Rapid deforestation of a coastal landscape driven by sea‐level rise and extreme events

Abstract

Climate change is driving ecological shifts in coastal regions of the world, where low topographic relief makes ecosystems particularly vulnerable to sea‐level rise, salinization, storm surge, and other effects of global climate change. The consequences of rising water tables and salinity can penetrate well inland, and lead to particularly dramatic changes in freshwater forested wetlands dominated by tree species with low salt tolerance. The resulting loss of coastal forests could have significant implications to the coastal carbon cycle. We quantified the rates of vegetation change including land loss, forest loss, and shrubland expansion in North Carolina’s largest coastal wildlife refuge over 35 yr. Despite its protected status, and in the absence of any active forest management, 32% (31,600 hectares) of the refuge area has changed landcover classification during the study period. A total of 1,151 hectares of land was lost to the sea and ~19,300 hectares of coastal forest habitat was converted to shrubland or marsh habitat. As much as 11% of all forested cover in the refuge transitioned to a unique land cover type—“ghost forest”—characterized by standing dead trees and fallen tree trunks. The formation of this ghost forest transition state peaked prominently between 2011 and 2012, following Hurricane Irene and a 5‐yr drought, with 4,500 ± 990 hectares of ghost forest forming during that year alone. This is the first attempt to map and quantify coastal ghost forests using remote sensing. Forest losses were greatest in the eastern portion of the refuge closest to the Croatan and Pamlico Sounds, but also occurred much further inland in low‐elevation areas and alongside major canals. These unprecedented rates of deforestation and land cover change due to climate change may become the status quo for coastal regions worldwide, with implications for wetland function, wildlife habitat, and global carbon cycling.

Full text

Rapid deforestation of a coastal landscape driven by sea-level rise
and extreme events
EMILY A. URY ,
1,3
XIYANG ,
2
JUSTIN P. W RIGHT ,
1
AND EMILY S. BERNHARDT
1
1
Department of Biology, Duke University, Box 90338, Durham, North Carolina 27708 USA
2
Department of Environmental Sciences, University of Virginia, Charlottesville, Virginia 22904 USA
Citation: Ury, E. A., X. Yang, J. P. Wright, and E. S. Bernhardt. 2021. Rapid deforestation of a coastal
landscape driven by sea-level rise and extreme events. Ecological Applications 00(00):e02339. 10.1002/eap.
2339
Abstract. Climate change is driving ecological shifts in coastal regions of the world, where
low topographic relief makes ecosystems particularly vulnerable to sea-level rise, salinization,
storm surge, and other effects of global climate change. The consequences of rising water tables
and salinity can penetrate well inland, and lead to particularly dramatic changes in freshwater
forested wetlands dominated by tree species with low salt tolerance. The resulting loss of
coastal forests could have significant implications to the coastal carbon cycle. We quantified
the rates of vegetation change including land loss, forest loss, and shrubland expansion in
North Carolina’s largest coastal wildlife refuge over 35 yr. Despite its protected status, and in
the absence of any active forest management, 32% (31,600 hectares) of the refuge area has
changed landcover classification during the study period. A total of 1,151 hectares of land was
lost to the sea and ~19,300 hectares of coastal forest habitat was converted to shrubland or
marsh habitat. As much as 11% of all forested cover in the refuge transitioned to a unique land
cover type—“ghost forest”—characterized by standing dead trees and fallen tree trunks. The
formation of this ghost forest transition state peaked prominently between 2011 and 2012, fol-
lowing Hurricane Irene and a 5-yr drought, with 4,500 990 hectares of ghost forest forming
during that year alone. This is the first attempt to map and quantify coastal ghost forests using
remote sensing. Forest losses were greatest in the eastern portion of the refuge closest to the
Croatan and Pamlico Sounds, but also occurred much further inland in low-elevation areas
and alongside major canals. These unprecedented rates of deforestation and land cover change
due to climate change may become the status quo for coastal regions worldwide, with implica-
tions for wetland function, wildlife habitat, and global carbon cycling.
Key words: climate change; coastal forests; deforestation; ghost forest; remote sensing; salinization;
saltwater intrusion; sea-level rise; wetlands.
INTRODUCTION
Climate change is transforming landscapes around the
world, and in the case of many coastal regions, change is
occurring faster than ecosystems are able to adapt. Sea-
level rise may inundate up to 76,000 km
2
of land in the
conterminous United States alone (Haer et al. 2013).
Around the world, coastal landscapes are experiencing
inundation, saltwater intrusion, coastal storms, and
other extreme events associated with anthropogenically
driven climate change (Meehl et al. 2007, Church et al.
2013, Cloern et al. 2016). Shoreline loss and marsh
migration are well-documented ecosystem responses to
sea-level rise (Wasson et al. 2013, Fagherazzi et al.
2019); however, in low-lying coastal plain landscapes,
the impacts of sea-level rise and saltwater intrusion are
not confined to the coastal margins (Manda et al. 2014,
White and Kaplan 2017, Bhattachan et al. 2018). Low-
lying coastal landscapes often support forested wetland
vegetation; these tree-dominated communities are not
well adapted to permanently saline conditions and their
response to sea-level rise is not well understood (Munns
and Tester 2008). The Atlantic Coastal Plain of North
America is characterized by bottomland forested wet-
lands and fringing marshes (Brinson 1991). These wet-
lands are important wildlife habitat, provide valuable
ecosystem services, and are globally important carbon
sinks (Brinson 1991, Duarte et al. 2013, Spivak et al.
2019).
Historic land use of coastal plain regions has resulted
in a landscape that is particularly vulnerable to sea-level
rise. The construction of drainage ditches for agriculture
and channelization for navigation has increased the con-
nectivity of the landscape interior to saline coastal
waters (Bhattachan et al. 2018). Salt moves up-gradient
because of diffusion, and its effects on vegetation often
precede other visible evidence of sea-level rise (Tully
et al. 2019). Land drainage for agriculture and forestry
Manuscript received 6 March 2020; revised 16 September
2020; accepted 27 October 2020. Corresponding Editor: John
Christopher Stella.
