Effective hyperspectral band selection and multispectral sensing based data reduction and applications in food analysis

This thesis focuses on the development of new band selection algorithms and sensing technology to accommodate for the growing demand in commercial and industrial hyperspectral imaging applications. Hyperspectral imaging is inherently complex as it combines two-dimensional spatial image information with highly resolved spectral measurements in a three-dimensional data structure. With rising dimensionality of the data, the required complexity of mathematical models increases exponentially. Dimensionality reduction in the form of band selection is analysed in this thesis, as it preserves physical interpretability of the original data. Two separate approaches are developed. The first one defines an optimisation heuristic for a criterion for band subsets carrying most information and least redundancies. The second approach embeds the identification of the most relevant bands into the reconstruction of the whole dataset. Both approaches successfully reduce the data amount while maintaining the classification accuracy. This thesis further develops a prototype for a cost-effective, portable and easy-to-use hyperspectral imaging system. RGB LED based time-multiplexed illumination enables rapid acquisition of multi-channel images that are subsequently used to reconstruct spectral signatures from hyperspectral prior via a multilayer perceptron. It is shown that a very accurate reconstruction is possible for a limited number of desired signatures that are sufficiently distinct. Lastly, the thesis focuses on industrial case studies of hyperspectral imaging in the context of food quality monitoring. The capabilities of SWIR and UV imaging were compared to estimate the smokiness of Scotch Whisky from the concentration of phenolic flavour compounds on barley malt used for distillation. UV imaging shows potential to estimate the concentration but is in its current state of technological development not ready for implementation, as opposed to SWIR, which due to its maturity, has the potential to be implemented in a production environment. Problems arising when imaging food products are examined on the example of salmon fillets.This thesis focuses on the development of new band selection algorithms and sensing technology to accommodate for the growing demand in commercial and industrial hyperspectral imaging applications. Hyperspectral imaging is inherently complex as it combines two-dimensional spatial image information with highly resolved spectral measurements in a three-dimensional data structure. With rising dimensionality of the data, the required complexity of mathematical models increases exponentially. Dimensionality reduction in the form of band selection is analysed in this thesis, as it preserves physical interpretability of the original data. Two separate approaches are developed. The first one defines an optimisation heuristic for a criterion for band subsets carrying most information and least redundancies. The second approach embeds the identification of the most relevant bands into the reconstruction of the whole dataset. Both approaches successfully reduce the data amount while maintaining the classification accuracy. This thesis further develops a prototype for a cost-effective, portable and easy-to-use hyperspectral imaging system. RGB LED based time-multiplexed illumination enables rapid acquisition of multi-channel images that are subsequently used to reconstruct spectral signatures from hyperspectral prior via a multilayer perceptron. It is shown that a very accurate reconstruction is possible for a limited number of desired signatures that are sufficiently distinct. Lastly, the thesis focuses on industrial case studies of hyperspectral imaging in the context of food quality monitoring. The capabilities of SWIR and UV imaging were compared to estimate the smokiness of Scotch Whisky from the concentration of phenolic flavour compounds on barley malt used for distillation. UV imaging shows potential to estimate the concentration but is in its current state of technological development not ready for implementation, as opposed to SWIR, which due to its maturity, has the potential to be implemented in a production environment. Problems arising when imaging food products are examined on the example of salmon fillets