Journal of
Marine Science
and Engineering
Article
Topological and Morphological Controls on Morphodynamics
of Salt Marsh Interiors
Ben R. Evans 1 , Iris Möller 1,2
1
2
*
Citation: Evans, B.R.; Möller, I.;
Spencer, T. Topological and
Morphological Controls on
Morphodynamics of Salt Marsh
Interiors. J. Mar. Sci. Eng. 2021, 9, 311.
https://doi.org/10.3390/jmse9030311
and Tom Spencer 1, *
Cambridge Coastal Research Unit, Department of Geography, University of Cambridge, Downing Place,
Cambridge CB2 3EN, UK; bre24@cam.ac.uk (B.R.E.); moelleri@tcd.ie (I.M.)
Department of Geography, Museum Building, Trinity College Dublin, Dublin 2, Ireland
Correspondence: ts111@cam.ac.uk
Abstract: Salt marshes are important coastal environments and provide multiple benefits to society.
They are considered to be declining in extent globally, including on the UK east coast. The dynamics
and characteristics of interior parts of salt marsh systems are spatially variable and can fundamentally
affect biotic distributions and the way in which the landscape delivers ecosystem services. It is
therefore important to understand, and be able to predict, how these landscape configurations may
evolve over time and where the greatest dynamism will occur. This study estimates morphodynamic
changes in salt marsh areas for a regional domain over a multi-decadal timescale. We demonstrate
at a landscape scale that relationships exist between the topology and morphology of a salt marsh
and changes in its condition over time. We present an inherently scalable satellite-derived measure
of change in marsh platform integrity that allows the monitoring of changes in marsh condition.
We then demonstrate that easily derived geospatial and morphometric parameters can be used
to determine the probability of marsh degradation. We draw comparisons with previous work
conducted on the east coast of the USA, finding differences in marsh responses according to their
position within the wider coastal system between the two regions, but relatively consistent in
relation to the within-marsh situation. We describe the sub-pixel-scale marsh morphometry using a
morphological segmentation algorithm applied to 25 cm-resolution maps of vegetated marsh surface.
We also find strong relationships between morphometric indices and change in marsh platform
integrity which allow for the inference of past dynamism but also suggest that current morphology
may be predictive of future change. We thus provide insight into the factors governing marsh
degradation that will assist the anticipation of adverse changes to the attributes and functions of
these critical coastal environments and inform ongoing ecogeomorphic modelling developments.
Academic Editor: Achilleas Samaras
Received: 16 February 2021
Accepted: 2 March 2021
Keywords: wetland; salt marsh; degradation; satellite time series; self-organisation; morphodynamic
feedback; geospatial
Published: 11 March 2021
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affiliations.
Copyright: © 2021 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
1. Introduction
Salt marshes represent a major component of low-lying sedimentary coastal systems
and occur across the world [1]. Over recent decades, salt marshes have attracted increasing
attention, with much research being focused on the services and functions they provide [2].
Salt marshes exhibit an extremely high biodiversity [3] and primary productivity [4]. They
attenuate wave energy and contribute significantly to the protection provided by natural foreshores from high waves and water levels threatening coastal communities [5,6].
Marshes are a sink for atmospheric carbon [7–9], while providing a habitat for many endangered or threatened species and nursery areas for commercial fish stock species [10,11].
They also have cultural value as areas for recreation and tourism. In the UK, over the
last 150 years map and aerial imagery suggest an expansion of marsh areas in northern
England, as against areal loss in the south, attributed to regional variations in sediment
supply [12]. It has, however, proven difficult to ascertain patterns of marsh areal change
and controlling factors over the latter half of the twentieth century. The East Anglian
J. Mar. Sci. Eng. 2021, 9, 311. https://doi.org/10.3390/jmse9030311
https://www.mdpi.com/journal/jmse
J. Mar. Sci. Eng. 2021, 9, 311
2 of 22
coast (Figure 1) is thought to be a region with high rates of wetland loss (e.g., [13]), but
in reality rates and types of marsh loss have exhibited great spatial variability (e.g., [14]).
Furthermore, achieving precise estimates of changes in marsh area has been shown to be
challenging by the few large-scale UK inventories attempted [15–17].
The position of a marsh system within a broader context of the coastal zone exerts
controls on factors such as sediment or nutrient import and export, tidal flushing and
residence times, and forces exerted by tidal currents. The connectivity between intertidal
wetland areas and offshore deep channel zones is crucial to water, sediment, and nutrient
exchange and thus to the morphological evolution of marshes [18].
The position of a marsh parcel within the surrounding system, and the position of
a point on the marsh within individual parcels, modulate local-scale responses. There
has long been recognition that elevation–sedimentation relationships vary with scale. For
example, single-tide sediment deposition decreases with distance from tidal channels,
while surface elevation provides an important marsh-wide control over annual to decadal
timescales [19]. Kearney and Rogers [20] previously used logistic regression to predict
internal platform integrity changes in marshes at a regional scale in Chesapeake and
Delaware Bays, USA. They demonstrated empirical relationships between changes in
marsh surface condition and factors such as distance up-estuary and position within a
marsh parcel. The marsh systems on the UK East coast have a different context (tidal range,
position within tidal frame, sedimentology, vegetation community) to those studied by
Kearney and Rogers [20] but we are able to draw comparisons between the topological
relationships presented here and their findings.
The morphology of the marsh, described by the spatial distribution of landscape units
(at a scale of metres) such as vegetated platforms, salt pans, creeks, and large channels, is
the integrative result of historic morphodynamics [21]. Morphology is also thought to exert
a control on future changes through biogeomorphic feedback [22], while the interactions
between topography and hydrodynamic forces have been extensively explored from a
numerical perspective [23]. From an empirical perspective, a relationship between the
functional form of marsh margins and erosion rates has been demonstrated [24].
Of particular, and most immediate, interest for landscape management are the loci of
greatest dynamism; the most important to understand when considering future ecosystem
service provision are those exhibiting (or likely to exhibit) erosion. This study aims to
assess the decadal morphodynamics of salt marsh systems on the east coast of the UK,
evaluate the role of topological and morphological factors in determining the observed
changes, and provide understanding of the controls on salt marsh morphological evolution
to support ecosystem management and the development of models that combine ecological
and physical functioning. We present landscape-scale statistical models relating such
changes to easily derived spatial parameters at scales from the regional (tens of kilometres)
to local (metres). Such understanding will help with the prediction of locations likely to
exhibit degradation accompanied by concomitant losses of ecosystem services. This will
thereby facilitate targeted management interventions to protect the ecological and physical
functioning of these important coastal ecosystems.
2. Materials and Methods
2.1. Study Area
The coastline of East Anglia, UK, is bounded by the Humber estuary to the north and
the Thames estuary to the south (Figure 1). The region is, in many parts, densely populated.
The population in the coastal districts of Suffolk and Norfolk exceeds 600,000 (Office for
National Statistics GB 2016 census data—www.ons.gov.uk/census accessed on 23 October
2018). A considerable amount of variability occurs along this stretch of coast in terms of
the hydrodynamic and sedimentary contexts. The mean spring tidal range (MSTR) varies
from 6.18 m at Immingham on the Humber to a minimum of 1.94 m at Lowestoft, Suffolk,
before increasing again further south (ntslf.org accessed on 10 December 2020).
J. Mar. Sci. Eng. 2021, 9, 311
3 of 22
Figure 1. Study region of East Anglia, UK, showing areas where marsh is present [17] and indicating
locations referenced in the text. Blue dashed lines denote the boundaries of ISCE units after [25]. Red
dots denote standard ports with tide gauge records.
Pethick and Leggett [25] partitioned the region into three Integrated Scale Coastal
Evolution (ISCE) units. The Northern ISCE includes the eroding glacial cliffs north of the
Humber, which provide sediment inputs for the sandy shorelines in Lincolnshire and the
infilling embayment of The Wash. The ISCE includes the spit and barrier island system
of the north Norfolk coast as far east as Cromer. The second ISCE is dominated by cliffs
composed of glacial sands and gravels and lies between Cromer and Thorpeness. The third,
southerly, unit comprises numerous estuaries and inlets sitting within a large embayment
and characterised by silt or clay sediments [26]. The region contains some 14,406 ha of
salt marsh [17] of diverse character and setting; open coast, embayment, back-barrier, and
estuarine marshes are represented [27]. Suffolk and Essex have both experienced a net loss
of marsh area in the second half of the twentieth century [15] while The Wash embayment,
between Norfolk and Lincolnshire, represents a long-term sediment sink [28] and continues
to infill [26], with attendant increases in marsh area [24]. At the regional scale, none of
these systems can be thought of as ’sediment starved’.
2.2. Morphological Change
We address controls on morphodynamic change within marsh interiors comprising the
complex morphologies of vegetated surfaces, creeks, pools, and pans that lie landward of
the seaward margin of the marsh as a whole. We use satellite imagery to estimate changes
in vegetation distributions within these areas, from which we infer morphological changes.
This inference is based on well-established elevational controls on intertidal vegetation
establishment [29], which have been thoroughly ground-referenced for NW Europe by
J. Mar. Sci. Eng. 2021, 9, 311
4 of 22
Suchrow and Jensen [30]. In inferring morphological change, we assume that all surfaces
of sufficient elevation to support vegetation become colonised, while those that are too low
do not.
