Estuarine, Coastal and Shelf Science 236 (2020) 106643

Contents lists available at ScienceDirect

Estuarine, Coastal and Shelf Science

journal homepage: http://www.elsevier.com/locate/ecss

Classification and mapping of saltmarsh vegetation combining

multispectral images with field data
Samantha Yeo a, 1, Virginie Lafon b, 2, Didier Alard a, C�ecile Curti b, 2, Aur�elie Dehouck b, 2,
Marie-Lise Benot a, *
a
Univ. Bordeaux, INRAE, BIOGECO, F-33600, Pessac, France
b
GEO-Transfert/ADERA, UMR 5805 EPOC, Universit�e de Bordeaux, All�ee Geoffroy Saint Hilaire, 33405, Talence, France

A R T I C L E I N F O A B S T R A C T

Keywords: Salt marshes are areas of high conservation value encompassing diverse ecological gradients responsible for
Arcachon bay creating unique vegetation structure and composition. In complement to the large body of studies developing
Multispectral data vegetation mapping methods through the use of remote sensing data, we tested for the possibility of developing a
Multi-temporal image classification
cost-effective method to map salt marsh vegetation as a basis for monitoring a French Nature Reserve. Using
Phytosociology
classical multivariate ordination and cluster analyses, accurate and ecologically relevant vegetation groups
Saltmarsh zonation
Southwestern France matching existing typologies were determined from a vegetation database collected for management and con­
Vegetation mapping servation rather than mapping purposes. This resulted in six distinct vegetation groups, which were mapped
Wetland remote sensing through the combination of the NIR spectral band and radiometric indices (NDVI and NDWI) derived from
multispectral 2 m-resolution satellite images (Pleiades images). The addition of a LiDAR-derived digital elevation
model (DEM) layer was also tested. The final classified map of the reserve based only on optical layers had an
overall accuracy of 75.5% (Kappa coefficient of 0.71), with varying success between the different vegetation
groups. The addition of the DEM slightly decreased map accuracy to 73.6% (Kappa of 0.68). Decreasing the
number of records used for map training had detectable negative effects on map accuracy. This study demon­
strated that using already existing field observations combined with only a few spectral bands and radiometric
indices from multi-temporal multispectral images with a high spatial resolution can be used as a basis to aid in
vegetation classification and mapping of saltmarsh habitats, with the goal of monitoring their dynamics.

1. Introduction sediment circulations (Belluco et al., 2006).

Lying at the interface between land and sea, the existence and
Saltmarshes, complex ecotones forming an interface between aquatic functioning of saltmarshes depend on the complex relationship of
and terrestrial habitats, are considered to be some of the most produc­ erosion and deposition processes, and are heavily influenced by their
tive ecosystems in the world (Boorman, 1995). They provide a wide biotic components (primarily vegetation), making saltmarshes highly
range of valuable ecosystem services (Kelly and Tuxen, 2009) including sensitive to environmental changes (Silvestri and Marani, 2004; Lefeu­
coastal protection and nature conservation, and offer various socioeco­ vre et al., 2003). Consequently, these coastal ecosystems face challenges
nomic activities such as fishing and recreation (Janssen, 2001; Belluco corresponding to increasing global changes, such as anthropogenic ac­
et al., 2006). These habitats, fringing fresh and saltwater ecosystems, are tivities and rising sea level. The latter has been identified as one of the
populated by halophytic vegetation adapted to saline and saturated soil greatest threats to saltmarsh communities and their ecosystem func­
conditions (Bertness and Ewenchuk, 2002; Boorman, 2003) exhibiting a tioning, impacting sediment retention and accumulation capacities
strong spatial zonation (Chapman, 1976; Adam, 1990; Marani et al., (Adam, 1990; Bromberg et al., 2009; Robins et al., 2016) and shifting
2003). Saltmarsh vegetation is a key element in the stability of these vegetation distribution, replacing high saltmarsh species with lower
areas for their role in the feedback cycle involving hydrodynamic and saltmarsh vegetation (Klemas, 2011). In a context of global change,

* Corresponding author.
E-mail address: marie-lise.benot@u-bordeaux.fr (M.-L. Benot).
1
Present address: Department of Geography, University of California, Los Angeles, CA 90095, USA.
2
Present address: I-SEA, Bordeaux TechnoWest, 25 rue Marcel Issartier, 33700 M� erignac, France.

https://doi.org/10.1016/j.ecss.2020.106643
Received 9 March 2019; Received in revised form 16 December 2019; Accepted 12 February 2020
Available online 17 February 2020
0272-7714/© 2020 Elsevier Ltd. All rights reserved.
S. Yeo et al. Estuarine, Coastal and Shelf Science 236 (2020) 106643