3
E-mail: ury.emily@gmail.com
Article e02339; page 1
Ecological Applications, 0(0), 2021, e02339
©2021 by the Ecological Society of America

has further accelerated the effects of sea-level rise by
inducing subsidence of the ground’s surface through oxi-
dation of previously anoxic soils (Holden et al. 2004). As
a result, modification of coastal plain landscapes has altered
the pathways and spatial patterns of ecological change.
Extreme events further complicate our attempts to
measure and predict salinization in coastal landscapes,
and increases in the number of extreme events (storms,
drought, fire, and flooding) are an additional impact of
climate change (Meehl et al. 2007). Without accounting
for disturbance events, elevation is the primary determi-
nant of the forest–marsh boundary and the rate of lat-
eral wetland retreat (Schieder and Kirwan 2019).
However, disturbance events, which are prevalent in
coastal environments, exert control on the timing and
pace of lateral retreat. Both coastal storms and droughts
can increase the inputs of saltwater into coastal interiors
(Tully et al. 2019). Major coastal hurricanes can reverse
the flow of water in coastal creeks and canals and deliver
sea spray far inland. Droughts in these low-elevation
landscapes can allow wind tides to mix saltwater miles
upstream through stagnant creeks and canals. These dis-
turbance-driven inputs of marine salts exacerbate the
gradual processes associated with sea-level rise, leading
to an “ecological ratchet”effect, whereby each salt-load-
ing disturbance enhances the potential for an ecosystem
state change to occur in response to chronic sea-level rise
(Fagherazzi et al. 2019, Hillebrand and Kunze 2020).
The location and timing of extreme events may thus
determine much of the spatiotemporal patterns of
coastal landscape change.
One of the most striking transitions observed in
coastal landscapes in recent decades is the rapid mortal-
ity of entire stands of trees, or “ghost forests,”character-
ized by a high density of standing dead trees and
indicating recent, rapid, and synchronous tree mortality
(Kirwan and Gedan 2019). Recent work in coastal for-
ests around the Chesapeake Bay describes a fringe of
ghost forest at the transition between healthy trees and
migrating salt marshes (Schieder and Kirwan 2019).
Indeed, we might expect this ghost forest fringe to be the
natural outcome of marsh migration in response to sea-
level rise marking the inland extent of saline conditions.
Coastal trees can withstand periodic salinity exposure,
but the osmotic stress of chronically saline conditions
will impede germination and eventually lead to tree mor-
tality (Munns and Tester 2008, Kirwan and Gedan 2019).
In this study, we seek to understand the rates and spatial
patterns of ghost forest formation and other pathways of
coastal vegetation change in service of understanding the
drivers and extent of change in coastal plain landscapes.
Remote sensing is valuable for understanding land-
scape-scale response to changing environmental condi-
tions. Numerous remote sensing platforms, sensors, and
techniques have been used to study vegetation and vege-
tation change. Among them, the Landsat satellite record
holds one of the longest-running data sets that can be
applied to look at change over time (Adam et al. 2010,
Kennedy et al. 2014). Here, we demonstrate the utility of
remote sensing time series for understanding ecosystem
responses to global climate change. We measured the
magnitude, the trajectory, and the spatial distribution of
land and forested wetland loss throughout this vulnera-
ble coastal landscape over the last 35 yr. We asked three
questions. First, what proportion of the land area and of
forested wetlands were lost? Second, to what extent do
losses of land and forest occur beyond the coastal fringe?
Third, are the temporal trends in land and forest loss
driven by gradual sea-level rise or by extreme events?
METHODS
Study location
The Atlantic Coastal Plain of the United States is
experiencing rapid rates of relative sea-level rise—esti-
mated 3–4 mm/yr compared with global mean sea-level
rise of 1–2 mm/yr (Douglas and Richard Peltier 2002,
Church et al. 2013). Protected by a string of barrier
islands, the Coastal Plain of North Carolina is charac-
terized by large areas of intact coastal forest at low ele-
vation and with little topographic relief. The Alligator
River National Wildlife Refuge (NWR) is the second
largest protect area in the state of North Carolina and
an ideal place to study the pace and drivers of vegetation
change because of the distinct lack of development, for-
estry, and other anthropogenic activities. The median
elevation of our study area is less than 0.5 m above sea
level. Soils are largely peat based and poorly drained
with water table depth ranging from 20 cm to +10 cm
above the ground (Poulter 2005). The region was also
extensively drained for agriculture and forestry during
the mid-20th century resulting in a dense network of
ditches and canals across the landscape (approximately
4 km of drainage length per square-kilometer area;
Poulter et al. 2008).
Vegetation includes high and low pocosins (poorly
drained peatlands of the southeastern United States
characterized by dense shrubby vegetation), freshwater
and brackish marshes (mainly Caladium jamaicense and
Spartina sp.), hardwood swamps (various Quercus spe-
cies, Acer rubrum and Liquidambar styraciflua), and bald
cypress (Nyssa sylvatica) (Brinson 1991). The Refuge
was established in 1984 to protect the unique pocosin
wetland habitat and associated wildlife, including several
endangered species: the red wolf, the red-cockaded
woodpecker, and the American alligator as well as the
globally threatened Atlantic white cedar (Chamaecyparis
thyoides) ecosystem. The specific region of interest
includes most of Alligator River NWR and surrounding
natural lands, but excludes major roads, water bodies,
agricultural land, and the Dare County Bombing Range
(Fig. 1). The study area was affected by the Pains Bay
Fire during the summer of 2011 and Hurricane Irene,
which made landfall in August of the same year. The fire
affected approximately 97 km
2
within the 615-km
2
Article e02339; page 2 EMILY A. URY ET AL. Ecological Applications
Vol. 0, No. 0

refuge (U.S. Fish and Wildlife Service Southeast Region
Fire Management Organization 2011) and the hurricane
brought a 2-m storm surge to the nearest gauge station
in the Pamlico Sound (National Weather Service
Advanced Hydrologic Prediction Service 2016). The
storm surge was said to have inundated the Refuge with
water cresting over Highway 264, more than 2 km inland
from the coast (S. Lanier, personal communication).
Image preprocessing
Google Earth Engine (GEE) was used to generate
annual composite images using scenes from the Landsat
Satellite record from the U.S. Geological Survey (USGS).
