Pattern recognition methods for the analysis of infrared imaging data and multivariate calibration standardization for near-infrared spectroscopy

Automated analysis methods for infrared imaging data were investigated in two applications. The first application involved a hyperspectral Fourier transform infrared spectroscopic imaging system. The objective was to differentiate human breast cells in normal and cancer states based on infrared imaging measurements combined with pattern recognition techniques. The Fourier transform imaging system consisted of a step-scan interferometer coupled to an infrared microscope. A mercury cadmium telluride detector with 64 x 64 array format was used. Cluster analysis and artificial neural networks were employed to develop automated classifiers for determining cell and non-cell pixels within the image and for predicting the disease state of the cell pixels. Satisfactory results were obtained and the results show that this approach provides a promising tool to aid the pathologist in the differentiation of normal and diseased human tissues.; The second application involved a multispectral remote sensing imaging system mounted on an aircraft. The objective was to develop automated methods for the remote detection of atmospheric chemical species from airborne multispectral infrared imaging data. The imager used in this study was a multispectral infrared line scanner based on 14 spectral bands. The stack emissions from an ammonia plant within a nitrogen fertilizer facility were investigated in this study. Cluster analysis and piecewise linear discriminant analysis were used in training data selection and classifier building, respectively. Satisfactory results were obtained for the classification of image pixels into plume/non-plume and CO₂/non-CO₂ categories.; Multivariate calibration has been widely used in analytical chemistry for decades. However, there are many situations when a multivariate calibration model may become invalid. To avoid the overhead of a full recalibration, multivariate calibration standardization is necessary. A new multivariate calibration standardization approach is developed based on the idea of building robust calibration models. The method is based on calibration sample selection and a weighting procedure. The main advantage of this approach is that compared to standardization methods such as direct standardization (DS) or piecewise direct standardization (PDS), the identical set of samples need not be measured with both primary and secondary instruments.; In an initial study, three data sets with the same constituents collected on the same instrument over a period of six years were used. The results showed that the proposed method significantly improved prediction performance in the new situation. A second study further investigated calibration standardization across three different instruments (one primary and two secondary instruments). The performance of the proposed algorithm, DS, and PDS were compared in this study. The results showed that the proposed algorithm outperformed both DS and PDS.