171
Picture 5: zoom into a result of a hard classification
Snow extension is calculated in the following way:
UJ 0% snow
IÉÉ 20% snow
Q 50% snow
ill 80% snow
1Ü1 00% snow
Picture 6: zoom into a map of snow presence obtained by
a soft classification
Snow area — pixel size • Npj xe | 0 f snow class
where:
Snow area is the snow-covered area estimated.
Pixel size is the terrain area corresponding to the
sensor IFOV.
Np^i o* snow class is the number of pixels assigned to
snow class by the classifier.
Snow covered area extension can be valued regarding to
snow pure pixels only and so it is underestimated
because ali partial contributions are neglected.
In this supervised Fuzzy-statistic classifier does soft
classifications. This classier requires pure training set
depicting pure situations and mixed training set depicting
image mixture too. Mixed pixels true composition is hard
to be valued because sub-pixel information is required.
The simplest way to evaluate mixture composition is to
overlay a contemporary higher spatial resolution on the
image to be classified. The overlap of different spatial
resolution images is a hard topic because the same item
is represented in different ways that are hard to agree.
Reference mistakes are very dangerous because they
produce wrong composition evaluations and sc wrong
training set. Next picture 7 show' the same area in two
different spatial resolution images.
5. SNOW COVERED AREA EXTIMATION BY
AUTOMATIC SOFT CLASSIFICATION
Automatic soft classifiers are able to describe mixture
because they define membership degrees of all image
pixels to ail the classes. Each membership degree
represents how a class is present into a pixel. Empirical
test showed that membership degrees are link to the
percentage of presence of a class into a pixel. In this way
pure pixels have one membership degree equal to 1 and
ail the other ones equal to 0, the mixed pixeis instead
have more than one degree greater then 0 and their sum
is 1. A soft classification does not produce a single
thematic map but as maps as the number of classes and
each map describes the presence of single class by
different percentages, in this way soft classifiers are able
to correctly describe transition zone since they are able to
solve mixture problem.
Next picture 6 is a zoom into a map of snow presence
obtained by a soft classification
