292
B. Igne et al.
• Optimizing the existing model (preprocessing, variable range, number of latent
variables) to accommodate for the new sources of variability
• Update the calibration set to add a limited number of samples representing the
missing sources of variability
• Re-build the model
In situations where the process is transferred to a different facility and/or analytical instrument, ensuring that the model performance before and after transfer is
practically equivalent is critical. This can be done as part of method development by
including samples from multiple instruments in the calibration, test and validation
sets (robust models across units) or through the use of calibration transfer methods
[21]. If calibration transfer cannot be achieved (resulting error on the secondary unit
is too high), model re-build may be necessary.
12.5 Additional Considerations for Multipoint Systems
All the concepts discussed in this chapter are applicable to any NIR spectroscopic
system. However, some additional considerations should be noted for multipoint
(imaging) systems. Without discussing the details of how images are constructed,
NIR imaging systems usually consist of arrays of sensing elements that can be
considered as juxtaposed single point spectrometers. There are numerous NIR
systems where the sample is measured at multiple points, but the resulting light is
usually directed toward a single detector channel. On the contrary, imaging systems
or specialized systems such as those employed in spatially resolved spectroscopy
will detect light coming back to different detector elements, as being measured at
different locations on the sample. This spatial information is key to imaging but also
brings some additional points to consider regarding sample membership and how to
construct a representative calibration dataset.
A single-point spectrometer assumes that the collected spectrum is representative
of the area interrogated (scale of scrutiny). When collecting the equivalent sample for
reference analysis, care is taken to ensure that the volume sampled for reference analysis corresponds to the variability measured by the NIR measurement. In multipoint
systems, this assumption is not correct. While there will be a degree of correlation between neighboring sampling points, homogeneity of the sample cannot be
assumed at the scale of measurement, particularly when using microscopic or wideangle lenses. It is actually the desired intent: to investigate the distribution of physical
and/or chemical properties within the area of interest via spectral information from
pixel to pixel (or sampling location to sampling location).
What makes imaging an ideal tool for the investigation of sample spatial distribution also makes it a challenge for model development. In many cases, it is no longer
correct to assume that the percentage of light reflected by a 99% reflectance standard
is actually homogeneous across the detector array; it is no longer correct to assume
that a wavelength standard will have the same composition at all pixel locations; and
B. Igne et al.
• Optimizing the existing model (preprocessing, variable range, number of latent
variables) to accommodate for the new sources of variability
• Update the calibration set to add a limited number of samples representing the
missing sources of variability
• Re-build the model
In situations where the process is transferred to a different facility and/or analytical instrument, ensuring that the model performance before and after transfer is
practically equivalent is critical. This can be done as part of method development by
including samples from multiple instruments in the calibration, test and validation
sets (robust models across units) or through the use of calibration transfer methods
[21]. If calibration transfer cannot be achieved (resulting error on the secondary unit
is too high), model re-build may be necessary.
12.5 Additional Considerations for Multipoint Systems
All the concepts discussed in this chapter are applicable to any NIR spectroscopic
system. However, some additional considerations should be noted for multipoint
(imaging) systems. Without discussing the details of how images are constructed,
NIR imaging systems usually consist of arrays of sensing elements that can be
considered as juxtaposed single point spectrometers. There are numerous NIR
systems where the sample is measured at multiple points, but the resulting light is
usually directed toward a single detector channel. On the contrary, imaging systems
or specialized systems such as those employed in spatially resolved spectroscopy
will detect light coming back to different detector elements, as being measured at
different locations on the sample. This spatial information is key to imaging but also
brings some additional points to consider regarding sample membership and how to
construct a representative calibration dataset.
A single-point spectrometer assumes that the collected spectrum is representative
of the area interrogated (scale of scrutiny). When collecting the equivalent sample for
reference analysis, care is taken to ensure that the volume sampled for reference analysis corresponds to the variability measured by the NIR measurement. In multipoint
systems, this assumption is not correct. While there will be a degree of correlation between neighboring sampling points, homogeneity of the sample cannot be
assumed at the scale of measurement, particularly when using microscopic or wideangle lenses. It is actually the desired intent: to investigate the distribution of physical
and/or chemical properties within the area of interest via spectral information from
pixel to pixel (or sampling location to sampling location).
What makes imaging an ideal tool for the investigation of sample spatial distribution also makes it a challenge for model development. In many cases, it is no longer
correct to assume that the percentage of light reflected by a 99% reflectance standard
is actually homogeneous across the detector array; it is no longer correct to assume
that a wavelength standard will have the same composition at all pixel locations; and