Salt marshes of the northeastern United States are dynamic landscapes where tidal flooding regime creates patterns of plant zonation based on differences in elevation, salinity, and local hydrology. These patterns of zonation can change quickly due to both natural and anthropogenic stressors, making them vulnerable to degradation and loss. We compared several remote sensing techniques to develop a remote sensing tool that accurately maps high- and low-marsh zonation to use in management and conservation planning for this ecosystem in the northeast USA. We found that random forests (RF) outperformed other classifier tools when applied to the most recent National Agricultural Imagery Program (NAIP) imagery, NAIP derivatives, and elevation data between coastal Maine and Virginia, USA. We then used the RF method to classify a 500-m buffer around coastal marsh delineated in the National Wetland Inventory. We found classification accuracies between 1.4% – 16.3% for high marsh and 13.7% – 29.6% for low marsh zones, varying by sub-region. The detailed output is a 3-m resolution continuous map of tidal marsh vegetation communities and cover classes that can be used in habitat modeling of marsh-obligate species or to monitor changes in marsh plant communities over time.
Coastal marshes are among the world’s most productive ecosystems and provide significant services to humans across the globe. These marshes serve as a gateway between land and sea for humans and wildlife alike, act as a buffer against coastal storms, and provide critical nutrients to marine food webs (Barbier et al., 2011). Tidal marshes also support and protect biodiversity by providing habitat to marine and estuarine fish, crustacean populations, and migratory birds (Boesch and Turner 1984; Master 1992; Brown et al. 2002).
Within the world’s tidal marsh systems, those located along the Atlantic coast of the United States support the highest number of habitat specialists described worldwide (Greenberg, Maldonado, Droege, & McDonald, 2006). This suite of species includes fish and mammals, but the bulk of vertebrate species specialized to northeastern tidal marshes are birds. Several species are limited completely to these marshes during the breeding season, several of which are in decline (Correll et al. 2017), with extinction predicted for the saltmarsh sparrow within 50 years (Correll et al. 2017a; Correll et al. 2017b; Field et al. 2017).
These declining species nest predominantly within the high-marsh zone of coastal marshes. High marsh differs from other marsh areas in elevation, salinity, and frequency of inundation (Bertness & Ellison, 1987; Ewanchuk & Bertness, 2004; Pennings & Callaway, 1992) and is characterized by flooding during spring tides linked to the lunar cycle. The plant species S. patens, short-form S. alterniflora, Distichlis spicata, and Juncus gerardii characterize high-marsh zones, which also include Salicornia spp., Glaux maritima, and Solidago sempervirens (Nixon and Oviatt 1973; Bertness 1991; Emery et al. 2001, Ewanchuk and Bertness 2004). Conversely, low marsh is characterized by daily tidal flooding and is a near monoculture of tall form S. alternilfora. The surrounding terrestrial border experiences infrequent inundation by salt water during extreme tides and storms, and is characterized by a more diverse flora that is often dominated by Iva frutescens and Typha spp. (Miller and Egler 1950; Ewanchuk and Bertness 2004). Introduced Phragmites australis also occurs within tidal marshes, especially around the borders of disturbed marshes (Chambers et al. 1999; Philipp and Field 2005).
These plant community zones within tidal marsh ecosystems can be quickly altered by both natural and anthropogenic stressors such as sea-level rise (SLR), nutrient run-off from adjacent uplands, and the spread of introduced species (Day et al., 2008). The significant increase in sea level during recent decades poses one of the largest threats to these marsh ecosystems. As sea levels encroach on the marshes’ seaward side and upland marsh migration is limited by human-developed coastal infrastructure (Field et al. 2017b) and upland habitats (Field et al. 2016), a “pinching effect” can occur, resulting in marsh loss. Coastal marshes can grow vertically through accretion (Kirwan et al. 2016), but when the rate of SLR exceeds the rate of accretion, marsh area will decline (Crosby et al. 2016). Invasion of high marsh areas with flood tolerant low marsh species (Donnelly and Bertness 2001; Field et al. 2016) can also cause a transition from high into low marsh (Kirwan et al. 2016), however this pattern is not ubiquitous to all marshes (Kirwan & Guntenspergen, 2010; Wilson et al., 2014). In addition to sea level rise, extreme storm events that flood the coastline have been shown to permanently alter marsh structure within days (Day et al., 2008) and can have a lasting effect on plant community structure and saltmarsh degradation.
