MONITORING SALT MARSH CONDITION AND CHANGE WITH SATELLITE REMOTE SENSING

Salt marshes are a frontline of climate change providing a bulwark against sea level rise, an interface between aquatic and terrestrial habitat, important nursery grounds for many species, a buffer against extreme storm impacts, and vast blue carbon repositories. Since the 1700s salt marshes have been in flux due to anthropogenic actions, such as reclamation for development causing loss and an influx of sediment from land clearing leading to marsh expansion. The Clean Water Act of 1972 provides legal protections for wetlands, limiting wetland reclamation and requiring that impacts be offset. However, salt marshes continue to change rapidly due to anthropogenic stressors including elevated rates of Sea Level Rise (SLR) due to climate change, herbivory driven by overfishing, droughts, and eutrophication. Salt marsh monitoring across large spatial extents requires remote sensing. This dissertation’s objectives include: Developing methods for monitoring how midAtlantic salt marsh ecosystems are changing and where, determining how restoration and Hurricane Sandy affected Jamaica Bay’s salt marshes, and quantifying the effect of the tidal stage at the time of acquisition on very high spatial resolution (<1 m) salt marsh mapping. This dissertation is composed of three chapters in the format of published and prepared manuscripts for professional journals. In chapter/manuscript 1, a methodology for monitoring salt marsh with very high resolution imagery was developed and applied to the Jamaica Bay Unit of Gateway National Recreation Area. Jamaica Bay’s salt marshes were mapped using object-based image analysis (OBIA), random forest classifier, and a diverse set of data including high spatial resolution (<1 m pixel size) satellite imagery. Change analysis was conducted at Gateway National Recreation Area with satellite imagery collected in 2003, 2008, 2012, and 2013. All classifications achieved >85% overall accuracies. In Jamaica Bay, from 2012 to 2013, restoration efforts resulted in an increase of 10.6 ha of salt marsh. Natural salt marshes within the Bay demonstrated a decreasing trend of loss. Larger salt marshes in 2012 tended to increased vegetation extent in 2013 F(4, 6) = 13.93, p = 0.0357 and R 2 = 0.90). In chapter/manuscript 2, the effect of the tidal stage on salt marsh mapping was modeled using topobathymetric LiDAR and VDatum. Verification of the tidal effect on very high resolution imagery was explored within Jamaica Bay using bathtub models derived from topobathymetric LiDAR and imagery data collected at a range of tidal stages. The effect of the tidal stage was minimal at 0.6 m above MLW, only 3.5% of S. alterniflora was inundated. This varied greatly between salt marsh islands within the Bay. In chapter/manuscript 3, salt marshes change across seven HUC-8 mid-Atlantic watersheds was mapped from 1999 to 2018 using time series analysis of the Landsat 7 and 8 archives with Google Earth Engine. Back-barrier salt marshes are integral to the barrier systems function and their long-term resilience in the face of SLR and future extreme storms. This analysis included watersheds across Maryland, Delaware, northern North Carolina, Virginia, New York, and New Jersey. Aboveground green biomass across the mid-Atlantic declined by an average of -68 g m. The Landsat derived estimates of aboveground green biomass were an indicator of salt marsh vegetation extent within a pixel (F(1165,1)=1316, p < 0.001) and R =0.53 Salt marsh environments along the mid-Atlantic coast are in decline and projected to suffer more losses due to SLR. These changes are evident with both localized mapping and regional assessments. Satellite remote sensing monitoring provides the spatial context necessary for successful salt marsh management. The response of salt marshes to SLR is uncertain, where will migration, persistence, and loss occur? Satellite remote sensing of salt marsh change is necessary for the appropriate management of these ecosystems. The synergistic stressors that are driving loss require both in situ monitoring to determine change and remote sensing to expand these analysis beyond a singular location.

. Accuracy assessment analysis (producer's, user's, and overall accuracy) 26 Table 3. Change between 2003 and 2013 (ha). Areas that had no change between the two dates are in grey. 27 Table 4. Change of land cover classes between 2012 and 2013 for West Pond area. 28   Table I. Tidal stage at time of Worldview-2 (WV-2) and Quickbird-2 (QB-2) image acquisition for the data utilized 73 Table II. Accuracy assessment conducted with stratified random selection of 765 points. Producers, users and overall accuracy were calculated for the 2013 classification [6]. Land cover classes are abbreviated as MUD=Mudflat, Sand, WK=Wrack, SA=S. alterniflora, PSA= Patchy S. alterniflora, HM= High Marsh, PHG= Phragmites, WTR= Water, UP= Upland, UA = Users Accuracy, PA = Producers Accuracy, OA=Overall Accuracy 73 Table III: modeled and classified impact of tidal stage on NDVI for elders point east 73 Table IV. ANCOVAs results comparing inundation between islands for S. alterniflora and NDVI. 74 Chapter 3 x Table 1. The percentage of change, total area, and mean trend of E2EM1P, E2EM1N, E2EM1Pd and, E2EM1Nd classes from 1999 to 2018. 111 Table 2. The results of the Moran's I test of spatial autocorrelation for each of the watersheds. 112 Table 3.

