APPLICABILITY OF NEURAL NETWORK ARCHITECTURE BY A FUZZY MODEL TO IDENTIFY
NATURAL VEGETATION REGROWTH IN BRAZILIAN AMAZONIA
Joäo Roberto dos Santos !
Adriano Venturieri ?
Ricardo José Machado '
Frederico dos Santos Liporace 3
' Instituto Nacional de Pesquisas Espaciais - INPE, Sáo José dos Campos, Brasil
jroberto@ltid.inpe.br
? Centro de Pesquisa Agroflorestal da Amazónia Oriental-EMBRAPA, Brasil
? [BM-Centro Científico Rio, Rio de Janeiro, Brasil
. Commission VII, Working Group 3
KEY WORDS: Vegetation, Monitoring, Network, Segmentation, Regrowth, Amazonia.
ABSTRACT
The objective of this study is to present the results of preparation of a neural network trained by a backpropagation
algorithm, in order to process TM-Landsat images, used to identify areas under secondary succession, among other
landuse classes in Amazonia. Image segmentation techniques by an algorithm of region growth and labelling of these
segments by fuzzy-logic were also used for the preparation of the start-up network.The performance of the network to
delineate the “initial” and “advanced” regrowth areas, was obtained using both sensitivity and specificity indices and of
the MSE (Mean Square Error). Generally, it appears that both spectral, textural (entropy and correlation) and contextual
descriptors are effective for the identification of regrowth areas in Amazonia.
1. INTRODUCTION
The changes of the Amazonian landscape due to human
interference, have been the basis of discussion by the
Brazilian scientific community. The dynamics of land
occupation, the degradation of natural resources as well
as the private interest to follow the process of ecological
succession, they all deserve a research effort with remote
sensing techniques, taking into account specially the
integration of image segmentation techniques and neural
networks, which would contribute to reduce the
acquisition time of information.
Information extraction from large areas using an analogic
method is based primarily on the experience of visual
analysis of images, where a satellite image is classified in
homogeneous regions based on spectral, textural,
geometric and contextual characteristics. When using
conventional digital classification procedures, i.e. the
algorithms for supervised and non-supervised
classification, an image is stratified by its' spectral values
according to a defined statistical concept.
The first step of the method used in this study is the TM-
Landsat. image segmentation in spectral homogeneous
regions, aiming to identify specially those areas under
natural regrowth. This image stratification is based on the
region growth algorithm. Each segment is labeled in
thematic classes, representing the variations of landuse,
called “basic categories”. The cloud and shadow classes
204
were also identified and named "interfering categories".
This labeling, which is made by the analyst and is based
also on field information, is defined by a fuzzy-logic
approach, with degrees of partial membership that are
attributed to each segment. Based on these segments, the
start-up network is established (training procedure) by
the backpropagation algorithm. While the start-up
network is setup - when special features of the training
area, such as spectral, textural and geometrical features
are detected - the recording procedure is running, and
new areas can be classified, to monitor the landscape of a
certain region.
2. DESCRIPTION OF THE AREA UNDER STUDY
The area under study is located at SE Pará State, in
Amazonia, at the following geographical coordinates:
3° 30' to S 4? 30’ and W 49° 30’ to W 50° 30°. This
region is close to the Tucurui Hydroelectric Power Plant,
and it has a traditional land occupation pattern along the
road Transamazônica. The tropical vegetation is
classified as Ombrophilous Dense Forest, over deep soils
with medium to highly clay texture. The soils are
permeable and present low fertility. This region presents
a tropical humid climate with an average annual rainfall
of 1,300 to 3,000 mm. The drought period is 3 months
long. The area under study is covered by LANDSAT-TM
path 224 and row 63.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