We used images from three of Landsat’s instruments,
Landsat 5 Thematic Mapper (TM), Landsat 7 Enhanced
Thematic Mapper Plus (ETM+), and Landsat 8 Opera-
tional Land Imager (OLI) to span the period from 1985
to 2019. For each imager, only the atmospherically cor-
rected Tier 1 Surface Reflectance product was used. All
available scenes of the study region were clipped to apply
a water mask generated from the same year’s annual
water occurrence values from the JCR Yearly Water
Classification History (Pekel et al. 2016). For each year
we used both a winter (November–March) and summer
(May–September) composite to generate the input to
our classifier. The composite for each season is gener-
ated from the stack of images within the seasonal win-
dow, clouds and cloud shadow pixels were masked out,
and pixels were reduced to the median reflectance of
each band. Each composite was inspected for clouds and
other artifacts or distortions, and if present, the con-
tributing images were removed individually. For some
years, a suitable composite was not achieved and thus
left out of the analysis. The two seasonal composites
were combined such that the input for the classifier con-
tains 12 bands in total, 6 (blue, green, red, NIR, SWIR1,
and SWIR2) from each season.
Land cover classification
Training data for vegetation classes were collected
from Google Earth imagery (collected 24 March 2017,
10 cm RGB [Red, Green, Blue]) and from North
10 km
FIG. 1. Location of study region (shaded in green) in Eastern North Carolina.
Xxxxx 2021 RAPID COASTAL WETLAND VEGETATION CHANGE Article e02339; page 3

Carolina orthophotos available from the USGS (col-
lected 28 February 2012 and 11 April 2010, 15 cm
RGB). Randomly generated validation points (n=500)
were classified by hand with additional points added to
improve spatial coverage of all classes. Vegetation was
classified into the following categories: pine, deciduous,
shrub, marsh, and ghost forest. Ghost forest pixels were
characterized by the density of visibly dead trees present
in the high-resolution imagery. Ghost forest stands are
easily distinguished due to the lack of branches on snags
and the presence of fallen tree trunks. Because the
orthophotos were collected in the winter, it was easy to
distinguish between the pine and deciduous classes (see
Fig. 2). For some analyses, pine and deciduous classes
are combined into one class: “forested wetland.”There is
no need to distinguish between forested wetland and
upland forest in this study because the entire region is
considered wetland (either woody or emergent herba-
ceous) according to the USGS National Land Cover
Database (Yang et al. 2018). The shrubland class con-
sists of mixed scrub-shrub type vegetation, with some
herbaceous vegetation visible and large trees mostly
absent. The marsh class consists of either freshwater or
saltwater herbaceous species. Because of the Pains Bay
Fire of 2011 that affected a large portion (16%) of the
study area, a burn category was also included in the
2012 imagery demarcating areas that burned beyond rea-
sonable recognition of the vegetation. In 2012, large
areas of pine trees exhibited browning needles, most
likely because of drought or fire. These trees largely
recovered in subsequent years, but the unusual coloring
reduced the accuracy of the classifier, so an addition
class was added in 2012 (called “dry pine”), which was
then later merged back into the pine class for analysis.
We used a random forest decision tree algorithm
(Breiman 2001) to classify the composited inputs for
each year of training data sets (2010, 2012, and 2017).
Our use of spectral signatures for detecting and classify-
ing vegetation follows standard approaches of remote
sensing land cover land use classification (Asner 1998,
Ustin et al. 2009, Ustin 2013). The reflectance of bare
tree trunks is distinct from both live trees and under-
story vegetation, allowing us to detect ghost forests as a
unique class. Accuracy assessment was performed on
these years using k-fold cross-validation with five
groups. The classifier that was trained on 2017 data was
then applied to the other annual composites from Land-
sat 8 OLM (2015–2019). The classifier that was trained
on 2010 data was used to classify the composites from
Landsat 5 TM (1985–2011).
Statistical analysis
Statistical analyses were run in the R statistical com-
puting environment (R Development Core Team 2013).
Classification was done using the R package “ran-
domForest”(Liaw and Wiener 2002) with Ntree and
mtry default values. Other R packages used for data
transformation and visualization included “raster”(Hij-
mans 2019), “rgdal”(Bivand et al. 2019), “alluvial”
(Bojanowski and Edwards 2016), and the Tidyverse
(Wickham et al. 2019). We used bivariate logistic regres-
sion (glm() function, family =binomial) to determine
the influence of topographic characteristics on forest
loss and ghost forest formation. In these models, forest
loss is a binary response variable indicating if a pixel was
classified as forest at the start of the study period but
not at the end. Environmental predictors for channels
and coastline were generated by calculating a simple
linear distance raster in R using tools from the “raster”
package and derived from the National Hydrography
Data Set (NHD) flow lines for the State of North
Carolina and 30-m resolution Digital Elevation Model
(DEM) from the USGS. A bootstrapping approach
with random draws of 1,000 pixels using a 500-m
exclusionary buffer was used to generate model input
data as a means of reducing spatial autocorrelation
among covariates.
Training and validation
ShrublandMarsh
Pine Deciduous
Landsat composite Aerial imagery, Google (10-cm resoluon)
Ghost forest
Classified map
Burn
Input Output
FIG. 2. Classification schematic and examples of each vegetation class as seen in Google Earth imagery.