Marsh degradation and rapid change is likely to continue into the future due to the paired effects of climate change and human development. Sea levels are expected to rise substantially between 2013 and 2100 (IPCC, 2013), and more frequent storm events affecting coastal regions are also predicted. The future distribution of high- and low-marsh habitat therefore remains uncertain (Chu-Agor, Muñoz-Carpena, Kiker, Emanuelsson, & Linkov, 2011; Kirwan et al., 2016). It is therefore essential to develop tools to identify coastal marsh plant communities, particularly high marsh, on a biologically relevant timescale to protect existing ecosystem services and to inform the adaptive management of coastal wetlands as habitat for high-marsh specialist species.
Remote sensing allows for a repeatable analysis of marsh plant communities if the image resolution captures the spatial vegetation pattern and the spectral resolution allows differentiating between the vegetation communities. Several studies have previously demonstrated distinct spectral differences between tidal marsh species using hyperspectral imagery (Belluco et al., 2006; Rosso, Ustin, & Hastings, 2005; Yang, 2009). Hyperspectral imagery is costly, especially across large landscapes and Belluco et al. (2006) compared several aerial and satellite sensors with changing spatial and spectral resolution and although the hyperspectral imagery performed slightly better than the multispectral, spatial resolution was the most important. Belluco et al. (2006) recommend the use of multispectral satellite imagery for the mapping of marsh vegetation. Beside the visible spectrum (RGB), multispectral imagery should include also infrared (IR) to allow for discerning between vegetation types, calculation of vegetation indices (like NDVI) and detecting differences in soil moisture (Jin & Sader, 2005; Pettorelli et al., 2005), particularly in tidal wetlands (Klemas, 2011), and has previously been used as a tool to predict tidal marsh communities both in smaller regions within the northeastern United States (Gilmore et al., 2008; Hoover, Civco, & Whelchel, 2010; Meiman, Civco, Holsinger, & Elphick, 2012) and elsewhere (Isacch et al., 2006; Liu et al., 2010). National Aeronautic and Space Administration (NASA) Landsat satellite imagery is free multispectral imagery commonly useful for classifying coarse cover types, although heterogeneity in tidal marsh vegetation often occurs at scales smaller than 30 m pixels, and classifications of tidal marsh plant communities using this imagery across large landscapes have so far been unfruitful (e.g. Correll 2015). There is thus a clear need for an alternative path to create a regional classification of tidal marsh vegetation.
Recent advances in high-resolution airborne imagery provide new opportunities to develop large-scale classifications of coastal plant communities in the northeast. The National Agricultural Imagery Program (NAIP) from the United States Department of Agriculture (USDA) captures 3-band, high-resolution orthophotos (Red, Green, and Blue, or RGB) during the continental-growing season. Since 2007 most states have added a Near-InfraRed (NIR) band to the image requirements, important for an accurate classification of vegetative cover. The image resolution is 1 m with 6-m horizontal accuracy and a maximum of 10% cloud cover. The imagery, freely available for governmental agencies and public, is an affordable alternative to commercial aerial and high-resolution multispectral satellite imagery. Recent applications of NAIP imagery include mapping of tree cover (Davies et al., 2010), forest clearings (Baker, Warner, Conley, & McNeil, 2013), isolated trees (Meneguzzo, Liknes, & Nelson, 2013), land cover classification (Baker et al., 2013) and mining activity (Maxwell, Strager, Warner, Zégre, & Yuill, 2014).