Changes in the extent of vegetation within marsh interiors were estimated using a
modified trend analysis of the Landsat archive, which represents the longest appropriate
satellite time series available (1984–present). Imagery dating between 1985 and 2016 was
used in this study. The exact time period over which the metric is calculated will vary
slightly between pixels due to the different acquisition dates of imagery over certain areas
and cloud cover precluding the use of some pixels (see https://osf.io/mgsyz/).
A metric was derived, denoted δPveg , that can be interpreted as representing the percentage change in vegetation cover within any given pixel over the timeframe of the satellite
observations. As such, it reflects the temporal variations in the areal unvegetated–vegetated
marsh ratio (UVVR) within each 30 m by 30 m pixel. The UVVR could be considered a
geomorphic metric that has itself been shown to be related to marsh vulnerability [31]. The
methodological workflow we use is summarised in Figure 2 and is detailed extensively in
Appendix A.
Figure 2. Summary of workflow used to derive δPveg metric.
In Google Earth Engine [32], the Normalised Difference Vegetation Index (NDVI), a
proxy for the amount of chlorophyll (and therefore vegetation) present, was computed for
all scenes within the Landsat archive for the study region up to 2016. Conceptually, the
NDVI for a pixel in a satellite image can be considered to be a function of the percentage
of that pixel covered by vegetation and the nature of the vegetation within that pixel.
A change in the NDVI reflects a change in the percentage vegetation cover within the
pixel, with the signal potentially being modulated by any attendant changes in community
composition (and therefore spectral signature) and the vigour of the vegetation present.
Lopes et al. [33] evaluted a number of vegetation indices for the monitoring of salt marsh
extent and condition in Portugal based on the Landsat archive and concluded that NDVI
performed best, with a seasonally varying goodness-of-fit [34] typically exceeding 0.9 at
one location and 0.75 at another, where a perfect classification would result in a value of 1.
NDVI values were cross-calibrated to account for differences in radiomemtric response
between different Landsat sensors before a linear trend was computed over the time series
for each pixel. A relationship between the coefficients of the trend and the mean NDVI over
J. Mar. Sci. Eng. 2021, 9, 311
5 of 22
the time series was found, which was removed by taking the residuals of the regression fit
between the slope of the trend line and the mean NDVI (R2 = 0.22, p ≤ 0.001). The resulting
residuals were standardised and calibrated against 25cm Environment Agency aerial
photography from 1992 and 2013/14. Calibration was achieved by the visual assessment of
the aerial photography subsets extracted for a random sample of satellite pixels (n = 83)
stratified to represent the full range of observed standardised residual trend coefficents.
For each satellite pixel area, 50 points were randomly distributed and each was manually
classified as either coinciding with vegetated or unvegetated surfaces in both the 1992
and 2013/14 aerial photography. For each year, the percentage of satellite pixel area that
was vegetated in the aerial photography was then calculated based on the counts in each
surface cover class. A strong relationship was found between the satellite-derived residuals
and the photography-based estimates of change in vegetation cover (R2 = 0.87, p ≤ 0.001),
implying that changes in vegetation extent dominate changes in δPveg . The calibration
sample was drawn from the entire study region. δPveg varies in the range −100% to
+100%, where −100% represents a change from fully vegetated to fully bare and +100% the
opposite. The method was validated against a sample of pixels representative of the full
range of calibrated estimates (n = 100) but drawn from a local subset area (Hamford Water,
Essex) to obviate the effect of any large-scale (e.g., latitudinal) signal that may have been
present within the calibration. The δPveg metric performed well, with an RMSE of 11.9% of
full scale (±100%).
2.3. Topological and Morphological Metrics
Two metrics were derived to describe the topology of a location within a marsh and
two morphological metrics were developed to describe the distribution of landscape units
within a single Landsat pixel area.
2.3.1. Geomorphic Setting
The geomorphic setting of the marsh, meaning its context within the wider coastal
system, was represented as a cost function, denoted Cost Distance (CD), describing how
difficult it would be for a parcel of water originating offshore to reach any given location
within a marsh. Thus, for example, interior marsh areas towards the head of estuaries are
more ’costly’ to reach (less well connected to the offshore) than the seaward margins of
open coast marshes. Without a full simulation of tidal exchanges against which to calibrate
flow pathways, the CD metric is not expected to behave in an isomorphic manner with
actual tidal exchanges. Rather, CD values are internally consistent and represent a scale of
relative connectivity within the study domain.
The area between the 10 km offshore limit and the land was represented on a 10 m grid
and divided into four zones. For ease, we denote these ‘subtidal’, ‘intertidal’, ’supratidal’,
and ’terrestrial’, while recognising that they do not conform strictly to these descriptions.
Intertidal, for example, would ordinarily refer to elevations between the lowest and highest
tides experienced. In the UK, this range would extend well below 0 m Ordnance Datum
Newlyn (ODN), which approximates mean sea level but is used here as the lower bound of
the ‘intertidal’ zone (see below). The landward limit was defined by the UK Environment
Agency Second Generation Shoreline Management Plan as segments vector layer(SMP2),
which typically reflects the line of engineered defences. Costs to traverse each cell of the
grid were based on distances, with the cells in each zone given different weighting factors.
Areas below 0 m ODN were denoted ’subtidal’ and were assigned a cost according to their
euclidean distance in metres from the offshore boundary.
An ‘intertidal’ zone was defined as areas between 0 m ODN and the level of the
highest astronomical tide (HAT). Elevations were derived from the Environment Agency 2
m-resolution LiDAR composite product (2008) resampled to the 10 m grid. Since spatially
resolved data describing HAT were not available, modelled MSTR values [35] were extrapolated to intersect the shoreline and an 0.7× MSTR was used to approximate the level of
HAT. Our comparison of the levels thus estimated and known levels at the four standard
J. Mar. Sci. Eng. 2021, 9, 311
6 of 22
ports in the study region (Immingham, Cromer, Lowestoft, and Felixstowe) suggests that
this approximation tends to slightly overestimate (order of 10 cm) the level of HAT. The
dependency of intertidal wetland development on tidal range is well documented, with
equivalent landforms occurring at higher levels relative to the mean sea level in macrotidal
settings than in microtidal ones [27,36]. The elevation range available for the development
of intertidal landforms can be described as an ‘accommodation range’. The DEM elevations
above 0 m ODN were therefore normalised to reflect the spatial variability in accommodation range, which was itself approximated as 0.5× MSTR. This produced a normalised
elevation (NE) raster that did not exhibit substantial dependency on tidal range and was
used to scale the resulting cost functions for the intertidal zone.
A ‘supratidal’ zone was defined as areas above HAT but seaward of the SMP2 vector
marking the terrestrial limit. Such areas may still be inundated and therefore permit water
flow during exceptional events, so this zone was given a uniform but very high cell cost
relative to the other zones. A ‘terrestrial’ zone landward of engineered defences was
modelled as impassable for tidal waters.
The zones were combined to produce an overall cost surface raster which was converted to a cost-distance raster where each cell value represents the cumulative cost to
reach it by the least-cost path from the 10 km offshore limit of the domain. The weightings
for each zone were manually adjusted to produce the best achievable visual replication of
expected flow routes for tidal waters to reach a given location through complex channel
networks. Figure 3 shows a schematic cross-shore transect from the engineered defence
to the 10 km offshore limit of the domain (top panel). Per-cell and cumulative costs are
plotted in red, with the levels of 0 m ODN and HAT also indicated. The four zones upon
which the overall cost surface was based are indicated along with their final weighting
functions (where D is the horizontal distance in metres and NE is the normalised elevation).
The lower panel of Figure 3 shows examples of the least-cost paths calculated for random
pixels superimposed on both the local per-pixel cost surface and aerial photography to
demonstrate the ability of the method to reproduce expected flow routes.
Terrestrial
Impassable (NaN)
Supratidal
Intertidal
Subtidal
D x NE x 500
D x 0.001
,
Figure 3. Schematic representation of topography and associated costs (top). The four zones and
their associated cost weighting functions are identified. Example from Hamford Water, Essex, of
per-cell costs (bottom left) and aerial photograph for reference (bottom right). Pink lines on maps
are example least-cost paths derived for a random selection of pixels. See Figure 1 for location.
J. Mar. Sci. Eng. 2021, 9, 311
7 of 22
2.3.2. Distance from Creek/Edge of Marsh Parcel
This measure represents the well-established tendency for sediments to be deposited
rapidly once creek banks become overtopped during inundation, leading to the development of creek levees and limited deposition rates in areas away from cliffed seaward
margins [37,38]. To represent this dynamic, a measure of distance from a creek or the
edge of a marsh parcel was used. To facilitate analysis at scales commensurate with the
size of Landsat pixels used for other variables (30 m by 30 m), a reduced resolution product representing offshore or large-creek margins of marsh parcels was derived from the
Environment Agency’s salt marsh extent layer [17]. Large creeks were defined as those
wider than the diagonal of a Landsat pixel (ca. 42.5 m), since these are the units for which
internal morphological changes were measured. Creeks and pools narrower than 42.5 m
were removed using a 21.25 m outer buffer on the polygons of the salt marsh extents layer,
followed by the dissolving of any overlaps thus created and 21.25 m inner buffer operations.
