EAS 
ON 
ion of 
ch/87, 
ion of 
nages 
sed on 
' 1070 
ibined 
.Asa 
areas 
mean 
| areas 
tted to 
aps of 
1, Was 
everal 
'alized 
ons to 
s been 
ind of 
due to 
mined 
howed 
lateau. 
lso be 
dy we 
tor the 
|, from 
tate of 
being 
14°20’ 
Chris Bellman 
north latitude. The region has a tropical climate with a dry season from October to March, and a rainy season from May 
to September. Mean annual precipitation varies from 1,750 to 2,000 mm, with daytime mean temperatures generally 
higher than 30” C during the entire year. 
The Serra Tepequém, comprised by Middle Proterozoic sandstones and conglomerates of the Roraima Formation 
(Borges and D’Antona 1988), constitutes an isolated plateau bounded by abrupt sloping erosion scarps, standing up to 
1,000 meters above the surrounding metavolcanic country rocks. Mainly savanna open grass fields constitute the 
nn cover in the plateau, whereas surrounding area is covered by a tropical rain forest (see color composite in 
Figure 2). 
  
    
  
lp ——— 
64930' W 
VENEZUELA 
   
  
W 
— 4930” N 
        
  
S Did ; f t im study AYea f 
Rs. c.r. n au m IN MR m m uh n rj 
E e den m aui el Le iu ime v. "e : ul 29000* N 
  
  
  
Figure 1. Location of the study area. 
3. LANDSAT-TM DATA AND DIGITAL PROCESSING 
In this study we used all the available free-cloud Landsat-TM scenes covering the study area (path 233, row 57), 
acquired in March/87, October/91, November/94, January/96, and January/99. Images were converted to the UTM 
coordinate system, using common control points extracted from a topographic map at the scale of 1:100,000. This 
procedure yielded a registration accuracy equivalent to 0.8 pixel. Before processing, digital numbers were converted to 
reflectance values, and a 1% linear contrast stretch was applied to all the images to adjust grayscale ranges. As an 
attempt to improve performance of image segmentation, a median filter was applied to smooth the images. 
The digital image processing was carried out in a geographic information and image processing system (SPRING) 
developed by INPE (Cámara Neto et al., 1996). Image segmentation and region-classification techniques were the 
digital image processing techniques applied to generate thematic maps of the erosion features. Image segmentation 
permits partitioning off images into homogeneous regions, which may have particular common attributes such as gray 
level means. The used approach, based on region growing technique, can be described by the following sequential steps 
(a) segmentation of the image into regions (one or more pixels); (b) comparison of each segment with its neighbors and 
merging it with the statistically more similar segment, updating the mean gray level; (c) segment continues growing by 
comparison with the new neighbors until to have no more joinable segments, when a completed region is labeled; and 
(d) the process moves to the next uncompleted cell, repeating the entire sequence until all cells are labeled (Almeida- 
Filho et al., 1977). 
Landsat-TM band 3 (0.63-0.69 um) of the different years were selected for segmentation, due to the better enhancement 
between eroded areas and sparsely surrounding vegetated terrain. Best definition of the target areas in segmented 
images were obtained using a similarity threshold (distance to the center of the classes) and an area threshold (number 
of pixels) of 20 and 10, respectively. 
A region classifier algorithm based on clustering techniques was used to classify segmented image. The approach can 
be described by the following steps: (a) firstly it is created a list of regions ordered by size (number of pixels); (b) the 
classes present in the regions are detected; and (c) a k-means algorithm is applied to reclassify the regions, based on the 
class means. In the first step, the regions are ordered in a decreasing order by size. In the second step, the detection of 
classes is accomplished as follows: the statistical parameters (mean vector and covariance matrix) of the first region are 
taken as the initial parameters for the class. This region is then removed from the list. Next, given a chi-square threshold 
and the statistical parameters of the class, a decision surface is defined (in this case an hyperellipsoid). All the regions in 
the list with mean vector lying inside this hyperellipsoid are marked as not eligible for class detection. Finally, the mean 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 67