Article e02339; page 4 EMILY A. URY ET AL. Ecological Applications
Vol. 0, No. 0

RESULTS
The results from our classification demonstrate that it
is possible to detect coastal ghost forests using Landsat
imagery. The overall accuracy for all five classes was
92.2% for the Landsat 8 classifier, 82.5% for the Landsat
7 classifier, and 82.3% for Landsat 5 classifier. The aver-
age producer accuracy and user accuracy for the ghost
forest class is 88.8% and 85.5%, respectively. In the
orthophotos used for training data, each point charac-
terized as ghost forest contained approximately 20–40
visible snags or fallen tree trunks per Landsat pixel
(about 220–440 dead trees per hectare). In a random
sample of Landsat pixels classified as ghost forest, the
mean number of snags and fallen trunks per pixel was
lower (mean =13, n=50), which is likely because of
bare ground reflectance, which also contributes to the
spectral characteristics of a ghost forest pixel. This may
suggest a slight overfitting of the ghost forest class; how-
ever, the overall accuracy for this class is still quite good
(see Appendix S1: Table S1). We assume that the overall
accuracies are a good approximation of the accuracy of
each map produced with the respective classifier. Lower
accuracy from the older imagery is due in part to a lack
of validation points within the ghost forest class (there
simply were not as many ghost forests present at that
time on which the classifier could be trained). We found
that using both a winter and summer composite image
greatly enhanced the classifier’s ability to distinguish
between the pine, deciduous, and ghost forest categories;
in summer the pine/deciduous distinction is more diffi-
cult to parse and in winter the deciduous/ghost forest
signatures are similar, but combining summer and winter
reflectance enables us to distinguish between all three.
Overall, pine, deciduous, and marsh were the most suc-
cessfully distinguishable classes for the classifiers and
shrub and ghost forest were more difficult to parse out
(see complete confusion matrices in Appendix S1: Tables
S1–S4).
In this 99,347-hectare (ha) study area, 77%
(76,575 ha) was forested wetland in 1985. Over 35 yr,
the refuge has lost 1,151 ha of land (1.2%) to open water
and had a net loss in forest cover of 15,811 ha (Table 1).
Most strikingly, of the 76,575 ha that were forest in
1985, 21,097 ha (27.6%) transitioned to either shrub,
marsh, ghost forest, or open water by 2019. In fact, only
69.2% percent of this protected wildlife refuge has not
experienced a vegetation transition over the last 35 yr
(shown in light gray in Fig. 3A). Of the forest classes
that transitioned to nonforest classes, 40.8% (8,616 ha)
went through a detectable ghost forest stage at some
point during the study period (these are the pathways
shown in red in Fig. 3B). The majority (17,115 ha, 81%)
of the forested area that transitioned is now categorized
as shrubland. Parts of the Refuge have experienced for-
est regrowth (mainly from shrubland) totaling 5,287 ha.
The marsh class has grown about 7% over the study per-
iod; however, its position on the landscape has shifted:
marshes are moving inland to areas that were shrub and
forest as the shoreline retreats. The conversion to marsh
is occurring both with and without a ghost forest transi-
tion state (Fig. 3B), which may indicate the relative con-
tributions of multiple mechanisms driving vegetation
change.
The spatial distribution of change within the refuge is
most heavily concentrated along the coasts, and the east-
ern side along the sound is more severely affected than
the western coast on the Alligator River (see the map in
Fig. 3A and histogram in Fig. 4D). The distribution of
forest loss is concentrated on the eastern side of the
refuge; however, 59% of forest loss is occurring in the
interior part of the landscape (more than 1,000 m from
either coast). The distribution of forest loss is more con-
centrated at low elevations (Fig. 4E). The distribution of
forested pixels and forest loss based on distance to chan-
nel is heavily right skewed, owing to the prevalence of
channels in this landscape (Fig. 4F).
Each of the topographic predictors in the spatial mod-
els of forest loss and ghost forest formation (binary
response variables) are statistically significant (Pvalues
<0.001) drivers of these transitions except for distance to
channel on ghost forest formation (Pvalue =0.11)
(Fig. 5). The spatial autocorrelation of some variables
(e.g., topography) violates the assumption of indepen-
dent observations, and reduces the effective sample size.
However, using random subsets of the data reduces the
effect of autocorrelation (Moran’s I is reduced for each
model from 0.86/0.71 to 0.39/0.21), and the subsequent
impact on the Pvalues is negligible. Distance to the
sound (the eastern shore of the refuge) was the strongest
predictor of both forest loss and ghost forest formation
compared with elevation and distance to a channel. We
standardized the predictor variables in the linear model
such that the slope coefficients may be compared to
determine the relative strength of each predictor on the
outcome of a forested pixel. The coefficient estimate for
distance to the sound was 0.839 (95% confidence inter-
val [CI] 1.01 to 0.669) for forest loss and 0.906
(95% CI 1.12 to 0.689) for ghost forest formation.
Negative coefficient estimates indicate that forests closer
to the coast are more likely to undergo these types of
transitions. The coefficient estimate for elevation was
0.673 (95% CI 0.862 to 0.485) for forest loss and
0.677 (95% CI 0.928 to 0.426) for ghost forest
TABLE 1. Summary of the net change for each vegetation class
(in hectares).
Class 1985 2019 Net change
Pine 27,483 20,205 7,279
Deciduous 49,092 40,560 8,532
Ghost forest 66 2,051 +1,985
Shrub 13,798 25,850 +12,052
Marsh 8,907 9,530 +623
Open water 0 1,151 +1,151
Xxxxx 2021 RAPID COASTAL WETLAND VEGETATION CHANGE Article e02339; page 5

formation; forests at lower elevation are more likely to
transition. The coefficient estimate for distance to a
channel was +0.239 (95% CI 0.098–0.380) for forest loss
and 0.170 (95% CI 0.357 to 0.018) for ghost forest
formation. The effects of proximity to a channel are of
lesser magnitude and are inconsistent: forested wet-
lands are less likely to be lost alongside channels, but
the effect on ghost forest formation is not statistically
significant. We speculate that adjacency to a channel
may be associated with less drought risk for much of
the area and for the possibility of event-based saliniza-
tion and mortality for trees alongside canals nearest to
the sound. The ghost forest response variable indicates
a ghost forest was detected at any time during the
time series but modeling each year individually shows
that the relative importance and even the direction of
these relationships vary across the time series (see
Appendix S1: Fig. S1).
The changing areas of Alligator River NWR are
shown in Fig. 6A for the entire period of study. The
steady loss of land to open water is apparent in the blue
wedge along the x-axis. In contrast to the linear pace of
land conversion to open water, the vegetation classes
fluctuate over time (10–20% of this fluctuation maybe
attributed to classification error). The forested area (pine
and deciduous classes combined) is declining over time
and the shrubland is expanding. The formation of ghost
forest happened most dramatically between 2011 and
2012, and although the rates of ghost forest formation
remain high (Fig. 6B), the total area of ghost forest on
the landscape declines after peaking in 2012. We address
the possible reasons for the observed timing of ghost for-
est formation in the discussion, but the direct hit of Hur-
ricane Irene in the study region is a notable factor.
