tanbul 2004
LAND USE/LAND COVER MAPPING IN BRAZILIAN AMAZON USING
NEURAL NETWORK WITH ASTER/TERRA DATA
E. H. Mendoza ' *, J. R. Santos !, A. N. C. Santa Rosa?, N. C. Silva?
' INPE, National Institute for Space Research, Av. dos Astronautas, 1758 - CEP 12.227-010,
Säo José dos Campos (SP), Brasil - (eddy, jroberto)(Qltid.inpe.br
* UnB, University of Brasilia, Geoscience Institute - GRM Campus Universitário — 70910-900,
Brasilia (DF), Brasil — (nunos, niltoncs)@unb.br
Commission TS, WG VII/2
KEY WORDS: Forestry, Land cover, Land use, Monitoring, Networks, Neural, ART2, Mapping.
ABSTRACT:
The objective of this study is to show the applicability of the genetic synthesis of the unsupervised artificial neural network ART2
(Adaptive Resonance Theory) in the classification of ASTER image for land use/land cover mapping. The area under study is
located in northern Mato Grosso State, Brazil, and is characterized by a strong human occupation process, which caused intensive
changes at the landscape, by deforestation, selective logging and agriculture. Field data were acquired in May/June 2003. The use of
ASTER data allowed an improvement of the analysis of the occupation process in tropical forest areas. ASTER images have
adequate spatial and spectral resolution and are an alternative to the remaining remote sensing data available. The data had a
correction of the cross-talk problem, after realized a resampling from SWIR bands (spatial resolution 30 to 15 m), a atmospheric
correction and rectification of ASTER images from both data sets 2002 e 2003. The input parameters for the neural network ART?
were optimized by genetic algorithm and the neural network was evaluated by a comparison of classification results with field data.
The evaluation of accuracy was done using Kappa statistics. The results of the classification were of satisfactory quality. ASTER
bands 2 (630-690 nm), 3 (760-860 nm) and 4 (1600-1700 nm) allowed an increased differentiation of classes, while bands 8 (2295-
2365 nm) and 6 (2185-2225 nm) were complementary for the identification of classes. The main land use changes that occurred
between 2002 and 2003 were related to deforestation, since many areas of tropical forest were replaced by agriculture and pastures.
1. INTRODUCTION Apart the innovation of the product-sensor in this study, there is
also a preoccupation on the evaluation of a new form of digital
The National Institute for Space Research — INPE/Brazil, has image processing. There is an interest for the use of Artificial
done several research and application studies with remote Neural Networks applied to the reconnaissance of patterns.
sensing data in the forestry area. Among the most important Within this perspective, Lee et al. (1990) report that neural
studies one can mention the project "Monitoring of the networks seem perform image classification better than the
Brazilian Amazonian Forest by Satellite" (PRODES), an traditional statistical techniques, since they don't require
estimation of the extent and rate of deforestation using TM explicitly that data that will be classified requires a parametric
Landsat 5 images. More recently, due to the degradation of the distribution. There are several models of artificial neural
TM-Landsat 5 signal and the non-availability of TM-Landsat 7 networks which were developed for the most different
data, other sensors have been used to get subsidies needed by applications, such as: perception by layers, self-organizing, by
the Brazilian Government to control and monitor the Amazon learning, and backpropagation, etc.
environment.
: s : Within this context, the general objective of this study is to use
The sensor system TERRA/ASTER has a great potential 0 4 genetic synthesis of the unsupervised neural network ART?
follow on these environmental studies, due to its ^ (Adaptative Resonance Theory) to map land use/land cover of a
spatial, spectral and radiometric resolution characteristics. Itis section from northern Mato Grosso Statosusinæ images from
also a complementary source of data and information derived Terra/ASTER, at the visible, near (VNIR) and middle (SWIR)
from other sensors which arc now operational | infrared bands. Furthermore a temporal analysis of ASTER
(NOAA/AVHRR, AQUA/MODIS, SPOT-S/HRV, sensor data was done in this region for the years 2002 and 2003
RADARSAT, IRS and fore recently. CBERS-2/CCD), as 4 in order to investigate land use/land cover changes.
subsidy to studies which involve different evaluations in
regional and local scales. Until now there are no works The specific objectives are: (1) to analyze ASTER VNIR and
using ASTER sensor data for studies on forest formations, types SWIR bands for the spectral discrimination of land use/land
of vegetation cover and land use in the Amazon, which justifies cover classes; (2) to evaluate the performance of initially given
the interest of this study. parameters for the genetic optimization of the neural network
ART?2 for the reconnaissance of patterns of ASTER data; (3) to
Corresponding author. This is useful to know for communication with the appropriate person in cases with more than one author.
123
