de Carvalho, Luis
Two algorithms will be used to implement this filter bank. The first follows a pyramidal scheme where only one sample
out of two is kept after the filtering process (i. e. downsampling) (Burt and Adelson, 1983). The resolution decreases by
a factor of two at each decomposition level. No information is lost by this procedure and perfect reconstruction of the
original signal is still possible as long as the chosen filters allow it. Pyramidal transforms are particularly useful for data
compression as well as for resolution reduction.
The second is known as the à trous algorithm (Holschneider et al. 1989) where all output samples are kept. This means
redundancy as we could have ignored samples without losing information. Even then, for remote sensing applications
this redundancy can be of much use. One drawback is that after each decomposition level we double the memory
required for storage. On the other hand, feature extraction is improved by this kind of algorithm.
3 STUDY SITE AND DATA
An area around the city of Sáo Tomé das Letras in the
state of Minas Gerais, south-eastern Brazil, was chosen as
case study (Figure 2). The site is used for agriculture,
Minas Gerais
cerrado (Brazilian savanna), rocky fields and semi-
deciduous Atlantic forests. Climate is Cwb (Kópen's
classification), characterised by dry winter and wet
summers. In the last 20 years there was an increase of
mining activities and losses of forest cover among
periodic changes due to agricultural activities. South-
eastern Brazil is the most populated region of the country.
Sáo Tomé das Letras Pressure over natural resources has been very intense
since the last century resulting on a highly fragmented
area.
Figure 2. Location of the Study Site.
One Landsat MSS image from July 1981 and two Landsat TM images from November 1985 and August 1998 were
used in this research. Each pixel of the Landsat TM and MSS images covers a ground area of about 900m? and 3600m?
respectively. Landsat TM bands 2 (520-600 nm), 3 (630-690 nm), 4 (760-900 nm) and Landsat MSS bands 1 (500-590
nm), 2 (610-680 nm), 3 (790-890 nm) were chosen to perform this experiment because they cover comparable portions
of the electromagnetic spectrum (Buiten and Clevers 1996). In addition, also phenological conditions are different
within this data set (Figure 3).
Ancillary data comprised aerial photographs from 1979 and 1984 as well as field visits during the summer of 1999.
= ORNA LS anb mn
Figure 3. Landsat MSS from 1981 (a) bands 321 in RGB. Landsat TM from 1985 (b) and from
1998 (c) bands 432 in RGB reduced to 60 m of ground resolution. Using this colour combination
forest areas appear in red tones while mining sites appear in white.
rock exploitation and for the protection of remnants of.
342 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000.
ENG
a-^ "A ou. uo Ju Fb
