performed with assistance of the Technological
Foundation of Acre (FUNTAC) and University of Acre
(UFAC). During the field campaign, several land use
classes were observed, and several vegetation profiles
(arboreal and bush individuals along sections of 60 m x 2
m size) were made at natural vegetation regrowth
including: the description and collection of flora
composition, DBH above 3 cm, total height, percentage
of crown cover. In this study, the general allometric
equation for secondary forest,
In Y 7 -2.17 * 1.02 In (DBH) + 0.39 x In Height
according to Uhl et al. (1988), was used to estimate the
biomass values, mainly of regrowth areas. An overview
of the entire area, as well as 35 mm photos were obtained
during an overflight.
The TM images were analyzed considering the following
steps: atmospheric and geometric corrections, image
segmentation based on the algorithm for the growth of
regions (similarity threshold = 6; area threshold = 10);
labeling of segment samples of thematic classes: Forest
(F) Initial (IR) and Intermediate (AR) Regrowth,
Overgrown Pasture (OP), Fresh Pasture (FP), Pasture
with Bare Soil (PS); application of the ISOSEG
Classifier; and generation of a thematic map. The SIR-C
data were analyzed according to the following procedure:
speckle reduction filtering (MAP filter) and merging of
SAR data at different polarizations (HH, VV, HV, LL
and Total Power) with TM images, application of a new
version of the EBIS texture classifier (Evidence Based
Interpretation of Satellite Images, developed at DLR) and
also plotting of the mean backscatter values as a function
of wavelength and polarization for each land cover
identified, and generation of a thematic map. EBIS is a
algorithm used for classifying textures, based on co-
occurrence feature vectors, that are modeled as
multinomial density functions (Lohmann, 1991).
Additionally to the classes mentioned above, for the
SIR-C data, the floodplain forest (V) was included.
4. RESULTS
Figures 1 and 2 are histograms showing the behavior of
land use classes at both SIR-C/L-band and TM-Landsat
data. In the signature plots of these figures, the original
data have been used in association with the variance
values of the training areas. The combination of bands
TM 3 and 5 present the best performance, as compared to
the other bands for the general discrimination of the set
of thematic classes, i.e. discrimination of forest, regrowth
and pasture areas. During the thematic classification, we
could observe that TM images alone did not allow the
discrimination of “Terra Firme” (Uplands) from
“Varzea” (Floodplain) Forests. In contrast to that, L-band
images, due to the several polarizations available and
textural characteristics, allowed a very good
210
discrimination among these two important forest types. It
is known that it is possible to separate different forest
types as well as logged forest using texture information,
and in this case, the different moisture and relief
conditions of these two environments, -are the main
reasons for its’ discrimination.
Regrowth areas are best discriminated with TM band 4
data, in the succession “Initial” and “Intermediate”, while
SIR-C/L-band data does not allow this discrimination.
The same finding applies for the discrimination among
"regrowth" and "overgrown pasture" classes. For the
identification “initial” regrowth and overgrown pasture,
TM band 3 was a better indication. The misclassification
in SIR-C data (L band) is mainly due to larger amounts
of shrubs and palms, like Maximiliana maripa and
Orbignia martiniana, which abound in these former
pastures and lead to higher backscatter values (sometimes
as corner reflectors) as well as to texture variances.
For the discrimination among "pasture with bare soil"
and “fresh pasture” classes, the L-band image presented a
similar performance as TM-Landsat data. The soil type
and the physical-structural conditions are also an
important factor for this discrimination analysis.
Generally, optical sensors are most sensitive to plant
structure at micrometer scales, whereas radar interacts
mainly at centimeter scales. Of special interest to radar
are the vertical stems of plants and the trunks of trees,
because wave propagation and backscattering through
these media are polarization-dependent (NASA, 1989).
Being so, the different polarizations provide different
views of the canopy’s structure. From another overview,
Figure 1, shows that the HV polarization of L-band is
more sensitive to the vegetation growth and it is
appropriated for land occupation/management studies.
As far as backscattering from vegetated surface is
concerned, three scattering components must be
considered: vegetation, soil and also the interaction
component of the L-band data studied here. Leaves and
stalks of vegetation are relatively transparent to radiation
and so the significance of the soil and the interaction
component predominate. The variability found in the
sample areas in each of the thematic classes of L-band
data, shows a certain sensitivity of this sensor to the
horizontal and vertical polarizations, but a lower
sensitivity to Total Power (TP) polarization.
Taking into account the actual interest of the scientific
community on specific studies of regrowth areas, it is fo
say that at these areas, the contribution of the horizontal
polarization is moderately higher than the vertical one.
The first one is of interest to define parameters such as
the spectral and textural amplitude of the physiognomic
structural variations of secondary succession areas
(Figure 3). Considering the vegetation data inventoried
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996