Predicting Palustrine Wetland Probability Using
Random Forest Machine Learning and
Digital Elevation Data-Derived Terrain Variables
Aaron E. Maxwell, Timothy A. Warner, and Michael P. Strager
Abstract
The probability of palustrine wetland occurrence in the state of
West Virginia,
USA
, was mapped based on topographic vari-
ables and using random forests (
RF
) machine learning. Models
were developed for both selected ecological subregions and the
entire state. The models were first trained using pixels random-
ly selected from the United States National Wetland Inventory
(
NWI
) dataset and were tested using a separate random subset
from the
NWI
and a database of wetlands not found in the
NWI
provided by the West Virginia Division of Natural Resources
(
WVDNR
). The models produced area under the curve (
AUC
) val-
ues in excess of 0.90, and as high as 0.998. Models developed
in one ecological subregion of the state produced significantly
different
AUC
values when applied to other subregions, indicat-
ing that the topographical models should be extrapolated to
new physiographic regions with caution. Several previously
unexplored
DEM
-derived terrain variables were found to be of
value, including distance from water bodies, roughness, and
dissection. Non-
NWI
wetlands were mapped with an
AUC
value
of 0.956, indicating that the probability maps may be useful for
finding potential palustrine wetlands not found in the
NWI
.
Introduction
Increasing recognition of the ecological importance of wet-
lands has resulted in growing interest in mapping their distri-
bution. Wetlands provide key ecosystem services, including
cleaning polluted waters, protecting shorelines, recharging
groundwater, and mitigating both flooding and drought (Geor-
giou and Turner, 2012). Wetlands also serve important roles
in natural chemical cycles as both sources and sinks, and in
particular, play a major role in global carbon cycling (Mitsch
et al
., 2013). Wetlands absorb and bind heavy metals (Sheo-
ran and Sheoran, 2006), and due to their role in transforming
chemical species have been described as “the kidneys of the
landscape” (Mitch and Gosselink, 2007). Wetlands have also
been described as “ecological supermarkets” because they
generally foster high biodiversity (Mitch and Gosselink, 2007).
The United States’ National Wetland Inventory (
NWI
) is an
example of a nation-wide wetland mapping effort.
NWI
data
are publically available and produced mainly from manual
interpretation of color-infrared and black-and-white aerial
photographs (Tiner, 1997). The
NWI
classifies wetlands using
the Cowardin
et al
. (1974) system, which distinguishes three
types of freshwater wetlands: palustrine, riverine, and lacus-
trine. Palustrine wetlands are freshwater, vegetated wetlands
that may include small water bodies. Palustrine wetlands are
further classified based on dominant vegetation and substrate
composition as palustrine aquatic bed (
PAB
), emergent (
PEM
),
forested (
PFO
), scrub-shrub (
PSS
), and unconsolidated shore
(
PUS
) wetlands.
PEM
wetlands are dominated by erect, rooted
herbaceous hydrophytes,
PSS
wetlands are dominated by
woody vegetation less than 6 m tall, and
PFO
wetlands are
dominated by woody vegetation taller than 6 m (Cowardin
et
al
., 1974). The
NWI
dataset was created with a defined target
mapping unit (
TMU
). This is not the smallest wetland that is
mapped; but rather the size class of the smallest group of wet-
lands the
NWI
attempts to map consistently (Tiner, 1997).
Many studies have evaluated the accuracy of the
NWI
(for
example, Swarthout
et al
., 1981; Crowley
et al
., 1988; Kuzila
et al
., 1991; Nichols, 1994; Stolt and Baker, 1995; Kudray and
Gale, 2000), and found the product to be notably reliable. For
example, Swarthout
et al
. (1981) report accuracies greater
than 95 percent for differentiating wetlands and uplands
using the
NWI
in Massachusetts. Despite this overall high ac-
curacy, there are limitations in the
NWI
dataset. In a study in
the Upper Great Lakes Region, USA, Kudray and Gale (2000),
observed 100 percent accuracy of
NWI
non-forested wetlands,
but only 91 percent of forested wetlands were found to be cor-
rectly mapped. Stolt and Baker (1995) documented low com-
mission errors (i.e., almost all
NWI
wetlands were indeed wet-
lands) but a larger number of omission errors (wetlands that
were not mapped) for palustrine wetlands in the Blue Ridge
Mountains of Virginia. Tiner (1997) supports this observation,
suggesting that, overall, omission error is more pronounced
than commission error in the
NWI
. Tiner (1997) further notes
that the manual image interpretation used in producing the
NWI
is likely to miss many wetlands, and favor larger and
wetter wetlands, as well as open water bodies. Other poten-
tial problems include an evolving understanding of wetlands
throughout the creation of the
NWI
, resulting in inconsistency
and changes in quality control and field review, exacerbated
by the limited quality of the images used and the complex-
ity of the upland-wetland interface. Certain wetland types
are also more difficult to map, such as those that are forested,
small, narrow, or farmed. It is more difficult to map wetland
extents in certain landscapes, such as in flat terrain. It should
also be noted that the accuracy of the
NWI
is highly variable
across the country, as some inventories are decades old and
made use of black-and-white aerial photography (Tiner, 1997).
Given the limitations of
NWI
, and in particular the apparent
Aaron E. Maxwell is with Alderson Broaddus University, 101
College Hill Drive Philippi, West Virginia 26416
(
).
Timothy A. Warner is with West Virginia University,
Department of Geology and Geography, West Virginia
University Morgantown, West Virginia 26506.
Michael P. Strager is with the West Virginia University
Division of Resource Management, West Virginia University
Morgantown, West Virginia 26506.
Photogrammetric Engineering & Remote Sensing
Vol. 82, No. 6, June 2016, pp. 437–447.
0099-1112/16/437–447
© 2016 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.82.6.437
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
June 2016
437