appropriately monitoring saltmarshes and having quantitative and ac­ sampling to cover environmental gradients, Kulawardhana et al., 2014,
curate observations of vegetation and their distribution is a critical 2015; stratified sampling from unsupervised classification, Rapinel
component for managing these ecosystems and preserving the dynamics et al., 2018, etc.). On the contrary, field data formerly acquired for other
of natural communities (Belluco et al., 2006). To reach these conser­ purposes may not perfectly match remote sensing requirements, and
vation objectives, large data-sets of ground data are usually acquired, their relevant coupling is still challenging (Luque et al., 2018).
which first enable the establishment of a vegetation classification and Thus, the present study addresses the possibility of using an existing
habitat typology. Such intensive field work is however time-consuming vegetation data-set resulting from an intensive field campaign designed
and thus hard to frequently reiterate. Developing operational tools to for in situ temporal vegetation survey to create a habitat map of the
create accurate maps of saltmarsh vegetation based on the finest National Nature Reserve (NNR) of Ar�es and L�ege Cap-Ferret (South­
possible habitat typology, while limiting the cost and sampling effort, is western Atlantic coast of France) through the combination of multi-
thus indispensable for documenting and managing these natural areas. temporal and multispectral imagery. Our objective was to test the pos­
There has been a growing interest in applying remote sensing tech­ sibility of developing a simple but accurate method to map plant com­
niques to mapping saltmarsh vegetation, and studies have shown that it munities that could reach local managers’ requirements in terms of
is an effective means of collecting spectral information at a large scale in vegetation typology. For that purpose, existing vegetation records were
a noninvasive and cost effective manner (Belluco et al., 2006; Tuxen used for the identification and classification of plant communities by the
et al., 2008; Hladik et al., 2013). Additionally, it has been used to map means of classical ordination and cluster analyses, and their unique
detailed vegetation patterns, distribution, and detect vegetation change spectral signatures from satellite images were retrieved through a
(Belluco et al., 2006; Hirano et al., 2003; Silvestri and Marani, 2004; Maximum Likelihood Classification (MLC). Possible improvements
Wang et al., 2007; Klemas, 2011) through synthesizing temporal and derived from the use of additional LiDAR data were searched for.
spatial dynamics of vegetation patterns (Kelly and Tuxen, 2009). Finally, the effect of a decrease in the sampling effort of the field data
Remote sensing techniques have long been used to monitor saltmarsh and a reduction of ecological information by aggregating vegetation
vegetation, and the availability of high resolution spectral imagery groups into broader classes was evaluated.
coupled with improved image processing and classification techniques
make mapping increasingly accurate (Kelly and Tuxen, 2009). 2. Material and methods
Fine-scale representation of wetland plant species has proved to be
challenging because of the high level of spectral confusion between 2.1. Study area
species (Ozesmi and Bauer, 2002; Schmidt et al., 2004; Tuxen et al.,
2011; Hladik et al., 2013). Additionally, wetlands exhibit high spectral This study was conducted in the National Nature Reserve (NNR)
and spatial variability which can also complicate mapping vegetation FR3600065 of Ar� es and L� ege-Cap Ferret saltmarshes located in the
(Wang et al., 2007; Belluco et al., 2006). It is therefore recommended to northern tip of the Arcachon bay (France, 44� 440 24.0600 N; 1�
have a combination of field data (McRoberts et al., 2002; Klemas, 2011; 90 17.7800 W), a macrotidal coastal lagoon and the only major indentation
Tuxen et al., 2011) in addition to multi-temporal satellite images in in the Southwest coastline of France (Fig. 1). The Arcachon bay repre­
order to acquire information from different vegetation development sents one of the most important oyster farming areas and tourist desti­
stages to increase mapping reliability (Belluco et al., 2006). Many nations in France. The Reserve occupies an area of 320 ha and is home to
studies have previously focused on sites with relatively limited species the largest saltmarsh area in Southwestern France. Hydrologically, the
diversity (see Zhang et al., 1997, 2011), on the quantification of biomass Reserve is under the dual influence of saltwater from the Atlantic Ocean
and carbon stocks (see Byrd et al., 2014, 2018) or on the discrimination and freshwater from upstream freshwater lakes connected to the bay by
of broad vegetation communities or general cover classes (see Kelly, a canal (“Canal des Etangs”), creating a gradient of salinity from the sea
2001; Andresen et al., 2002; Hirano et al., 2003; Stankiewicz et al., moving progressively inland. This factor contributes to the organization
2003; Smith et al., 1998; Wang et al., 2007; Klemas, 2011; Zhang et al., of plant communities ranging from the slikke (mudflats) to the schorre
2011). In line with the link between saltmarsh vegetation zonation and (saltmarsh) and finally transitioning to the foot of the sand dune where
surface elevation (Pennings and Callaways, 1992; Kulawardhana et al., woody communities can be found. In addition to vegetated areas, the
2015 and references therein), several studies demonstrate a promising saltmarsh of the Reserve includes areas of bare sand, mudflats and
potential of using Digital Elevation Models (DEM) and/or two-frequency artificially created hunting ponds.
laser LiDAR imagery to derive vegetation maps in coastal wetlands The Reserve is part of the European Natura 2000 site FR7200679
(Rosso et al., 2006; Chust et al., 2008; Collin et al., 2010; Owers et al., “Bassin d’Arcachon et Cap Ferret” (Reveillas et al., 2012). This site was
2016). In particular, the combination of multispectral images and LiDAR chosen because it is one of the few saltmarshes in France with a diversity
can be an effective way to quantify wetland aboveground biomass and of habitats containing the complete gradient from slikke to schorre.
carbon stocks (e.g., Kulawardhana et al., 2014, 2015) or to map wetland
vegetation classes (e.g., Rapinel et al., 2015; Ballanti et al., 2017). 2.2. Field data and treatment
Additionally, due to the complexity of wetland landscapes, many studies
use hyperspectral imagery due to its advantage in spectral resolution Fieldwork was conducted between the months of June and
(see Hirano et al., 2003; Rosso et al., 2005; Hladik et al., 2013). How­ September in 2011 on the saltmarsh of the Reserve (Reveillas et al.,
ever, these sensors can be impractical as their deployment and the 2012). The sampling strategy was designed to build a temporal in situ
analysis of hyperspectral data can be complex and highly expensive survey of vegetation dynamics based on a spatially explicit method
(Klemas, 2011; Zhang et al., 2011). Moreover, studies have shown that covering the extent of the Reserve under tidal influence. Survey sites
spatial resolution has a higher impact than spectral resolution (see e.g., were regularly distributed based on a 50 m grid superimposed on the
Belluco et al., 2006). entire site extent, with each floristic record performed at the intersection
Remote sensing has been extensively used for coastal wetland of the axes of the grid. This resulted in a total of 832 survey points across
vegetation mapping (Ozsemi and Bauer, 2002). However, it still remains 193.43 ha covering the elevation gradient from the slikke to the schorre.
challenging to identify a cost-effective method, which can meet local The survey points were comprised of 16 m2 records with their x and y
managers’ expectations both in terms of accuracy and ecological rele­ coordinates defined a priori and found on site during the survey using a
vance. One of the possible limits to using the remote sensing approach is GPS with about 5 m average accuracy (Garmin® 62 S). Each record
that it requires spatially and temporally accurate field data both for map contained five 1 m2 quadrats: one in each corner and one in the center.
training and validation. Several strategies of in situ data collection have In each quadrat, coefficients of abundance and cover for each detected
been used to match mapping purposes (e.g., systematic random plant species were noted. By averaging the values in each of the five 1 m2

2
S. Yeo et al. Estuarine, Coastal and Shelf Science 236 (2020) 106643

Fig. 1. Map of the Arcachon Bay (left) showing the location of the study area in Southwestern France. The study site is shown outlined in white (right). Image source:
IGN BD ORTHO ® 50cm.

quadrats, species’ abundance was then evaluated for the entire record. removed for data treatment. Then, floristically heterogeneous records
The cover of bare soil was also taken into account in each of the 1 m2 were removed as they were expected to have a fuzzy spectral signature
quadrats for each record. The notation used for species abundance and that would negatively impact pixel classification. In that purpose, a
bare soil corresponds to the dominance scale (Hill et al., 2005) and Bray-Curtis (BC) dissimilarity index was calculated for each record as
varies from one (less than 4% of vegetation cover or bare soil) to 10 the average Bray-Curtis distance among each pair of quadrats contained
(91–100% vegetation cover or bare soil). in a record (five quadrats per record leading to ten Bray-Curtis dis­
The original dataset was treated following several steps in order to tances). This index was used to determine the floristic dissimilarity
keep only floristically homogenous records suitable for the mapping amongst the quadrats in a record based upon species composition, in
purpose. First, out of the 832 vegetation records initially planned, 156 order to determine floristically homogeneous or heterogeneous records.
were not visited in the field as they were positioned on bare soil, The index ranges between 0 and 1; 0 being exactly similar intra-record
resulting in a total of 676 vegetation records in the database containing species composition and 1 meaning that the quadrats within a record
68 plant species (see Table S1 for the complete species list). Corre­ share no species. A threshold of 0.1 was established as a limit of ho­
spondence analyses (CAs) were used to identify the main gradients that mogeneity, and records with a BC score superior to 0.1 were not retained
structure species’ assemblages and were followed by hierarchical clus­ (194 records). Finally, low frequency species (species present in only
tering (HAC, Ward’s method using Euclidean distance among records one or two records, 27 species) were removed. All of these manipula­
calculated from coordinates along the first three CA axes) to discrimi­ tions led to a final database of 452 records containing a total of 35
nate vegetation groups along these gradients. A first analysis allowed 12 species (Table 1; see also Table S1 for the list of the remaining species).
groups of records to be discriminated on the basis of their floristic A final CA was applied to the filtered database, followed by a hier­
composition. Six of these groups were removed (representing 30 re­ archical clustering of the records based on their scores along the CA1
cords) as they either had too few records or were considered to be and CA2 axes (Ward’s method). Vegetation groups were determined by
outliers on the gradient due to the presence of rare species. Plant species cutting the dendrogram initially to form six groups. Then to progres­
that were not representative of Atlantic and Continental Saltmarshes sively merge the vegetation groups, the dendrogram was cut at higher
(Natura, 2000 codes 1310–1330; six species, see Table S1) were also levels forming five and four groups (Fig. S2). The Indicator Value

