bands were highly correlated. The unsupervised classi-
eas. While reclassifying clouds and shadows did create
fication was stratified by ecodistrict to maximize sepa-
some error, we did this to provide complete coverage
ration of spectral signatures within areas of similar
of ecotypes. Areas occluded by smoke were reclassi-
physiography and ecological characteristics. A total of
fied on the basis of relationships particular to the physi-
18 alpine, 29 highland, 28 lowland, and 11 riverine spec-
ographic district where the smoke occurred. Differences
tral classes was generated.
in gravelly and loamy ecotypes were differentiated us-
For field verification of spectral classes, we checked
ing the geomorphic units. Overall, dozens of rules were
ground truth in 17 areas distributed within the various
created using input from the spectral classes, DEM,
ecodistricts. In each area, 1020 points were sampled
DLG, ecosection (geomorphic units), and ecodistrict
along a meandering route (35 km long) designed to
(physiography) layers. After we initially developed the
sample all the spectral classes in the area. Points ini-
decision rules, we visually evaluated the resulting map
tially were selected from the classified image in the
to determine whether the rules were suitable for the
office and were chosen to fall within patches that had
scene as a whole. We then changed the rules as neces-
at least 3 3 cells of the same class. GPS points were
sary through several iterations before the modeled map
generated for the centers of these patches. In addition,
was made final.
the classified image was copied onto an acetate over-
Eleven ecotypes (i.e., Lowland Dwarf Scrub Bog,
lay for the aerial photo for each area. The GPS coordi-
Riverine Wet Meadow) could not be mapped because
nates, aerial photographs, and acetate overlays were
they were relatively rare, occurred in small patches, or
used to find the selected points in the field.
their spectral signatures were not sufficiently distinct.
We initially classified the spectral signatures by cor-
Classes that could not be distinguished with the satel-
relating spectral classes with field-determined ecotypes.
lite imagery were included as errors within the other
For classes that did not have sufficient field data, we
classes. Inherent to this approach is that the classifica-
used the large- and intermediate-scale color photogra-
tion was driven by the ground data and what could be
phy to interpret the ecotype represented by each spec-
distinguished on the ground, not by what could be dis-
tral class. The large-scale photography, obtained when
tinguished on the imagery. To facilitate use of the map
vegetation was in fall color, was particularly useful in
for management, the classified image was filtered to
determining vegetation structure, and the dominant
eliminate most small patches (13 cells).
plant species usually could be identified by their unique
We assessed the accuracy of the final ecotype map
fall colors.
by comparing the ecotypes of original ground-reference
To improve the classification of spectral classes, we
sites with their final map classes, because funding con-
used a rule-based approach that incorporated ancillary
straints did not allow the additional fieldwork that would
data and conceptual models of ecological relationships
have been required to collect independent data. While
to separate classes that had similar spectral signatures
this is not a truly valid assessment of the accuracy be-
(Hutchinson 1982, Satterwhite et al. 1984, Joria and
cause the data were not independent of those used to
Jorgenson 1996). The conceptual model was based on
create the map, it does provide an indication of map
an ecological relationships matrix that identified asso-
accuracy. Plots for which the ground-determination was
ciations among climatic zones, elevation, physiographic
not a mapped class were excluded, leaving 332 plots
districts, geomorphic units, slope, moisture, vegetation
for the analysis. Omission and commission errors were
structure, and vegetation composition. For the most part,
summarized by ecotype.
the ecotypes were mapped by associating physiographic
Ecosections and ecodistricts
and geomorphic characteristics obtained from the
ecodistrict and ecosection maps with the vegetation
Ecosection maps were based on photo-interpretation
structure obtained from the classified Landsat image.
of landform characteristics on 1:63,000-scale color-
For example, dwarf scrub types in alpine areas and
infrared photography taken in 1979 and 1980. Bound-
floodplains were differentiated using the physiographic
aries were delineated on 1:100,000-scale prints of a
map to create Alpine Rocky Dry Dwarf Scrub and Riv-
false-color composite of the georeferenced Landsat
erine Gravelly Dry Dwarf Scrub. Confusion between
image. The boundaries were digitized and codes en-
dark waterbodies, closed spruce forests, and steep,
tered for each polygon.
north-facing slopes was eliminated by buffering around
Ecodistricts were delineated on a 1:300,000-scale
lakes and ponds in the USGS DLG layer, and classify-
print of the Landsat image. During this process, we re-
ing only those pixels that occurred within the buffers
ferred to the map of land resource areas used in the
as lakes and ponds. Areas with clouds and shadows were
exploratory soil survey of Alaska (Rieger et al. 1979)
reclassified on the basis of elevation relationships and
and the map of ecoregions of Alaska (Gallant et al. 1995)
what ecotypes were most abundant in the affected ar-
to try to provide consistency in boundaries.
12
Previous Page