In this study we compare several remote sensing techniques applied to NAIP imagery, available elevation data from the National Elevation Dataset (NED, https://nationalmap.gov/elevation.html), and local tidal information records from the National Oceanic and Atmospheric Administration (NOAA, https://tidesandcurrents.noaa.gov) to develop an affordable tool capable of repeated classification of high-marsh zones in tidal marshes in the northeastern United States. We then use the best-performing classifier to categorize marsh vegetation communities with a 3-m resolution from coastal Maine to Virginia, USA.
Methods
Study site and community types
Our marsh-mapping effort encompasses all salt marshes of the Northeast Atlantic coast of the USA, from northern Maine to Virginia. To define our classification extent we applied a 500 m buffer to all coastal, tidal marsh as delineated by the National Wetland Inventory (NWI, https://www.fws.gov/wetlands/index.html) estuarine emergent wetland (E2EM) layer. The study site is further split into 8 subzones (Figure 1), to accommodate data management and processing.
These coastal marshes vary substantially from north to south. Due to local bathymetric structure, tidal amplitudes in the Gulf of Maine are among the highest in the world (Garrett, 1972), while those farther south experience much less variation between high and low tides. Similarly, a preponderance of rocky or highly sloped shorelines in the north limits marshes to small (~10 – 100 ha) patches, while southern marshes form larger patches of marsh along the coast. Across our study area, however, tidal marsh ecosystems can be reliably separated into eight distinct cover types, in addition to two bordering cover types not directly included in tidal marshes. We classified the following cover types in our marsh mapping effort:
1. High marsh: Area flooded during spring tides related to the lunar cycle and dominated by Spartina patens, Distichlis spicata, Juncus gerardii, and short form S. alterniflora. In addition, Juncus roemerianus, Scirpus pungens, Scirpus robustus, Limonium nashii, Aster tenuifolius, and Triglochin maritima are secondary cover.
2. Low marsh: Area flooded regularly by daily tides and dominated by tall form S. alterniflora.
3. Salt pools/pannes: Depressed, bare areas with sparse vegetation cover and extreme high soil salinities. Generally, pools retain water between high tides while pannes do not.
4. Terrestrial border: Area infrequently flooded by storm and spring tides and can include areas of marsh with fresh/brackish water due to a high water table and/or runoff from impervious surfaces. Typical plant species include T. angustifolia, I. frutescens, Baccharis halimifolia, S. sempervirens, virgatum, S. robustus, and S. pectinata.
5. Phragmites australis: A species of considerable management interest due to the invasive nature of an introduced form (Saltonstall 2002), especially in marshes with freshwater input, upland development, and/or increased nutrients (Dreyer & Niering, 1995, Bertness et al. 2002, Silliman and Bertness 2004).
6. Mudflat: Exposed muddy areas free of vegetation.
7. Open water (bordering cover type): Channels and bays leading to open ocean included within the 500 m buffer.
8. Upland (bordering cover type): All non-marsh terrestrial cover included within the 500 m buffer.
Data sources
We collected training data for marsh vegetation classes both in the field and remotely using aerial imagery, depending on the cover type. We collected training polygons for high marsh, low marsh, and P. australis using a GEO 7X Trimble GPS between May and August of 2015 and 2016. All technicians collecting training polygon data were collectively trained at the beginning of the season in salt marsh vegetation identification.
We used a generalized random tessellation stratified (GRTS) sampling framework outlined in Wiest et al. (2016) to select delineation sites for the collection of training data across our entire study area. Once technicians navigated to a selected survey point, we located a contiguous patch of high marsh, low marsh, mixed marsh, and P. australis larger than 10 X 10 m in area. We then placed a stake flag or other highly visible marker on the ground to indicate the beginning of polygon delineation. We used a Trimble GEO7X unit to delineate the outer boundary of contiguous patches. Delineations excluded any large pools, pannes, or channels. We collected training data for open water, pools and pannes, and mudflat cover classes using manual digitization of 2014-2015 1 m NAIP imagery using ArcGIS 10.3 since these cover classes are easily identifiable in visible wavelength imagery.
