The main disadvantage to this integrated approach
ographic and soil attributes associated with geomor-
is that physiography or slope position is scale depen-
phology, along with the plant structure, to help differ-
dent (e.g., a small raised area seen on the ground may
entiate species composition.
function as an upland even though it occurs within a
Accuracy assessment
broad lowland area), and this contributes to uncertainty
in classification and mapping. This problem with dif-
Our assessment revealed that the overall accuracy
ferentiation of physiography is similar to that associ-
of the final ecotype map with 37 classes was 47% (Ap-
ated with the hydrogeomorphic classes (e.g., slopes,
pendix C). The errors fell into three major categories.
depressions, flats) developed by Brinson (1993). A sec-
First, there was substantial confusion between Low-
ond disadvantage is that the grouping of the many eco-
land Wet Low Scrub and Lowland Tussock Scrub Bog
logical components can generate a large number of
because spectral classification discriminated poorly
classes. For practical purposes, the number of classes
between low scrub alone and low scrub with tussocks,
needs to be reduced by combining similar characteris-
and because differentiation also was difficult on the
tics and ignoring unusual plots that do not fit the sim-
ground. In our study, we used a cutpoint of 20% cover
plified trends.
of Eriophorum vaginatum for the tussock class, though
others use cover values as low as 12%. Secondly, it
Ecotypes
was difficult to assign upland and lowland physiogra-
Classification and mapping
phy to map classes. This confusion is largely a ques-
Field data from ground-reference plots were used to
tion of scale; small patches of upland within a larger
identify 48 ecotype classes within Fort Greely (Table
lowland region often were called upland on the ground,
4, Fig. 8 and 9, Appendix B). Of these, 37 classes were
but mapped as lowland. Finally, the last type of large
differentiated in the final map (Fig. 10). The 11 classes
error was attributable to a lack of data describing soil
that were not mapped were not spectrally distinct
texture and moisture. While including these descrip-
enough or large enough to map reliably. For example,
tors as part of the ecotype classification provides valu-
low and tall scrub classes were merged for mapping in
able information, it adds complexity to the classifica-
some upland areas, and Lowland Fen Meadow, Lacus-
tion that can only be reduced by a very large ground
trine Fen Meadow, and Lowland Dwarf Scrub Bog were
verification effort.
merged with the nearest similar class because they were
Because the accuracy was poor for the map
not spectrally distinct.
with the 37 ecotypes, we derived another map with 20
The map revealed a high diversity of ecotypes re-
classes by aggregating similar classes that were
sulting from the strong elevation gradient and diversity
prone to large error (Appendix D). Map accuracy
of geomorphic processes. Overall, the most abundant
for the 20 aggregated ecotypes was 70% (Table A7).
ecotypes were Lowland Tussock Scrub Bog, Lowland
This aggregated map still includes sufficient discrimi-
Wet Low Scrub, and Lowland Wet Needleleaf Forest
nation of ecosystem properties for many management
(Table 5). Unusual ecotypes found on Fort Greely that
objectives, and also provides an example of the deriva-
are relatively rare elsewhere in interior Alaska included
tive products possible through manipulation of the map
Lowland Gravelly Dry Broadleaf Forest, Riverine Grav-
database.
elly Dry Dwarf Scrub, and Riverine Gravely Dry
This accuracy assessment, however, does not repre-
Meadow, which were associated with dry outwash grav-
sent the "true" map accuracy because the comparison
els; Upland Rocky Dry Low Scrub, which was associ-
was made with ground-reference data used in the map
ated with dry gravelly moraines; and Lowland Dwarf
production and not with independent data. Thus, it may
Scrub Bog, which was associated with thick organic
be biased toward a result of higher accuracy because
deposits.
the plots were used to develop map classes. Conversely,
Although we initially used the Alaska Vegetation
the results may be biased toward a poorer accuracy
Classification (AVC) for vegetation types, it generated
because ground plots often were located in small
a large number of classes because of changes in the
ecotype patches and this probably increased errors as-
canopy coverage (open, closed, and woodland) of trees
sociated with co-registration of plot and map data. Most
and shrubs. In many cases, such as for black spruce
accuracy assessments focus the sampling on large ho-
types, the understory vegetation was similar among
mogenous patches, which tends to artificially increase
classes. Similarly, small changes in tree composition
map accuracy. Without an independent assessment, the
generated numerous deciduous and mixed forest classes.
true accuracy is unknown, yet we believe that our
We avoided this proliferation of vegetation classes pri-
pseudo-accuracy results are consistent with our knowl-
marily by relying on the upper levels of the AVC that
edge of the study area and the problems we encoun-
characterize plant structure. We then relied on physi-
tered during mapping.
22
Previous Page