DISCUSSION
Despite its protected status and the absence of any
timber management, more than 19,000 ha of freshwater
wetland forests have been lost from the Alligator River
NWR since 1985. We determined that 30.8% of the land-
scape in our study region has experienced a transition of
vegetation type within the last 35 yr. The magnitude of
forest cover lost is far greater than the amount of land
lost outright to open water over the same time period
(~1,150 ha). Forest loss was concentrated in a band
approximately 1 km inland from the coastal margin,
and was greater in extent on the eastern, more saline
shores of the Albemarle-Croatan-Pamlico Sound than
on the western side along the Alligator River estuary.
This forested wetland loss was not just on the shoreline,
with more than half of all forest loss occurring in the
interior of the refuge. Although the trajectory of land
loss in the study area was linear, forest loss was not.
Instead, most of the ghost forest formation occurred
within the last decade following a series of extreme
events including 5 yr of drought (2007–2011), a major
fire, and Hurricane Irene (both in 2011).
This analysis is the first reported attempt to use
remote sensing to map the spatial and temporal distribu-
tion of coastal ghost forests. We were successful at train-
ing a classification algorithm to detect this unique
transition state. Our analysis shows that ghost forests
are both a prevalent and a transient feature in the
Coastal Plain of North Carolina. We were surprised by
the rapidity with which formerly forested patches transi-
tioned through the ghost forest phase into shrublands or
marsh, with pixels typically classified as ghost forests for
no more than a few years. A portion (26.1%) of the
Deciduous
Pine
Shrub
Marsh
Other
Ghost
forest
Deciduous
Pine
Ghost f.
Shrub
Marsh
Water
1985 Transition
state
2019
(A) (B)
FIG. 3. (A) Map of study region areas that have undergone change since 1985. Blue denotes the conversion from land to open
water, red indicates transition through ghost forest at any point in time, dark gray are all other changing areas, and light gray indi-
cates no change. (B) For the changing pixels only, the alluvial plot shows the relative frequency of the various transition pathways.
Article e02339; page 6 EMILY A. URY ET AL. Ecological Applications
Vol. 0, No. 0

ghost forests detected earlier in the record (before 2010)
have returned to a forested state, but for most pixels the
ghost forest stage appears to represent a transition to
nonforested vegetation cover. Most of the ghost forests
have become shrublands. We do not yet know whether
the resulting increase in scrub-shrub wetlands represents
an expansion of native plant communities that are char-
acteristic of this habitat type or the creation of novel or
random assemblages. The impact of these widespread
ecosystem transitions on biodiversity and community
composition deserves further study. A portion of forests
becoming scrub-shrub does not transition through a
ghost forest stage, which we attribute to gradual tree
mortality and forest thinning without replacement
rather than mass tree mortality events that lead to large
numbers of standing dead trees. The shrubland land
cover type is rapidly expanding in area and represents a
major shift in ecosystem structure for which the conse-
quences for carbon storage and other ecosystem services
have not yet been quantified.
FIG. 4. Maps of spatial predictors (A) distance to the sound, (B) elevation above mean sea level, and (C) distance to a channel.
Histograms of the distribution of forested pixels (white), pixels where forest cover was lost (gray), and forest pixels that went
through a ghost forest state (red) across the same topographical parameters are shown in (D)–(F).
Xxxxx 2021 RAPID COASTAL WETLAND VEGETATION CHANGE Article e02339; page 7

If sea-level rise were the primary driver of forested
wetland loss, we would expect to see a linear rate of veg-
etation change and confinement of forest loss to shore-
lines. Instead, the spatial distribution of forested wetland
loss appears more consistent with saltwater intrusion,
and the timing suggests that extreme events are leading
to forest mortality. In this landscape of low topographic
relief and shallow groundwater, all hydrologically con-
nected areas are potentially vulnerable to saltwater
intrusion (Bhattachan et al. 2018, Zhang et al. 2018,
FIG. 5. Distance to the sound is the dominant controlling factor of ghost forest formation (red triangles) and forest loss (black
dots) followed by elevation and distance to channel (coefficient estimates shown with 95% confidence interval). Predictors with neg-
ative coefficients increase the riskof forest loss and ghost forest formation.
FIG. 6. (A) Changing areas of the Alligator River National Wildlife Refuge (pine and deciduous classes were combined to “for-
est,”and data were interpolated linearly across years missing data). The formation of ghost forest (red) shows a marked increase fol-
lowing Hurricane Irene. The Pains Bay Fire (summer 2011) left a large burn scar on the land that was visible in the 2012 imagery.
(B) Extent of ghost forest observed each year with the hashed area representing areas not classified as ghost forest in the previous
year and considered newly formed.
Article e02339; page 8 EMILY A. URY ET AL. Ecological Applications
Vol. 0, No. 0

2019). Marine salts move inland as a result of diffusion
and wind tides during droughts as well as during hurri-
cane storm surges (Manda et al. 2014, 2018, Tully et al.
2019). The greatest period of ghost forest formation
(2011–2012) followed Hurricane Irene, which arrived
after nearly 5 yr of drought. Although vegetation may
still be in the recovery phase following these more recent
disturbances, some areas may not recover, having been
pushed too far into a new state. The combination of
these two stressors may have been responsible for the
increase in ghost forest formation, an interaction effect
that may become more prevalent in the future as climate
change is leading to an increased frequency of extreme
events (Meehl et al. 2007, Miao et al. 2009).
Our study area represents only a small fraction of the
97,000 km
2
of palustrine forested wetland in the coastal
plain of eastern North America (National Oceanic and
Atmospheric Administration [NOAA] 2010). By defini-
tion, these are low-lying coastal forests that are vulnera-
ble to rising sea levels, saltwater intrusion, and coastal
storm damage. If the rates of change observed in this
study are consistent across the entire North Atlantic
Coastal Plain, climate deforestation could have already
resulted in a tremendous loss of these unique ecosystems
from North America. Field surveys have already docu-
mented dramatic losses of freshwater forested wetlands
in the region (Desantis et al. 2007, White and Kaplan
2017, Kirwan 2019, Taillie et al. 2019a, Ury et al. 2020).
