PSS
wetlands, masked to the extent of woody vegetation cover
in the state. The
AUC
values are listed for all models applied to
all five ecological subregions for
PEM
wetlands in Table 2 and
for
PFO
/
PSS
wetlands in Table 3. For
PEM
models trained and
applied within the same subregion (the diagonals in the table),
AUC
values are generally 0.94 or higher (Table 2), and for
PFO
/
PSS
models,
AUC
values were generally above 0.99 (Table 3),
with the exception of the Great Valley of Virginia, where the
AUC
was 0.96. We attribute the lower
AUC
value for the Great
Valley model to its comparatively subdued topography; previ-
ous research has indicated that wetlands are harder to dis-
tinguish in flat areas (Tiner, 1997). Overall, these
AUC
results
indicate excellent model performance (Hanley and McNeil,
1982; Swets
et al
., 2000; Fawcett, 2007). The fact that the
PFO
/
PSS
models have such high
AUC
values is particularly notable,
considering that prior research specifically identified wetlands
in forested regions as having higher omission errors than other
wetland types in the
NWI
(Tiner, 1997; Kudray and Gale, 2000).
Generally, models trained and validated in the same subre-
gion had significantly higher
AUC
values than models trained
in one subregion and extrapolated to another. This was espe-
cially true for the
PEM
models (Table 2), in which all extrapo-
lated models performed statistically less accurately than mod-
els trained within the region in which they were validated
.
Generally,
AUC
values were higher for the
PFO
/
PSS
models
(Table 3) than the
PEM
models (Table 2). We attribute this to the
confounding effects of topography and land cover.
PEM
wet-
lands are associated with grasslands, and in West Virginia grass-
lands are commonly found in floodplains and valley bottoms,
topographic locations where wetlands may be hard to differen-
tiate from adjacent, upland sites. In contrast,
PFO
/
PSS
wetlands
are associated with forest and shrub vegetation, typically found
in steeper topography, such as sideslopes and ridges, and
poorly drained areas where wetlands typically occur are likely
to be more distinctive topographically. For these reasons it may
be more difficult to map
PEM
wetlands than
PFO
/
PSS
wetlands
.
The statewide model was generally statistically less ac-
curate than the individual models, where those models were
trained and applied in the same subregion. This was especial-
ly true for
PEM
wetlands. The statewide
PFO
/
PSS
model (Table
3) is somewhat anomalous, as it showed a slightly higher
AUC
for the Pittsburgh Low Plateau than the model trained within
this subregion; however this difference was small and not
statistically significant.
Relative Importance of Terrain Variables
Figure 1 shows the relative conditional importance of the 21
terrain predictor variables used to generate the models for the
individual ecological subregions and for the entire state for
Figure 1. Conditional predictor variable importance for modeling PEM wetlands as estimated using OOB mean decrease in accuracy for
individual models for each ecological subregion (A through E) and a model for the entire state (F).
442
June 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