Table 1
Phytosociological categories (alliance or association) and Natura 2000 habitat codes (N2000 Code) for the identified vegetation groups. Group codes and generic
names are provided for each determined thematic class in addition to the total number of ground records and the maximal number of records remaining for map
training.
Group Generic name Alliance Association N2000 Total number of Maximal number of
code Code records training records

Six groups
A Top level Atlantic saltmarsh and Agropyrion pungentis Atriplici hastatae-Agropyretum pungentis 1330–5 42 22
grass communities
B High schorre communities Agropyrion pungentis Agropyro pungentis-Inletum crithmoidis 1330–5 95 75
C Mid schorre communities Halimionion Halimionetum portulacoidis 1330–2 104 84
portulacoidis
D High Atlantic slikke communities Salicornion Astero tripolii-Suaedetum maritimae/ 1310–1 51 31
dolichostachyae Salicornietum obscurae
E High slikke Spartina maritima Spartinion anglicae Spartinetum maritimae 1320–1 72 52
swards
F High slikke Spartina anglica pioneer Spartinion anglicae Spartinetum anglicae – 88 68
communities
Five groups
CD Low-mid schorre communities Halimionion – 1330 155 135
portulacoidis
Four groups
AB Top level and high schorre Agropyrion pungentis – 1330–5 137 117
communities

Total 452 -

3
S. Yeo et al. Estuarine, Coastal and Shelf Science 236 (2020) 106643

(IndVal), a quantitative index of the fidelity and specificity of a species 2.5. Classification method
to a group, was calculated for each species in the determined vegetation
types (Dufr^ene and Legendre, 1997). Only indicator species that were A Maximum Likelihood Classification (MLC) was then applied on the
statistically significant (p-value < 0.05) with an IndVal greater than 0.2 false color composite image with and without the DEM layer in order to
were retained. On the basis of indicator species, each vegetation group create spectral signatures for each vegetation group defined by the ty­
was then identified and classified within a phytosociological typology pology and to assign each pixel to a given vegetation group. For the
(alliances and associations) following vegetation syntax described by the classification, two subset groups of the records were created (Wang
Provisional Reference Typology for Natural and Semi Natural Saltmarsh et al., 2007): one to be used as a training group and the other as a
Habitats in Aquitaine by the French National Botanic Conservatory of validation group. The training and validation samples for the classifi­
South-Aquitaine (“Conservatoire Botanique National Sud-Atlantique”, cation were defined by the range of pixel values contained in each re­
CBNSA). When relevant, the N2000 habitat codes were also documented cord, which were then assigned to the vegetation group to which the
(Table 1). corresponding record belonged.
All statistical analyses were carried out using R 2.15.2 (R Develop­ For the validation group, 20 records were randomly selected from
ment Core Team, 2016; http://www.r-project.org/). Packages used each defined vegetation group to test the accuracy of the map. This
included vegan (Oksanen et al., 2016), Ade4 (Dray and Dufour, 2007), number remained constant for each group in each classification. The
labdsv (Roberts, 2016), and agricolae (de Mendiburu, 2016). records selected were evenly spaced according to the ecological gradient
created by the CA (CA1 axis) and dispersed in the study zone in order to
2.3. Multispectral image acquisition, processing and spectral band represent the full range of intra- and inter-group variability. This num­
selection ber was obtained for the sample size from Hay (1979) who concluded
that a minimal sample size of 50–100 pixels per established group is
Very high resolution multispectral images were acquired by the recommended for accurate validation (Janssen, 2001). Each 16 m2
Pleiades Satellite 1 B via a research dedicated program developed by the vegetation record contains four pixels, making a total of 80 pixels for
French National Center for Space Studies (CNES) for the site of the validation per group.
reserve on four different dates: April 14th and 25th, August 21st, and For the training samples, the remaining records independent of the
December 4th in 2013. Each image provides 2 m ground resolution and validation groups were selected at 10% intervals ranging from 100% (all
4-band multispectral (blue, green, red, and near-infrared/NIR) products, the remaining available records of each vegetation group) to 10% in
making a total of four multispectral images with 16 combined bands. order to assess the effect the sample effort on the overall accuracy. These
The images for August and both dates in April were georeferenced by records were randomly reselected for each group for each percentile of
IGN (French National Institute of Geographic and forest information) to degradation in order to define a range of variance of pixels representa­
the French Geodetic Network (RGF93) system and the Lambert-93 tive of each vegetation group to be used in the image classification. The
projection. The image for December was georeferenced by hand using training records were also evenly spaced along the ecological gradient
well distributed ground control points (GCPs). No atmospheric correc­ and dispersed throughout the study site. As no record was acquired in
tions were made on the images, and all four images were cloud free. Two situ for sampling sites containing no vegetation, training samples of
indicators, the Normalized Difference Vegetation Index (NDVI, Rouse water (including channels and hunting ponds), bare mudflat, and sand
et al., 1974) and Normalized Difference Water Index (NDWI, McFeeters, were manually delineated by photo-interpretation and were included in
1996), were calculated for each date of the satellite images to facilitate the classification in order to increase spectral differentiation of the
the detection of vegetation and water in the study area, creating an Reserve. The data-set did not include any validation samples for these
additional eight bands. A Kruskall Wallis rank sum test, with the vege­ habitats.
tation groups as the main factor, was used to determine the bands that A total of 60 classifications were examined for six, five, and four
most easily discriminated the groups according to their numeric pixel vegetation groups, both with and without the DEM data, with the
values. The subset bands for which the Kruskall Wallis rank sum test training sample effort diminishing in 10% intervals. A post-classification
detected a significant difference (p-value < 0.05) between vegetation Majority Filter using the eight nearest neighbors was applied on the
groups were selected. The values of the four 2 � 2m pixels representa­ classified habitat maps in order to smooth out erroneously classified
tive of the vegetation records were extracted for each band by creating pixels and reduce the “speckle” effect. The validation samples were then
16 m2 square buffers around each GPS point. Finally, only NIR, NDVI used to test the accuracy of the smoothed maps in predicting the vege­
and NDWI bands were retained for the four dates as they were statisti­ tation groups across the study site.
cally found to discriminate the best vegetation groups (see Fig. S1 for a ArcMap version 10.2.1 software program (http://www.esri.com)
visual representation of the retained bands) and combined into a single was used for all image analyses, indicator calculations, and classification
false color composite image. processes.

2.4. LiDAR data acquisition and processing

2.6. Accuracy assessment
In December 2011, the Intercommunal Union for the management of
the Gironde pond watershed (“Syndicat intercommunal d’am� enagement Quantitative standard accuracy measurements to assess the classified
des eaux du bassin versant des �
etangs du littoral girondin”, SIAEBVELG) maps is based on the Confusion Matrix, in which the Kappa coefficient
acquired the LiDAR data for the study site. Data was collected with a (K), overall accuracy, producer’s accuracy, and user’s accuracy for each
LiteMapper 6800 mounted in a Cessna 206 (F-GDAP) flown at an alti­ vegetation group were calculated using the error matrix method on the
tude of 500 m above ground level. The survey was conducted with a final smoothed images (Congalton, 1991). The Kappa coefficient (K), a
laser PRF of 200 kHz with a total field of view of 60� and a target point discrete multivariate technique used in accuracy assessment, is a mea­
density of 3.8 laser points per m2 (Parallele 45 G�eom� etres-Experts sure of accuracy in which 0 means an accuracy of less than what would
Associ�es, Lacanau, France). From this data, a digital elevation model be expected by chance and 1 is perfect accuracy (Viera and Garrett,
(DEM) was derived (2 m pixel resolution), creating an additional band 2005). Overall accuracy (A) is computed by dividing the sum of the total
(25 bands in total). correct classified pixels by the total number of pixels in the vegetation
groups. For each classified map, the percent of the standard error for the
overall accuracy (and the Kappa coefficient) was calculated, using the
formula:

