containing all the thematic classes within the scene. The
analyst determined the decision tree used to combine the
separate classifications to insure that all deforestation and
secondary growth areas were properly classified. As a result of
this methodology and due to the nature of digital remote
sensing, the output classification was not entirely accurate.
Therefore, the classification must be edited. The editing is
performed manually using the project GIS methods described
below.
ISODATA. This classification technique was used for our
analysis of MSS and TM scenes in Southeast Asia. We found
that the thresholding technique employed in the Amazon MSS
analysis was not an effective first pass classification
technique due to several factors, including increased
topographic effects and the continuum of land cover types
typically found within individual scenes in Southeast Asia. We
have the following 5 output classes for our analyses in this
region: forest, non-forest, water, cloud, and cloud shadow.
Creation of Output Classes: We first created a new band which
was the ratio of band 4 (NIR) to band 2 (RED). We ran an
unsupervised classification using only band 2 (RED) and this
new band. We applied the ERDAS routine ISODATA to locate
45 output clusters. This ISODATA algorithm uses the statistics
of the data to determine the initial clusters arbitrarily. Then
the arbitrary means of the clusters are shifted to locate the
actual spectral statistics for each cluster. This shifting step
was repeated until at least 95% of the pixel were not reassigned
to new clusters during subsequent iterations that shifted the
cluster means..
Assignment of Final Classes: A minimum distance decision
rule was then applied to the entire image to assign all pixels to
one of the 45 output classes.
Recoding output classes into project classes: The output
classification was recoded into a new classification containing
up to 5 classes. This was done by overlaying the 45 class
classification on the original imagery to aggregate the 45
classes into 5.
Conversion from Raster to Vector Format. The
output classification was converted from ERDAS file format to
an ARC/INFO GRID file. The GRID file was then converted to
vector format using the ARC/INFO GRIDPOLY routine.
Manual Editing of Vector Product. The vector product
was plotted at 1:250,000 scale on vellum using an
electrostatic plotter or large format inkjet printer. The vellum
plot was then overlaid on our 1:250,000 scale colorfire
photoproduct of the Landsat image. Polygons that were
misclassified were identified and edited. The vector coverage
was then re-plotted and checked for further editing. These
plotting and editing steps were repeated until the
classification was completed.
482
Registration to Base Map. Because the original digital
data set was only system corrected there were misregistration
errors up to several kilometers. However, since the original
data was map projected an affine shift was sufficient to register
the vector coverage to a UTM base map. The LP project used
the World Data Bank II boundaries as our base map. Once the
first scene was registered to the base map subsequent scenes
were registered and edge matched to other registered scenes.
We found that for the 1980s Amazon assessment we moved the
scenes on average 2539 meters in the X direction and 904
meters in the Y direction. For the 1970s Southeast Asia
assessment on average we shifted the scenes 2082 meters in X
and 4667 meters in Y. The offsets for these two regions were
comparable in X, but not in Y. The difference is probably
attributable to the source of the data. The 1980s data came
from the EDIPS format, while the 1970s data was from the
older X format. The referencing to actual Earth coordinates was
finalized by using field derived control points to constrain the
WDBII positional information.
Field Validation. To obtain an estimate of the accuracy of
our final analysis, the project used a field-based accuracy
assessment program. Our objectives were to quantify the
thematic and positional variance. We did this at three levels of
analysis:
1. We conferred with experts in each region, gaining
insights from their extensive knowledge of
conditions. To facilitate a close working relationship
with experts in the countries and regions in which we
were studying, we had a visiting scientists program which
provided support to colleagues from tropical countries to
spend time in residence at our research labs. Visiting
scientists spent anywhere from 2 to 12 months in
residence.
2. We conducted preliminary and cursory field excursions to
various areas, where we obtained a good on-the-ground
sense of conditions and established initial classification '
rules and procedures.
3. We conducted systematic field validation exercises, where
points on the field were selected and measurements were
made using a Global Positioning System. The results of
these field exercises were used to develop a statistical
accuracy assessment using standard methods of
presentation ‚in contingency tables. In these field
exercises we tested two aspects of accuracy. The first was
thematic, assessing errors of omission and commission
in classification of the images. The second, using the
GPS and obvious features, we assessed the geometric and
positional accuracy of the image registration.
The project established approximately 22 field test sites
throughout the tropics. At each test site we collected as much
ancillary data as possible as well as other sources of remote
sensing data including Spot XS&PAN and various SAR data.
Field measurements were conducted at each test site. We also
International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
local :
used
helpe
the
varial
labor
classi
perfo!
prelir
Amaz
Some
land
asses!
varial
obser
and a
practi
acros:
more
assoc.
Resul
manu:
For t
preser
there :
area
an ani
sq. kn
3096
secon
forest
cloud
these
cells ¢
that h
NASA
Partne
(TRFI
Throu
comm
create
The fc
extent
1986
areas
thousa
prelim
compe
rates ;
signifi
results
due of
propor
Visual