There is no reason to expect the east coast of North
America to be an outlier in the vulnerability of its
coastal forested wetlands. Indeed, a recent analysis sug-
gested that parts of northern Europe, Chile, Australia,
and particularly southeast Asia will experience the most
severe effects of climate-driven losses of coastal plain
wetlands (Blankespoor et al. 2014).
The extensive transformation of coastal plain vegetation
we observed has taken place despite the protected status of
the refuge and active management to maintain endemic
and rare plant communities (Laderman 1989, Poulter et al.
2008). In many other coastal plains, historic drainage and
deforestation have confined endemic forested wetlands to
the limited areas where drainage is very costly. These rem-
nant wetland forests, which maybe the last stronghold and
seedbanks for threatened species such as the Atlantic white
cedar (Chamaecyparis thyoides) (Laderman 1989), may
now be especially vulnerable because of their landscape
position. Already the loss of forested freshwater wetlands
in coastal North Carolina has been linked to declines in
bird diversity (Taillie et al. 2019b).
Because freshwater forested wetlands are known to be
important global carbon sinks (Brinson 1991), their con-
version to shrublands may create positive feedbacks to
climate change (Duarte et al. 2013, Spivak et al. 2019).
Forested wetlands have far higher aboveground biomass
than the shrublands that replace them (Kirwan and
Gedan 2019). The effects of salinization on the carbon
stocks stored in freshwater wetland sediments are less
certain. Recent experiments in the Florida Everglades
showed that salinization led to rapid declines in soil ele-
vation and large soil carbon losses in freshwater and
brackish wetlands (Charles et al. 2019), and gradient
studies in North Carolina documented reductions in soil
respiration and methane emissions during periods of
saltwater intrusion (Helton et al. 2019).
We suggest that coastal forested wetland loss repre-
sents an important new class of climate change–driven
ecological shifts akin to the desertification of drylands
(Huang et al. 2016, Berdugo et al. 2020) and the shrubi-
fication of arctic tundra (Myers-Smith and Hik 2018,
Pastick et al. 2019). Like these other examples, rising
water tables and salinization ultimately will shift these
formerly freshwater forests into a new ecosystem state
(sensu Scheffer 2009), because salinization and sulfida-
tion of soils prevent the regeneration of shrubland back
into coastal forests (Munns and Tester 2008, Fagherazzi
et al. 2019).
Efforts to forecast the potential for global coastal wet-
land loss have focused almost entirely on sea-level rise.
Schuerch et al. (2018) estimated up to 30% of all coastal
wetlands could be subsumed by the sea by 2100. Other
recent papers suggest that an even larger proportion of
coastal wetlands are vulnerable (Kirwan et al. 2016,
Spencer et al. 2016). When we consider the likelihood of
wetland loss in the coastal plain interior due to ecosys-
tem salinization, these estimates are likely to increase in
magnitude. Perhaps more importantly, considering both
salinization and saltwater intrusion will require us to
expand our perception of the range of coastal wetland
types that are vulnerable to climate change. Our analysis
shows that ghost forest formation can be associated with
both episodic disturbances (i.e., hurricanes) and the
gradual progression of sea-level rise and that landscape
features such as drainage ditches can increase local vul-
nerability. We suggest that an analysis of coastal freshwa-
ter wetland vulnerability at a global scale is urgently
needed. In addition to determining what proportion of
interior freshwater forested wetland habitats have already
been lost, new remote sensing technologies may enable us
to detect vegetation stress and vulnerability prior to forest
mortality, that could guide more effective and strategic
protection and management.
ACKNOWLEDGMENTS
Funding for this research was provided by a NASA Earth
and Space Science Fellowship (grant 80NSSC17K0355), the
North Carolina Sea Grant/Space Grant (project R/MG-1806),
NSF Coastal SEES Collaborative Research (award 1426802),
and by Duke University Data+Program through the Rhodes
Information Initiative. We thank D. Urban for help with spatial
statistics; our student interns and research assistants, K. Chang,
H. Lynch, J. Reeves, and A. Kosinski; and our collaborators at
the Alligator River National Wildlife Refuge, especially S.
Lanier. Author contributions: EU, JW, and EB conceived of the
ideas and project design, EU and XY performed the remote
sensing analysis, EU and EB led the writing of the manuscript,
and all authors contributed substantially to revising the writing
and gave final approval for publication.
Xxxxx 2021 RAPID COASTAL WETLAND VEGETATION CHANGE Article e02339; page 9

LITERATURE CITED
Adam, E., O. Mutanga, and D. Rugege. 2010. Multispectral
and hyperspectral remote sensing for identification and map-
ping of wetland vegetation: A review. Wetlands Ecology and
Management 18:281–296.
Asner, G. P. 1998. Biophysical and biochemical sources of vari-
ability in canopy reflectance. Remote Sensing of the Environ-
ment 64:234–253.
Berdugo, M., et al. 2020. Global ecosystem thresholds driven
by aridity. Science 367:787–790.
Bhattachan, A., R. E. Emanuel, M. Ard
on, E. S. Bernhardt, S.
M. Anderson, M. G. Stillwagon, E. A. Ury, T. K. BenDor,
and J. P. Wright. 2018. Evaluating the effects of land-use
change and future climate change on vulnerability of coastal
landscapes to saltwater intrusion. Elementa: Science of the
Anthropocene 6:62.
Bivand, R., T. Keitt, and B. Rowlingson 2019. rgdal: bindings
for the ‘geospatial’data abstraction library. R package ver-
sion 1.4-4. https://CRAN.R-project.org/package=rgdal
Blankespoor, B., S. Dasgupta, and B. Laplante. 2014. Sea-level
rise and coastal wetlands. Ambio 43:996–1005.
Bojanowski, M., and R. Edwards. 2016. alluvial: R package for
creating alluvial diagrams. R package version 0.1-2. https://
github.com/mbojan/alluvial
Breiman, L. 2001. Random forests. Machine Learning 45:5–32.