4
S. Yeo et al. Estuarine, Coastal and Shelf Science 236 (2020) 106643

rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pð%Þqð%Þ (K ¼ 0.71) for the optical and DEM classification (Table 2A). User’s
SEð%Þ ¼
n accuracy was the highest for vegetation group C (91%), intermediate for
groups A, E and F (77%–84%) and the lowest for groups B and D (62%
in which p is the overall accuracy (or the Kappa coefficient), q is 1-p, and and 68%). Variations in Producer’s accuracy were different with the
n is the total number of pixels in the classification. The percent of the highest values obtained for group B (88%), E and F (86%) and the lowest
standard error was calculated for each classification map, and all errors values for groups A, C and D (64–66%; Table 2A). These values were
were rounded to whole numbers or to the nearest tenth. Producer’s overall lower when DEM was included in the classification (Table 2B).
accuracy (or omission error) indicates the probability of a pixel being The best classification results were obtained for the five vegetation
correctly classified, and is calculated by the total number of correct groups, when 100% of classification records were used, having overall
pixels found in a vegetation group divided by the total number of pixels accuracies of 83.6% (K ¼ 0.8) and 82.4% (K ¼ 0.78) for the optical and
of that group as derived from the ground data. Finally, the user’s ac­ DEM classifications, respectively. The results show that including the
curacy (or commission error), a measure indicative of the probability DEM layer in the classification systematically results in lower A and K
that a pixel classified on the map actually represents that category ac­ values for all groups in all categories, with the exceptions of when 10%
cording to the field data, is calculated by dividing the total number of and 30% of records were selected for the five and six vegetation groups
correct pixels in a vegetation group by the total number of pixels that in which the A and/or K values are slightly higher after the addition of
were classified in that group (Congalton, 1991). the DEM layer (Table 3, Fig. 4, Appendix S2).
Generally, the results show higher A and K values with an increasing
3. Results percentage of classification records used, coupled with decreasing the
number of vegetation groups, despite a few exceptions (Table 3, Fig. 4).
3.1. Vegetation records and groups For the six and five vegetation group categories, the trends in the A
and K values for both optical and DEM classifications follow the same
Using the nomenclature provided by the CBNSA, the six distinct fluctuational tendencies (Table 3, Fig. 4). The K values for these two
vegetation groups defined from using the 452 vegetation records and 35 groups decrease when reducing the number of records used for map
species of the final data table and their corresponding indicator species training from 100 to 90%, remained rather constant from 90 to 60%,
(Fig. 2, see Table S1 for corresponding names to species codes) were then steadily decrease with reducing the sample effort to 20%. The MLC
classified into five phytosociological alliances and six associations accuracy performed with four vegetation groups however, does not
(Table 1). The vegetation groups present a clear spatial gradient follow this general tendency. The K values are stable when the sample
permitting to visualize the typical zonation patterns of saltmarsh plant effort ranges between 100 and 50%, and slightly decrease when
communities from the high schorre featuring top level Atlantic and high reducing the sample effort to 20%. Regardless of the number of vege­
schorre communities (designated as groups A and B) to the slikke, tation groups, decreasing the percentage of classification records down
containing Spartina maritima swards and Spartina anglica pioneer com­ to 10% corresponds in a very small number of records for some classes,
munities (groups E and F) (Table 1, Fig. 3). When the dendrogram was ranging between 2 records (8 pixels) for class A and 14 records (56
cut at higher classification levels, groups C (mid schorre communities) pixels) for class CD (Table 1). This has severe consequences on the map
and D (high Atlantic slikke communities) were first merged, leading to a training, as shown by collapses in overall accuracy and null to negative
rougher group. Then, both A and B groups were merged (Table 1, Kappa values (Table 3, Fig. 4), and the failure to map some of the
Fig. S2). vegetation groups (Appendix S2). Reducing typology accuracy to only
Most of the 196 vegetation records considered floristically hetero­ four vegetation groups prevented this dramatic effect (Table 3, Fig. 4).
geneous according to the Bray-Curtis dissimilarity index were located (i)
within vegetation groups B, C, and D (high schorre, intermediate 3.3. Qualitative map accuracy
schorre, and high slikke respectively), (ii) at the interface between these
groups or (iii) at the interface between these groups and groups A (high For purposes of brevity, nine out of the sixty classifications were
schorre) or E (high slikke) (Figs. S3 and S4). chosen to demonstrate general trends for the classification results
(100%, 50%, and 20% intervals for the three group categories for the
3.2. Quantitative map accuracy optical classification; Fig. 5). From visual interpretation of the vegeta­
tion groups, all classifications demonstrated the distribution of vegeta­
When 100% of the classification records were used, the map repre­ tion groups in the final classified products to be generally consistent
senting the six vegetation groups reached an overall accuracy of 75.5% with typical zonation patterns of saltmarshes and more specifically

Fig. 2. (A) Hierarchical Classification performed on

the [452 records x 35 species] data table. The red
line indicates the level where the dendrogram was
cut to define six clusters (named A-F) used for the
associations. (B) Correspondence analysis on the
F1–F2 factorial plan of the records showing the
vegetation groups (defined in the Hierarchical Clas­
sification). Habitat turnover from high to low marsh
is visualized by the arrow. (For interpretation of the
references to color in this figure legend, the reader is
referred to the Web version of this article.)

5
S. Yeo et al. Estuarine, Coastal and Shelf Science 236 (2020) 106643

Fig. 3. Maximum likelihood classification (MLC) maps (after application of the post majority filter), obtained using 100% of the remaining available records for the
training groups (A) only with optical layers (overall accuracy: 75.5%) and (B) with optical and digital elevation model (DEM) layers (overall accuracy: 73.6%). The
six vegetation groups in addition to water bodies (channels and hunting ponds), sand, and mudflats are featured. No LiDAR data were available for water bodies and
mudflats, which do not appear on the resulting map (B).

Table 2
Interactive Supervised Classification (ISC) error matrix of the final map using 100% of the remaining
available records for the training groups (A) only with optical layers and (B) with optical and digital
elevation model (DEM) layers. Columns represent the reference data (what the pixel actually was
based on validation data) and rows represent the image data (what the pixel was classified as).
Shaded cells show the correctly classified pixel observations for each association. Results are in total
number of pixels, while the producer and user’s accuracy are in percentage.

represented by the field records, showing the gradation from high records used in the training sample, certain spectral groups become
schorre to slikke (Fig. 5), with exception to the 10% category (not more dominant while others become less prominent, or completely
shown) for five and six vegetation groups. In reducing the percentage of disappear (e.g., group A with 20% of training samples and six vegetation

6
S. Yeo et al. Estuarine, Coastal and Shelf Science 236 (2020) 106643

Table 3
Overall accuracy assessment (%) of the classification results for six, five, and four vegetation groups, with and without digital elevation model (DEM) layers.
Input Variables Remaining Records Used in Training Sample (%)

100 90 80 70 60 50 40 30 20 10

Six Groups
Optical 75.5 67.0 65.9 68.4 68.2 62.1 60.8 53.3 49.9 9.6
Optical þ DEM 73.6 62.0 58.4 61.1 64.1 58.0 52.8 52.3 47.9 18.6
Five Groups
Optical 83.6 73.5 72.6 76.0 78.0 70.6 68.2 58.7 57.5 17.1
Optical þ DEM 82.4 70.0 67.2 67.1 73.2 66.1 63.9 59.0 50.4 21.2
Four Groups
Optical 80.1 81.0 81.0 80.4 80.7 81.1 77.5 75.7 75.8 57.5
Optical þ DEM 76.7 72.2 76.7 70.6 76.3 77.7 69.2 70.0 73.4 54.2

For each vegetation group, the validation sample comprised 20 vegetation records.

