*2.2. Data*

Satellite imageries were collected on 4/15/2015 and 5/25/2015 with Worldview-3 and Worldview-2 sensors, respectively. Four-band National Agriculture Imagery Program (NAIP) data acquired between 2011 and 2017 and the true color National Aerial Photography Program (NAPP) data acquired in 1994 were used to classify the salt marshes (Table 1). The aerial imageries were collected at a range of tidal stages (Table 1). Change analysis was also conducted with a 1997 classification of FIIS performed with true color aerial photos for the island and verified with in situ field assessment with a highest achieved overall accuracy of 87.5% [45,46].


**Table 1.** The description of data used by acquisition date, spectral resolution (band wavelength when available), sensor or program, and spatial resolution. R = red, G = green, B = blue, NIR = near infrared, CB = coastal blue, Y = yellow, RE = Red edge.

\* Average tidal stage across the acquisition as derived from USGS tidal gauge 01309225 [47].

#### *2.3. Tidal Stage E*ff*ects*

The effects of the tidal stage at the time of imagery acquisition on mapped salt marsh extent has long been recognized e.g., [48,49]. In this study, the tidal stage could impact the edge erosion calculations and the panne analysis. Therefore, the highest tidal stage out of all the imageries was analyzed using topobathymetric LiDAR, bathtub models, and the 2015 FIIS classification [50]. The method utilized the topobathymetric LIDAR-derived Digital Elevation Model (DEM) to create a bathtub model, i.e., a binary raster of inundated and non-inundated pixels, at the target tidal stage. The highest tidal stage of our images was 35.66 cm MLLW or 14.3 cm above the North American Vertical Datum 1988 (NAVD 88) occurring in 2017. The bathtub model was then used to determine areas mapped as vegetated in 2015 that were likely inundated at the 2017 image's tidal stage. The method has been applied in Jamaica Bay, New York, and was utilized in this study to understand the potential impact tidal stage had on the aerial image classifications. These data provide an understanding of the uncertainty derived from inundation that could occur in this analysis.

#### *2.4. Object-Based Image Analysis*

OBIA begins with an unsupervised classification or segmentation dividing the image into areas with similar spectral characteristics and spatial proximity [51]. This method used mean shift segmentation: a hierarchical segmentation with demonstrated success in remote sensing and other disciplines [52,53]. OBIA allows for the combination of spectral, spatial, and ancillary data and has been shown to increase classification accuracy compared to pixel methodologies when using VHR satellite imagery [54–56]. In this study, a multiscale segmentation approach was used, selecting under segmented areas and resegmenting them at a finer segmentation scale [57,58] (Figure 2). This study's final segmentation was dual scale with 80% of objects segmented at a spectral radius of 13 and minimum size of five pixels. The other 20% were segmented at a spectral radius of 8 and minimum size of five pixels. Data processing and segmentation were conducted with Python 2.7 [59] and Orfeo Toolbox 5.2 [60].

**Figure 2.** The data processing and classification workflow for classification of the Worldview-2 and Worldview-3 imagery.

The classification was composed of 10 categories including *S. alterniflora*, patchy *S. alterniflora*, high marsh, upland, dune vegetation, sand, mudflat, water, *Phragmites*, and wrack. A one-thousand-nine-hundred-and-thirteen 1-m<sup>2</sup> vegetation plot data were adapted to create training data. The training data were composed of plots with a Braun–Blanquet percent cover greater than ≥50%. Objects that intersected training points were selected, resulting in a total of 1964 training samples. The species included in the high marsh category were *S. patens*, *D. spicata*, *I. frutescens*, and *J. gerardii*. Percent cover differentiated the two *S. alterniflora* classes with the patchy *S. alterniflora* class being between 49 and 10% cover, and the *S. alterniflora* class being ≥ 50% cover. The vegetation plots were predominantly within the salt marsh environment leading to water, sand and upland classes being trained from samples gathered by visual interpretation and field knowledge. The Random Forest (RF) classifier was used to classify the 2015 Worldview-2/Worldview-3 image data for vegetation mapping and the panne and edge classifications.

Accuracy assessment was conducted for the 2015 Worldview-2/Worldview-3 image classification using a subset of the total training data. The data were randomly split 60% training and 40% for testing. The training data were used to train the RF model. An error matrix including kappa, producers, users, and overall accuracies was computed using the testing data.

The 2015 image classification was compared with a 1997 classification based on aerial imagery [45]. The 1997 classes of Reed grass marsh, high marsh, low marsh, and mosquito ditches were compared with the 2015 classes of Phragmites, high marsh, and *S. alterniflora* classes accordingly. Mosquito ditches in 1997 were included as vegetated area due to their small average width reported from 25.4 to 50.8 cm [61]. This width is below the minimum mapping unit of 0.25 ha for the 1997 classification [45]

and the three-pixel width of the VHR classification. Change rate was calculated between 2015 and the 1997 salt marsh.

#### *2.5. Change Analysis (1994–2017)*

The panne and pool analysis, subsequently referred to as panne analysis, was conducted on imageries from 1994 to 2017. The edge erosion change analysis was conducted for 1994 to 2011 and 2011 to 2017. Each image collection was segmented at a spectral radius of 10 and shape radius of 5 with mean shift segmentation. This segmentation scale was adequate given the lower spectral resolution of the aerial images. ArcMap 10.5 [62] was used to select segments which intersected the interior mud or water areas of 2015 Worldview-2/Worldview-3 classifications, these segments comprised the 2015 pannes. The segments were then merged, creating a multitemporal segmentation. The classification parameters included mean, median, standard deviation, simple indices (i.e., red band/blue band), normalized difference vegetation index (NDVI) for those years with NIR, and the difference of each band for the year of interest and subsequent year. The panne and edge classifications were trained with objects that were either vegetated or non-vegetated. Their accuracies were verified with 522 randomly selected points, which were assessed as vegetated or non-vegetated for each time period.
