vay at IPARDES 
oils, hydrology, 
themes into a 
GIS). Although 
ped at the same 
at these physical 
changeable over 
ficiently current. 
ctive method of 
nd cover, which 
which produced 
porated into the 
)ES, SPVS, and 
ver in the region 
1sing supervised 
re. National map 
followed, and 
1 the field would 
tion and allow 
Land cover data 
ized roads and 
ade available to 
(Environmental 
stor and ERDAS 
its. 
e created using 
S programs we 
DOS-based GIS 
ty) would be 
lyses. Copies of 
training to both 
je. 
offered several 
ing its powerful 
solution, ready 
le pricing. Some 
ess at mapping 
milar ecoregions 
93; King, 1994; 
1ayas and Pope, 
this study the 
fication in this 
s still untested. 
il 1994 from the 
search (INPE), 
field survey was 
eavy cloud cover 
nber 1993, four 
a second image 
etection analysis 
(not yet attempted). Bands 2, 3, 4, 5 and 7 were used 
for our analyses. 
We chose to utilize supervised classification 
methods because of several apparent advantages over 
unsupervised or manual approaches. Supervised 
methods: 
e encouraged compatibility with existing ecologic 
literature and mapping efforts in the ecosystem, 
e facilitated objective classification and 
compatibility with future classifications, and 
e ensured that ground truthing would play an 
integral role in the study. 
3.1 GPS and Other Field Measurements 
Global positioning system (GPS) receivers 
collected ground location information during field 
work, which was conducted during January and 
February 1994. Three types of location data were 
needed: 
® control points surrounding the study area for 
georeferencing the images, 
© training plot locations for the supervised 
classification, and 
€ road and tourist attraction locations for map 
production and orientation purposes and for the 
accessibility-deforestation analysis. 
Collection of the control points, road, and 
tourist attraction information was done using a six 
channel Pathfinder Pro (Trimble Navigation Ltd.) 
GPS receiver and data logger unit, which stored up 
to 1 megabyte of readings exportable as geographic 
objects into GIS format. This unit was brought from 
the United States. A similar six channel unit without 
the data logger (Trimble Pathfinder Basic Plus) was 
used for the training site locations, where simple 
point locations were sufficient. This receiver was 
rented for the field season from a vendor in Rio de 
Janeiro. 
All readings were differentially corrected 
after downloading to a personal computer, using data 
from a survey-grade base station set up at FUPR in 
Curitiba especially for this field work. 
Two forestry students were hired for January 
and February, when university classes were closed 
for the summer, to collect and enter the field plot 
data. Their work was supplemented by two 
additional employees of SPVS who kept the rented 
GPS receiver in use during weekends and holidays. 
Two employees from the Superior National 
Forest in Minnesota experienced in GPS surveying 
joined the team for the first half of the field season to 
establish GPS procedures and train the field crews on 
their use. They also were responsible for establishing 
73 
the locations of the control points, roads, and most of 
the tourist attractions. Notes and sketches from field 
books were supplemented by brief notations typed 
into the data logger. These provided digital 
documentation of the data collection process which 
was carried through into the GIS as comment fields. 
With the GPS antenna mounted on a 
vehicle's roof, we digitized road segments in 
streaming mode by driving them and collecting 
readings every 10 seconds. Each road segment was 
terminated by a point feature to ensure solid location 
readings at intersections and road ends. Roads were 
classified by surface quality, and time stamps 
embedded in the GPS readings provided indications 
of relative travel speed. 
Eight useable control points were established 
around the APA, principally road intersections or 
bridges, and three more were added later from 
topographic maps. As with other point features, each 
location consisted of at least 180 individual readings, 
which after differential correction were averaged 
together. All GPS output files were plotted and 
visually examined for outliers which were deleted, 
although such outliers were rare. Differential 
correction and other GPS  postprocessing was 
performed using the PC program PFINDER (Trimble 
Navigation, Ltd.). 
For the training plots we designed a data 
entry form using a simple database program, 
providing both paper-based forms for field use and 
an electronic form for data entry and analysis. 
Several types of measurements were taken at each 
site, including variables describing basic descriptive 
information, attractiveness for ecotourism, and 
reliability for image classification. This last group 
included present cover type, nearby cover types in 
each of four directions, and indications of recent 
change in cover type. Estimated locations and times 
were also recorded to provide redundancy with GPS 
data for later error checking. Forms were tested in 
the field and revised several times. 
Prior to field work we selected twenty 
categories of land cover, based on the literature 
(Roderjan and Kuniyoshi, 1988), expert judgment, 
and preliminary field work. These included an urban 
category, seven agricultural categories, six stages of 
dense forest (floresta ombrofila densa) regrowth or 
disturbance, plus six additional categories of forest 
or herbaceous communities. We intentionally 
selected more categories than we expected to 
ultimately resolve from the images to encourage 
collection of a diversity of conditions and to allow 
subsequent flexibility in combining categories during 
the classification process.