'IFY
pagation
ng other
of these
twork to
es and of
ontextual
egories".
is based
zZy-logic
that are
ents, the
dure) by
start-up
training
features
ing, and
cape of a
STUDY
State, in
rdinates:
30°. This
'er Plant,
dong the
ation is
eep soils
soils are
presents
] rainfall
| months
SAT-TM
3. MATERIALS AND METHODS
LANDSAT-TM images of bands 1, 2, 3, 4, 5 and 7 were
used. Topographic maps (scale 1:100,000) sheets SA.22-
ZC-VI (Tucuruf), SA.22-ZD-IV (Goianésia), SB-22-XA-
II (Maracajá) and SB-22-XA-III (Repartimento) were
used for georeferencing orbital data. The software
packages SITIM and SPRING, developed at INPE, as
well as the IBM-developed package NICE (Neural Image
Classification Environment) were used. Figure 1 presents
a flow diagram that summarizes the methodological
procedures, including segmentation and classification by
neural networks.
Due to the complexity of land occupation, including as
well small agricultural settlements, timber exploration
and large pasture areas, the field verification was made in
previously selected sample areas. The area under study
was stratified as a grid and the sample areas correspond
to the UTM geographical coordinates of the topographic
maps used. The sample areas (n-138) were randomly
defined, taking into account the themes identified at the
image. A Global Positional System was used to locate the
sampling areas. During field work, the following
thematic legend was defined: Forest (F), Advanced (A)
and Initial (T) Secondary Succession, Clean Pasture (P),
Overgrown Pasture (O), Crops (C), Urban Zone (U),
Water (W), Clouds (N) and Shadows (S).
The image segmentation procedure was based on the
algorithm for the growth of regions, that generates,
according to a pixel (i, j), a region containing (i, j) that
includes an average gray level close to that of (i, j).
During the segmentation procedure, five values of
similarity thresholds (6, 8, 10, 12 and 14) were tested.
Each change of the similarity values caused variations of
the computer work to process the images, since the
similarity degree was defined by the tolerance parameter
t, represented by the Euclidean Distance among the
vectors associated to each segment.
Considering the distance (Ri, Rj) as a defined measure of
similarity among regions Ri, Rj, that increase
proportionally to the differentiation between Ri and Rj, it
was necessary to define the value A, a constant that
determines the minimum size for each region in the
segmentation process (Liporace, 1994). In this case, the
smaller region presented a minimum area of 10 pixels.
After the segmentation was made, the label of each
segment follows the fuzzy-logic, that allows the analyzer
to assign to each class, total or partial degrees of
membership (Barbosa, et al, 1993). The label of a
segment consists of a vector [0,1]" , with a dimension
associated to each thematic class. The value 1 indicates
total membership of the segment to a certain class.
Intermediate values, that correspond to probabilities of
0.75, 0.50 and 0.25, were associated by the interpreter to
segments representing partial degrees of membership. As
205
a rule during the procedure of labeling, it was established
that the total summation of the weights given to each
segment, without interferences such as clouds and
shadows, should be equivalent of 1, and that if there are
interferences, the segments could have a total summation
above 1. The main advantage of the use of fuzzy-logic is
that it allows to model transition phenomena (like the
stages of secondary vegetation and the’ conditions of
pasture areas), or phenomena at a border position, among
pixels of distinct classes, that could have the
characteristics of both classes, considering the sensor
resolution.
This labeling phase was prepared for “training” of the
neural network, in order to establish the knowledge base
and to test the neural network, i.e. to monitor its’
performance. During the establishment of the start-up
network, a backpropagation algorithm was applied for
training. At this start-up network both spectral (average
of gray values for each band) and textural (variance,
correlation and entropy) descriptors were used. During
the phase of monitoring of the network, for each thematic
class, the mean square error (MSE) and the indices of
sensitivity and specificity for a given set of segments
were analyzed, within a certain acceptance threshold.
After this phase, which was one of the objectives of this
study, either the classification can start or the procedure
where the network is equalized, in order to have a better
performance. From this point on, one has a central
network, with the same set of descriptors from the start-
up network, including also a neighborhood descriptor,
allowing the obtainment of a thematic map.
4. RESULTS
During the segmentation procedure of the Landsat/TM
image, among the thresholds tested, that one with value
10 allowed the best discrimination of thematic classes
found in the area under study. Lower similarity values
(6 and 8) presented an excessive fragmentation, while
higher ones (12 and 14) grouped in a same segment,
showed spectrally distinct areas. The computational effort
spent during the segmentation procedure is directly
proportional to the number of TM bands and inversely
proportional to the similarity threshold used. For the
segmentation of each image module (1025 rows and 763
lines), and considering the 6 optical bands of TM,
approximately 45 minutes processing time are needed at
a SUN-Sparc 10 workstation (Venturieri, 1996).
The basis of knowledge needed to train the neural
network, labeled by fuzzy logic, were 11,697 segments
(including totally 322,100 pixels), of which 1,146
belonged to the thematic class Advanced Secondary
Succession, found in the area under study. The use of
fuzzy-logic allowed several combinations of pertinence
for each class, indicating transitions among "neighbor"
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996