ed
in
of
at
ns
in
of
he
or
ce
et
is
he
on
ed
is
"
en
or
in
M)
in
the new coordinate system , the separability of the
classes is described in terms of the different
proportions of primary components that these
classes present.
4. EXPERIMENTAL RESULTS
4.1 Methodology
The experiment was performed over an area
("ITAPEVA") that was reforested with Eucalyptus,
with different growing stages, in the state of Mato
Grosso do Sul, Brazil.
A Landsat TM image, taken on July gc 1984, with
bands 1, 2, 3, 4, 5 and 7 was used. The digital
numbers in the original bands were converted to
reflectance, according to the procedure proposed by
Markham and Baker (1986). Figures 4.1 displays the
resulting bands, called respectively, 1R, 2R, 3R,
4R, 5R and 7R.
Aerial photographs taken approximately one month
before the orbital coverage were available, The
experiment was performed over a test area of 161
rows by 161 columns, corresponding to the area
covered by the aerial photographs.
The first phase of the experiment was the synthetic
bands generation, through the application of the
computational methods that were mentioned in
Section 2: Constrained Least Squares (CLS) and
Weighted Least Squares (WLS). ^
The choice of the primary components was based on
the work of Shimabukuro (1987). For this area,
three components were considered: vegetation
(eucalyptus), soil and shadow. The reflectance
values of the vegetation and soil components were
extracted from the image by Shimabukuro (1987),
through the sample selection based on the available
aerial photographs and reforestation map. The
reflectance values of the shadow component were
also obtained by Shimabukuro (1987), through the
experiments performed by Heimes (1977), Figure 4.2
presents the reflectance curves of the components.
The analysis of the results obtained in this phase
was qualitatively performed through the work
previously performed by Shimabukuro (1987), since
it is very difficult to obtain quantitative
information from field work.
The second phase of the work consisted of the
comparative analysis of the Maximum Likelihood
classifier under the gaussian assumption, through
the conventional feature reduction methods
described in Section 3 and the Mixing Model.
Classification results were analyzed through the
classification matrices generated from training
samples. It is well known that the average
classification performance estimated over these
samples is optimistic but, since the objective of
the present analysis is the comparison between
different feature reduction methods, this fact was
disregarded. Thematic images generated by the
classification procedure were analyzed and
qualitatively compared by using the available
information. Unfortunately, it was not possible to
reproduce here the thematic images, so they are not
present in this work. They will be displayed in the
poster presentation and in a future, work in
preparation.
4,2 Results
261
Phase 1 : Use of proportion estimators
Synthetic bands derived from the proportions of
eucalyptus (Vegetation Band), soil (Soil Band) and
shadow (Shadow Band) generated by the CLS method
are presented in Figure 4.3. Figure 4.4 presents
the synthetic bands generated by the WLS method.
Both methods present very similar results, are
compatible with the available ground truth and are
similar to the results obtained by Shimabukuro
(1987). Therefore, it was decided to follow the
experiments using the synthetic bands generated by
the CLS method.
Visually, one can notice using the synthetic bands
a more clear distinction between two types of
eucalyptus. According to Shimabukuro (1987), this
difference is due to age variations of the
eucalyptus plantation. By analyzing the Shadow Band
(Figure 4.3.c), it is possible to notice that one
of this areas presents a higher shadow proportion,
what means less uniformity and a higher age.
On the basis of this results, the selected classes
for the Maximum Likelihood classification were:
New E.: reforestation with eucalyptus, with age
between 8 months and 2 years;
Old E.: reforestation with eucalyptus, with age
greater than 2 years; and
Soil : exposed soil.
These classes can be discriminated on the basis of
the different proportions of primary components,
which indicate the structural characteristics of
each class. As it can be observed in Figure 4.3, in
class New E., the pixels present a greater
proportion of the Vegetation component. On the
other hand, in class Old E., which is less uniform
due to age, one can notice a larger influence of
the Soil and Shadow components. Class Soil, as it
was expected, is basically due to Soil component.
Phase 2: Comparison between the Mixing Model and
the Conventional feature reduction methods.
Tables 4.1 to 4.4 present the classification
matrices, when the following bands are used:
a. The first three components (Cl, C2 and C3)
generated by the Principal Components
Transformation;
b. The two components (Cl and C2) generated by the
Canonical Analysis procedure;
c. The three original bands (3R, 4R, 5R) selected
by J-M Distance; and
d. Synthetic bands derived from the proportions of
primary components (Vegetation, Soil and Shadow
Bands).
In terms of average performance, the best results
were obtained through the Canonical Analysis and
the Mixing Model procedures. However, it should be
considered that all methods present high values
(and optimistic) for estimated average performance
(higher than 92Z in all cases) and small variations
from one case to another. Therefore, it is not
possible to present a definite conclusion about the
comparison in terms of classification performance.
Through the qualitative analysis of the thematic