Introduction
Jamaica Bay, an estuary within the New York City (NYC) limits, is heavily influenced by urbanization. The salt marshes serve as an interface between the Bay and surrounding urban areas. Currently, over a dozen marsh islands span the Bay.
Their landscapes are composed of mudflats, a variety of salt marsh plant species, sediment deposited to rebuild drowning salt marsh, transitional vegetation denoting the shift to upland, and human created upland areas. Salt marshes provide numerous ecological benefits such as high biodiversity, improved water quality, flood reduction, and carbon sequestration [1]. The wetland ecosystems of New York State, including salt marshes, were reduced by 60% from 1780 to 1980 [2]. Nationally, salt marshes have been under particular stress with increasing rates of loss from 2004 to 2009 caused in part by coastal storms [3]. In the past, these trends were exacerbated in the urban-impacted Jamaica Bay.
Jamaica Bay's salt marsh loss is severe. Since 1951, approximately 60% of the Bay's salt marsh has converted into mudflats due to a combination of factors including a reduction in sediment supply, changes in tidal regime, nutrient enrichment and increased hydrogen sulfide concentrations [4]. This estimate does not include areas of wetlands around the estuary lost to land filling and urbanization. From 1989-2003, Jamaica Bay's salt marshes were in rapid decline losing 13.4 ha/year [5]. The nitrogen load of the Bay is one factor that may contribute to this high rate of loss [6].
Remote sensing is uniquely suited for monitoring coastal environments, due to the difficulty of in situ access and the high temporal resolution required to understand these dynamic landscapes [7]. Remote sensing monitoring of the salt marsh landscape can be used to determine vegetation trends for the entire bay and individual islands, facilitating an assessment of restoration impacts. Remote sensing is an important tool for furthering our understanding of how Jamaica Bay's salt marshes are affected by anthropogenic and natural factors [8,9]. This study used imagery data spanning a decade and two high resolution sensor systems.
In October 2012, Hurricane Sandy impacted the coast of New York and surrounding states with high winds and storm surge. It was a 1 in 500-year storm surge event at the Manhattan Battery [10]. The boroughs of Brooklyn and Queens directly surrounding Jamaica Bay were inundated; the storm caused 2 million New Yorkers to lose power [11]. This study seeks to understand the impact of Hurricane Sandy on salt detrimental effects on coastal salt marshes [12]; thereby, enhancing the need for accurate remote sensing monitoring and assessment of coastal wetlands to inform decision-makers.

Study Area
Jamaica Bay is an urban estuary residing within the New York City boroughs of Brooklyn and Queens. Kings County, synonymous with Brooklyn, is the most populated county in New York State [13]. Approximately 3,704 ha of the Bay are managed by the National Park Service as Jamaica Bay National Wildlife Refuge, a subunit of Gateway National Recreation Area ( Figure 1). The region has a humid continental climate with a mean temperature of approximately 10 °C. Over the last 150 years, anthropogenic impacts to Jamaica Bay have been extensive. The Bay's volume has increased 350% while surface area fell by approximately 4,856 ha [14]. In 2005, Waste Water Treatment Plants serving 1,610,990 people discharged into Jamaica Bay [15]. Beginning in 2003, salt marsh islands including Big Egg, Yellow Bar, Rulers Bar, Black Wall, Elders Point East and West ( Figure 1) have undergone salt marsh restoration. After restoration, sites were monitored in situ for 5 years [4]. These marsh restoration projects involved the deposition of dredge sediment from channels in the Bay onto the marsh surface then the transplanting and seeding of salt marsh vegetation [16].

Remote Sensing Data
High spatial resolution Quickbird-2 and Worldview-2 data were employed for salt marsh mapping and change analysis. This study uses object-based image analysis (OBIA) which first divides an image into objects, using a segmentation algorithm, and then classifies those objects based on their spectral and spatial attributes [18]. Object-based change detection (OBCD) utilizes image objects to conduct a change analysis between multiple time periods. The change analysis can be conducted with object attributes, classified objects, multi-temporal image objects, or a hybrid of these techniques [19]. This study   compared the classified 2003, 2008, 2012 and 2013 objects to understand restoration and Hurricane Sandy's impact on wetlands within Jamaica Bay.