This spatial resolution limit, imposed by the buffering process, effectively disregards the
smaller creeks, which nevertheless perform important system functions. This fact must
be taken into account when interpreting the findings, and we comment further on this
aspect in the discussion. The other metrics used, including δPveg , are not sensitive to spatial
resolution limits in the same manner.
Landward limits of the marsh polygons were removed by extracting only those portions of perimeter lines that were greater than 30 m (1 landsat pixel) from the SMP2 vector,
ensuring that distances do not represent those from the landward limit. For each parcel,
the euclidean distance from the edge was then calculated on a 30 m grid aligned to that
used for δPveg estimation. The metric is henceforth referred to as Euclidean Distance (ED).
2.3.3. Integrity of Marsh Platform
The first morphological metric describes the integrity of the marsh platform by the
proportion of each 30 m satellite pixel’s area that contains undissected vegetated marsh
areas. It is referred to as pCore.
The Morphological Spatial Pattern Analysis (MSPA—[39]) tool, supplied within the
GUIDOS toolbox [40], is a morphological segmentation algorithm designed to assess
habitat connectivity and fragmentation—for example, in forest ecosystems [41,42] or for
planning of green infrastructure [43]. MSPA operates on a binary raster of ‘foreground’ and
‘background’ pixels, where foreground represents the habitat or land cover type of interest.
It classifies foreground pixels according to how many other foreground pixels they are
adjacent to and a distance parameter controlling the width of areas considered to be ‘egde’
because they are close to background pixels. Pixels are allocated to one of seven foreground
classes (core, edge, perforation, bridge, loop, branch, and islet) and a background class.
Each of these classes can also have an attribute describing whether it is entirely surrounded
by other foreground classes (internal) or with connectivity through adjacent background
classes to the edge of the raster (external). Loops and bridges can additionally be defined
as appearing independently or within edges or perforations. The result is 22 possible
types of feature that are extracted from the binary raster. In this context, foreground
pixels represent vegetated marsh platform areas while background pixels correspond to
bare sediments, channels, and pools. Although these classes were originally designed to
describe habitat connectivity features, analogues can be readily defined for most classes
in the context of fine-scale saltmarsh morphology (Table 1). MSPA was applied to a 25
cm-resolution binary raster derived from the Environment Agency’s salt marsh extents
layer [17], which describes recent (2008–2010) marsh morphology. Core areas were defined
as parts of vegetated marsh greater than the MSPA edge width parameter (2.5 m) from
a background area. The pCore metric was the aggregation of both internal and external
MSPA ‘core’ classes for a given satellite pixel area.
J. Mar. Sci. Eng. 2021, 9, 311
8 of 22
Table 1. Outline of the landform classes produced by MSPA and their interpretation in the context
of fine-scale salt marsh morphology (internal: no background connectivity to raster edge; external:
background connectivity to raster edge).
Feature
Classification Rule
Interpretation
Background
Background pixels
Salt pans or pools if internal,
channel or mudflat if external
Core
Areas greater than the edge distance from nearest background
pixel
Large, coherent marsh areas
Edge
Areas within the edge distance
of background pixels
Margins of core areas with
channel connectivity
Perforation
Edge pixels entirely enclosing
an area of background pixels
Margins of salt pans or pools
without channel connectivity
Branch
Strip less than twice the edge
distance wide that joins to a
core area at one end
Narrow extensions from core
areas
Islet
Foreground area too small to
contain any core area
Small, isolated marsh fragments
Bridge
Strip less than twice the edge
distance wide that joins two
core areas
Narrow causeway between
larger coherent marsh areas
Bridge in Edge
Bridge joining two areas of
edge class
As above but for edge areas
Bridge in Perforation
Bridge joining two areas of perforation class
Narrow strip of marsh separating marsh areas that are
distinct but entirely contained
within a salt pan or pool
Loop
Strip less than twice the edge
distance wide that joins to the
same core area at both ends
Narrow strip of marsh enclosing a salt pan or pool
Loop in Edge
Strip less than twice the edge
distance wide that joins to the
same edge area at both ends
Narrow strip of marsh enclosing a salt pan or pool
Loop in Perforation
Strip less than twice the edge
distance wide that joins to the
same perforation area at both
ends
Narrow strip of marsh enclosing a smaller salt pan or pool
within a larger one
2.3.4. Limitations to Tidal Connectivity
The second morphological metric describes the proportion of pixel area comprised of
unvegetated areas and their periphery that are unconnected to the drainage network; it is
denoted pUncon. It is also derived from the MSPA segmentation and reflects pools or pans
and their periphery that do not have connectivity via channels to the offshore domain. The
pUncon metric was computed as the aggregation of the ‘internal’ MSPA classes, including
unvegetated areas.
An example area of the MSPA segmentation for six adjacent satellite pixels is shown
in Figure 4, with the respective values for the pCore and pUncon metrics. These two
morphological metrics are interpreted as describing the marsh morphology within a dataspace that separates, at its extremes, between contiguous, undissected marsh platform
(high pCore), marsh areas that are heavily fragmented by creek networks (low pCore,
low pUncon), and areas that have a high density of salt pans and pools (low pCore, high
pUncon). The inset graph on Figure 4 illustrates the position of the depicted pixels within
this data-space. The processes leading to these morphologies are likely to be different, as
J. Mar. Sci. Eng. 2021, 9, 311
9 of 22
are their likely future evolutionary trajectories. Creek networks, particularly low-order
ones, are typically stable in the platform once established [44], and maintain a dynamic
equlibrium with the tidal prism [45]. Interior areas of intact marsh platforms may remain
so or become punctuated by pools or pans [46–48], with the density of pans being related
to the tidal range and sediment type at a national scale in the UK [49] and the elevation
and distance from marsh edge at a site scale [47].
Figure 4. TMSPA segmentation for six adjacent 30 m satellite pixels showing the aggregated classes
used to compute the pCore and pUncon metrics. Metric values of each pixel are provided and plotted
within the data-space in the inset graph. Pixels are numbered 1–3 across the top row and 4–6 across
the bottom row.
2.4. Probability of Observing Marsh Degradation
Marsh degradation was considered to have occurred for pixels where the δPveg was
negative. For each of the topological and morphological metrics, the probability of a location exhibiting degradation was computed from the proportion of pixels where δPveg < 0
in each of 100 equal intervals across the observed ranges of the four topological or morphological metrics described above (CD, ED, pCore, pUncon). Least-squares regression
analyses were used to establish relationships for each of the metrics between the observed
probabilities of degradation and interval centre points.
3. Results
3.1. Morphodynamic Change
δPveg was estimated for approximately 180,000 30 m pixels, amounting to approximately 16,200 ha of marsh. An example of the resulting metric for about 600 ha of salt
marsh in Hamford Water, Essex, is presented in Figure 5 to illustrate the degree of spatial
variability observed in changes to marsh integrity over relatively small spatial scales. The
inset histogram shows the distribution of estimates across the entire study domain from
the Humber to the Thames, which can be seen to be normal but with a mean somewhat
above zero, suggesting a dominance of vegetation establishment within the domain, even
though this is not necessarily evident within the Hamford Water subset depicted. The
overall probability of degradation for the entire dataset was 0.144 for the approximate time
period 1985–2016.
J. Mar. Sci. Eng. 2021, 9, 311
10 of 22
Figure 5. δPveg estimates for approximately 600 ha of marsh in Hamford Water, Essex, between 1985
and 2016. Inset histogram shows the distribution of estimates across the entire study domain between
the Humber and the Thames (16,200 ha). Red circles highlight the locations of pixels (black) excluded
from the regression analysis of probability of degradation against Cost Distance (CD).
3.2. Topological and Morphological Relationships
No significant regression relationship was found between Cost Distance (CD) and
the probability of degradation when using the entire dataset. The CD parameter, however,
has very long tails to its distribution, particularly at the upper end of the scale. At the
lower end, these pixels represent the seaward limit of open-coast marsh areas, while at
the upper end they are mainly high, supratidal regions towards the heads of estuaries or
around islands of high ground. These locations may not be expected to experience the
same controls on marsh morphodynamics as the majority of interior and estuarine marsh
areas, so might confound the regression analysis for such areas. Areas with exceptionally
low CD values coincide with the seaward margins of open coast marshes, which have the
potential to retreat [14] and advance rapidly [24]. Values of δPveg in such areas therefore
tend towards extreme values and reflect changes in marginal position more than changes in
marsh platform integrity. Where CD is exceptionally high, δPveg values are associated with
the upland boundary and may again not reflect changes in marsh platform integrity but
rather shifts between terrestrial and halophytic vegetation communities, which may have
completely different reflectances. The extremities of the CD distribution were discarded
where CD ≤ 500,000 and CD ≥ 1,400,000. 91% of the population was retained, and all
the locations excluded within Hamford Water are marked, for illustration purposes, as
black pixels within the red circles on Figure 5. With the extremities of the distribution
excluded, a significant positive linear relationship between CD and the probability of
degradation was found (Adjusted R2 = 0.37, p ≤ 0.001—see Figure 6, Panel A). The
low positive coefficient suggests that, as CD increases—implying a reduction in tidal
connectivity to offshore waters—the probability of marsh degradation increases slightly.
