f important
of this work
ropriate for
"fish-bone"
some larger
| analysis of
hasis on the
IM 231/067
km? and is
delimited by coordinates 9? 27' S, 61? 08' W; 11? 04' S, 61? 29'
W; 10? 50' S, 63? 08' W; and 9? 10' S, 62? 47' W. The BR-364
federal highway (Rondónia's most important road, linking the
State capital to the Southern part of Brazil) traverses the area in
the SE-NW direction; most of the developed properties are
located along the secondary roads that derive from the BR-364
highway.
Following FIBGE (1992), the native vegetation in the region
corresponds to the "Floresta Estacional Semidecidual
Submontana" class. This tropical seasonal semi-deciduous forest
covers areas with elevations ranging from 100 m to 600 m, and a
4 to 6-month dry season.
The majority of the properties have small areas (around 100 ha);
some cattle-raising farms have 1,500 ha or more. A number of
cacao and rubber plantations, funded by subsidies programs in
the 1980's, have areas between 200 and 500 ha.
MATERIALS & METHODS
Satellite Data and Image Processing System
This study used Landsat Thematic Mapper bands 3, 4 and 5 for
dates 07/15/94 and 08/05/90 (the area of work was defined
considering five more images for years 1995, 1992, 1988, 1986
and 1985, that are being used for further work). TM data
acquired and processed by INPE was provided in CDs.
The image processing system SPRING (Cámara et al. 1992) was
used in this work. This INPE-developed software runs on UNIX
workstations, and incorporates the functionalities mentioned in
the following section.
Image Processing Procedure
The technique adopted in this work follows the procedure
described by Alves ef al. (1996) for image co-registration,
segmentation and classification.
Six Landsat TM scenes (corresponding to the period 1985-1994)
were initially read from the CDs. The 1994 image was geo-
referenced using control points acquired with a Global
Positioning System device (GPS) during a May 1995 visit to the
region; the remaining 5 scenes were co-registered to the 1994
scene and then resampled to 120 m resolution (an August 1995
scene was later introduced into the data base).
The co-registered data were processed as described by Alves ef
al. (1996) to assure that all three bands for each image have the
same variance and then segmented using the region-growing
algorithm described by Bins et al. (1996). The result of the
segmentation algorithm consists of an image of labeled regions
that is subsequently classified by means of a region-oriented
unsupervised classifier.
15
The segmentation method allows the user to define the minimum
size of the areas and a minimum distance in digital levels for
region growing. The un-supervised clustering algorithm classifies
the regions, merging those that are closer than a specified
threshold. After clustering, an interactive interpretation
procedure is performed to assign clusters to one of the following
classes:
e forest;
e use (agriculture and pasture);
© abandonment;
e water (rivers, reservoirs and ponds);
e shadows;
e undetermined (typically "noises" from different sources
such as relief or forest texture)
Segmentation and classification errors are corrected using
SPRING editing functions. Areas erroneously classified are
corrected either by masking operations in the raster format or by
editing polygon attributes after a raster-vector conversion. Areas
missed by the segmentation procedure are hand-digitized at the
end of the procedure.
After processing, the images for years 1990 and 1994 were
analyzed using SPRING's raster analysis tools and the fractions of
abandonment for each year and the areas classified as abandoned
in both images were determined.
RESULTS AND DISCUSSION
Segmentation and Classification Results
As described by Alves ef al. (1996), the segmentation results are
visually more agreeable than pixel-by-pixel classifications (see
figure 2). The classified images presented less edge effects and
unclassified pixels than the pixel approach. As a result, the
authors' experience indicates that complete classifications
(including corrections) can be performed in an easier and more
satisfactory way than pixel-based classifiers.
The adopted unsupervised classifier presented reasonable results
for the initial discrimination of forest, use (agriculture and
pastures) and abandonment. This can be partially explained by
the relatively high contrast among these classes on TM bands 3, 4
and 5, that leads to good segmentation and classification results.
The results, however, did not differentiate abandonment from
some perennial plantations, like cacao and rubber, occurring in
the region; it can be noted that the authors are not certain, at the
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996