Segmentation
An important component of OBIA classifications is the determination of segmentation scale, which determines the size and similarity of resulting image objects, and parametrization i.e. the inclusion of texture [20]. Texture is the use of a moving window to quantify measures that represent ideas such as coarseness and roughness [19]. This study arrived at an appropriate segmentation scale with the comparison of multiple segmentation scales for each time period to maximize intrasegment homogeneity and intersegment heterogeneity [21,22]. The parametrization of the resulting image objects included spectral values, texture, geospatial attributes, upland data, vegetation indices, and neighborhood and scene difference attributes (Described in Section 2.4). Segmentation scale is the key to accurately mapping a landscape. Scale parameters can be arrived at through "trial-and-error". However, this method risks determining an inappropriate segmentation scale. Over or under segmenting an image can result in lower classification accuracy [23]. In addition, segmentation scale can impact the land cover classes that can be accurately mapped [20]. This study used the mean shift clustering approach to determine segmentation.
Mean shift is a non-parametric segmentation algorithm which groups pixels based on their spectral mean in a feature space. The algorithm has improved accuracy when compared to other clustering techniques [24,25]. Mean shift considers a spectral radius in the feature space as the scale parameter, which results in a hierarchical relationship between segmentation scales [26]. These factors make the algorithm suitable for multiscale segmentation.
There are different methods for assessing the quality of segmentation. This study assessed segmentation scales with an index of intra-segment homogeneity and intersegment heterogeneity [21]. Intersegment heterogeneity was assessed through computation of Global Moran's I that were normalized and then combined with the intra-segment homogeneity, as determined by normalized area controlled variance, to create a single parameter measuring segmentation quality [22]. The mean shift segmentation parameters that were determined were minimum size and spectral radius.
Minimum size refers to the fewest number of pixels that can compose a segment, and spectral radius is the distance in the feature which a pixel must be within to merge into the segment. Each image date was tested with the parameters from 5-50 spectral radii in increments of 1 and minimum size from 5-50 in increments of 5. Appropriate segmentation scale for the Worldview-2 2012 data was determined to be a spectral radius of 15 and a minimum size of 5 pixels. The 25% most over segmented objects were segmented again at a quantitatively determined appropriate scale of spectral radius 20 and minimum size 5 pixels. The same was done for 25% most under segmented objects, for which the appropriate scale was spectral radius 6 and minimum size 5 pixels. The appropriate scale for the Worldview-2 2013 data was determined to be spectral radius 22 and minimum size 20 pixels. The 25% most over segmented objects were segmented again at a quantitatively determined appropriate scale of spectral radius 27 and minimum size 5 pixels. The 25% most under segmented objects were re-segmented at a scale of spectral radius 7 and minimum size 5 pixels. The Quickbird-2 data were segmented at a spectral radius of 8 and a minimum size of 20 pixels. No additional levels of segmentation were done as this scale adequately captured the landscapes and spectral complexity of the Quickbird-2 data.
The classification was conducted with the Random Forest classifier. Random Forest is a non-parametric ensemble learning algorithm that has been demonstrated to achieve appropriate classification accuracy in a variety of landscapes [25, 28, and 29].
The 9 classes used in this study included 6 from a previous study of the Bay [8]. These classes included water, mudflat, sand, high marsh, patchy S. alterniflora, and S. alterniflora (≥50% vegetation cover). The two S. alterniflora classes were based on percent cover with patchy being between 10%-49% vegetation cover and S. alterniflora (≥50% vegetation cover) being above 50%. Salicornia species are present within the Bay as a small component of the salt marsh [30], and were not prevalent enough to classify on their own. Additional classes included in this study are wrack, upland vegetation, Phragmites, and shadow, however shadow was removed with a decision tree post-classification. The 2003 classification did not include wrack due to the limited separability of the class in those images. These additional classes were included to expand our understanding of the Bay and inform management decisions.