We used the most recent digital ortho-photography (RGB and NIR) available from the NAIP collected during the growing season from 2014 or 2015 as imagery predictor data. We resampled raw 1 m NAIP imagery to 3 m to match the NED, which was used as the digital elevation model (DEM) for this analysis. All NAIP imagery derivatives were calculated in ArcMap 10.3 using the raw band values. We refer to them as ‘pseudo’-vegetation indices because we use the raw band values instead of reflectance values. Using these data we also calculated six derived variables often used for classifying vegetative communities: the Normalized Difference Vegetation Index (Rouse, Haas, & Schell, 1974), the Normalized Difference Water Index (McFeeters, 1996), the Difference Vegetation Index (Richardson & Wiegand, 1977), and the first three principle components from a principal components analysis (Fung & Ledrew, 1987) of the four NAIP bands, which collectively explained > 95% of the variance.
Elevation is indicative of the tidal flooding frequency and thus influences plant species zonation (Silvestri, Defina, & Marani, 2005). Consequently, different marsh habitats are influenced by their elevation. We used the NED for all elevation predictor data. The NED is derived from different contributed datasets and then processed by the USGS into a near-continuous DEM at various resolutions across the US. We used 1/9-second (roughly 3 m resolution) data when available for our classification. When no 1/9 arc-second imagery was available, we used 1/3 arc-second data (~10 m resolution). To account for the large differences in tidal inundation across our study area, tidal data for the study area was collected from the closest NOAA tidal gauge station creating 29 different tidal zones (https://tidesandcurrents.noaa.gov/stations.html?type=Water+Levels, appendix 1). For each of the stations we collected the following tidal datums: HAT, MHHW, MHW, MSL, NAVD88 and Max (Table 1). We resampled the NED data to an exact 3 m resolution to match NAIP imagery and clipped the resulting imagery with the 500 m buffer of all coastal tidal marsh in the NWI. We further clipped the NED by the 29 tidal gauge zones of the study area, and rescaled each zone to the NAVD88 datum using the NOAA tidal amplitude data. To calibrate the DEM across the entire study site we used the Mean High Tide (MHT) divided by Mean Highest High Tide (MHHT) value for each tidal gauge zone.
We classified the NED of each tidal gauge zone based on the tidal amplitude data for that zone. To do this, we rescaled NOAA tidal amplitude data to the NAVD88 datum used in the elevation dataset, and defined elevation limits for the water, high- and low marsh, terrestrial border, and upland class based on flooding history (Figure 2). We then used these thresholds in conjunction with NAIP imagery reflectance values and derivatives to identify water, high marsh, low marsh, and P. australis cover types. We used only elevation thresholds to define the terrestrial border and upland cover types. In rare cases in small areas along the coast where there was no NED layer available, we classified the marsh communities without the DEM data input. In these cases there are no terrestrial border or upland defined.
Data analysis
We compared three classification methods to delineate tidal marsh cover types. We conducted the comparison of classifiers in the center of our study area (Delaware Bay, subregion 6, Figure 1) to maximize utility both to the north and south. First, we used classification and regression trees (CART), which are a flexible, rule-based classifier that is computationally fast and makes no statistical assumptions regarding the distribution of the data (Otukei & Blaschke, 2010). CART methods are particularly useful when integrating environmental variables with different measurement scales and are robust for large datasets. Post-hoc pruning removes nodes that do not provide enough accuracy reducing overfitting. We used the R package rpart (Therneau, Atkinson, Ripley, & Ripley, 2015), which implements the CART methods described by Breiman et al. (1984). Second, we used random forests, which are decision tree ensembles that improve the accuracy and stability of a single decision tree (Breiman, 2001). RFs perform well with small training sample sets, offer a cross-validation-like accuracy measure through the out-of-bag (OOB) error estimates, and assign variables’ importance by assessing accuracy loss when feature values are randomly permuted (Breiman, 2001). We used the R package RandomForest (Liaw & Wiener, 2002) for all RF analyses. Finally, we used support vector machines (SVM), which are non-parametric classifiers that use risk minimization to calculate a hyperplane that separates two classes with a maximum margin defined by the ‘support vectors’ – that is, the points that lie closest to the splitting hyperplane. Success depends on how well the process is trained and if the classes are linearly separable, but an SVM-kernel could be applied if the classes are not linearly separable. In general, SVM offers high training performance versus low generalizing errors, but is sensitive to over-fitting, especially with noisy and unbalanced data. We used the svm function in the R package e1071 (Dimitriadou, Hornik, Leisch, & Meyer, 2006) with the default values for C and Y. To select the best remote sensing classifier for the final marsh layer, classification and validation is based on a randomly selected independent training and validation subset (66%-33%) of the first year of polygon training data from zone 6.
