of the
NWI
wetlands on aerial orthophotography from 1996 to
1998 and more recent leaf-off orthophotography from 2003.
The
NWI
polygons were visually compared to the aerial imagery
and moved if the locations did not appear to be accurate. The
shapes of the polygons were not altered and no new polygons
were digitized. Each wetland polygon in the edited
NWI
dataset
is labeled as freshwater emergent, freshwater forested/scrub-
shrub, freshwater pond, lake, riverine, or other. Only freshwa-
ter emergent (
PEM
) or freshwater forested/scrub-shrub (
PFO
/
PSS
) wetlands were used in this study, since it was assumed
that wetlands associated with open water could be identified
reliably from the aerial image interpretation methods used in
the
NWI
.
PFO
, and
PSS
wetlands were combined in this study
due to the limited number of
PSS
wetlands in the state and also
because the edited
NWI
data provided by
NRAC
did not distin-
guish between these types of wetlands. The combination could
be thought of as a woody wetland class. How combining these
classes may impact the accuracy of the model is not clear; how-
ever, we are making the assumption that
PSS
and
PFO
wetlands
should have a similar topographic signature in this landscape
.
The wetland presence training data were created by randomly
selecting pixels within mapped
PEM
and
PFO
/
PSS
wetland extents.
First, the updated
NWI
polygons were rasterized on a 9 m grid to
match the cell resolution of the terrain data used in this study.
Random samples of these pixels were then drawn as needed.
The
RF
algorithm requires both presence and absence data to
create a prediction (Breiman, 2001; Evans and Kiesecker, 2014;
Strager
et al
., 2015). Pseudo absence data were generated as ran-
dom pixels throughout the state, but only within areas that were
mapped as grass or woody cover in a previous mapping project
(Maxwell, 2013), and not within the
NWI
polygons. Although
it is possible that the pseudo absence points may have been
placed within some unmapped wetlands, it was assumed that
such points would be rare, because wetlands generally make up
a small proportion of the state (about 1 percent), and unmapped
wetlands likely comprise an even smaller percentage of the state
.
Two separate training and validation datasets were produced.
The first set was used to evaluate wetland models produced in
the selected subregions. For each ecological subregion, 1,000
random presence and 1,000 random pseudo absence pixels were
selected. This process was repeated five times to create five
separate sets of 2,000 points within each of the five ecological
subregions in order to produce five separate
RF
models that were
later combined to produce a single predictive model. Thus, for
the final model, a total of 10,000 training points were used. This
approach with five separate training sets was chosen to increase
the variability of the individual classification trees in the
RF
model, and thereby reduce generalization error (Strager
et al
.,
2015). A sixth random set of 1,000 presences and 1,000 pseudo
absence points that did not overlap with the training data were
created using the same methods, and withheld for validation.
Separate sets were produced for
PEM
and
PFO
/
PSS
wetlands.
The general approach used to develop training and validation
data for the ecological subregions was then repeated at the state
level. A total of 1,000 random presence pixels and 1,000 random
pseudo absence pixels were selected, repeated five times to pro-
duce five independent datasets (i.e., a total of 10,000 pixels). A
sixth dataset of 1,000 presence and 1,000 absence pixels was also
created and withheld for validation to assess the statewide result
.
The statewide wetland probability map was also evaluated
with respect to its potential for identifying wetlands outside
those of the
NWI
by using wetland point data provided by the
West Virginia Division of Natural Resources (WVNDR). The
WVD-
NR
data comprised 1,232 mapped palustrine wetland locations,
of which 567 were outside of the mapped
NWI
wetlands. These
567 points (hereafter referred to simply as non-
NWI
wetlands)
were combined with a set of 567 pseudo absence points, selected
using the method described above, to produce a second valida-
tion dataset. Due to the limited number of points in this dataset,
we created a single model for
PEM
and
PFO
/
PSS
wetland types
combined, as opposed to separate models at the state level for
each class.
In order to test the differences between the
NWI
and non-
NWI
data, we also produced a model trained on the non-
NWI
wetland data, essentially reversing the previous experiment.
The 567 non-
NWI
wetlands were split into two, with 283
presence samples combined with 283 randomly selected
pseudo-absence pixels. The remaining 284 non-
NWI
points
were paired with 284 pseudo-absence points and used for
validation. A second validation set was generated from with
284
NWI
presence and 284 pseudo-absence samples.
Terrain Variables
The
DEM
used in this study was produced using photogrammet-
ric methods. During the spring of 2003 and 2004, leaf-off, 0.61
m (2 ft) true color imagery was collected for the entire state of
West Virginia. Break lines and elevation mass points were gen-
erated using photogrammetric methods at a 3 m interval with a
vertical accuracy of ±10 ft. The final 3 m
DEM
has a tested verti-
cal accuracy of 0.209 m (Fedorko, 2005). For this study, the
original 3 m
DEM
data were mosaicked to create a single
DEM
for
the entire state, and the data were resampled to a 9 m resolu-
tion using pixel aggregation. This scale was chosen because it
provided an optimal tradeoff between a manageable file size
and a resolution that preserved fine-scale hydrological features.
From the statewide
DEM
, a variety of terrain derivatives were
produced as described in Table 1. The only variable unique
to this study is the distance from waterbodies and streams
weighted by slope (for brevity’s sake this variable will be re-
ferred to as just distance from waterbodies weighted by slope).
This variable was created using the Cost Distance tool in the Ar-
cMap
®
Spatial Analyst Extension (Esri, 2012). For this variable,
waterbodies and streams were obtained from the West Virginia
T
able
1. D
escription
of
DEM-
derived
V
ariables
U
sed
to
P
redict
the
T
opographic
P
robability
of
P
alustrine
W
etland
O
ccurrence
Measure
Description
Reference
Number
of Variables
Distance from
Waterbodies
Euclidean distance from
water-bodies weighted by
slope.
N/A (This paper)
1
Slope (°)
Slope (gradient or rate
of maximum change in
Z
)
atan
(
)
(
)
Rise
Run
2
2
× 57.29578
Burrough and
McDonell, 1998
1
Surface
Curvature
Second derivative
of slope
Moore
et al
., 1991;
Zeverbergen and
Thorne, 1987
1
Plan
Curvature
Curvature
perpendicular to slope
Moore
et al
., 1991;
Zeverbergen and
Thorne, 1987
1
Profile
Curvature
Curvature in
direction of slope
Moore
et al
., 1991;
Zeverbergen and
Thorne, 1987
1
Compound
topographic
moisture
index (CTMI)
Measure of steady state
wetness as estimated
from terrain characteristics
ln(
(
)
(
)
Upstream contibuting area
tan Slope
)
Gessler
et al
.,
1995;
Moore
et al
.,
1993
1
Slope
position
Scalable slope position
Z
–
Z
mean
Berry, 2002
5
Roughness
Roughness or terrain
complexity index
Z
standard deviation
_
Riley
et al
., 1999;
Blaszczynski,1997
5
Dissection
Dissection of
landscape index
(
)
(
)
Z Z
Z
Z
−
−
minimum
maximum minimum
Evans, 1972
5
440
June 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