Object Attributes
Spectral attributes included the mean and standard deviation of all available spectral bands. The spatial variables computed were perimeter, area, and nodes. The panchromatic band was utilized to create Grey-Level Co-Occurrence Matrix (GLCM) textural measurements, including inverse difference moment, entropy, contrast, correlation, and uniformity. GLCM and other texture measures have been shown to improve classification accuracies in both Very High Resolution image classification [28] and object-based wetland classification [32]. Red Edge-based vegetation indices, have been shown to more accurately discern differences between high density vegetation species [33]. In this study, Worldview Vegetation Index (WVVI), Worldview Water Index (WVWI), Red Edge-based NDVI, NDVI, and Soil Adjusted Vegetation Index (SAVI) were calculated after pan-sharpening due to its benefits for detecting small vegetation patches (formulas in Table 1) [34]. Ancillary data included an upland GIS layer created from a geomorphological map of Jamaica Bay [35] and Digital Elevation Model (DEM) derived from 2014 Topo-bathymetric Light Detection and Ranging [36].
Object neighborhoods, those objects that share a border with an object, and weights were calculated to determine the neighborhood difference of the mean spectral, textural and vegetation index attributes giving additional spatial context to the data [33]. The final Worldview-2 image objects had 79 attributes including 3 spatial attributes, 18 texture attributes, 32 spectrally derived attributes, 7 elevation based, 18 vegetation index, and a binary upland variable (Table A3). The Quickbird-2 image objects had additional attributes including tasseled cap values but no Red Edge based NDVI.

Accuracy Assessment
The accuracy assessments were conducted for each classification by generating equalized random points. The number of points to generate was calculated with following equation [34].
Where is the Chi-squared distribution with 1 degree of freedom for the target error divided by the number of classes, ∏ is the percent land cover of the most prevalent class and is the desired confidence interval of that class.

Wetland Change
The 2003 and 2013 classifications were compared to determine change between all classes (Table 3).  (Figures 1 and 2).
These islands were not being actively restored during the 2012-2013 period, however they did increase in salt marsh extent (Table A1)

Impact of Hurricane Sandy
West Pond ( Figure 2) is a retention pond created during the construction of the Cross Bay Boulevard and an important resource for migratory birds [42] (Figure 4).
Hurricane Sandy breached West Pond, resulting in salt water intrusion into the fresh water environment [43]. Prior to this breach, West Pond's wetlands were dominated by Phragmites australis. The area represents the most drastic change from Hurricane Sandy; alterations to the upland and freshwater wetlands are evident ( Figure 4, Table   4).
Between 2003 and 2013, the JoCo site lost salt marsh vegetation going from 131.2 ha to 127.6 ha. However, from 2012 to 2013 vegetation increased (Table A1).
This increase in vegetation was accompanied by a reduction in wrack across the Bay Since the mid-2000s, the Bay has had a 30% reduction in nitrogen load [44].
Nutrient enrichment in salt marsh systems can lead to creek bank collapse and conversion to mudflat [6]. The Waste Water Treatment Plants in Jamaica Bay account for 89% of all nitrogen inputs into the Bay; due to the Bay's currents, the highest nitrogen concentrations were in the south and eastern sides of JoCo [15]. The different responses of salt marshes in the Bay to nutrient enrichment was partly explained by lower elevation marshes having longer periods of inundation increasing decomposition and loss of organic matter [45]. The nitrogen load reduction coincided with the slowing of salt marsh loss, however, the impact is unknown and in situ analysis would be necessary to explore this possible connection.

Restoration
In 2003, the first salt marsh restoration in Jamaica Bay began at Big Egg. The project utilized dredge sediment to increase marsh elevation and then S. alterniflora plugs were planted 50 cm apart [17]. In 2006, Elders Point East's elevation was increased with dredge sediment and then vegetated with both plugs and hummock relocation, the removal followed by placement of the entire salt marsh platform on areas of restored elevation [4].

Hurricane Sandy
The response of salt marshes to storm events vary and include net elevation

Wrack
Wrack Throughout the Bay most wrack became S. alterniflora, capturing the removal of wrack and regrowth of impacted salt marsh vegetation in the following growing season. These findings suggest recovery from wrack can be rapid, with storm events as a major driver in the deposition and distribution of the material throughout Jamaica Bay.

Long-term monitoring;
The two most prevalent mapping protocols for wetland change analysis are the National Wetland Inventory Salt marsh losses are increasingly driven by sea level rise and high water events causing migration of S. alterniflora into areas previously composed of high marsh [60]. In order to understand these shifts between vegetation communities, a specialized high resolution classification is necessary. When conducting analysis over large areas C-CAP and NWI programs are invaluable. However, a specialized protocol is preferable when presented with single study site and unique management issues.
The regular collection of satellite imagery is necessary for long-term monitoring. This can have a prohibitive cost, when using very high resolution satellite data. This study's five-year data collection interval and additional data collected following the storm event was adequate for understanding both the decadal trends and Hurricane Sandy's impact. Jamaica Bay is representative of the future for increasingly populated coastal communities worldwide, necessitating continued remote sensing monitoring of the impact of urbanization on the Bay's salt marsh. Long-term monitoring requires additional exploration of the impact that multiple sensors have on change analyses. The switch from Quickbird-2 to Worldview-2 could be partly responsible for the change seen from 2008 to 2012. Quantifying this impact is a necessary step as we proceed into the third decade of commercially available very high resolution satellite imagery.