CD values do not represent real-world physical quantities. For instance, a doubling of CD
could occur because of multiple different combinations of increased horizontal distance
and/or elevation. The precise coefficients of this relationship are therefore not readily
interpretable beyond their sign and approximate magnitude.
A second-order polynomial relationship was found between Euclidean Distance
(ED) and the probability of degradation, whereby areas with a low ED exhibit moderate
J. Mar. Sci. Eng. 2021, 9, 311
11 of 22
probabilities of degradation in the order of 0.1, with this probability declining to a minimum
of less than 0.05 at around 350 m before increasing again with ED to values in excess of 0.3
at distances of 1200 m. The relationship is significant at p ≤ 0.001, and the high adjusted
R2 of 0.76 suggests that ED is an important determinant of changes in condition (Figure 6,
Panel B). The ED metric is directly comparable to that used by Kearney and Rogers [20] to
predict changes in marsh condition, and the regression fit from their study is superimposed
on the figure for comparison.
Significant relationships were also found between both pCore and pUncon and the
probability of degradation (Figure 6, panels C and D, respectively). The former relationship
implies a decrease in the probability of degradation with increasing values of the pCore,
although very high values of pCore become associated with increased probabilities of
degradation again. With high adjusted R2 values of 0.63 for pCore, the proportion of a pixel
that is core vegetated marsh area is strongly related to changes in marsh condition. Pixels
containing salt pan features were relatively rare within the dataset, with n = 13,157, representing 9.2% of all pixels. The regression only included those pixels where pUncon > 0.
The relationship is significant with an R2 value of 0.63, suggesting a strong positive relationship between the proportion of an area that comprises salt pans and their peripheral
marsh features and the history of morphological evolution.
Summaries of all four regression relationships are provided in Table 2.
Figure 6. Probability of degradation against topological and morphological metrics. CD panel (A),
ED panel (B), pCore panel (C), and pUncon panel (D). Relationship observed for ED by Kearney and
Rogers [20] depicted on panel (B) for comparison. All relationships are significant at p ≤ 0.001.
Table 2. Summary of regression models between topological or morphological metrics and the
probability of marsh degradation.
CD
ED
pCore
pUncon
Adjusted R2
x
x2
f-Stat
p-Value
0.37
0.76
0.63
0.63
7.3 × 10−8
−2.9 × 10−4
−2.6 × 10−3
−2.0 × 10−3
N/A
4.0 × 10−7
1.9 × 10−5
−4.2 × 10−5
17.4
147
85.7
83.9
≤0.001
≤0.001
≤0.001
≤0.001
J. Mar. Sci. Eng. 2021, 9, 311
12 of 22
4. Discussion
4.1. Morphodynamic Change
The regional data for the Humber to Thames study area show an overall trend towards
increasing marsh platform integrity over time, with rapid increases in integrity being much
more common than rapid losses (inset to Figure 5). This pattern is, in part, a result of
the rapid infilling and marsh advance that occurred in The Wash embayment during the
observation period (1985–2016), where rates of margin advance up to 75 m per year were
observed between 1992 and 2014 [24]. This established a large area of marsh that is included
in the analysis and is associated with very high values of δPveg . The pattern of increase and
decrease in internal platform integrity is much patchier elsewhere, as seen for Hamford
Water (Figure 5).
A degree of bias towards positive values of δPveg was introduced by the choice of a
marsh presence/absence mask from relatively late in the study period. Phelan et al. [17]
base their mapping on data from 2008 to 2010. This implies that areas that may have
contained marsh early in the study period, but which became completely unvegetated
prior to 2008 and would produce negative δPveg , will have been excluded from the analysis.
A visual assessment of the aerial photography from 2013/14 and 1992 suggests that this
phenomenon was rare within the region. The scale of negative changes in the overall
horizontal extent of the marshes over the study period, relative to the Landsat pixel size,
is typically small. Only isolated pixels at the seaward limits of some open-coast marshes
are likely to have been affected by this bias [24]. Our findings support the idea that
vegetated areas of marsh platform have the potential to increase in platform integrity
much more rapidly than they deteriorate. This is likely the result of the potential for the
rapid colonisation of large horizontal extents of unvegetated surfaces that accrete to a
critical elevation to support seedling survival [29] near-simultaneously because of their
very low gradients. Conversely, during erosive phases sediment loss is likely to be more
gradual due to biostabilisation [50,51], and is also likely to be more localised in areas of
high hydrodynamic stress. Assuming that morphological change can be inferred from
vegetation extent change, the data suggest that, in marsh interiors at least, accretionary
processes tend to outpace erosional ones.
We address change over a single time period only. Our methodology is not, therefore,
able to detect regime shifts or abrupt changes as distinct from more gradual trends that
result in the same δPveg value. A natural development of our work, as the data archives
become longer and denser, would be to incorporate temporal segmentation approaches
similar to those used by the LandTrendr algorithm [52,53]. This would allow us to investigate whether the topological and morphological controls on abrupt regime shifts differ
from those on gradual changes. Such work is beyond the scope of the current study.
4.2. Geomorphic Setting (CD)
No significant relationship was found when considering probability of degradation
across the entire range of CD. This is likely because CD has a large range between 3.9 × 105
and 3.1 × 106 , with over 90% of values falling in the interval 5.0 × 105 and 1.4 × 106 . The
upper end of the distribution in particular is therefore very sparse. This leads to very small
numbers of pixels being represented in each of the 100 intervals across the CD distribution.
With the relatively low prevalence of degradation in the order of 10% within the entire
dataset, these small samples at high CD values often fail to represent any degraded pixels,
resulting in probabilities of degradation of zero. Even if some degraded pixels are sampled,
the resulting probability estimate is unlikely to be usefully precise while the sample size
remains below about 50 pixels. Furthermore, the areas where CD > 1.4 × 106 are typically
high-elevation areas which are likely to be supratidal and may therefore be expected to
be responding to different processes controlling their dynamics when compared to the
intertidal marsh areas that are of principal interest here. Similarly, the data density is very
low where CD < 5.0 × 105 , as these areas represent the seaward limits of open-coast
marshes. The observations in these areas are likely to be dominated by the wave-driven
J. Mar. Sci. Eng. 2021, 9, 311
13 of 22
erosion of the marsh margin rather than the internal marsh changes that δPveg represents.
The probabilities of degradation in these areas are relatively high, aligning well with the
observations of the widespread retreat of open coast marsh margins in the region [24].
Excluding such areas from the analysis is therefore justifiable, as it enables insight into
dynamics in the intertidal zone that are otherwise masked by variance in other areas that
are either artefactual or likely to arise from a different suite of processes to those types of
marsh interior processes that this study seeks to assess.
When the range of CD was limited to the interval 5.0 × 105 and 1.4 × 106 , the
regression analysis showed a significant positive linear relationship between CD and the
probability of a pixel showing degradation. Low CD values are associated with minimum
probabilities of degradation (Panel A, Figure 6). This implies that the interiors of open-coast
marshes experience little degradation, while those towards the head of estuaries or tidal
inlets are more susceptible. This aligns well with the conceptualisation of sediment sources
for this study region, which is dominated by minerogenic marshes [26] and offshore tidal
waters are the primary source of allochthonous inputto support marsh elevations [54].
Those areas of marsh that are least tidally connected (high CD) are shown to be the most
vulnerable, since the delivery of sediments is most restricted in these areas. This finding
contrasts with those of Kearney and Rogers [20], who observe an increasing probability of
degradation at shorter distances up-estuary. This is because the domain that they studied
on the east coast of the US has much more significant riverine inputs and the locations
of sedimentation are therefore fluvially, rather than tidally, dominated. Additionally, the
Chesapeake Bay (US) marshes are highly organogenic, with the upper marsh sediment total
organic content being reported to be around 80% at Rhode River [55], which implies that
autochthonous accumulation dominates. By comparison, the UK east coast marshes rely
more heavily on allochthonous inputs with much lower total organic contents than those
of the US, at around 15% [56]. The relationship we find between CD and the probability
of degradation is relatively weak (R2 = 0.37), reflecting the fact that other factors, likely
occurring at smaller scales, may be more significant. These may include the controls
represented by the other ED, pCore, and pUncon metrics in this study, but also factors
such as proximity and the magnitude of fluvial sediment loads that we have been unable
to address here. The strength of the relationship may also arise, in part, from the failures
of the simplified CD metric to perfectly represent tidal connectivity within the domain.
Nevertheless, given the size of the region considered, the heterogeneity of marsh settings
represented and the methodological limitations, the fact that any relationship emerges
implies that the position of a marsh within the wider coastal setting is an important control
on its vulnerability to sediment starvation and subsequent degradation. The contrast
between our findings and those of Kearney and Rogers [20] suggests that the marsh
position is not, however, a diagnostic parameter, since its effect varies between regions.