Fig. 4. Comparison of obtained K values for the six, five, and four vegetation groups with (dashed lines) and without (solid lines) digital elevation model layer. Bars
represent the standard error of the K values for the classifications.

groups, Fig. 5G, Appendix S2). However, regardless of the quantity of Kappa interpretation scale). Both producer’s and user’s accuracies var­
records, the maps produced exhibit an overall agreement with the ied according to the vegetation group considered. The producer’s ac­
general spatial patterns of the observed records. curacy was the highest for groups B (high schorre communities), E, and F
(high slikke Spartina maritima swards and S. anglica pioneer commu­
4. Discussion nities), ranging from 86 to 88%, with group B being the best classified.
From a user’s perspective, the most accurate groups were groups A (top
Complementary to previous studies aiming at mapping saltmarsh level Atlantic saltmarsh and grass swards), C (mid schorre commu­
plant communities, we developed an approach combining robust but nities), and E (high slikke S. maritima swards), with a user’s accuracy
already existing field data collected for other purposes than vegetation reaching 91% for group C. These results imply that the detailed spectral
mapping with multi-date multispectral imagery. This approach proved information provided by the multispectral image data may be particu­
efficient in capturing the full gradient of saltmarsh vegetation, from the larly effective for differentiating vegetation groups in the upper and
slikke to the schorre, deriving a complete vegetation typology and lower marsh. However, in the intermediate marsh (mostly low schorre
mapping of these vegetation groups. and high slikke), the classification was less accurate. Vegetation group D
The classification results show that the mapping method devised (high Atlantic slikke occupied by Salicornia species) had the lowest
from hierarchically classifying vegetation into groups derived from a scores (user’s and producer’s accuracies of 64 and 68% respectively),
complete habitat typology is generally in quantitative and qualitative showing that there is significant spectral confusion in discriminating this
agreement with observed field data. The overall classification accuracy cluster from other clusters. Using hierarchical methods to classify
of the most complete map (i.e., using 100% of available records for map vegetation groups from a robust data-set spanning the entire study site
training and 6 vegetation groups) was rather high (75.5%) with a allowed us to build a thorough habitat typology, therefore reducing
moderate Kappa coefficient of 0.71 (see Viera and Garrett, 2005 for uncertainties of the data. Vegetation typology and the way by which it is

7
S. Yeo et al. Estuarine, Coastal and Shelf Science 236 (2020) 106643

Fig. 5. Maximum likelihood classification maps demonstrating results for 100% (top), 50% (middle), and 20% (bottom) of the remaining available records for the
training groups for the six (A, D, G), five (B, E, H), and four (C, F, I) thematic classes. Only classifications including optical layers are shown.

built may affect vegetation group mapping. In wet grasslands, Rapinel vegetation groups, thereby increasing their spectral discrimination.
et al. (2018) demonstrated that a compromise of vegetation typology Increasing spatial and spectral resolutions have been proposed to
based on both floristic and spectral information provided a better ac­ increase mapping efficiency. Using images with higher spatial resolution
curacy than classic phytosociological typology. In the present study, presents the advantage of having a larger number of reference pixels to
only floristic composition was used to classify vegetation composition, be used in the classifier training, and therefore increasing the spectral
using classical ordination and cluster analyses. Although this may have separability of classes (Janssen, 2001; Hladik, 2012). This is a particu­
decreased mapping accuracy, as suggested by Rapinel et al. (2018), our larly important advantage in the case of highly heterogeneous vegeta­
approach enables the ability to build a typology that can match local tion communities (Belluco et al., 2006) where under-represented classes
managers’ knowledge and conservation needs, independently of the can be lost in the mixed pixels when performing a supervised classifi­
acquisition of remote sensing data. Other studies on supervised image cation (Hladik, 2012). In our study, all of the four pixels constituting a
classification for precisely defined vegetation classes such as phytoso­ vegetation record were assigned to the vegetation group defined at the
ciological associations are rare, however one study combining hyper­ scale of this record. Thus, floristically heterogeneous records, which
spectral imagery with field samples collected using the Braun-Blanquet represented almost 30% of the 676 records containing vegetation, were
method to classify and map vegetation groups had an overall accuracy of removed from the analysis as they were thought to have fuzzy spectral
40% with a Kappa coefficient of 0.35 (Schmidt et al., 2004). This same signatures that could have impacted classification accuracies. Floristi­
study also showed that by combining ancillary data to the classification, cally heterogeneous records indicating small scale (i.e., less than 16 m2)
such as height, slope, aspect, and terrain position, the classification heterogeneity were found all along the slikke to schorre gradient, both
accuracy improved to 66% with a Kappa coefficient of 0.64. Using su­ within vegetation groups and at the boundaries between vegetation
pervised classification for general cover classes based on four saltmarsh groups (Figs. S3 and S4). As it was removed from the analysis, such
dominant species in the Lagoon of Venis, Belluco et al. (2006) reported heterogeneity cannot explain differences in map accuracy. For instance,
accuracies ranging from 74.6 to 96.1% with Kappa coefficients ranging group C (mid schorre) appears to be the one in which the largest number
from 0.67 to 0.86. These high accuracies are typical with broad cover of heterogeneous records are found, but it is also the group with the
classes, as supervised classifications have been shown to have higher highest user’s accuracy. Similarly, group B (high schorre), for which we
performance when fewer classes are distinguished (Janssen, 2001). In obtained the highest producer’s accuracy, also contained several het­
our study, the use of multi-temporal, although opportunistically ac­ erogeneous records. Thus, for these groups, small-scale floristic het­
quired, Pleiades images certainly contributed to the improved map ac­ erogeneity (i.e., at the record scale) did not affect classification quality.
curacy as it enables the integration of phenological differences between On the contrary, group D (high Atlantic slikke) was the least well

