classes. At the Landsat scene studied, this refers to a
transition process of secondary vegetation growth, as
well as to a certain spectral non-definition of those pixels
that form a certain group of segments. The descriptor that
defined certain segments of the advanced succession
includes, within its space of spectral attributes, some
features of forests and of initial succession (Figure 2).
During the analysis of spectral identification of the class
“secondary succession", within the context of degrees of
membership, TM band 4 (near infrared) allows a better
separability. Considering all 10 classes labeled in the area
under study, at Figure 3, the percentage of segments,
where combinations of membership (total and partial
degrees) occur, are presented. One can observe that the
classes “overgrown pasture” and “initial and advanced
regrowth” present a higher level of combinations,
explaining its’ transition character. For the labeling
procedure of the “test image” of the neural network,
3,300 segments were used, corresponding to 153,000
pixels.
As for the performance of training of the neural network,
associated to the 10 thematic classes identified, the Mean
Square Error (MSE) was minimized according to its’
training time. There is a further reduction of the MSE for
the class “Advanced Succession” as compared to “Initial
Succession” (Figure 4), which his coherent with the
percentage of combinations of the degrees of partial
pertinence presented by these classes.
At the present stage of our studies, one.can have an idea
of the capacity of detection or rejection of each theme by
the neural network, taking into account the evolution of
the Cartesian distance of classes, associated to the
relationship among the indices of sensitivity and
specificity (Table 1). In order to monitor this distance,
the class “Initial Succession” shows already a better
performance of the neural network, while for the case of
“Advanced Succession”, there is a tendency to minimize
this distance, with the increase of sensitivity, during the
training.
Table 1 - Neural Network Performance related with the
Sensitivity and Specificity indices.
Classes Sensitivity/Specificity| Distance
Forest 0.5977 | 0.9964 | 0.4023
Advanced Regrowth | 0.2970 | 0.9910 | 0.7030
Initial Regrowth 0.0019 | 09982 | 0.3931
Clean Pasture 0.5314 | 0.9977 | 0.4686
Overgrown Pasture 0.4499 | 0.9671 0.5511
Crop 0.5833 1 099007 | 04167
Urban Area 0.6884 | 0.9968 | 0.3116
Water 0.8725 44.0.9700. |... 0.1308
Clouds 0.5397 | 0.9818 | 0.4607
Shadow 0.8848 | 0.9880 | 0.1158
206
For those cases where the MSE values and sensitivity and
specificity measures are not considered satisfactory to
perform the classification, it will be necessary to made an
equalization of the neural network, creating afterwards a
central network, that will use the contextual
characteristics of each segment for ‘the thematic
classification.
5. CONCLUSION
The use of fuzzy-logic to label image segments allowed
the stratification of different levels .of secondary
succession, where the segmentation approach for the
growth of regions is becoming an operational issue for
Amazonia. It is expected that those vegetation cover
classes that are physiognomically and structurally more
complex, such as the succession stages, regarding the
different land management practices, will need a longer
processing time in order to obtain an adequate
performance of the neural network. For the identification
of deforested areas in Amazonia, the neural network
approach presents already a very high (92%)
performance. The aggregation of these landuse classes
within the central network is being performed and the
previous analysis has shown a large potential use of
neural network, even taking into account the complexity
of land occupation in Amazonia.
6. REFERENCES
Barbosa, V.C.; Machado, R.J.; Liporace, F.S., 1993. A
neural system for deforestation monitoring on Landsat
images of the Amazon region. Technical Report CCR-
157, IBM, Rio Scientific Center, Brazil. 38p.
Liporace, F.S., 19940: Umd- sistema neural para
monitoraçäo do desflorestamento na regiäo Amazônica
utilizando imagens do Landsat. Master’s Thesis,
Universidade Federal do Rio de Janeiro, Brasil. 123p.
Venturieri, A., 1996. Segmentacáo de imagens e lógica
nebulosa para treinamento de uma rede neural artificial
na caracterizaçäo do uso da terra na regiäo de Tucurui
(PA). Master’s Thesis, Instituto Nacional de Pesquisas
Espaciais - INPE, Säo José dos Campos, Brasil. 116p.
Santos, J.R.; Venturieri, A.; Machado, R.J.; Liporace, F.
S., 1995. Monitoring land use in Amazonia based on
image segmentation and neural networks. In:
International ~~ Geoscience and Remote Sensing
Symposium, Firenze, Italy. vol. I, pp. 108-111.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