Brinson, M. M. 1991. Landscape properties of pocosins and
associated wetlands. Wetlands 1:441–465.
Charles, S. P., et al. 2019. Experimental saltwater intrusion
drives rapid soil elevation and carbon loss in freshwater and
brackish everglades marshes. Estuaries and Coasts 42:1868–
1881.
Church, J. A., et al. 2013. Sea level change. Pages 895–900 in T.
F. Stocker, D. Qin, G.-K. Plattner, M. Tignor, S. K. Allen, J.
Boschung, A. Nauels, Y. Xia, V. Bex, and P. M. Midgley, edi-
tors. Climate change 2013: the physical science basis. Contri-
bution of working group I to the fifth assessment report of
the intergovernmental panel on climate change. Cambridge
University Press, Cambridge, UK.
Cloern, J. E., et al. 2016. Human activities and climate variabil-
ity drive fast-paced change across the world’s estuarine–
coastal ecosystems. Global Change Biology 22:513–529.
Desantis, L. R. G., S. Bhotika, K. Williams, and F. E. Putz.
2007. Sea-level rise and drought interactions accelerate forest
decline on the Gulf Coast of Florida, USA. Global Change
Biology 13:2349–2360.
Douglas, B. C., and W. Richard Peltier. 2002. The puzzle of glo-
bal sea-level rise. Physics Today 55:35–40.
Duarte, C. M., I. J. Losada, I. E. Hendriks, I. Mazarrasa, and
N. Marb
a. 2013. The role of coastal plant communities for
climate change mitigation and adaptation. Nature Climate
Change 3:961–968.
Fagherazzi, S., S. C. Anisfeld, L. K. Blum, E. V. Long, R. A.
Feagin, A. Fernandes, W. S. Kearney, and K. Williams. 2019.
Sea level rise and the dynamics of the marsh-upland bound-
ary. Frontiers in Environmental Science 7:1–18.
Haer, T., E. Kalnay, M. Kearney, and H. Moll. 2013. Relative sea-
level rise and the conterminous United States: Consequences of
potential land inundation in terms of population at risk and
GDP loss. Global Environmental Change 23:1627–1636.
Helton, A. M., M. Ard
on, and E. S. Bernhardt. 2019. Hydro-
logic context alters greenhouse gas feedbacks of coastal wet-
land salinization. Ecosystems 22:1108–1125.
Hijmans, R. J. 2019. raster: Geographic data analysis and mod-
eling. R package version 3.0-2. https://CRAN.R-project.org/
package=raster
Hillebrand, H., and C. Kunze. 2020. Meta-analysis on pulse dis-
turbances reveals differences in functional and compositional
recovery across ecosystems. Ecology Letters 23:575–585.
Holden, J., P. J. Chapman, and J. C. Labadz. 2004. Artificial
drainage of peatlands: hydrological and hydrochemical pro-
cess and wetland restoration. Progress in Physical Geography:
Earth and Environment 28:95–123.
Huang, J., H. Yu, X. Guan, G. Wang, and R. Guo. 2016. Accel-
erated dryland expansion under climate change. Nature Cli-
mate Change 6:166–171.
Kennedy, R. E., et al. 2014. Bringing an ecological view of
change to Landsat-based remote sensing. Frontiers in Ecol-
ogy and the Environment 12:339–346.
Kirwan, M. L., and K. B. Gedan. 2019. Sea-level driven land
conversion and the formation of ghost forests. Nature Cli-
mate Change 9:450–457.
Kirwan, M. L., S. Temmerman, E. E. Skeehan, G. R. Guntensper-
gen, and S. Fagherazzi. 2016. Overestimation of marsh vulnera-
bility to sea level rise. Nature Climate Change 6:253–260.
Laderman, A. D. 1989. The ecology of Atlantic white cedar
wetlands: a community profile. Fish and Wildlife Service,
U.S. Department of the Interior, Washington, DC, USA.
Liaw, A., and M. Wiener. 2002. Classification and regression by
randomForest. R News 2:18–22.
Manda, A. K., A. S. Giuliano, and T. R. Allen. 2014. Influence
of artificial channels on the source and extent of saline water
intrusion in the wind tide dominated wetlands of the southern
Albemarle estuarine system (USA). Environmental Earth
Sciences 71:4409–4419.
Manda, A. K., E. Reyes, and J. M. Pitt. 2018. In situ measure-
ments of wind-driven salt fluxes through constructed chan-
nels in a coastal wetland ecosystem. Hydrological Processes
32:636–643.
Meehl, G. A., J. M. Gregory, A. Kitoh, R. Knutti, J. M. Mur-
phy, A. Noda, and Z.-C. Zhao. 2007. Global climate projec-
tions. Pages 747–845 in S. Solomon, D. Qin, M. Manning, Z.
Chen, M. Marquis, K. B. Averyt, M. Tignor, and H. L.
Miller, editors. Climate change 2007: the physical science
basis. Contribution of working group I to the fourth assess-
ment report of the intergovernmental panel on climate
change. Cambridge University Press, Cambridge, UK. http://
hdl.handle.net/102.100.100/124551?index=1
Miao, S., C. B. Zou, and D. D. Breshears. 2009. Vegetation
responses to extreme hydrological events: sequence matters.
American Naturalist 173:113–118.
Munns, R., and M. Tester. 2008. Mechanisms of salinity toler-
ance. Annual Review of Plant Biology 59:651–681.
Myers-Smith, I. H., and D. S. Hik. 2018. Climate warming as a
driver of tundra shrubline advance. Journal of Ecology
106:547–560.
National Oceanic and Atmospheric Administration, Office for
Coastal Management. 2010. “C-CAP regional land cover and
change”coastal change analysis program (C-CAP) high-reso-
lution land cover. NOAA Office for Coastal Management,
Charleston, South Carolina. www.coast.noaa.gov/ccapftp
National Weather Service Advanced Hydrologic Prediction Ser-
vice. 2016. Pamlico Sound at Oregon Inlet Marina. https://wa
ter.weather.gov/ahps2/hydrograph.php?wfo=mhx&gage=orin7
Pastick, N. J., M. T. Jorgenson, S. J. Goetz, B. M. Jones, B. K.