We applied the best performing classifier, RF, to all biogeographic zones from Maine to Virginia. For the final classification, we used the out-of-bag (OOB) error estimates produced by the RF algorithm to measure accuracy of our classification by zone and across all zones (Table 4). We then clipped the resulting marsh classification by the DEM-based cover types (upland, terrestrial border), if the NED layer was available.
Due to variable image quality and/or due to high tide during image acquisition, we sometimes encountered artefacts in the imagery that affected the accuracy of our classification, particularly in Zones 1 and 6. Sun glitter or high mud content in open water sometimes caused misclassification of this cover type as low marsh. We used the RF probability scores for the open water class to better represent the actual water cover and then updated the open water classification in Zones 1 and 6 to improve overall accuracy.
All methods and datasets involved in our study are freely available to the public, and our analyses are limited to tools available through ArcGIS, a commonly-used GIS in federal, state, and private conservation organizations, or simple Program R code, which is freely available to the public. Our manuscript provides code in Program R to complete analyses described which are not available through the ArcGIS interface.
Results
We found that RF methods outperformed the other two classifier tools in classification of tidal marsh plant communities (Table 3). Classification of high and low marsh cover types, returned accuracy rates ranging between 73% and 88% across all classifiers. For all classifiers, error was high for the P. australis cover type, ranging from 45-68%.
In our final data layer, we used the random forest method to classify 16,014 km2 of tidal marsh and bordering communities within the 500-m buffer at a 3-m resolution (Fig. 3). Almost 4000km2 of the data layer is covered by tidal marsh: high marsh (36%), low marsh (21%), mudflats (7%), phragmites (7%), pools and pannes (5%), and terrestrial border (24%) – (Table 5). The amount of high and low marsh cover varies from north to south with an overall increase in low marsh area. Similarly, the percentage cover for Phragmites increases when moving southwards. Differently, the percentage mudflats is increasing northwards. Mean classification accuracies range from 99.9% for open water to 25.1% for P. australis (Table 4). Mean classification accuracies varied among cover types. Open water, mudflat, and pools/pannes were classified with high accuracies across all regions (>95% accuracy in 23 of 24 cases: Table 4). Classification of the three vegetation types was less accurate and varied among cover classes. High marsh was classified with a mean accuracy of 94.2%, but with clear regional variation. In the regions from New Jersey north and in the inner portions of Chespeake Bay, accuracy was generally greater than 95%. In contrast, accuracy from Delaware Bay to Virginia was lower with only 83-87% of high marsh correctly classified. Classification of low marsh was generally less accurate than high marsh, with an overall accuracy of 77.9%. Again, there was regional variation, but unlike high marsh, accuracy tended to be higher in the more southern regions, especially the inner portions of Delaware and Chesapeake Bays, where accuracies were >80%. Classification of P. australis showed the greatest variation across regions. Overall, accuracy was ~75%, but in most regions >80% of sites were correctly classified >80% of the time, and in coastal Massachusetts accuracy approached 95%. By contrast, accuracy in coastal New Jersey was only 67%, and in our northernmost region (coastal New Hampshire and Maine) accuracy was as low as 20%.
Discussion
Tidal marshes of the northeastern U.S. are critical pieces of the coastal landscape, providing key habitat and ecosystem services to humans and wildlife. We developed a repeatable method to classify plant communities within this important ecosystem from Maine to Virginia; the resulting data layer is the first of its kind to classify this marsh ecosystem at such a high spatial resolution (3 m) regionally.
