International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
2. STUDY AREA
The study area is located on the northern fringe of the
department of Guaviare in the Colombian Amazon. It extends
from the capital of the department, San Jose del Guaviare, 30
km south to El Retorno from latitude 2°35' to 2°20' north and
from longitude 72°47' to 72?35' west (Bijker, 1997).
ATLANTIC OCEAN
: Pahama v, Venezuon
* Colombia
hs
&
G eta
o
S "
& - WL
c { i Sand dat Gratin
^ dt^ E
fcusdor ^
Brazfi
d esf
ir Perti
Figure 1. Location of study area
2.1 Land cover
Before the 1950s, tropical evergreen rain forests covered a
larger part of the area. Due to the process of colonization,
extensive parts have been deforested and replaced by crops,
pastures or dense secondary vegetation. Besides the evergreen
rain forest and the human influenced vegetation, savannahs also
exist in the study area.
2.2 Land use
In the uplands pasture for cattle breeding is the dominant land
use in cleared areas. The following process will mostly be
applied to convert the forest into pastures: first the primary or
secondary forest will be cut and burned, after that perennial
crops (e.g. cacao, plantain) and annual crops (e.g. maize, rice)
are planted. A smaller part of the area will be planted with more
permanent crops like rubber, fruit trees, sugarcane and coca.
The number of years a field is used for crops can vary, but after
a view years of cultivation it will be left fallow or turned into
pastures (Bijker, 1997).
[n the alluvial plain of the Guaviare river agriculture is the main
land use. The soils are more fertile than in the uplands,
therefore the crop yields are higher in the alluvial plain. The
main cultivated crops are bananas, cacao, maize, soybean,
cotton, sesame, cassava and sugarcane (Bijker, 1997).
3. DATA AND METHODOLOGY
The input data in this research consist of one C-, L-, P-band
polarimetric AIRSAR image of May 1993 with a spatial
resolution of 6 m (Hoekman et al., 2000) and a time series of
nine ERS-1 images from May 1992 until September 1994 with
a spatial resolution of 12.5 m as well as the land cover maps
derived from these images (Bijker, 1997). The land cover maps
were based on all ERS-1 images available till that date.
Common land cover classes are needed to facilitate the
comparison between the classifications of the high spatial
resolution image and the lower spatial resolution images. In this
research four land cover classes were used: primary forest,
secondary forest, pastures and recently cut areas.
Furthermore, spatial scaling is needed. Spatial scaling takes
information at one scale and uses it to derive processes at
another scale (Jarvis, 1995). This can be either upscaling, where
information at a higher spatial resolution is taken and
transformed to the lower spatial resolution or downscaling,
which works in opposite direction (Jarvis, 1995). In this case
upscaling of the high spatial resolution AIRSAR land cover
map was applied.
Two processes will be applied to the high resolution AIRSAR
data. First, the classification of these data, as made by Quifiones
(1995; Hoekman et al., 2000) will be upscaled. The outcome is
referred to as AIRSAR land cover map 1 later on. Secondly, the
original AIRSAR image will be upscaled first and subsequently
classified. This result is referred to as AIRSAR land cover map
a
Two different approaches for upscaling were evaluated in order
to determine their effect on proportional area of land cover
classes and to detect changes in the number and size of the
patches of the AIRSAR land cover map during the upscaling
process as well as to select the most eligible procedure to
upscale the AIRSAR data to 12.5 m, the resolution of the ERS-
1 land cover maps. Finally, the classified ERS-1 images of
Bijker (1997) were compared with the upscaled AIRSAR land
cover maps to assess their conformity.
This research refers to pixels. According to Bian (1997), only
objects that operate at a scale larger than the size of the pixel
can be revealed during upscaling. But it also needs to be
mentioned, that small objects, which have a high contrast with
their surrounding area, may be detectable even if they are
smaller than the pixel size. However, at a high spatial resolution
pixel sizes are mostly smaller than objects of interest.
Therefore, the neighbouring pixels are highly correlated and a
low variance exists among them. With an increase in the pixel
size, the similarity decreases and the variance increases
(Rahman et al., 2003; Woodcock et al., 1987).
4. RESULTS
4.1 Evaluation of different upscaling approaches
The first approach implies stepwise upscaling in half-meter
steps, beginning with the AIRSAR land cover map 1 of 6 m
resolution until the resolution of 12.5 m was obtained. The
stepwise upscaling is based on each previous interim result. The
second approach used direct upscaling to different levels of
spatial resolution each based on the 6 m resolution AIRSAR
image. The two approaches resulted both in 14 upscaled
AIRSAR land cover maps of different spatial resolutions.
In order to calculate the new output pixels, the nearest
neighbour resampling method was applied, because it uses the
nearest pixel without any interpolation to create the resampled
image. The original pixels are simply relocated onto 2
geometrically correct map grid.
4.1.1 Changes in proportion of land cover classes
Research regarding problems of upscaling high resolution
remote sensing data showed, that it can be assumed that in
general the proportion of land cover classes will decrease with
958
Intern
A ia
co
pr
20
We
ho
ap
ca
thi
ari
th
th
lai