The application of Partial Least Squares regression (PLS) and PLS Discriminate Analysis (PLS-DA) to NIR spectra for quantification and classification purposes are well established and proven technologies. However the creation and validation of calibration models requires a known Y value for every spectrum included in the model training and test sets — either a quantity or class member value. This makes extension of PLS to hyperspectral imaging difficult, since accurate Y values for each pixel based spectrum are typically not known.

In conventional spectroscopy it is common to use a single bulk sample value — an average value representing a homogenous sample, matched to a single sample spectrum. In the case of hyperspectral imaging, how should the individual pixel spectra be used? A single spectrum can be computed from the set of spectra selected within a defined region of interest (ROI), but is this the optimal choice? Should the average spectra from multiple ROI's be used? Or perhaps sample variation should be incorporated into the calibration model by including all spectra within a single ROI, mapped to a single Y value. This poster presents various sampling strategies used for pixel spectra selection to optimize PLS and PLS-DA hyperspectral image model building.