · c
(7.4)
where A λ is the absorbance at wavelength λ, T λ is ratio of light transmitted through the
sample (a.k.a. transmittance) at wavelength λ, ε λ is the wavelength-dependent molar
absorptivity for the chemical constituent of interest, l is the effective path length of