Rather, it modulates the impacts of larger-scale geomorphological contexts (that may be
considered as boundary conditions at the scale of the analysis presented here), such as
sediment sources and delivery pathways. This finding implies that empirical attempts to
predict morphological change in coastal wetlands must be nested within a hierarchy of
process understanding in order to ensure that outcomes are appropriate for a particular
location. Not all regions behave equally.
4.3. Distance from Creek/Edge of Marsh Parcel (ED)
The relationship established between Euclidean Distance (ED) and probability of
degradation is significant and stronger (R2 = 0.76) than that found for CD. It is also nonlinear with initially high probabilities at short distances that decline to a minimum around
350 m before increasing dramatically. There is a notable similarity with the relationship
observed in Chesapeake and Delaware Bays (black dashed line on Figure 6). The initial
decline observed here is somewhat less rapid than that found by Kearney and Rogers [20],
and beyond 400 m there is a strong increase in the probability of degradation with increasing
ED at a rate that exceeds that found by Kearney and Rogers [20].
J. Mar. Sci. Eng. 2021, 9, 311
14 of 22
The inter-regional similarity observable regarding ED suggests that topological factors
controlling marsh stability at scales below those represented by CD (i.e., within-marsh
scales rather than within-estuary scales) may be operating in very similar fashions between
regions, although the differences in the rates of decline and increases in degradation
probability around the 350 m minima suggest a possible dependence on the scale/size of
the marsh systems being observed. Marsh parcels are substantially smaller in the UK than
in the USA example, where ED values of up to about 100 km were observed, although the
goodness of fit above about 1 km declines. Differences in drainage network morphologies
may also contribute. For the context of the east coast of the UK, the initial decline in
probabilities can be interpreted, in part, as a function of the scale of channel identified in
the basemap used for the distance calculations. Only major creeks exceeding 42.5 m width
were included. Where the ED values are very low, the degradation measured by δPveg
could have a component that reflects the marginal erosion of the banks of large channels,
where fetch is sufficient for wave-driven erosion to become significant in determining
the spectral change within the pixel over the study period. Thereafter, however, it still
appears that marsh vulnerability decreases with increasing distance from these major
channels. This could be associated with the increasing drainage density of smaller creeks
(not resolved by the ED metric used here), leading to shorter unchannelised lengths,
a greater preponderance of levee-effects [19], and ultimately a greater potential for the
drainage of and allochthonous sediment import to marsh areas within 350 m of a major
channel. Beyond this distance, the probability of degradation increases again, which we
attribute to declining drainage density and efficiency, leading to decreasing potential for
tidal flushing, drainage, and sediment import as distances from the primary tidal channels
continue to increase. Internal marsh areas without efficient drainage and a long way from
major channels tend to experience relatively little elevation gain from external sources [57]
and, particularly in the context of relative sea level rise, vegetation experiences increasing
water and salt stress, potentially leading to die-back and substrate collapse [58]. This
phenomenon is observable within the region and is illustrated by the hotspots of negative
δPveg in marsh interior areas shown in Figure A1 within the Appendix A.
4.4. Percentage Core Areas (pCore)
pCore is based on a 25 cm spatial resolution and therefore resolves differences in
marsh morphology at a sub-pixel scale in terms of δPveg . The same is true for pUncon. The
relationship observed whereby the probability of degradation decreases to a minimum
at pCore values of around 70% before levelling off or increasing slightly implies that
fragmented areas of marsh (low pCore) reflect a history of greater marsh degradation. This
implies that more degraded areas of marsh may continue to degrade at a faster rate than
areas of platform with a high integrity. As a marsh platform becomes fragmented, pCore
declines and the perimeter–area ratio of the remaining marsh parcels increases, providing a
greater effective marginal length that may be vulnerable to erosion by hydrodynamic forces.
The stability of the morphological state therefore decreases with increasing fragmentation,
representing a positive morphodynamic feedback controlling marsh degradation. The
slight increase in probability at very high pCore values may reflect the tendency for marsh
areas that are separated from minor creeks by only a few metres to begin to degrade as
a result of water stress [58], which Ursino et al. [59] showed can increase rapidly away
from channels. Additional contributors to marsh degradation occurring at high pCore
values may be reduced sediment inputs [19,57] or increased salt stress [60] in the interiors
of marsh parcels.
4.5. Percentage of Unconnected Areas (pUncon)
The relationship between pUncon and the probability of degradation is significant,
with a R2 value of 0.63. The form of the relationship takes an almost opposite form to
that for pCore, with higher values of pUncon being associated with higher probabilities
of degradation. pCore and pUncon are somewhat negatively correlated to each other
J. Mar. Sci. Eng. 2021, 9, 311
15 of 22
(R = −0.11, p ≤ 0.001), so this inverse relationship may in part reflect the same processes
as are discussed with reference to pCore. The analysis for pUncon, however, incorporates
only those marsh pixels where pans or pools are present, which represents a small subset of
the dataset (9.2%). As such, this analysis allows for the isolation of behaviours exclusively
in areas where pans are observed. Areas with a higher pUncon and therefore a greater
proportion of pan-related features exhibit higher probabilities of degradation. Notably, for
fragmentation identified by the pUncon metric, the maximum probabilities of degradation
observed are much higher (≈0.5) than those calculated on the larger dataset of pCore (≈0.2),
where low metric values also indicate fragmentation. Low pCore values, by contrast,
largely reflect fragmentation by creek networks and bare surfaces connected to them.
The difference in magnitudes of probability identified by the two regression relationships
implies that the presence of pans indicates a much more vulnerable landscape configuration
than one fragmented by creeks. Neumeier et al. [61] identified the life cycle of pans in
Canadian marshes whereby they undergo a phase of active expansion before achieving
maturity. The data presented here also support the conclusion that pans have a tendency
to expand over decadal timescales. The analysis cannot identify causality, however. It is
possible that the presence of pans establishes a feedback encouraging their subsequent
expansion as the marsh develops [48]. Questions also remain over the causes of initial pan
formation, with some theories suggesting that event-based phenomena, such as smothering
of vegetation by rafted debris, may cause pans to develop. French et al. [62] noted that
pan densities are lower on the more enclosed of the backbarrier marshes in Norfolk, where
less debris may be rafted into the marshes over shorter seaward margins. Their finding is
correlative but Pethick [47] argues that a mechanism for pan initiation on mature marshes
must exist. Whatever the cause of pan initiation, our findings suggest that, once established,
marsh areas containing pans are highly vulnerable compared to other areas. It is possible,
therefore, that stochastic events, such as debris rafting, could initiate a long-term landscape
vulnerability. In contrast, pans may simply be diagnostic of other conditions such as relative
sea level changes, tidal range, or coastal configuration, rather than morphodynamic drivers
in their own right [49]. If pans are not, in themselves, morphodynamic drivers, then our
findings would suggest that the external controls causing pan formation continue to cause
subsequent expansion and the conclusion that panned landscapes are vulnerable stands.
Further work is needed to address the question of controls on salt pan development to
better understand the fundamental causes of this vulnerability.
5. Conclusions
Our findings show that there exist certain overarching controls on the vulnerability
of salt marshes to degradation at a range of scales. We demonstrate this observation
through the application of a systematic spatial analysis of a simple, satellite-image-derived
measure of change. By applying this methodology at the regional scale and across a range
of estuarine, backbarrier, embayment, and open coast settings, we illustrate the nature of
these controls more clearly than has hitherto been possible.
This work has presented an inherently scalable method for monitoring the platform
integrity of salt marsh surfaces at a sub-pixel scale that is easily adapted to incorporate
current and future data sources (such as imagery from the more recently launched Sentinel
2 satellites). The δPveg metric presented here is a continuous variable representing change
in marsh surface integrity over a given time period. It therefore provides substantially
more statistical possibilities than the MSCI [20] or than we have explored in this study.
We express topological and morphological factors at a range of scales as metrics
describing the connectivity of tidal waters, position within marsh, marsh platform integrity, and the prevalence of salt pans or pools. We demonstrate that all of these are
significantly related to the morphodynamic evolution of marshes. This spatial hierarchy
of morphological controls may continue to be applicable at both larger and smaller scales
than are addressed here. The presence of controls at larger scales is shown by the finding of
contrasting relationships between geomorphic setting and marsh degradation in this study
J. Mar. Sci. Eng. 2021, 9, 311
16 of 22
and that of [20]. We hypothesise that this difference arises from regional-scale contrasts in
climate, biota, sediment supply, and sea-level history that cause fundamental differences
in the processes by which the marsh systems function and evolve. At a whole-marsh
scale, however, responses to the distance from the marsh margin seem to be fairly consistent between regions, suggesting commonality in some of the processes governing marsh
dynamics. At a smaller scale than is considered here, a variety of factors may become
important, such as the spatial distribution of vegetation types [46,63] or phenotypes [64],
which may determine sediment erodibility.