8
S. Yeo et al. Estuarine, Coastal and Shelf Science 236 (2020) 106643

represented group. Rather than small-scale floristic heterogeneity, a sort show that incorporating elevation data improves the accuracy of salt­
of “ecological heterogeneity” within records located at the interface marsh aboveground biomass or carbon stock quantification (Kula­
between two vegetation groups could be responsible for misclassifica­ wardhana et al., 2014, 2015), in addition to saltmarsh (Klemas, 2011;
tion. Several factors other than floristic composition, which we used for Van Beijma et al., 2014; Owers et al., 2016), as well as other wetland
vegetation typology (e.g., proposition of bare soil, water content, habitat mapping (Rapinel et al., 2015, 2018; Ballanti et al., 2017). On
biomass), can lead a single vegetation group to be characterized by the contrary, our results show that adding LiDAR-derived DEM layer to
several spectral signatures (Rapinel et al., 2018). Indeed group D is a optical information generally decreased overall map accuracy. This
spatially complex mosaic narrow habitat at the transition between suggests that LiDAR data used in the present study were not accurate
groups C (mid schorre) and E (high slikke S. maritima swards), repre­ enough to detect sharp variations in altitude. In our study, LiDAR data
sented by the smallest number of vegetation records. Defining a clear were used to derive a DEM, while classification improvements provided
signature from this group proved challenging. Other studies have also by LiDAR data do not strictly rely on the accurate description of eleva­
shown that, in heterogeneous areas of saltmarshes, supervised classifi­ tion, but also on the use of LiDAR data to characterize three-dimensional
cations are generally unable to classify under-represented classes and vegetation structure (Kulawardhana et al., 2017), which may help in
mixed pixels (Belluco et al., 2006). Increasing sampling effort within this defining vegetation groups. Moreover, several factors can affect LiDAR
vegetation group would thus be required to try and increase the accu­ accuracy such as sensor resolution or vegetation density, which may
racy of its mapping. Additionally, positional discrepancies in the refer­ limit laser penetration (Hladik, 2012). DEM derived from raw LiDAR
enced data due to accuracy errors of the GPS should also be considered. data may not be accurate enough for flat vegetated habitats such as
These differences could have resulted in a slight shift in the position of those targeted in the present study (Chust et al., 2008; Fernandez-Nunez
the records on the map, leading to a misrepresentation of pixels. et al., 2017). Indeed, poorer vertical accuracy is found for taller vege­
Increasing spectral resolution of the satellite images by using tation (Hladik and Alber, 2012). The application of species or
hyperspectral imagery instead of multispectral imagery could also habitat-specific correction factors (e.g., vegetation height or above­
improve the identification of saltmarsh associations. Hyperspectral im­ ground biomass density) greatly improves the accuracy of the
ages are able to collect a high number of continuous spectral bands LIDAR-derived DEM (Hladik and Alber, 2012; Fernandez-Nunez et al.,
(sometimes greater than 200) and allow for better species differentiation 2017). Corrections are thus generally recommended, without what
when compared to multispectral imagery (Belluco et al., 2006; Hladik, using LiDAR data for saltmarsh vegetation mapping may be useless
2012). However, studies have shown that there is a redundancy in the (Hladik and Alber, 2012). In the present study, the way LiDAR data were
information of hyperspectral bands, creating a tradeoff between the processed is thus the most probable explanation for the negative effect of
number of spectral bands chosen and the accuracy of the classifier, adding the LiDAR-derived DEM band on map accuracy.
therefore the number of spectral bands chosen must be limited. Addi­ One way to improve the overall accuracy of the classification in this
tionally, although higher spectral resolution yields more accurate re­ study would be to merge training samples into broader vegetation
sults, studies have shown that having higher spatial resolution is more classes, defining the groups at a higher hierarchical level in the typology.
important than higher spectral resolution (Belluco et al., 2006). Merged samples are known to give a better spectral representation of the
In the present study, the Pleiades images used were opportunistically vegetation class as a whole (Janssen, 2001), as shown by the high
recorded and dates were not chosen with a specific aim. Another study in classification accuracies in Belluco et al. (2006). This, however, would
the Arcachon bay defined the potential of using long term time-series of be at the expense of important information in regards to the vegetation
satellite imagery to derive exhaustive maps of intertidal and saltmarsh groups. Accordingly, merging vegetation records into coarser groups (i.
habitats (Lafon et al. 2014), demonstrating that good quality maps were e., C-D low-mid schorre communities for the 5-group, and A-B top level
obtained by fusing the results of the Mahalanobis Distance classifier and high schorre communities for the 4 group mapping) overall
applied to SPOT5 imagery acquired from winter to summer 2011 to enhanced map accuracies. In addition, combining clusters E (high slikke
Iterated Conditional Modes (ICM) classification results derived from Spartina maritima swards) and F (high slikke Spartina anglica pioneer
Synthetic Aperture Radar (SAR) seasonal acquisition. They also showed communities) could further decrease confusion amongst pixels and in­
that imagery obtained in the springtime is particularly advantageous in crease spectral differentiation and accuracy mapping of the slikke from
distinguishing species and assemblages, therefore, Pleiades images from the intermediate levels of the saltmarsh, at the expense of distinguishing
April are of great interest. It is also important to consider the difference between the groups E and F. The general trend of the classification maps
in the dates of data acquisition, as the remotely sensed imagery was shows that the accuracy results become highly degraded when finer
taken two years after the field surveys were conducted. As saltmarsh vegetation groups are distinguished coupled with reduced ground in­
vegetation is dynamic, vegetation patterns are subject to change rather formation. As finer vegetation groups may occupy a smaller area on the
quickly (Janssen, 2001). Using synchronized field samples and satellite study site, our systematic field sampling design resulted in a smaller
images could have potentially different classification results. number of records within these vegetation groups. For instance, vege­
Identifying boundaries in which one vegetation group stops and tation groups A and D were respectively represented by 42 and 51 re­
another begins is often difficult, especially in saltmarsh communities cords, leading to a maximum number of 22 and 31 records for map
which exhibit mixed vegetation types blending into one another (Kelly training after the removal of the 20 validation record. A 10% sub-sample
and Tuxen, 2009). In our study, this heterogeneity was detected in of training records is thus insufficient to map these vegetation groups. As
almost 30% of the vegetation records, suggesting that about one third of suggested above, a stratified field sampling i.e., sampling of previously
the vegetation surface is likely to be hardly assigned to one vegetation defined vegetation groups, would help increase the sampling effort in
group or another. Previous research has shown that it is difficult to vegetation groups occupying small areas but could lead to the omission
classify saltmarsh plant communities by purely using their spectral of unknown vegetation types.
signatures (Zhang et al., 2011) therefore many studies use a combina­
tion of satellite images with topographic and edaphic data to increase 5. Conclusion and perspectives
classification accuracies (see Schmidt et al., 2004; Hladik, 2012; Hladik
et al., 2013) and verify the accuracies afterwards with a select number of Although the reliability of the final map can be improved by several
field samples. Saltmarsh vegetation exhibits a strong spatial structura­ methods, the approach developed in this study demonstrates that
tion (Chapman, 1976; Adam, 1990; Marani et al., 2003) and it is well vegetation data-sets that are not initially designed in mapping purpose
established that inundation frequency and elevation are the primary (i.e. systematic rather than stratified sampling, ca. 5 m-accurate record
determining factors for these patterns (Adam, 1990; Boorman, 2003; positioning) can be used in combination with rather easily acquired and
Silvestri and Marani, 2004; Wang et al., 2007; Bilodeau, 2010). Studies processed multi-temporal multispectral imagery to create adequate