Conclusions
This study reiterates the importance of continuing salt marsh monitoring with high spatial resolution satellite data within Jamaica Bay

SUPPLEMENTARY MATERIALS
Appendix A Table A1. Land cover extent of salt marsh islands (ha).       [1], [2]. Salt marsh losses in Jamaica Bay, an estuary within New York City, are driven primarily by nutrient enrichment, an increased tidal range, a lack of sediment, and increased sulfide concentrations [3]. Jamaica Bay has a long history of salt marsh mapping and monitoring using remote sensing. Salt marshes mapped from aerial photographs acquired in the 1950s demonstrated significant losses [4]. Since 2003, very high-spatial-resolution satellites have been used to monitor and determine the change in the bay [5], [6]. An object-oriented classification using the Worldview-2 satellite imagery has been used to map the salt marsh extent and the change caused by a storm event and restoration activities in the bay [6].

Marsh
The accurate determination of the salt marsh extent and the change by remote sensing is impacted by the tidal stage at the time of image acquisition. When mapping a vegetation change in tidal environments, differences in the tidal stage can lead to an erroneous identification of change [7]. The influence of the tidal stage on salt marsh vegetation mapping is a topic that has been addressed infrequently in the literature.
Salt marsh vegetation zonation and extent are dependent on many factors driven by tidal inundation. For example, the lower bound of the growth range of smooth cordgrass, S. alterniflora, is limited by physical stress from abiotic factors [8]. A tidal stage above mean low water (MLW) can reduce the extent of vegetation mapped; an imagery acquired above mean highest high water corresponded with a 40% reduction in the mapped vegetation extent [9]. That study led to the recommendation that when mapping salt marsh, an imagery should be acquired within 0-0.6 to a maximum of 0.9 m above MLW. These guidelines have been applied to the Coastal Change Analysis Program protocol and other salt marsh mapping projects [10], [11]. The spatial resolution of remote-sensing data can influence many aspects of image classification and the coastal change analysis [12]. A variety of high-spatial-resolution imageries, including Worldview-2, Quickbird-2, orthoimagery, and historic imageries, have been utilized for mapping salt marshes [13]. Therefore, understanding the impact of the tidal stage on a very high resolution (VHR) imagery in coastal mapping is necessary.
In this study, impact is defined as an increase in misclassification of salt marsh vegetation due to tidal inundation muting spectral differences. The study quantifies this as those areas with normalized difference vegetation index (NDVI) < 0 in the imagery and inundated areas in the models.
There have been several approaches to quantifying and accounting for tidal uncertainty in remote-sensing classifications. In situ measurements and the Quickbird-2 satellite-obtained spectra have been found to be similar despite a variety of tidal stages [14]. For a medium-resolution imagery, a digital elevation model (DEM) in combination with a satellite imagery has been used to quantify and limit the impact of the tidal stage on vegetation mapping [15]. In this paper, we explored a novel approach to understand the impact of the tidal stage on the vegetation extent using VHR satellite remote-sensing data and topobathymetric light detection and ranging (LiDAR).

LiDAR is often incorporated into salt marsh classifications with the creation of
LiDAR-derived vegetation indices [16] or LiDAR-derived elevation to augment spectral classifications [17]. The limited penetration of LiDAR into the salt marsh canopy can result in a bias toward higher elevations [18]. However, areas of dense canopy are minimally impacted by tidal inundation unless completely submerged. This makes the bias toward including the salt marsh vegetation height in ground elevations within salt marshes a minor concern for this paper.
Bathtub models are a method to determine inundation. A DEM is used to determine whether a pixel is inundated or not at a certain tidal stage or flood elevation.
Additional nuance can be added by adjacency rules, i.e., a number of adjacent pixels must be inundated before a pixel is considered inundated [18]. Bathtub models have been used to determine SLR [19] and storm surge impacts [20] for coastal landscapes.
Inundation has been shown to increase with the spatial resolution of the DEM [19].
Local tides can influence these predictions, and tidal variation can be included in bathtub models by converting elevation data to a tidal datum with software, such as VDatum [21]. Bathtub models are commonly used to assess SLR and have yet to be utilized to understand tidal impacts on VHR salt marsh mapping.
This paper seeks to understand the relationship between the elevation and the salt marsh vegetation extent within Jamaica Bay by modeling the tidal stage impact on NDVI and classified S. alterniflora from MLW to mean high water (MHW). This  were used to understand the tidal variation across Jamaica Bay [27], [28] (Fig. 1). The 2013 Worldview-2 imagery was coregistered to the LiDAR generated DEM. All other satellite imageries used in the analysis were coregistered to the 2013 Worldview-2 imagery (Table I).