The high classification accuracies in the high marsh zone make this data layer particularly helpful for use by marsh managers, researchers, and planners. Tracking marsh vegetation change can be used to help better understand how SLR is changing marsh conditions, and explore use of this vegetative change as an early warning of marsh loss. Quantifying loss of marsh can also help predict changes in coastal ocean factors that can affect the flooding potential of coastal properties. Identification of marshes where rate of change is low could also help identify the best areas of marsh to protect into the future, for conservation planning for marsh-obligate species that use the high marsh zone as their primary breeding or foraging habitat. As sea levels continue to rise and human development continues to alter marsh hydrology and accretion (Day et al., 2008), a method to repeatedly classify coastal marsh cover types will be integral to measuring the amount and distribution of available habitat and change in habitat over time. The small spatial resolution and high horizontal accuracy of this data layer also allows it to be used across varying spatial extents, from local municipality borders to multi-state regions.
The NAIP dataset used in this classification offers a high-resolution, low-cost set of multispectral imagery with a refresh rate of 3 years. This dataset, however, has limitations when used over large areas found by our study as well as others (Meneguzzo et al., 2013), and potential users should carefully consider the pros and cons of this dataset before setting out on a similar classification effort. Variation in the time and day of image acquisition across the NAIP dataset results in different tidal stages, plant phenology, and illumination across images. The NAIP post-hoc color balancing applied to these images by each contractor (7 contractors across our study area) does not correct for differences in illumination and atmosphere in a standardized way. This results in a radiometric imbalance across the spatial extent of the dataset, limiting the utility of the NAIP dataset in large-scale classification efforts. Further, the temporal resolution of the NAIP dataset (one set of images per year, flown during “greenup” between June and August) limits analyses of NAIP data specific to this time period/season. Since the timing of high and low tide changes daily, this temporal resolution also results in imagery taken at different parts of the daily tidal cycle, and low marsh or mudflat areas were likely partially flooded when most imagery was collected.
The RF methods used in our final classification work well to incorporate different data sources with low threat of over-fitting. Classification success presented in table 3 shows an average high accuracy; accuracies for the high marsh cover type, for example, range from 1.2% to 16.3%, depending on biogeographic zone. These OOB error estimates are a consistent measure of RF accuracy, however OOB error estimates do not systematically evaluate classification accuracy outside the training polygons (e.g. Fry et al. 2011). Additional work to collect independent data and compare it to this layer’s predictions would be a valuable next step would make clear where remote sensing research for this ecosystem type should proceed next. Particularly, the OOB differences among zones suggest additional tests should be geographically broad to further assess on-the ground accuracy across the entire region.
Any generalized messages we can draw that go beyond just our data set???
Although the primary focus of our study was to distinguish between high and low marsh, the importance of invasive P. australis to many management decisions led us to also consider this cover type. Unfortunately, Phragmites proved particularly difficult to classify (Table 4). The smaller amount of training data available for this cover type (n = 148 polygons, table 3) compared to the other cover classes likely contributed to this error. Secondly, the classification algorithm for P. australis likely confounded P. australis with terrestrial border species not included in the training data, particularly with stands of Typha spp, which is similar in structure to P. australis. This combined with the inconsistent ground elevation at which P. australis can be found likely further confounded the classification. After our classification effort, a strong need remains for development of a method for large-scale classification of this invasive species for use in monitoring and management along the coast. Further study is necessary across latitudes, flooding regimes, and imagery sources to develop an accurate classification algorithm for P. australis across this region to improve on the methods and results presented in this paper.
Conclusion
Repeated, large-scale classification of coastal vegetation communities is urgently needed to help inform conservation of this rapidly shrinking ecosystem. We present a classification at a 3-m resolution of distinct cover types within tidal marshes to serve 1) as a vegetation community delineation for use in management and conservation decision-making, 2) as a layer for local and regional analyses of this biological community, and 3) a base layer for use in comparisons into the future to measure community change over time. These actions will all be integral for the long-term preservation of tidal marshes and the species they support as climate change and further human influence continue to affect this ecosystem.