We have presented aggregated relationships that emerge at a regional scale. Overall,
these suggest that marsh areas that are already exhibiting some form of fragmentation that
lie far from the nearest creek and towards the heads of estuaries and inlets are the most
likely to exhibit degradation. The prediction of changes in a particular location remains
challenging. No relationships were found between the continuous values of δPveg itself
and any of the predictors presented here using basic statistical techniques. Only when
the dimensionality of the dataset is reduced to probabilities do aggregated relationships
emerge. The large number of parameters influencing marsh dynamics, the multiple scales
of processes involved, and the complex interactions between them means that statistical prediction of internal marsh changes will require methods capable of capturing the
interactions and high dimensionality that are inherent to such natural systems. Further
work is needed to explore the potential of more sophisticated statistical techniques drawn
from machine learning domains to synthesise datasets such as the one presented here into
location-specific values of parameters describing the sign and magnitude of predicted
marsh evolution.
Author Contributions: Conceptualization, B.R.E.; methodology, B.R.E., I.M. and T.S.; software,
B.R.E.; validation, B.R.E.; formal analysis, B.R.E.; investigation, B.R.E.; resources, B.R.E., I.M. and T.S.;
data curation, B.R.E.; writing—original draft preparation, B.R.E.; writing—review and editing, B.R.E.,
I.M. and T.S.; visualization, B.R.E..; supervision, I.M. and T.S.; project administration, I.M. and T.S.;
funding acquisition, I.M. and T.S. All authors have read and agreed to the published version of the
manuscript.
Funding: This research was funded by European Commission’s Seventh Framework Programme
(grant number 607131) and was conducted within the Foreshore Assessment using Space Technology
(FAST) project. Further funds have been provided by the Isaac Newton Trust and the UKRI Natural
Environment Research Council (NERC) grant RESIST-UK (grant number NE/R01082X/1). The APC
was funded by the Cambridge Coastal Research Unit.
Data Availability Statement: The data supporting the analyses presented here are available on an
Open Science Framework repository. Doi:10.17605/OSF.IO/MGSYZ.
Acknowledgments: We would like to thank our colleagues on the FAST project for their assistance,
particularly that of Edward Morris. We are also grateful for the support and equipment provided by
the UKRI NERC Field Spectroscopy Facility in Edinburgh.
Conflicts of Interest: The authors declare no conflict of interest.
Abbreviations
The following abbreviations are used in this manuscript:
CD
ED
HAT
ISCE
L5
L7
L8
Cost distance
Euclidean distance
Highest astronomical tide
Integrated scale coastal evolution
Landsat 5
Landsat 7
Landsat 8
J. Mar. Sci. Eng. 2021, 9, 311
17 of 22
MSCI
MSPA
MSTR
NDVI
NE
ODN
pCore
pUncon
RMSE
SMP2
UVVR
δPveg
Marsh surface condition index
Morphological spatial pattern analysis
Mean spring tidal range
Normalised difference vegetation index
Normalised elevation
Ordanace datum Newlyn
Percentage ‘core’ areas
Percentage ‘unconnected’ areas
Root-mean-squared error
Second-generation shoreline management plan
Unvegetated–vegetated ratio
Change in percentage vegetation cover within pixel
Appendix A. Methodology for Estimation of Marsh Platform Integrity Changes from
Satellite Observations
The Normalised Difference Vegetation Index (NDVI), a dimensionless index taking
values between −1 and 1, is a proxy for the amount of chlorophyll (and therefore vegetation) present, with higher values indicating more chlorophyll. The NDVI is the normalised
ratio between reflectance in the red and near-infrared bands. Healthy vegetation produces
positive index values (approaching one) [65]. Conceptually, the NDVI for a pixel in a
satellite image can be considered to be a function of the percentage of that pixel covered by
vegetation and the nature of the vegetation within that pixel. Applying an assumption of
approximate stationarity in the reflectance of the bare sediments, changes in the percentage vegetation cover within the pixel are reflected in a change in NDVI, with the signal
potentially being modulated by any attendant changes in the community composition (and
therefore spectral signature) and vigour of the vegetation present.
Summer NDVI is relatively insensitive to the successional stage of the vegetation
within this study region, while being lower for pioneer stages than mature stages during
winter. Field spectroscopy was conducted at Tillingham, Essex, throughout the period
2015–2016. A total of 58 pioneer spectra and 103 mature marsh spectra were collected
using an SVC HR1024i spectroradiometer and calibrated reference panel. The resulting
hemispherical-conical reflectance factors were convolved to Landsat-8 band responses,
from which NDVI was calculated. Pioneer vegetation (dominated by Salicornia and Spartina
species) has similar Landsat-8 NDVI in summer (June-October) to the perennial marsh
canopy (t-test, p ≥ 0.05). However, the near-complete dieback of pioneer vegetation leads
to the exposure of bare sediment and therefore lower winter NDVIs.
In Google Earth Engine [32], all scenes intersecting the study area were extracted from
the Landsat archives for Landsat-5, -7, and -8 at a spatial resolution of 30 m, providing
coverage between 1985 and 2018 (excluding Landsat-7 images subsequent to the failure of
the scan line corrector on 31 May 2003). Level 1 Top-of-Atmosphere images were imported
and filtered for cloud cover based on their metadata cloud score and subsequently on a
per-pixel basis using the Simple Cloud Score algorithm. Thresholds for excluding data
were set at 20% for both stages. NDVI was computed for all the remaining pixels of all
images.
Differences in sensor response, and the NDVI values derived from them, are widely
recognised [66] and have been shown to produce bias in time series analysis if combined
without corrections being applied [67]. NDVI images were corrected for different sensor responses to produce a consistent time series of Landsat-7-equivalent NDVI values.
Landsat-8 scenes were corrected using the cross-calibration derived by Roy et al. [68]. No
existing cross-calibration was available for Landsat 5. To calculate a calibration function
pixels were identified that fell within large areas of established marsh (those exceeding
1500 m2 in the UK salt marsh extent map [17]) and more than 42.4 m (the diagonal of a
30 m Landsat pixel) from the edge of the marsh area. These were considered likely to be
‘pure’, fully vegetated locations. All the Landsat 5 and Landsat-7 scenes were selected for
the period between the launch of Landsat-7 on 15 April 1999 and 31 May 2003. This process
J. Mar. Sci. Eng. 2021, 9, 311
18 of 22
resulted in a total of 3747 pixels being selected for the cross-calibration using 5 Landsat-5
scenes and 40 Landsat-7 scenes. The per-pixel mean NDVI values were compared using a
second-order polynomial, since a degree of saturation occurred in the Landsat-5 sensor
when compared to Landsat-7 (p ≤ 0.0001, R2 = 0.756). The cross-calibration functions
applied are detailed in Table A1.
Table A1. Cross-calibration functions to correct NDVI from Landsat-5 (L5 ) and Landsat-8 (L8 ) [68] to
Landsat-7 (L7 ) equivalence prior to time series analysis.
Transformation
Function
R2
L5 to L7
L8 to L7
L7 = 0.085 + (0.460L5 ) + (0.247L25 )
L7 = −0.011 + 0.969L8
0.756
0.906
The corrected Landsat-7-equivalent time series of NDVI images was processed in the
Google Earth Engine to produce per-pixel linear trend coefficients. These trend coefficients
were weakly related to the mean NDVI for the pixel over the entire time series (R2 = 0.22,
p ≤ 0.001). This was interpreted as arising either from a CO2 fertilisation signal causing
vegetation in general to become ‘greener’ throughout the time series. Alternatively, it may
reflect a pixel-scale morphological control on change in vegetation distribution, whereby
pixels with higher platform integrity tend to increase their integrity more than those with
lower integrity. Either of these effects would confound the investigation of topological
controls. The relationship was therefore removed by using the residuals of the linear fit in
Equation (A1) where δN/δt denotes the NDVI trend and N̄ denotes the mean NDVI. The
residuals were subsequently standardised.
δN/δt = 0.011254 N̄ − 0.001430
(A1)
Standardised residuals were converted into sub-pixel estimates of percentage change
in vegetation cover (henceforth δPveg ) by calibration against geo-referenced vertical aerial
photography supplied by the UK Environment Agency. The observed range of the standardised residuals was divided into 100 equal intervals and one pixel was selected at
random within each interval. Some intervals towards the tails of the distribution contained
no pixels, precluding their inclusion. In total, 83 pixels were selected. Pixels were clipped
from the earliest available aerial photography (1992, panchromatic, 25 cm resolution) and
from colour photography from 2013/2014 (20 cm resolution).
In Matlab, 50 random points were generated per Landsat pixel image pair (1992
and 2013/14), and sequentially superimposed on each image. A single operator visually
assessed the surface cover and allocated each point as either vegetated or not vegetated.
Manual assessment and attribution is assumed to provide the highest achievable accuracy
in this context. A total of 4150 points were manually attributed for each year. From the
proportions of vegetated and unvegetated points in an individual Landsat pixel area, the
percentage vegetation cover was estimated. The difference between the percentage of
the pixel area that was vegetated in 2013/14 compared to 1992 was calculated. A linear
relationship was found between the standardised residuals and the change in vegetation
cover. The linear relationship between the change in percentage vegetation cover between
1992 and 2013/14 estimated from aerial photography and the standardised residuals of
the NDVI trend analysis (Rtrend ) is given in Equation (A2). The R-squared value of 0.87
implies that the change in vegetated area is the major contributor to changes in the residual
trend. The unexplained variance is likely the result of other factors not accounted for by
vegetation extent alone, such as species compositional change.