C. Object-Oriented Classification
An object-oriented classification approach was used, which begins with segmentation, i.e., dividing an image into spectrally similar patches. Objects were then classified giving a greater geospatial context and addressing many limitations of pixelbased methods [29]. Jamaica Bay's salt marsh islands were segmented using mean shift segmentation at multiple scales; the random forest classifier and a diverse set of parameters, including neighborhood differences, gray level co-occurrence matrix texture, and vegetation indices, were used in the classification [5]. The classification scheme included nine classes, Spartina alterniflora, Patchy S. alterniflora, Phragmites, upland, mudflat, water, high marsh, wrack, and sand. The Patchy S. alterniflora classes were those objects with 10%-49% cover, and S. alterniflora were those segments with ≥50% vegetation cover. A multiscale segmentation approach was implemented using local Moran's I and variance to determine which objects were under-and oversegmented and resegment those objects at a more appropriate scale [30]. The classification excluded DEMs to remain independent of the bathtub models which used the topobathymetric LiDAR. The classification results from September 19, 2013 Worldview-2 data were used as a baseline for analysis due to a tidal stage near MLW, and temporal proximity to the topobathymetric LiDAR collection date.

D. Topobathymetric LiDAR Data
Topobathymetric LiDAR systems collect both terrestrial and nearshore elevation simultaneously. The topobathymetric data were collected from January 8, 2014 to May 22, 2014 and achieved submerged accuracy and terrestrial vertical accuracy of 0.062 and 0.214 m, respectively [31]. The LiDAR point cloud data were binned and averaged into a DEM with 0.5-m spatial resolution to match the spatial resolution of the pan sharpened Worldview-2 data.

E. Tidal Analysis
Elders Point East, a salt marsh island in the northern portion of the study area  (Table   I).
NDVI was used as a proxy for the vegetation extent. A threshold of NDVI > 0 was applied to each of the images, all areas with NDVI > 0 were determined to be potentially vegetated. Imageries from 2012 and 2013 were included in the analysis as the area experienced a little change. The largest land cover change from 2012 to 2013 for Elders Point East was the reduction in areas classified as wrack [6]. This should have minimal impact due to the inclusion of wrack in the NDVI threshold.
Stony Creek, a salt marsh island in the western side of the bay (Fig. 1), was selected to compare the NDVI of objects derived from Worldview-2 imagery data as

F. Bathtub Modeling of S. alterniflora
The bathtub models of the tidal stage went from MLW to MHW at 5-cm intervals to correspond with the growth range of S. alterniflora. The growth range of Spartina alterniflora varies in the region with a lower bound above MLW and an upper bound around MHW [32]. VDatum was used to convert the LiDAR data from NAVD 88 to MLW. VDatum has been evaluated for use at the study site finding in situ and modeled elevations differed by a mean of 6.4 cm [32]. However, the conversion did introduced areas of no data to several marsh islands due to the VDatum's conversion extent. The study used only salt marsh islands which were completely converted into the MLW tidal datum (Fig. 1). Salt marsh islands with more high-marsh and upland areas, such as JoCo and Black Bank, were not fully converted and therefore excluded. The tidal surfaces were utilized to simulate the impact of a tidal stage on the classified vegetation and the NDVI of the 2013 Worldview-2 imagery.

G. Statistical Analysis
The RMSE of the tidal inundation bathtub modeling was calculated using

A. Image Classifications for Salt Marsh Mapping
The 2013 salt marsh classification included all tidally influenced areas of the salt marsh islands (Fig. 3). The 2013 classification with and without a DEM was trained with the same data. The out of box overall accuracy, a subset of samples withheld during each iteration of the classifier, was compared finding overall accuracies of 94.4% and 92.5% with and without a DEM, respectively. The classification with a DEM achieved a 92.81% overall accuracy with an independent accuracy assessment (Table II) [5].

