found predominantly in forested or agricultural areas. In addi-
tion, the
NWI
data may not fully characterize the topographic
characteristics of unmapped wetlands in West Virginia. In
summary, wetlands not included in the
NWI
may be more dif-
ficult to map and have topographic signatures atypical of the
wetlands included in the
NWI
dataset.
Although these findings might seem to undermine the argu-
ment that the
NWI
-trained model can be used to predict loca-
tions of non-
NWI
wetlands, it is important to note that the
AUC
for the models trained with
NWI
data and trained with non-
NWI
data were the very similar, 0.957 and 0.956, when validated
with non-
NWI
data. Furthermore, no statistical difference was
observed (
p
= 0.970) between the
AUC
values. Thus, we infer
that although non-
NWI
wetlands differ slightly in characteris-
tics from
NWI
wetlands, and despite the fact non-
NWI
wetlands
are a little harder to map (i.e., have lower accuracy compared
to
NWI
wetlands), there is little to be gained from developing
non-
NWI
training data sets; equal results can be obtained by
simply training on
NWI
wetlands. This is a potentially impor-
tant finding for those wishing to replicate this study in areas
where non-
NWI
wetland data cannot be obtained.
The focus of this work was on developing a probabilistic
surface to represent the likelihood of wetland occurrence
based on topographic properties of the landscape. Because
wetlands are inherently fuzzy features, with gradational
boundaries that potentially vary over time, probabilistic map-
ping would appear to be a particularly useful approach to
represent these features in digital maps.
Producing “hard” classifications of binary wetland/up-
land maps from the probability data implies a loss of the rich
probabilistic information in the original data. Nevertheless,
although hard classification of wetlands was not the focus of
this work, such maps may be useful for many applications.
Generating such maps from the probabilistic surfaces could
be done by simply applying a single threshold for the prob-
ability surface, or alternatively, by selecting multiple thresh-
olds, multiple classes of varying confidence could be mapped.
Additional analysis and processing using remote sensing data
could be useful for refining products generated in this way,
for example, to screen out surfaces with spectral properties
that indicate they are unlikely to be wetlands, such as paved
surfaces or dry soil. Manual photographic interpretation, or
even field work, could be useful to further refine the mapped
sites. Alternatively, the probability values could be used as
a way to prioritize field checking of previously mapped
NWI
or non-
NWI
wetlands. Field checking is inherently expensive
and the wetland probability values could be a useful tool for
focusing fieldwork in areas to produce useful results.
Conclusions
DEM
-derived terrain variables and machine learning were used
to predict the probability of wetland occurrence with strong
model performance as indicated by
AUC
values generally above
0.90, and as high as 0.998. The
AUC
values for the
PFO
/
PSS
wetland models were generally higher than those of the
PEM
wetlands. For the statewide model, and all ecological subre-
gions tested, with the exception of the Great Valley of Virginia,
PFO
/
PSS
wetland models trained and applied to the same areas
produced
AUC
values of 0.99 or greater. This is very encourag-
ing, since forested wetlands have previously been specifically
noted as being a class with lower reliability in the
NWI
data-
base (Kudray and Gale, 2000). Thus, terrain variables may offer
a way to strengthen the
NWI
, which was developed primarily
from visual interpretation of aerial imagery (Tiner, 1997).
Models trained and validated within the same ecologi-
cal subregion generally outperformed models applied to
other subregions, and local models (i.e., those developed for
the ecological subregions) generally, but not always, out-
performed the statewide model when applied to a specific
region. These results suggest that the topographic expres-
sion of wetlands may vary between different physiographic
regions. Thus, it may be important to consider whether there
are distinct physiographic regions when mapping wetlands
throughout large geographic areas.
A variety of terrain variables were used to produce the
models. Some
DEM
-derivatives, which have not to our knowl-
edge been used before in mapping wetlands, such as distance
to water bodies weighted by slope, dissection, and roughness,
were of particular value for mapping wetlands, and notably
were found to be of greater importance than some commonly
used measures, such as surface curvature and
CTMI
. Additional
research should be carried out to explore the relative merit of
these terrain variables for mapping wetlands in other regions.
The statewide model of the probability of palustrine wet-
land occurrence, trained using a subset of
NWI
wetlands, was
generally more accurate for predicting
NWI
wetlands than for
predicting non-
NWI
wetlands. Nevertheless, the
NWI
statewide
model produced a high
AUC
value of 0.956 for the non-
NWI
wetlands, providing strong evidence that the topographic
modeling approach is useful for predicting potentially
unmapped palustrine wetlands outside of the current
NWI
extents. This is of specific interest given the higher omission
error rates associated with the
NWI
.
The findings of this research emphasize the value of
probabilistic mapping, as opposed to per-pixel classification,
for mapping features on the landscape that may not have
well defined boundaries, such as the gradational transition
between wetlands and uplands. The fact that the non-
NWI
wetlands generally have lower probabilities in the model than
the
NWI
wetlands may provide valuable information about the
differences between these different models. In future research,
we plan to investigate how the probability values relate to
wetland location and physical characteristics
.
The methods developed in this research may be of value
for updating and improving the
NWI
. Probability surfaces
could be of value for identifying potential sites for incorpora-
tion into the
NWI
and also as a screening tool to focus evalu-
ation, refinement, and updating of previously delineated
wetlands, for example by prioritizing field work or aerial
image interpretation.
Acknowledgments
Funding support for this study was provided by West Virginia
View and AmericaView. The project described in this publica-
tion was also supported in part by West Virginia View through
Grant Number G14AP00002 from the Department of the Interi-
or, United States Geologic Survey to AmericaView. Its contents
are solely the responsibility of the authors; the views and con-
clusions contained in this document are those of the authors
and should not be interpreted as representing the opinions
or policies of the US Government. Mention of trade names or
commercial products does not constitute their endorsement
by the US Government. We would also like to thank Elizabeth
Byers and the West Virginia Division of Natural Resources
(
WVDNR
) for providing the additional validation data. Last, we
would like to think three anonymous reviewers for their help-
ful comments that greatly improved the manuscript.
References
Baker, C., R. Lawrence, C. Montagne, and D. Patten, 2006. Mapping
wetlands and riparian areas using Landsat ETM+ imagery and
decision-tree-based models,
Wetlands
, 26(2):465–474.
Berry, J.K., 2002. Beyond mapping use surface area for realistic
calculations,
Geo World
, 15(9):20–21.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
June 2016
445
