International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
analyze the performance of the neural network ART2 in the
thematic classification as related to data of field work, using
Kappa statistics procedures; (4) to perform a temporal analysis
with data of the years 2002 and 2003, aiming to investigate the
dynamics of occupation from the physical space, evaluating the
changes at the thematic classes identified by the neural network
ART? for both years.
2. AREA UNDER STUDY
2.1 Localization
The area under study is located partially at the municipalities of
Sinop, Cláudia and Itauba, at northern Mato Grosso State,
Central- West Brazil, between geographical coordinates S
10?48'55"- S 12?00'46" and W 54?54'03" - W 55?46'53" (Figure
1). This area is situated along a section of highway BR-163
Cuiaba-Santarém, encompassing an area of 382,000 ha.
Be, La: 108210"
¥ Lag. 0: 5508205
um
“
Foil
ESTADO DO MAT
La.S: 1148313"
Long. 0: 555153951"
Figure 1. Location of the study area, in a mosaic of two images
ASTER (bands 4, 3 and 2).
3. DATA AND METHODS
3.1 ASTER images
Spectral data obtained by the ASTER sensor system were used
in this study (Table 1). The images referring to year 2002 were
acquired at NASA and EROS Data Center (USA), and those of
2003 at ERSDAC (Earth Remote Sensing Data Analysis
Center), Japan, under Agreement Nr. H140250. All scenes were
of type 1B, which are delivered to users with radiometric and
geometric calibrations. The images from 2002 were used
complementarily, aiming to investigate the dynamics of land
use/land cover in the area under study.
YEAR ID IMAGE
2002 pg-PR1B0000-2002050402 019 001
¢-PR1B0000-2002050402 181 001
2003 g-PR1B0000-2003090902 005 001
pg-PR1B0000-2003090902 167 001
Table 1. ASTER images used.
o
3.2 Methodology
Initially a pre-processing of data is done through correction of
the cross-talk problem, of resampling from SWIR bands (spatial
resolution: 30 to 15 m), of image registration, of atmospheric
correction and rectification of ASTER images from both data
124
sets 2002 and 2003. Afterwards the automatic image
classification is done using a non-supervised ART2 algorithm
considering Artificial Neural Network (ANN), followed by a
post-classification to minimize the presence of isolated pixels.
Finally the evaluation of accuracy is done, using Kappa
statistics, related to ground truth, with maps on land use/land
cover are generated for both data sets considered.
4. RESULTS
The cross-talk problem (Iwasaki et al., 2001) was solved using
a "cross talk correction" program developed by ERSDAC
(2001). Afterwards resampling of SWIR bands from 30 to 15 m
was done, in order to integrate with VNIR.
For the registration of the ASTER images, 20 control points
were identified, using a first grade polynomial and resampling
of pixels by the nearest neighbor algorithm. During the
evaluation of the exactness of registration using as reference the
image from 2003, a RMS error of 0,27 was obtained. This result
is considered as reasonable, because for a spatial resolution of
15 m of ASTER images, the internal error of points used for
mapping (4 m) is less than half a pixels' spatial resolution of the
images. As for the 2002 image, the RMS error of 0.05 was
considered excellent.
In order to obtain reflectance values on the surface, an
atmospheric correction was done using the ACORN 4.0
program, based on the radiation transference model
MODTRAN 4, which transforms radiance values of ASTER
images for surface reflectance values. Data of water vapor were
obtained from MODIS sensor data, located at the same Terra
satellite corresponding to the same dates of obtainment from
these ASTER images (MODIS, 2003).
To get uniform ASTER data for 2002 and 2003, the method of
pseudo-invariant targets present in the image were used (Hill &
Sturn, 1991). This procedure is used to eliminate those
differences caused by factors which affect the image
acquisition. The digital values obtained from these pseudo-
invariant targets (clear and dark) at both dates, were adjusted by
a linear regression to digital values of the reference image
corrected, for clear and dark targets (Mendoza, 2004).
To evaluate the discrimination between land use/land cover
classes, an analysis of the spectral behavior given by the
reflectance values and the wavelengths in the spectral range of
500-2500 nm was performed, which includes 9 bands of
ASTER. At Figure 2 a graph of the average reflectance of
typical samples of thematic classes defined in the area under
study is presented for the scene from June 27° 2002,
Internationa
UE
| Band 1
1
j| Bend 2
il
f
t
Reflectance
4— Virihle —]1—
Figure 2. Sj
At Figure 2
Primary Fo
formations,
present a tj
reflectance
spectrum (V
again a low
infrared, sp«
The classes
increase of
Primary For
water contei
SWIR occur
Classes ref
hirsutum), |
“dried” cor
depending c
increase of 1
3), is causec
because the
former figu
between the
band 4 (160:
On the othe
for all class
because the
other SWIR
talk phenom
The digita
unsupervise
program G
(SGRNA) p
network, di
2,3,4,51. [2
[3,4,6] and
classificatio
the networl
network we
Besides the
patterns in
thematic cl:
Forest/ rec«