9
S. Yeo et al. Estuarine, Coastal and Shelf Science 236 (2020) 106643

saltmarsh vegetation maps. The uniqueness of this study also lies within Bertness, M.D., Ewanchuk, P.J., 2002. Latitudinal and climate-driven variation in the
strength and nature of biological interactions in new england salt marshes.
the robust data-set of ground referenced data covering the entire site, a
Oecologia 132, 392–401.
rich source of information permitting to create a thorough vegetation Bilodeau, C., 2010. Apports du LIDAR � a l’�etude de la v�eg�etation des marais sal�es de la
typology (phytosociological associations) of the Reserve. Decreasing the Baie du Mont Saint Michel. Domain.stic.geop. Universit�e Paris-Est, 2010. French.
number of vegetation records altered classification accuracy, proving <NNT : 2012PEST1007>. <tel-00587416>.
Boorman, L.A., 1995. Sea level rise and the future of the British coast. Coast. Zone Topic.:
that large ground data-sets are required for such a mapping approach to Proc. Ecol. Manag. 1, 10–13.
be efficient. Field data resulting from the fine scale systematic sampling Boorman, L.A., 2003. Saltmarsh Review. An Overview of Coastal Saltmarshes, Their
scheme used in this study would hardly be acquired at larger scales. Dynamic and Sensitivity Characteristics for Conservation and Management. In: JNCC
Report, 334. JNCC, Peterborough, p. 114.
Nevertheless, providing that a sufficient amount of vegetation records Bromberg, G.K., Silliman, B.R., Bertness, M.D., 2009. Centuries of human-driven change
and corresponding pixels per vegetation group are available, our in salt marsh ecosystems. Ann. Rev. Mar. Sci. 1, 117–141.
approach can be transferred to other saltmarshes even on larger areas Byrd, K.B., O’Connell, J.L., Di Tommaso, S., Kelly, M., 2014. Evaluation of sensor types
and environmental controls on mapping biomass of coastal marsh emergent
such as the entire Arcachon Bay. Despite their lower spectral accuracy, vegetation. Remote Sens. Environ. 149, 166–180.
compared to radar or hyperspectral images, multispectral images Byrd, K.B., Ballanti, L., Thomas, N., Nguyen, D., Holmquist, J.R., Simard, M., Windham-
represent an interesting alternative providing that they are acquired at Myers, L., 2018. A remote sensing-based model of tidal marsh aboveground carbon
stocks for the conterminous United States. ISPRS J. Photogrammetry Remote Sens.
several dates. Our results also provide further evidence of the negative 139, 255–271.
effects of including LiDAR data without any correction for the mapping Chapman, V.J., 1976. Coastal Vegetation, second ed. Pergamon Press, Oxford, p. 292.
of vegetated saltmarsh habitats (Fernandez-Nunez et al., 2017). Chust, G., Galparsoro, I., Borja, A.,
� Franco, J., Uriarte, A., 2008. Coastal and estuarine
habitat mapping, using LIDAR height and intensity and multi-spectral imagery.
Quantitative, accurate, and repeatable observations of vegetation
Estuar. Coast Shelf Sci. 78, 633–643.
distribution over time in saltmarshes are highly important for moni­ Collin, A., Long, B., Archambault, P., 2010. Salt-marsh characterization, zonation
toring purposes. Such observations must cover spatial and temporal assessment and mapping through a dual-wavelength LiDAR. Remote Sens. Environ.
scales, and remote sensing techniques are therefore ideally suited for 114, 520–530.
Congalton, R.G., 1991. A review of assessing the accuracy of classification of remotely
this task. The ability for repeated data acquisition of remotely sensed sensed data. Remote Sens. Environ. 37, 35–46.
imaging coupled with the accurately determined spectral signatures of de Mendiburu, F., 2016. Agricolae: Statistical Procedures for Agricultural Research. R
vegetation groups can therefore be used to assess spatial changes in the package version 1.2-4. https://CRAN.R-project.org/package¼agricolae.
Dray, S., Dufour, A.B., 2007. The ade4 package: implementing the duality diagram for
distribution of vegetation over time. ecologists. J. Stat. Software 22 (4), 1–20.
Dufr^ene, M., Legendre, P., 1997. Species assemblages and indicator species: the need for
Declaration of competing interest a flexible asymmetrical approach. Ecol. Monogr. 67 (3), 345–366.
Fernandez-Nunez, M., Burningham, H., Ojeda Zujar, J., 2017. Improving accuracy of
LiDAR-derived digital terrain models for saltmarsh management. J. Coast Conserv.
The authors declare that they have no known competing financial 21, 209–222.
interests or personal relationships that could have appeared to influence Hay, A.M., 1979. Sampling designs to test land-use map accuracy. Photogramm. Eng.
Remote Sens. 45 (4), 529–533.
the work reported in this paper. Hill, D., Fasham, M., Tucker, G., Shewry, M., Shaw, P., 2005. Handbook of Biodiversity
Methods: Survey, Evaluation and Monitoring. Cambridge University Press.
CRediT authorship contribution statement Hirano, A., Madden, M., Welch, R., 2003. Hyperspectral image data for mapping wetland
vegetation. Wetlands 23, 436–448.
Hladik, C., 2012. Use of Remote Sensing Data for Evaluating Elevation and Plant
Samantha Yeo: Formal analysis, Writing - original draft. Virginie Distribution in a Southeastern Saltmarsh. Ph.D. Thesis. The University of Georgia,
Lafon: Formal analysis, Writing - original draft. Didier Alard: Formal Athens, Georgia.
Hladik, C., Alber, M., 2012. Accuracy assessment and correction of a LIDAR-derived salt
analysis, Data curation, Writing - original draft. Ce
�cile Curti: Formal
marsh digital elevation model. Remote Sens. Environ. 121, 224–235.
analysis, Writing - original draft. Aure
�lie Dehouck: Formal analysis, Hladik, C., Schalles, J., Alber, M., 2013. Salt marsh elevation and habitat mapping using
Writing - original draft. Marie-Lise Benot: Formal analysis, Writing - hyperspectral and LIDAR data. Remote Sens. Environ. 139, 318–330.
original draft. Janssen, J.A.M., 2001. Monitoring of Salt-Marsh Vegetation by Sequential Mapping. Ph.
D. Thesis. Faculty of Science, University of Amsterdam, Netherlands.
Kelly, N.M., 2001. Changes to the landscape pattern of coastal North Carolina wetlands
Acknowledgements under the Clean Water Act, 1984–1992. Landsc. Ecol. 16, 3–16.
Kelly, N.M., Tuxen, K., 2009. Ch. 15: remote sensing support for tidal wetland vegetation
research and management. In: Yang, X. (Ed.), Remote Sensing and Geospatial
This study was partly funded by the CNES (SYNIHAL project) and the Technologies for Coastal Ecosystem Assessment and Management. Springer-Verlag
National Nature Reserve FR3600065 “Pr�es sal� es d’Ar�
es et de L�ege Cap- Berlin Heidelberg, pp. 341–363.
Ferret”, France, and by the CarHab program (French Ministry for Ecol­ Klemas, V., 2011. Remote sensing of wetlands: case studies comparing practical
techniques. J. Coast Res. 27, 418–427.
ogy and Sustainable Development). Pleaides images were provided by Kulawardhana, R.W., Popescu, S.C., Feagin, R.A., 2014. Fusion of lidar and multispectral
the CNES/RTU program. The vegetation database was built by Mathieu data to quantify salt marsh carbon stocks. Remote Sens. Environ. 154, 345–357.
Reveillas, Pauline Hernandez, Brice Giffard, Romuald Chapelle, Roxane Kulawardhana, R.W., Feagin, R.A., Popescu, S.C., Boutton, T.W., Yeager, K.M.,
Bianchi, T.S., 2015. The role of elevation, relative sea-level history and vegetation
Brel and Sylvain Poulain. We acknowledge Marius Bottin for providing transition in determining carbon distribution in Spartina alterniflora dominated salt
the updated list of species names. marshes. Estuarine. Coast Shelf Sci. 154, 48e57.
Kulawardhana, R.W., Popescu, S.C., Feagin, R.A., 2017. Airborne lidar remote sensing
applications in non-forested short stature environments: a review. Ann. For. Res. 60,
Appendix A. Supplementary data 173–196.
Lafon, V., Dehouck, A., Clos, G., Sell�e, A., 2014. Multi-temporal and multi-sensor
Supplementary data to this article can be found online at https://doi. classification applied to intertidal flat mapping. In: EARSel 17-18 March, Berlin.
Lefeuvre, J.C., Laffaille, P., Feunteun, E., Bouchard, V., Radureau, A., 2003. Biodiversity
org/10.1016/j.ecss.2020.106643. in salt marshes: from the patrimonial value to the ecosystem functioning. The case
study of the Mont Saint-Michel bay. Comptes Rendus Biol. 326 (1), 125–131.
References Luque, S., Pettorelli, N., Vihervaara, P., Wegmann, M., 2018. Improving biodiversity
monitoring using satellite remote sensing to provide solutions towards the 2020
conservation targets. Method. Ecol. Evol. 9, 1784–1786.
Adam, P., 1990. Saltmarsh Ecology. Cambridge University Press, Cambridge, UK, p. 461.
Marani, M., Belluco, E., D’Alpaos, A., Defina, A., Lanzoni, S., Rinaldo, A., 2003. On the
Andresen, T., Mott, C., Zimmermann, S., Schneider, T., Melzer, A., 2002. Object-oriented
drainage density of tidal networks. Water Resour. Res. 39, 1040.
information extraction for the monitoring of sensitive aquatic environments. IEEE
McFeeters, S.K., 1996. The use of normalized difference water index (NDWI) in the
Int. 5, 3083–3085.
delineation of open water features. Int. J. Rem. Sens. 17, 1425–1432.
Ballanti, L., Byrd, K.B., Woo, I., Ellings, C., 2017. Remote sensing for wetland mapping
McRoberts, R.E., Wendt, D.G., Nelson, M.D., Hansen, M.H., 2002. Using a land cover
and historical change detection at the nisqually river delta. Sustainability 9, 1919.
classification based on satellite imagery to improve the precision of forest inventory
Belluco, E., Camuffo, M., Ferrari, S., Modenese, L., Silvestri, S., Marani, A., Marani, M.,
area estimates. Remote Sens. Environ. 81 (1), 36–44.
2006. Mapping salt-marsh vegetation by multispectral and hyperspectral remote
sensing. Remote Sens. Environ. 105, 54–67.