The method outlined above produces estimates of change in marsh platform integrity
as continuous values between ±100%. This offers many statistical possibilities that will be
explored in future work. For the purposes of the current study, estimates were thresholded
at zero to produce a binary gain/loss indicator allowing for the alignment of statistical
J. Mar. Sci. Eng. 2021, 9, 311
19 of 22
methods with the work of Kearney and Rogers [20], who discerned between pixels that
were either degraded or not degraded.
δPveg = 17.953Rtrend + 17.616
(A2)
to
to
to
to
to
to
to
Figure A1. Estimated changes in percentage vegetation cover derived from the Landsat time series
(top), with aerial photography from 2014 (middle) and 1992 (bottom) for comparison.
The performance of the calibrated δPveg method was evaluated at a more local scale
than the entire study domain in case large scale spatial dependency associated with,
for example, latitude was a significant factor affecting calibration. The area selected for
validation was Hamford Water, Essex, a tidal inlet containing approximately 600 ha of
marsh that has been shown to exhibit a wide variety of morphodynamic behaviours in
close spatial proximity to each other [69]. Analysis up to the point of manual calibration
was repeated using only Hamford Water as the domain to establish locally derived residual
J. Mar. Sci. Eng. 2021, 9, 311
20 of 22
trends which were classified into ten classes using a Jenks Natural Breaks algorithm.
The largest contiguous area for each class was identified and from that area ten pixels
were selected at random for validation, providing 100 pixels representing a range of
morphological behaviours. These were analysed by manual point attribution from aerial
photography following the same method as previously outlined. The change in percentage
vegetation cover from manual point attribution was compared to δPveg estimated from the
regional-scale analysis, producing a root mean square error (RMSE) of 11.9% of full range
(−100% to +100% change).
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
McOwen, C.; Weatherdon, L.; Bochove, J.W.; Sullivan, E.; Blyth, S.; Zockler, C.; Stanwell-Smith, D.; Kingston, N.; Martin, C.;
Spalding, M.; et al. A global map of saltmarshes. Biodivers. Data J. 2017, 5, e11764. [CrossRef]
Friess, D.A.; Yando, E.S.; Alemu I, J.B.; Wong, L.W.; Soto, S.D.; Bhatia, N. Ecosystem Serices and Dissserices of Mangrove
Forests and Salt Marshes. In Oceanography and Marine Biology: An Annual Review; CRC Press: Boca Raton, FL, USA, 2020;
pp. 58:107–58:142.
Mitsch, W.J.; Gosselink, J.G. Wetlands, 3rd ed.; Wiley: New York, NY, USA, 2000; p. 920.
McLeod, E.; Chmura, G.L.; Bouillon, S.; Salm, R.; Björk, M.; Duarte, C.M.; Lovelock, C.E.; Schlesinger, W.H.; Silliman, B.R. A
blueprint for blue carbon: Toward an improved understanding of the role of vegetated coastal habitats in sequestering CO2.
Front. Ecol. Environ. 2011, 9, 552–560. [CrossRef]
Möller, I.; Spencer, T.; French, J.R.; Leggett, D.J.; Dixon, M. Wave transformation over salt marshes: a field and numerical
modelling study from North Norfolk, England. Estuarine Coast. Shelf Sci. 1999, 49, 411–426. [CrossRef]
Möller, I. Quantifying saltmarsh vegetation and its effect on wave height dissipation: Results from a UK East coast saltmarsh.
Estuarine Coast. Shelf Sci. 2006, 69, 337–351. [CrossRef]
Burden, A.; Garbutt, R.; Evans, C.; Jones, D.; Cooper, D. Carbon sequestration and biogeochemical cycling in a saltmarsh subject
to coastal managed realignment. Estuarine Coast. Shelf Sci. 2013, 120, 12–20. [CrossRef]
Chmura, G.L. What do we need to assess the sustainability of the tidal salt marsh carbon sink? Ocean. Coast. Manag. 2013,
83, 25–31. [CrossRef]
Roner, M.; Alpaos, A.D.; Ghinassi, M.; Marani, M.; Silvestri, S.; Franceschinis, E. Spatial variation of salt-marsh organic and
inorganic deposition and organic carbon accumulation: Inferences from the Venice lagoon, Italy. Adv. Water Resour. 2016,
93, 276–287. [CrossRef]
Boesch, D.F.; Turner, R.E. Dependence of fishery species on salt marshes: The role of food and refuge. Estuaries 1984, 7, 460.
[CrossRef]
Costa, M.J.; Costa, J.; de Almeida, P.R.; Assis, C.A. Do eel grass beds and salt marsh borders act as preferential nurseries and
spawning grounds for fish? An example of the Mira estuary in Portugal. Ecol. Eng. 1994, 3, 187–195. [CrossRef]
Ladd, C.J.; Duggan-Edwards, M.F.; Bouma, T.J.; Pagès, J.F.; Skov, M.W. Sediment Supply Explains Long-Term and Large-Scale
Patterns in Salt Marsh Lateral Expansion and Erosion. Geophys. Res. Lett. 2019, 46, 11178–11187. [CrossRef]
Hughes, R.G.; Paramor, O.A.L. On the loss of saltmarshes in south-east England and. J. Appl. Ecol. 2004, 41, 440–448. [CrossRef]
van der Wal, D.; Pye, K. Patterns, rates and possible causes of saltmarsh erosion in the Greater Thames area (UK). Geomorphology
2004, 61, 373–391. [CrossRef]
Burd, F. Erosion and Vegetation Change on the Salt Marshes of Essex and North Kent between 1973 and 1988; Technical Report; Nature
Conservancy Council: Peterborough, UK, 1992.
Cooper, N.J.; Cooper, T.; Burd, F. 25 years of salt marsh erosion in Essex: Implications for coastal defence and nature conservation.
J. Coast. Conserv. 2001, 7, 31–40. [CrossRef]
Phelan, N.; Shaw, A.; Baylis, A. The Extent of Saltmarsh in England and Wales: 2006–2009; Environment Agency: Bristol, UK, 2011.
Vandenbruwaene, W.; Meire, P.; Temmerman, S. Formation and evolution of a tidal channel network within a constructed tidal
marsh. Geomorphology 2012, 151–152, 114–125. [CrossRef]
French, J.R.; Spencer, T. Dynamics of sedimentation in a tide-dominated backbarrier salt marsh, Norfolk, UK. Mar. Geol. 1993,
110, 315–331. [CrossRef]
Kearney, M.S.; Rogers, A.S. Forecasting sites of future coastal marsh loss using topographical relationships and logistic regression.
Wetl. Ecol. Manag. 2010, 18, 449–461. [CrossRef]
Allen, J. Simulation models of salt-marsh morphodynamics: Some implications for high-intertidal sediment couplets related to
sea-level change. Sediment. Geol. 1997, 113, 211–223. [CrossRef]
Da Lio, C.; D’Alpaos, A.; Marani, M. The secret gardener: Vegetation and the emergence of biogeomorphic patterns in tidal
environments. Philos. Trans. Ser. Math. Phys. Eng. Sci. 2013, 371. [CrossRef]
D’Alpaos, A.; Lanzoni, S.; Marani, M.; Rinaldo, A. Landscape evolution in tidal embayments: Modeling the interplay of erosion,
sedimentation, and vegetation dynamics. J. Geophys. Res. Earth Surf. 2007, 112, 1–17. [CrossRef]
Evans, B.R.; Möller, I.; Spencer, T.; Smith, G. Dynamics of salt marsh margins are related to their three-dimensional functional
form. Earth Surf. Process. Landforms 2019, 44, esp.4614. [CrossRef]
J. Mar. Sci. Eng. 2021, 9, 311
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
21 of 22
Pethick, J.; Leggett, D. The morphology of the Anglian coast. In Coastlines of the Southern North Sea; Hillen, R., Verhagen, H., Eds.;
American Society of Civil Engineers (ASCE): Reston, VA, USA, 1993; pp. 52–64.
Pye, K.; French, P. Erosion and Accretion Processes on British Saltmarshes: Volume Three; Technical Report; Cambridge Environmental
Research Consultants: Cambridge, UK, 1993.
Allen, J. Morphodynamics of Holocene salt marshes: A review sketch from the Atlantic and Southern North Sea coasts of Europe.
Quat. Sci. Rev. 2000, 19, 1155–1231. [CrossRef]
Kestner, F. The Old Coastline of the Wash. T Geogr. J. 1962, 128, 457–471. [CrossRef]
Balke, T.; Stock, M.; Jensen, K.; Bouma, T.J.; Kleyer, M. A global analysis of the seaward salt marsh extent: The importance of tidal
range. Water Resour. Res. 2016, 52, 3775–3786. [CrossRef]
Suchrow, S.; Jensen, K. Plant species responses to an elevational gradient in German North Sea salt marshes. Wetlands 2010,
30, 735–746. [CrossRef]
Ganju, N.K.; Defne, Z.; Kirwan, M.L.; Fagherazzi, S.; D’Alpaos, A.; Carniello, L. Spatially integrative metrics reveal hidden
vulnerability of microtidal salt marshes. Nat. Commun. 2017, 8, 14156. [CrossRef]
Gorelick, N.; Hancher, M.; Dixon, M.; Ilyushchenko, S.; Thau, D.; Moore, R. Google Earth Engine: Planetary-scale geospatial
analysis for everyone. Remote Sens. Environ. 2017, 202, 18–27. [CrossRef]
Lopes, C.L.; Mendes, R.; Caçador, I.; Dias, J.M. Assessing salt marsh extent and condition changes with 35 years of Landsat
imagery: Tagus Estuary case study. Remote Sens. Environ. 2020, 247, 111939. [CrossRef]
Hargrove, W.W.; Hoffman, F.M.; Hessburg, P.F. Mapcurves: A quantitative method for comparing categorical maps. J. Geogr. Syst.
