o expand each
| larger patch of
is was done by
\n unsupervised
ands yielded a
litated this task.
iich noted cover
ed procedures to
s but our initial
. One difficulty
5, particularly in
PA. The location
here is presently
Sao Paulo so we
lassification was
i. Unfortunately,
vegetation types
> APA: a “humid
ind in the upland
1e highly mesic
d the more xeric
eau.
identify training
signature of the
at of mangroves,
1g the APA into
ns enabled us to
arated out based
ordôes arenosos”
; near the coast
f dry upland and
"his near-shore
nd was readily
diversity in the
n to the Atlantic
of cover types:
in. The Atlantic
Yout the year, in
n annually. This
nimized drought
ation types and
intainous coastal
levations ranged
plex topography,
ted considerable
e more effective
untain shadows
require a digital elevation model (DEM), but
appropriate DEM’s were unavailable could not be
produced within our budget.
Given these complicating factors, many of
the original cover types had to be consolidated.
Some consolidation had been anticipated. For
example, the four month lag between image date and
field work meant some crops may have changed
between the image and field survey dates. We
combined the various crops into a single cultivated
class. We also combined the various forest regrowth
stages together. Lastly, the pasture and bare sand
categories were combined with the exposed soil
category, but crops and forest were still difficult to
separate, as were the mangrove and nearby restinga
(maritime forest) categories.
We took a pragmatic approach to our
classification difficulties which = combined
unsupervised, supervised, and manual interpretation
approaches to produce vector coverages of each
cover type. Unsupervised classification was used to
identify those land cover categories where it was
most effective: deep water, shallow or turbid water,
exposed soil, and beach. Once classified, each of
these categories was vectorized as a separate
coverage. The supervised classification image was
used to manually “photointerpret” the mangrove,
restinga, planalto, cordóes arenosos, and humid field
categories, using onscreen digitization of vector
polygons in Arc/Info. The ratio of bands 3 and 2
proved particularly useful for this task.
To provide maximum flexibility in sharing
data resulting data were made available in Arc/Info's
transport format for vector data and ERDAS *.IMG
format for the raster image (land cover only). Each
of the two source programs plus Idrisi offer
conversion routines for these formats.
3.4 Map Production
While digital coverages were the principal
product of the project, hardcopy maps were an
important adjunct product for distribution of study
results. These were produced using ERDAS. Two
map products were identified as early products: a
natural resource characterisation for land managers
and planners depicting cover types, and an orienting
map for ecotourists identifying attractions and routes
of access.
To produce the map for managers, the
various vector land cover layers were portrayed as
patterns and overlaid on a false color image of the
broader region along with the road network. Names
of major rivers and communities were added along
75
with scale, localization information, and other details
in a 1:100,000 full color map.
The map for tourists is still awaiting
production but is planned to be simpler and cheaper
to produce, and therefore will be a vector product at
1:200,000 scale. It will include communities,
attractions, and roads but not cover types or a
background image.
4. RESULTS
Results of the project include both GIS
products and the skills acquired in deriving them.
4.1 Field Work and Georeferencing
The GPS portion of the project mapped 368 km of
roads in and near the reserve as well as 13 points of
special touristic interest. Replicate readings at
intersections were typically within 10 m of each
other, although accuracy at intermediate points will
be less. Flooded fords prevented mapping of the
distal ends of some roads, and a few road segments
will need to be digitized from paper maps due to file
corruption or equipment failure.
Data on cover types were collected at over
111 field locations of diverse types although some
desired types were missed, particularly the humid
soil (upland wetland), planalto (upland mixed forest),
and cordóes arenosos (ancient dunes) categories. A
pleasant surprise was the success of the GPS receiver
in getting accurate readings even under heavy forest
canopy.
Georeferencing adjusted locations on the
image to within 0.9 pixels (27 m) and within the
accuracy needs for the project.
4.2 Image Classification
Image classification mapped 8 categories of
land cover: beach, restinga (maritime forest), cordôes
arenosos, dense forest, upland wetland, upland mixed
forest, mangrove, and exposed soil. Two categories
of water were also recognized: deep water and
shallow or turbid water. These categories met the
minimum project requirements for use in planning
and monitoring.
While SPVS experts on the region were
consulted on the classification, a quantitative
assessment of the classification's accuracy is helpful.
We compared the classification of the image with the
land cover assignments at the field plots to help
understand potential misclassifications as well as the
correspondence between the original categories and