10
S. Yeo et al. Estuarine, Coastal and Shelf Science 236 (2020) 106643

Oksanen, J., Blanchet, F.G., Kindt, R., Legendre, P., Minchin, P.R., O’Hara, R.B., Rouse, J.W., Haas, R.H., Scheel, J.A., Deering, D.W., 1974. Monitoring vegetation
Simpson, G.L., Solymos, P., Stevens, M.H.H., Wagner, H.H., 2016. vegan: community systems in the great plains with ERTS. In: Proceedings, 3rd Earth Resource
Ecology Package. R package version 2.3-5. https://CRAN.R-project.org/p Technology Satellite (ERTS) Symposium, 1, pp. 48–62.
ackage¼vegan. Schmidt, K.S., Skidmore, A.K., Kloosterman, E.H., van Oosten, H.H., Kumar, L.,
Owers, C., Rogers, K., Woodroffe, C.D., 2016. Identifying spatial variability and Janssen, J.A.M., 2004. Mapping coastal vegetation using an expert system and
complexity in wetland vegetation using an object-based approach. Int. J. Rem. Sens. hyperspectral imagery. Photogramm. Eng. Rem. Sens. 70 (6), 703–715.
37, 4296–4316. Silvestri, S., Marani, M., 2004. Salt-marsh vegetation and morphology: basic physiology,
Ozesmi, S.L., Bauer, M.E., 2002. Satellite remote sensing of wetlands. Wetl. Ecol. Manag. modelling and remote sensing observations. In: The Ecogeomorphology of Tidal
10, 381–402. Marshes Coastal and Estuarine Studies, pp. 5–24. Ch. 2.
Pennings, S.C., Callaway, R.M., 1992. Salt marsh plant zonation: the relative importance Smith, G., Spencer, T., Murray, A.L., French, J.R., 1998. Assessing seasonal vegetation
of competition and physical factors. Ecology 73, 681e690. change in coastal wetlands with airborne remote sensing: an outline methodology.
R Core Team, 2016. R: A Language and Environment for Statistical Computing. R Mangroves Salt Marshes 2, 15–28.
Foundation for Statistical Computing, Vienna, Austria. https://www.R-project.org/. Stankiewicz, K., Dabrowska-Zielinska, K., Gruszczynska, M., Hosciło, A., 2003. Mapping
Rapinel, S., Hubert-Moy, L., Cl� ement, B., 2015. Combined use of LiDAR data and vegetation of a wetland ecosystem by fuzzy classification of optical and microwave
multispectral earth observationimagery for wetland habitat mapping. Int. J. Appl. satellite images supported by various ancillary data. In: Owe, M., D’Urso, G.,
Earth Obs. Geoinf. 37, 56–64. Toulios, L. (Eds.), SPIE. Remote Sensing for Agriculture, Ecosystems, and Hydrology
Rapinel, S., Rossignol, N., Hubert-Moy, L., Bouzill� e, J.-B., Bonis, A., 2018. Mapping IV, pp. 351–361.
grassland plant communities using a fuzzy approach to address floristic and spectral Tuxen, K., Schile, L.M., Kelly, M., Siegel, S.W., 2008. Vegetation colonization in a
uncertainty. Appl. Veg. Sci. https://doi.org/10.1111/avsc.12396. restoring tidal marsh: a remote sensing approach. Restor. Ecol. 16 (2), 313–323.
Reveillas, M., Alfonsi, E., Alard, D., 2012. Typologie et cartographie des habitats naturels Tuxen, K., Schile, L., Stralberg, D., Siegel, S., Parker, T., Vasey, M., Callaway, J.,
du compartiment sous influence tidale –RNN des Pr�es sal� es d’Ar�es et de L�
ege Cap- Kelly, M., 2011. Mapping changes in tidal wetland vegetation composition and
Ferret. University of Bordeaux. ONCFS, p. 61. pattern across a salinity gradient using high spatial resolution imagery. Wetl. Ecol.
Roberts, D.W., 2016. Labdsv: Ordination and Multivariate Analysis for Ecology. R Manag. 19, 141–157.
Package Version 1.8-0. https://CRAN.R-project.org/package¼labdsv. Van Beijma, S., Comber, A., Lamb, A., 2014. Random forest classification of salt marsh
Robins, P.E., Skov, M.W., Lewis, M.J., Gim�enez, L., Davies, A.G., Malham, S.K., Neill, S. vegetation habitats using quad-polarimetric airborne SAR, elevation and optical RS
P., McDonald, J.E., Whitton, T.A., Jackson, S.E., Jago, C.F., 2016. Impact of climate data. Remote Sens. Environ. 149, 118–129.
change on UK estuaries: a review of past trends and potential projections. Estuar. Viera, A.J., Garrett, J.M., 2005. Understanding interobserver agreement: the Kappa
Coast Shelf Sci. 169, 119–135. statistic. Fam. Med. 37 (5), 360–363.
Rosso, P.H., Ustin, S.L., Hastings, A., 2005. Mapping marshland vegetation of San Wang, C., Menenti, M., Stoll, M.P., Belluco, E., Marani, M., 2007. Mapping mixed
Francisco Bay, California, using hyperspectral data. Int. J. Rem. Sens. 26, vegetation communities in salt marshes using airborne spectral data. Remote Sens.
5169–5191. Environ. 107, 559–570.
Rosso, P.H., Ustin, S.L., Hastings, A., 2006. Use of lidar to study changes associated with Zhang, M., Ustin, S.L., Rejmankova, E., Sanderson, E.W., 1997. Monitoring pacific coast
Spartina invasion in San Francisco Bay marshes. Remote Sens. Environ. 100, saltmarshes using remote sensing. Ecol. Appl. 7 (3), 1039–1053.
295–306. Zhang, Y., Lu, D., Yang, B., Sun, C., Sun, M., 2011. Coastal wetland vegetation
classification with a Landsat Thematic Mapper image. Int. J. Rem. Sens. 32 (2),
545–561.

11