——— MM —— ——————
images (not present in this work), one concludes
that the best results were obtained by Canonical
Analysis, followed by the Mixing Model. A clear
improvement in terms of non-classified pixels can
be observed on these two images.
By analyzing the thematic image that results from
the use of synthetic bands, it is possible to
notice the highest concentration of non-classified
and confusion between New E. and Old E. classes
occurs over a stripe located in the area where the
new eucalyptus are planted. As it can be observed
in Figure 4.3, these pixels present a greater
proportion of the Soil component. This fact makes
them similar to the pixels that belong to Old E.
class, in terms of proportions. This structural
characteristic is not so clearly perceived when
other methods are used.
The reason for this structural difference found
within the area of New E. class is beyond the scope
of the present work. However, a deeper study of
forest engineering could be performed, by using
additional information about the canopy structure
provided by the primary components proportions.
Table 4.1 Classification matrix obtained when the
first three Principal Components (Cl, C2, C3) were
used.
Not ;
Classified New BE. ^ Old F. Soil
New E. 3.7 96. 0.3 0.0
Old E. 2.7 . 95.7 0.7
Soil 5.9 1.9 92.2
7
Average Abstinence: 3
Average Confusion:
6.0
1.0
0.0
Average Performance: 94,5
4,1
1.3
Table 4.2 Classification matrix obtained when the
Canonical axes (Cl, C2) were used.
Not
Classified New E. Old E. Soil
New E. 3.3 96.7
Old E, 4,7 0.0
Soil 2.2 0.0
Average Performance: 96.0
3.3
0.5
Average Abstinence: .
Average Confusion:
TOM Mo
Table 4.3 Classification matrix obtained when the
original bands selected by the J-M distance (3R,
4R, 5R) were used.
Not
Classified New E. Old E. Soil
New E. 5.3 94.3 0.3 0.0
Old E. 3.0 0.3 96.3 0.3
Soil 5.6 0.0 2.2 92.2
Average Performance: 94.24
Average Abstinence: 4.67
Average Confusion: 1.09
Table 4.4 Classification matrix obtained when the
synthetic bands (Vegetation, Soil and Shadow Bands)
were used.
Not
Classified New E. Old E. Soil
New E. 2.5 95.6 1.9 0.0
Old E. 1.0 0.3 97.0 1.3
Soil 1.0 0.0 1.7 97.3
Average Performance: 96.74
Average Abstinence: 1.52
Average Confusion: 1.74
4.3 Discussion
The results that were obtained with the "ITAPEVA"
experiment demonstrate that the Mixing Model can be
used as an alternative method for the feature
reduction phase of the classification process: its
average performance was comparable to conventional
methods and, furthermore, additional information
about the structural characteristics of the classes
was obtained.
It is important to point out that the methods
compared in this work aim at different objectives,
as it is described in the following.
The most straightforward method is the feature
reduction by the J-M Distance. The advantage of
this method is to be an indicator of the
performance of the Maximum Likelihood classifier,
if the gaussian assumption is verified.
The Principal Components transformation aims at the
optimum representation of the mixture of the
classes in the mean square error sense. The usual
choice is for those axes that have the largest
variances. However, for classification, the main
objective is class discrimination and not
representation (Duda and Hart, 1973). According to
Richards (1986) the good classification results
obtained with the Principal Components
transformation can be credit to the fact that
Remote Sensing classes usually are spread over the
first principal components axis. This is
particularly so for soils and spectrally similar
cover types.
On the other hand, Canonical Analysis aims
specifically at the discrimination. If the number
of classes, k, is less then the original feature
dimensionality, p, then the maximum number of
features generated by this method is (k-1). The new
features can be used as more adequate inputs to the
classification process. Another important
application is the use of the color composite of
the canonical axis, which were found to be very
useful for the visual interpretation in
heterogeneous areas, since the new features tend to
maximize the differences between soil categories
(Mather, 1987).
The use of the synthetic bands can also be very
useful for the visual interpretation of class
distributions. Qualitatively, their color composite
(not present in this work) is even better than the
color composites obtained by the other methods for
visual discrimination between the classes.
Furthermore, with the use of the synthetic bands,
classes are not described as a function of their
spectral response (spectral signatures) but in