Abstract
SINTEF and Nofima have previously developed novel NIR instrument prototypes for non-contact, non-invasive, sub-surface quality measurements of complex food products based on chemical composition, including the benchtop high-performance SmartSensor and the more compact and portable MiniSmartSensor. The interactance geometry in these instruments relies on a certain degree of light scattering to transfer light from the field of illumination to the field of detection. In some applications, however, the NIR spectral variation for complex, structured, and highly scattering samples is dominated by physical properties rather than the chemical information. Such variations are commonly removed from the spectra through pre-processing, to emphasize the chemical information for analysis and minimize signal interference caused by production processes or environmental conditions. However, there are examples where physical variation correlates with the quality variable of interest and should not be removed, as doing so would reduce the accuracy of the prediction model.
Care must be taken to ensure that the inclusion of signal changes due to physical variations in the sample leads to robust and reliable prediction models. It is important to have a fundamental understanding of how the sample’s physical properties affect the measurements, how its optical properties relate to quality attributes, the mechanisms causing correlations with the variable of interest, and light-matter interaction. This helps us determine the limitations of the model, which is important for trustworthiness in commercial applications. It also provides a starting point for developing models that incorporate physical information in a robust way.
In our talk, we will present our approach for understanding NIR spectral measurements in complex, structured, highly scattering samples, using dried salt-cured cod (clipfish) as a case study. We measured the sample’s bulk optical properties (BOP), i.e. the scattering and absorption properties, using a double-integrating-sphere (DIS) setup and the inverse adding-doubling (IAD) algorithm. These lab measurements were compared with industry-suitable NIR interactance measurements to understand which variations are caused by the bulk optical properties and which can be explained by other influences. The corresponding light propagation model helped identify limitations and improve the robustness of the prediction model. Specifically, we investigated the effects and robustness of various optical geometries, gained an understanding of the mechanisms that cause correlations with quality attributes, and explored how sample properties of interest affected the signals and prediction model accuracy in relation to unwanted sources of variation. The results demonstrate how a fundamental understanding can improve the commercial robustness of spectroscopic measurement solutions for complex samples.
Care must be taken to ensure that the inclusion of signal changes due to physical variations in the sample leads to robust and reliable prediction models. It is important to have a fundamental understanding of how the sample’s physical properties affect the measurements, how its optical properties relate to quality attributes, the mechanisms causing correlations with the variable of interest, and light-matter interaction. This helps us determine the limitations of the model, which is important for trustworthiness in commercial applications. It also provides a starting point for developing models that incorporate physical information in a robust way.
In our talk, we will present our approach for understanding NIR spectral measurements in complex, structured, highly scattering samples, using dried salt-cured cod (clipfish) as a case study. We measured the sample’s bulk optical properties (BOP), i.e. the scattering and absorption properties, using a double-integrating-sphere (DIS) setup and the inverse adding-doubling (IAD) algorithm. These lab measurements were compared with industry-suitable NIR interactance measurements to understand which variations are caused by the bulk optical properties and which can be explained by other influences. The corresponding light propagation model helped identify limitations and improve the robustness of the prediction model. Specifically, we investigated the effects and robustness of various optical geometries, gained an understanding of the mechanisms that cause correlations with quality attributes, and explored how sample properties of interest affected the signals and prediction model accuracy in relation to unwanted sources of variation. The results demonstrate how a fundamental understanding can improve the commercial robustness of spectroscopic measurement solutions for complex samples.