B. Tidal Stage
The impact of the tidal stage on the NDVI for the Elders Point East site was determined for five dates of imagery across a tidal range of 162 cm. In the subset, areas with an NDVI > 0 were reduced by 82% (Table III)

C. Bathtub Modeling
Bathtub modeling of the tidal stage at 5-cm intervals was applied to an NDVI > 0 layer and a S. alterniflora classified layer for a subset of salt marsh islands. Bay wide inundation of the salt marsh vegetation was minimal before 0.6 m above MLW (Fig. 4). However, the salt marsh islands, including Black Wall, Rulers Bar, and Pumpkin Patch, had ∼20% of the S. alterniflora inundated at 0.6 m above MLW; these salt marshes had significantly different inundation regimes than other salt marshes (see Fig. 4 and were not reduced to zero NDVI and they were impacted by the tidal stage. S. alterniflora height increases with the depth of tidal inundation due to increased nutrients and reduced edaphic stress [34]. The relationship between S. alterniflora and tidal inundations gives the possibility that these bathtub models are applicable between study sites. The relationship between the tidal range and the lower bound of S. alterniflora has been quantified as zmin = 0.7167 * (TidalRange) − 0.0483 [35].
At the tidal station located at Sandy Hook, NJ, USA, the range from MLW to MHW is In Jamaica Bay, algae deposited on beaches and mudflats is common due to eutrophic conditions and could add to misclassifications when mapping salt marsh.
Algal blooms are common within the bay during the late spring and summer months [36]. Algal blooms result in algal deposition on mudflats which can be misclassified as vegetation due, in part, to the strong NIR value of the algae [7]. Algal mats on mudflats and beaches within Jamaica Bay create uncertainty in change between land cover classes and are difficult to include in the analysis due to their transience. The analysis of Elders Point East suggested that at 70.1 cm above MLW, the tidal stage had a reduction (1/2 hectares) in areas above the NDVI threshold, and a little S. alterniflora was impacted. However, due to the restoration activity on the island, this model was not representative of other salt marsh islands in the bay (Table IV) alterniflora. Inundated areas can still be mapped as vegetation, though it is likely that spectra will be altered leading to more variability in the spectral signature for the S.
The impact of the tidal stage on the VHR mapping of S. alterniflora was similar to past estimates with a medium resolution imagery [8]. The image analysis method is preferred for determining local tidal impacts. However, the acquisition of several VHR images is often prohibitively expensive, making the modeling approach reasonable for understanding local tidal characteristics. The tidal impact on S.
alterniflora was varied by a salt marsh island. These differences were due to Jamaica Bay's tidal variability, vegetation characteristics, and restoration actions. S. alterniflora marshes have an area of taller high biomass vegetation along the marsh edge [38]. The finer spatial resolution would pick up some of these differences between edge and interior salt marsh. In addition, ground elevation was used in this analysis not accounting for differences in the vegetation height. These taller edge areas were less impacted by canopy inundation than shorter interior S. alterniflora. The use of the VHR imagery did not, on its own, limit the impact of the tidal stage on the mapped vegetative extent. Salt marsh mapping requires accurate measure of fine-scale changes in land cover; therefore, even minor tidal impacts are of concern and should be quantified.
Previous classifications and the change analysis of the study site (2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013) used imagery acquired at a range of tidal stages [6]. The tidal stage of the 2003 imagery was 78.1 cm and outside the recommended 60 cm of MLW (Table I) observed rates and is likely to continue to increase due to global warming [39].
Regional SLR in the mid-Atlantic is projected to be between 30 and 50 cm greater than global SLR by 2100 [40]. SLR is a major concern for Jamaica Bay. A 30-cm SLR scenario is projected to cause extensive salt marsh loss in the western portion of the bay [41] and could be further exacerbated by eutrophication [42], [43]. This tidal inundation analysis can be utilized to understand the areas of potential salt marsh loss within Jamaica Bay. 1) Tidal stage is even more a concern for VHR coastal mapping due to the desire for fine-scale measurements.
2) Tidal stage variation throughout a study site can be modeled improving the estimates of uncertainty.
3) When mapping S. alterniflora, the lower growth range of the species can be used to ensure limited impact and allow for an understanding of tidal impacts in microtidal areas.         Therefore, the effect of tidal outliers is a concern in salt marsh environments. The tidal marsh inundation index (TMII) has been successfully used to identify inundated pixels

Study site
The mid-Atlantic coastal region has a variety of estuaries and bays including  (2017).

