Classifications of objects on hyperspectral images

Hyperspectral imaging is a modern analytical technique combining benefits of digital imaging and vibrational spectroscopy. It allows to reveal and visualise spatial distribution of various chemical components. In a hyperspectral image
every pixel is a spectrum (usually VNIR, SWIR or Raman) of a depicted area. Such
image can be represented as a cube or a set of 2D “slices” — one slice for each spectral band. It contains large amount of data and to reveal useful information
proper methods for processing and analysis are needed. Multivariate image analysis (MIA) is one of such methods widely spread among chemometicians.
In most of the cases MIA treats pixels as objects, so an image cube has to be
unfolded into a matrix, where rows represent pixels and columns — wavelengths.
So in fact, multivariate image analysis works with an image as with a large set of
spectra, without taking into account information about spatial relations of the
pixels. This works well in general, especially for exploratory analysis or multivariate curve resolution, but for some specific tasks it is not beneficial at
all. One of such tasks is classification or clustering of objects on hyperspectral
images. An object here means a set of connected pixels, fully or partly separated
from other objects. That could be, for example, tablets, cereals, biological cells,
etc. If objects from opposite classes are not absolutely different (e.g. there are
similar pixels) it can lead to a problem. For example, if two different tablets have
the same or similar excipient and different active ingredients, some of the pixels
chemically will be identical. But these similar pixels will be associated with different classes when a classification model is being calibrated. This can give
unstable model and poor classification results. In the present work a classification
method that combines classic image classification approach and MIA is proposed. The basic idea is to group all pixels and calculate spectral properties of
the pixel group to be used further as a vector of predictors for calibration and
class prediction. The grouping can be done with mathematical morphology methods applied to a score image where objects are well separated. In the case
of small overlapping a watershed transformation can be applied to disjoint the
objects. The method has been tested on several simulated and real cases and
showed good results and significant improvements in comparison with a standard
MIA approach. The results as well as method details will be reported.