Abstract:
|
It is well-known in image processing that, by varying the wavelength, any material reflects and absorbs
in a different way the solar radiation. This is registered by hyperspectral sensors, which collect
multivariate discrete images in a series of contiguous wavelength bands, providing the spectral curves,
which can distinguish between materials.
In order to partition a multivariate image in regions belonging to different materials, we need to
compare these regions which are previously modelled by using compositional data matrices, where the
entries in each row is a statistical discrete distribution of the radiance values (columns). These rows
correspond to distinct but contiguos wavelengths. Thus the distribution in a row is very similar to the
distribution in close rows. To measure this proximity, we use Hellinger distance between rows, which
provides a distance matrix.
Given two hyperspectral regions of an image providing two compositional data matrices, we obtain
the corresponding distance matrices and, by using metric multidimensional scaling, we compute
two sets of principal coordinates, which are related by a multivariate association measure based on
canonical correlations.
We ilustrate this approach comparing some multivariate regions of images captured by hyperspectral
remote sensors |