2006, 8, 187–208. [CrossRef]
ABPMer. Atlas of Marine Energy Resources. 2008. Available online: http://www.renewables-atlas.info/ (accessed on 5 May
2015).
French, J. Tidal marsh sedimentation and resilience to environmental change: Exploratory modelling of tidal, sea-level and
sediment supply forcing in predominantly allochthonous systems. Mar. Geol. 2006, 235, 119–136. [CrossRef]
French, J.R.; Spencer, T.; Murray, A.L.; Arnold, N.S. Geostatistical analysis of sediment deposition in two small tidal wetlands,
Norfolk, UK. J. Coast. Res. 1985, 11, 308–321.
Reed, D.J.; Spencer, T.; Murray, A.L.; French, J.R.; Leonard, L. Marsh surface sediment deposition and the role of tidal creeks:
Implications for created and managed coastal marshes. J. Coast. Conserv. 1999, 5, 81–90. [CrossRef]
Soille, P.; Vogt, P. Morphological segmentation of binary patterns. Pattern Recognit. Lett. 2009, 30, 456–459. [CrossRef]
Vogt, P. GUIDOS: Tools for the assessment of pattern, connectivity, and fragmentation. In Proceedings of the EGU General
Assembly 2013, Vienna, Austria, 7–12 April 2013; Volume 15, p. 13526.
Ostapowicz, K.; Vogt, P.; Riitters, K.H.; Kozak, J.; Estreguil, C. Impact of scale on morphological spatial pattern of forest. Landsc.
Ecol. 2008, 23, 1107–1117. [CrossRef]
Saura, S.; Vogt, P.; Velázquez, J.; Hernando, A.; Tejera, R. Key structural forest connectors can be identified by combining
landscape spatial pattern and network analyses. For. Ecol. Manag. 2011, 262, 150–160. [CrossRef]
Chang, Q.; Liu, X.; Wu, J.; He, P. MSPA-Based Urban Green Infrastructure Planning and Management Approach for Urban
Sustainability: Case Study of Longgang in China. J. Urban Plan. Dev. 2015, 141, A5014006. [CrossRef]
French, J.R.; Stoddart, D.R. Hydrodynamics of saltmarsh creek systems: Implications for marsh morphological development and
material exchange. Earth Surf. Process. Landforms 1992, 17, 235–252. [CrossRef]
Friedrichs, C.T.; Perry, J.E. Tidal salt marsh morphodynamics: A synthesis. J. Coast. Res. 2001, 27, 7–37.
Adam, P. Saltmarsh Ecology; Cambridge University Press: Cambridge, UK, 1990; p. 461.
Pethick, J. The distribution of salt pans on tidal salt marshes. J. Biogeogr. 1974, 1, 57–62. [CrossRef]
Yapp, R.H.; Johns, D.; Jones, O.T. The salt marshes of the Dovey Estuary. J. Ecol. 1917, 5, 65–103. [CrossRef]
Goudie, A. Characterising the distribution and morphology of creeks and pans on salt marshes in England and Wales using
Google Earth. Estuarine Coast. Shelf Sci. 2013, 129, 112–123. [CrossRef]
Murray, A.B.; Knaapen, M.A.; Tal, M.; Kirwan, M.L. Biomorphodynamics: Physical-biological feedbacks that shape landscapes.
Water Resour. Res. 2008, 44. [CrossRef]
Spencer, T.; Möller, I.; Rupprecht, F.; Bouma, T.J.; van Wesenbeeck, B.K.; Kudella, M.; Paul, M.; Jensen, K.; Wolters, G.; MirandaLange, M.; Schimmels, S. Salt marsh surface survives true-to-scale simulated storm surges. Earth Surf. Process. Landforms 2016, 41.
[CrossRef]
Kennedy, R.E.; Yang, Z.; Cohen, W.B. Detecting trends in forest disturbance and recovery using yearly Landsat time series: 1.
LandTrendr—Temporal segmentation algorithms. Remote Sens. Environ. 2010, 114, 2897–2910. [CrossRef]
Kennedy, R.; Yang, Z.; Gorelick, N.; Braaten, J.; Cavalcante, L.; Cohen, W.; Healey, S. Implementation of the LandTrendr Algorithm
on Google Earth Engine. Remote Sens. 2018, 10, 691. [CrossRef]
Reed, D.J. Sediment dynamics and deposition in a retreating coastal salt marsh. Estuarine Coast. Shelf Sci. 1988, 26, 67–79.
[CrossRef]
Kirwan, M.L.; Langley, J.A.; Guntenspergen, G.R.; Megonigal, J.P. The impact of sea-level rise on organic matter decay rates in
Chesapeake Bay brackish tidal marshes. Biogeosciences 2013, 10, 1869–1876. [CrossRef]
Reef, R.; Schuerch, M.; Christie, E.K.; Möller, I.; Spencer, T. The effect of vegetation height and biomass on the sediment budget of
a European saltmarsh. Estuarine Coast. Shelf Sci. 2018, 202, 125–133. [CrossRef]
J. Mar. Sci. Eng. 2021, 9, 311
57.
58.
59.
60.
61.
62.
63.
64.
65.
66.
67.
68.
69.
22 of 22
Temmerman, S.; Govers, G.; Wartel, S.; Meire, P. Spatial and temporal factors controlling short-term sedimentation in a salt and
freshwater tidal marsh, scheldt estuary, Belgium, SW Netherlands. Earth Surf. Process. Landforms 2003, 28, 739–755. [CrossRef]
Reed, D.J. The impact of sea-level rise on coastal salt marshes. Prog. Phys. Geogr. 1990, 14, 465–481. [CrossRef]
Ursino, N.; Silvestri, S.; Marani, M. Subsurface flow and vegetation patterns in tidal environments. Water Resour. Res. 2004, 40.
[CrossRef]
Shen, C.; Zhang, C.; Xin, P.; Kong, J.; Li, L. Salt Dynamics in Coastal Marshes: Formation of Hypersaline Zones. Water Resour. Res.
2018, 54, 3259–3276. [CrossRef]
Neumeier, U.; Poulin, P.; Roge, M.; Morisette, A.; Huard, A.M. Morphology and evolution of salt marsh pans in the lower St.
Lawrence Estuary. In Proceedings of the Coastal Dynamics, Arcachon, France, 24–28 June 2013, pp. 1275–1286.
French, J.; Spencer, T.; Stoddart, D.R. Backbarrier Salt Marshes of the North Norfolk Coast: Geomorphic Developments and Response to
Rising Sea-Levels; Discussion Papers in Conservation; Ecology and Conservation Unit, Univeristy College London: London, UK,
1990; Volume 54, pp. 1–3.
Ford, H.; Garbutt, A.; Ladd, C.; Malarkey, J.; Skov, M.W. Soil stabilization linked to plant diversity and environmental context in
coastal wetlands. J. Veg. Sci. 2016, 27, 259–268. [CrossRef]
Bernik, B.; Pardue, J.; Blum, M. Soil erodibility differs according to heritable trait variation and nutrient-induced plasticity in the
salt marsh engineer Spartina alterniflora. Mar. Ecol. Prog. Ser. 2018, 601, 1–14. [CrossRef]
Rouse, J.; Haas, R.; Schell, J.; Deering, D. Monitoring The vernal Advancement and Retrogradation (Green Wave Effect) of Natural
Vegetation; Progress Report RSC 1978-1; Remote Sensing Center, Texas A&M Univ: College Station, TX, USA, 1973; p. 93.
Ke, Y.; Im, J.; Lee, J.; Gong, H.; Ryu, Y. Characteristics of Landsat 8 OLI-derived NDVI by comparison with multiple satellite
sensors and in-situ observations. Remote Sens. Environ. 2015, 164, 298–313. [CrossRef]
Sulla-Menashe, D.; Friedl, M.A.; Woodcock, C.E. Sources of bias and variability in long-term Landsat time series over Canadian
boreal forests. Remote Sens. Environ. 2016, 177, 206–219. [CrossRef]
Roy, D.; Zhang, H.; Ju, J.; Gomez-Dans, J.; Lewis, P.; Schaaf, C.; Sun, Q.; Li, J.; Huang, H.; Kovalskyy, V. A general method to
normalize Landsat reflectance data to nadir BRDF adjusted reflectance. Remote Sens. Environ. 2016, 176, 255–271. [CrossRef]
Evans, B. Processes Governing Saltmarsh Morphodynamics Methodological Challenges and Spatial Variability. Master’s Thesis,
Department of Geography, University of Cambridge, Cambridge, UK, 2011.