Statistical Analysis
In this study, the time series were analyzed for breakpoints with the Breaks for

Biomass modeling and change
The ability of the time series trend component to reveal salt marsh change was evident in the identification of both losses and gains across the watersheds. Across the studied watersheds 52% of salt marsh experienced a decline in aboveground green biomass with an average reduction of -17 g m -2 (Table 1). In the Chincoteague watershed, declines were most common and interior loss along the back-barrier of Assateague Island National Seashore was apparent ( Figure 3). Increases in aboveground green biomass were most prominent in the prograding areas to the south of Assateague Island (Figure 3c) and on the overwash fans on northern Assateague Island (Figure 3b). In general, Chincoteague, Eastern Lower Delmarva, and Southern Long Island all had moderate declines in biomass (Table 1) (Figure 4). The trend maps reveal clustering of loss around landscape features such as ditches, inlets, and rivers even in stable watersheds ( Figure 6). Moran's I for each of the watershed confirmed clustering of salt marsh change (Table 2).
Kruskal-Wallis test was used to test the difference between dominant salt marsh types with each analysis finding significant differences (Table 3). Dunn's post hoc test determined that Chincoteague watersheds had no statistically significant difference between regularly and irregularly flooded salt marsh (Table 3).  not represent a permanent change (Figure 9). Spearman's rank correlation showed that in non-disturbed pixels average summer aboveground green biomass in 2012 was correlated with the summer 2018 average biomass (rτ=0.74, p < 0.001). Disturbance pixels had a smaller correlation with 2018 average biomass (rτ=0.54, p < 0.001). In the long-term change maps areas and types of change are identifiable for example interior loss ( Figure 10).

Verification
The

Discussion
Aboveground biomass declined throughout three of study watersheds. These for Chincoteague, suggests a relationship between these losses with SLR (Table 3).
Tidal loss corresponded with high magnitude disturbances, but were much less common ( Figure 11). Small declines (<100 g m -2 ) across the salt marsh were of little concern as they fall well within the uncertainty of this data. These areas are likely stable, however, if a dramatic increase in SLR or other stressors occur this could change, and all locations need monitoring. Due to the medium spatial resolution, used in this study, the cause of these minor changes is difficult to determine. Small declines

Tidal filtering
The use of all available data is vital for understanding seasonal and long-term vegetation trends . Keeping all quality data is especially important with Landsat time series given the limited temporal phases due to clouds, tides, 16-day revisit, and Landsat 7's shutter synchronization anomalies. The TMII filter is unique to the vegetation cover of a particular pixel. Therefore, it did not over filter those areas with frequent inundation. Adapting the index to Landsat posed several challenges, including different bandwidths and lower temporal resolution.
These issues were addressed with the conversion of rolling to monthly averages and substitution of bands with appropriate equivalents. The index could be further improved by considering a subset of a date's month for years directly preceding and following it. The filtering improved time series trend estimates ( Figure 5). The rarity of false positives limited any reduction of quality data while removing many suspect images. In this study, the amount of data was essential to ensure enough images were available to filter by tides, cloud cover, and data quality. Tidal filtering is necessary to improve time series modeling of salt marsh and in turn our understanding of long-term salt marsh change.

Salt marsh change
Persistence versus die-off of salt marshes has been attributed to a variety of

Disturbance
The BFAST algorithm detected many disturbances. However, a large number of these disturbances were brief which is to be expected in salt marsh environments i.e. high inundation event or algal deposition on mudflats. Positive disturbances were common. However, these did not correspond with long-term increases (Figure 9). Both the long-term trend analysis and disturbance analysis identified areas of loss ( Figure   11). The disturbance pixels had less correlation with 2018 aboveground biomass than 2012 biomass in non-disturbed pixels. This correlation suggests that disturbed areas were less stable areas of the salt marsh. These disturbances illustrate the highly dynamic nature of these systems and the importance of monitoring salt marshes with time series data. Disturbances with an increase in aboveground green biomass could correspond with increased vegetation, changes to vegetation composition, algal deposition on mudflats, or algal blooms in pools. Temporary decreases could correspond with droughts, which have been observed as a driver of temporary salt marsh die-off in the southern United States (Alber et al. 2008).

Conclusion
This study puts forth an approach for understanding salt marsh change with a combination of medium resolution imagery and time series analysis. Declines in aboveground green biomass across the study area were identified with a mean of -17 g m -2 (Table 1) The limiting factor for the process was exporting data from GEE to be further analyzed. The Landsat archive is the only option for decadal time series of salt marsh environments with medium spatial resolution and an extensive archive. This approach demonstrates a promising method for both historic assessment and continued monitoring. However, higher spatial resolution imagery is necessary to increase the sensitivity of this methodology to fine-scale change. Next steps include applying the method to compare a broader range of sites and mapping areas identified as clusters of change with high spatial resolution imagery. Biomass is an important indicator of salt marsh sustainability, tied to ecogeomorphic feedbacks that contribute to salt marsh resilience. The current analysis demonstrates the use of aboveground biomass estimates as an indicator of salt marsh change at the watershed scale.