Wylie, B. J. Minsley, H. Genet, J. F. Knight, D. K. Swanson,
and J. C. Jorgenson. 2019. Spatiotemporal remote sensing of
ecosystem change and causation across Alaska. Global
Change Biology 25:1171–1189.
Pekel, J. F., A. Cottam, N. Gorelick, and A. S. Belward. 2016.
High-resolution mapping of global surface water and its
long-term changes. Nature 540:418–422.
Article e02339; page 10 EMILY A. URY ET AL. Ecological Applications
Vol. 0, No. 0

Poulter, B. 2005. Interactions between landscape disturbance
and gradual environmental change: Plant community migra-
tion in response to fire and sea-level rise. Dissertation. Duke
University, Durham, North Carolina.
Poulter, B., J. L. Goodall, and P. N. Halpin. 2008. Applications
of network analysis for adaptive management of artificial
drainage systems in landscapes vulnerable to sea level rise.
Journal of Hydrology 357:207–217.
R Development Core Team 2013. R: A language and environ-
ment for statistical computing. R Foundation for Statistical
Computing, Vienna, Austria. http://www.R-project.org/
Scheffer, M. 2009. Critical transitions in nature and society.
Princeton University Press, Princeton, NJ, USA.
Schieder, N. W., and M. L. Kirwan. 2019. Sea-level driven accel-
eration in coastal forest retreat. Geology 47:1151–1155.
Schuerch, M., et al. 2018. Future response of global coastal
wetlands to sea-level rise. Nature 561:231–234.
Spencer, T., M. Schuerch, R. J. Nicholls, J. Hinkel, D. Lincke,
A. t. Vafeidis, R. Reef, L. McFadden, and S. Brown. 2016.
Global coastal wetland change under sea-level rise and
related stresses: The DIVA Wetland Change Model. Global
and Planetary Change 139:15–30.
Spivak, A. C., J. Sanderman, J. L. Bowen, E. A. Canuel, and C.
S. Hopkinson. 2019. Global-change controls on soil-carbon
accumulation and loss in coastal vegetated ecosystems. Nat-
ure Geoscience 12:685–692.
Taillie, P. J., C. E. Moorman, B. Poulter, M. Ard
on, and R. E.
Emanuel. 2019a. Decadal-scale vegetation change driven by salin-
ity at leading edge of rising sea level. Ecosystems 22:1918–1930.
Taillie, P. J., C. E. Moorman, L. S. Smart, and K. Pacifici.
2019b. Bird community shifts associated with saltwater expo-
sure in coastal forests at the leading edge of rising sea level.
PLoS One 14:e0216540.
Tully, K., et al. 2019. The invisible flood: the chemistry, ecology,
and social implications of coastal saltwater intrusion. BioS-
cience 69:368–378.
Ury, E. 2021. Rapid deforestation of a coastal landscape driven
by sea level rise and extreme events—classified_maps (Geo-
TIFF). Figshare, data set. https://doi.org/10.6084/m9.figsha
re.14249912.v1
Ury, E. A., S. M. Anderson, R. K. Peet, E. S. Bernhardt, and J.
P. Wright. 2020. Succession, regression and loss: does evi-
dence of saltwater exposure explain recent changes in the tree
communities of North Carolina’s Coastal Plain? Annals of
Botany 125:255–264.
U.S. Fish and Wildlife Service Southeast Region Fire Manage-
ment Organization. 2011. FY2011 fire division report. https://
www.fws.gov/southeast/pdf/report/fire-report-2011.pdf
U.S. Geological Survey. 2019. EROS archive—aerial photogra-
phy—high resolution orthoimagery (HRO). https://doi.org/
10.5066/F73X84W6
Ustin, S. L. 2013. Remote sensing of canopy chemistry. Pro-
ceedings of the National Academy of Sciences of the United
States of America 110:804–805.
Ustin, S. L., A. A. Gitelson, S. Jacquemoud, M. Schaepman, G.
P. Asner, J. A. Gamon, and P. Zarco-tejada. 2009. Remote
sensing of environment retrieval of foliar information about
plant pigment systems from high resolution spectroscopy.
Remote Sensing of Environment 113:S67–S77.
Wasson, K., A. Woolfolk, and C. Fresquez. 2013. Ecotones as
indicators of changing environmental conditions: rapid
migration of salt marsh-upland boundaries. Estuaries and
Coasts 36:654–664.
White, E., and D. Kaplan. 2017. Restore or retreat? Saltwater
intrusion and water management in coastal wetlands. Ecosys-
tem Health and Sustainability 3:e01258.
Wickham, H., et al. 2019. Welcome to the tidyverse. Journal of
Open Source Software 4:1686.
Yang, L., et al. 2018. A new generation of the United States
National Land Cover Database: Requirements, research
priorities, design, and implementation strategies. ISPRS
Journal of Photogrammetry and Remote Sensing 146:108–
123.
Zhang, Y., W. Li, G. Sun, and J. S. King. 2019. Coastal wetland
resilience to climate variability: A hydrologic perspective.
Journal of Hydrology 568:275–284.
Zhang, Y., W. Li, G. Sun, G. Miao, A. Noormets, R. Emanuel,
and J. S. King. 2018. Understanding coastal wetland hydrol-
ogy with a new regional-scale, process-based hydrological
model. Hydrological Processes 32:3158–3173.
SUPPORTING INFORMATION
Additional supporting information may be found online at: http://onlinelibrary.wiley.com/doi/10.1002/eap.2339/full
DATA AVAILABILITY
Data (Ury 2021) are available in Figshare. https://doi.org/10.6084/m9.figshare.14249912 Source Landsat images are available
from the Earth Engine Data Catalog. https://developers.google.com/earth-engine/datasets/catalog/LANDSAT_LT05_C01_T1_SR;
https://developers.google.com/earth-engine/datasets/catalog/LANDSAT_LE07_C01_T1_SR; https://developers.google.com/earth-
engine/datasets/catalog/LANDSAT_LC08_C01_T1_SR.
Xxxxx 2021 RAPID COASTAL WETLAND VEGETATION CHANGE Article e02339; page 11