Several finer categories of forest types were
dropped during the initial field testing because key
tree species were too difficult to determine quickly in
the field. Some less common categories were added
later based on descriptions on the field forms.
The field crew was instructed to collect as
many diverse examples of each cover type as
possible in their two month field season. Stratified-
random plot selection methods proved too difficult to
implement in the field, although geographic
dispersion and patch size constraints were
emphasized. We were pleased to find that the GPS
unit was able to get useable location readings even
under heavy forest canopies, but this typically
required elevating the GPS antenna on a 10 m
telescoping pole.
While scheduling field work during summer
vacation made hiring students easier, rainfall is
heaviest during these months. Such rains prevented
location readings and crossing of river fords common
on smaller roads, limiting efficiency of field plot
establishment.
3.2 Preprocessing and Georeferencing
We used ERDAS Imagine 8.1 software to
process the 1993 Landsat TM image, including scene
stitching, georeferencing, and finally classification.
While most of the subject area fell within a single
satellite path and scene, a small area at the east end
of the APA had to be stitched to the main scene
before subsequent steps could be performed. The
added area had similar radiometric characteristics to
the main scene but the stitched seam remained
faintly visible in subsequent images. No other
preprocessing steps were performed however, such
as contrast enhancements or destriping.
We georeferenced the images before
classification using maximum likelihood techniques.
Georeferencing the images before classification
simplified use of GPS locations for the training sites.
An alternative approach would have required
classification first followed by georeferencing
through convolution. Although this alternative
promises greater cartographic accuracy, our
comparisons found little difference between the two
approaches, and using the maximum likelihood
method avoided creation of mixed pixels in the
procedure.
Based on preliminary analyses and the
literature, we confined our analyses to bands 2, 3, 4
and 5 although in a few cases radiometric
interference was too severe to use band 2.
Before conducting the actual supervised
74
classification it was necessary to expand each
training site’s point location into a larger patch of
similar pixels surrounding it. This was done by
digitizing polygons on screen. An unsupervised
classification of the original bands yielded a
provisional image that greatly facilitated this task.
Also helpful were field records, which noted cover
type changes near the plot location.
3.3 Image Classification
We tried using parallel piped procedures to
distiguish the different cover types but our initial
efforts revealed several difficulties. One difficulty
was unanticipated vegetation types, particularly in
the plateau at the north end of the APA. The location
of the State and APA boundaries here is presently
disputed by the states of Paraná and Sáo Paulo so we
did not target field work there, but classification was
nonetheless attempted for the area. Unfortunately,
cover there is dominated by two vegetation types
which do not occur elsewhere in the APA: a “humid
soil” marshland type in the valleys, and in the upland
“planalto” a transition between the highly mesic
dense forest of the coastal region and the more xeric
Araucaria forests of the interior plateau.
Expert judgement helped to identify training
pixels for these northern types. The signature of the
planalto type was confused with that of mangroves,
located along the bay, but stratifying the APA into
northern (plateau) and southern regions enabled us to
distinguish the two types.
Another new type was separated out based
on examination of the image: the “cordôes arenosos”
type, a complex of ancient dunes near the coast
characterized by alternating strips of dry upland and
seasonally flooded wetlands. This near-shore
complex is visually distinctive and was readily
identifiable on the image.
In addition to the extreme diversity in the
region, two edaphic factors common to the Atlantic
Forest complicated discrimination of cover types:
mesic conditions and complex terrain. The Atlantic
Forest enjoys plentiful rain throughout the year, in
this region totalling about 1500 mm annually. This
well-watered condition probably minimized drought
stress among the different vegetation types and
resulting signature differences.
The Atlantic Forest is a mountainous coastal
system. In the APA Guaraquegaba, elevations ranged
from sea level to 1580 m. This complex topography,
coupled with a low solar angle, created considerable
mountain shadow in the image. The more effective
approaches for dealing with mountain shadows
require a
appropriate
produced w
Giv
the original
Some cons
example, the
field work
between th
combined tl
class. We al
stages toget
categories '
category, bi
separate, as
(maritime fc
We
classificatio:
unsupervise
approaches
cover type.
identify tho
most effecti
exposed so
these categ
coverage. T
used to m
restinga, pla
categories,
polygons in
proved parti
To
data resultin
transport foi
format for t
of the twc
conversion r
3.4 Map Pr
Wh
product of
important ac
results. The
map produc
natural reso
and planner:
map for eco!
of access.
To
various vect
patterns anc
broader regi
of major riv