s to the 
Linear 
PRING 
cessing 
  
VT 
ds 3, 4, 
ed to a 
of 3326 
cessing 
re used 
ENCED 
basic 
on' and 
; used 
31S and 
yriented 
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
data model, which provides integration of raster and vector data 
representations in a single environment. 
2.2 Image Processing Procedure 
2.2.1 Shade and Soil fraction 
There are several approaches to determine the fractional 
proportions of a multispectral image. The constrained least 
squares method was employed here using bands 3, 4, and 5 of 
the Landsat TM to generate the fraction images of vegetation, 
soil, and shade within the pixels. The mathematical description 
of the mixture model based on the constrained least squares 
method is detailed in Shimabukuro and Smith, (1991). The end 
members spectral signatures were collected directly from the 
Landsat TM images, using representative targets of green 
vegetation, soil, and shade. This was done interactively until it 
was verified that the component signatures were an adequate 
representation of the mixture components of the area analyzed, 
and therefore the constraint equations were strictly true. 
2.2.2 Image segmentation 
The procedure used for image segmentation was based on the 
region growing' algorithm where, a region is a set of 
homogeneous pixels connected according to their properties 
(Zucker, 1976). A detailed description of the segmentation 
procedure used can be found in Alves ef al. (1996). 
The segmentation algorithm was applied to the shade fraction 
image, and the proportion, the spatial and the contextual 
attributes of the segments (regions) were acquired for use in the 
classification procedure. 
Two parameters have to be set by the analyst: 
(1) similarity: the Euclidean distance between the mean digital 
numbers (here shade proportion) of two regions under which 
two regions are grouped together; and (2) area: minimum area 
lo be considered as a region, defined in number of pixels. 
2.2.3 Classification 
An unsupervised classification based on a clustering algorithm 
was applied to the segmented shade fraction image. Clustering 
techniques are largely known (Duda and Hart, 1973). The 
algorithm used in this experiment, named ISOSEG (Bins et al., 
1993) uses the covariance matrix and mean vector of the regions 
to estimate the means of the classes. The analyst can define an 
acceptance. threshold, which is the maximum  Mahalanobis 
distance that means digital numbers (here shade proportion) of 
regions can be far from the centre of a class to be considered as 
belonging to that class. 
First, the original TM bands 3, 4 and 5 were converted to 
vegetation, soil, and shade fraction images applying Linear 
Spectral Mixing Model. Next, the selected fraction images were 
segmented and  non-supervised classification per region 
algorithm was applied. The results of the classification were 
edited. 
To estimate the extension of deforested areas for the 1997 TM 
image, the shade fraction image was used, which enhances the 
difference between forest and deforested areas. To estimate the 
increment of deforested areas between 1997 and 2000, 2000 and 
2001, 2001 and 2002, 2002 and 2003, the soil fraction images 
N 
o 
were used, after the extension of deforested areas had been 
masked over the soil fraction image related to the following 
year. Soil fraction image enhances the difference between forest 
and recent clear cut areas. 
  
Figure 2 — Results of the segmentation algorithm shown for 
shade fraction image (red). The WRS 23167-TM/LANDSAT 
The methodology was applied for 182 WRSTM/LANDSAT 
Amazon Brazilian images acquired in 1997, 162 Landsat TM 
images acquired in 2000, 192 Landsat TM images acquired in 
2001 that cover the Amazon area, 156 Landsat TM images 
acquired in 2002, and 80 Landsat TM images acquired in 2003 
which cover the critical area in terms of deforestation. 
  
Figure 3 — Deforestation map of Brazilian Amazon showing the 
extension up to 2001 (yellow), 2002 increment (red). The 
forestry areas are showed in green, the no-forestry areas are 
showed in pink and the areas covered by clouds are showed in 
white. 
3. CONCLUSIONS 
The results clearly indicate the commitment of the BRAZILIAN 
GOVERNMENT 1o treat with transparency the issue of 
deforestation in the Amazon. Moreover, it gives parameters to 
understand the causes and to develop mechanisms to control the 
depletion of the forest resources. We presented an automated 
procedure to estimate deforestation including mapping and area 
extent estimations. The PRODES DIGITAL PROJECT data is 
stored in a GIS system, properly georeferenced, forming a 
database of multi-temporal layers with easy access and 
integrated with different information sources and or computer 
systems (www.obt.inpe.br/prodes).