3.1. Barter Island

Barter Island (fig. 2) is on northern Alaska’s Beaufort Sea coast and the boundary of the U.S. Arctic National Wildlife Refuge; it contains a highly dynamic barrier system and permafrost bluffs fronting the Alaska Native community of Kaktovik and Clear Space Force Station. The sand and gravel barrier islands trend roughly east-west in the region and can be overwashed or breached during high wave or water level events. The central 3 kilometers (km) of Barter Island’s northern coast is comprised of eroding coastal permafrost bluffs, mostly fronted by narrow (less than 20 m) sand and gravel beaches.

Most morphological change along this microtidal coast occurs during the ice-free months, typically July to October. The surface permafrost also thaws during these months. During this open-water season, atmospheric low-pressure systems can bring high winds, storm surges, and waves (Erikson and others, 2020). The region is also predominantly cloudy, which limits usable satellite imagery, particularly during these storm events. Because wave energy is limited to the short open-water season and is relatively lower than lower-latitude coasts, the magnitude of littoral transport is thought to be low (Hopkins and Hartz, 1978). The area has a mean diurnal tidal range of 18 centimeters (cm; National Oceanic and Atmospheric Administration, 2025), though local winds and storm systems can change water levels by several meters.

Barter Island was selected for this study because it is one of the few locations in the Arctic where a relative abundance of historical imagery exists, including very-high resolution (less than 5 m ground sample distance) satellite and aerial photography on a near annual basis since 2000. Historical shoreline and bluff positions were delineated manually (Gibbs and others, 2021) from a combination of historical maps, aerial photography, declassified satellite photography, and very-high resolution satellite imagery for a dataset of morphological evolution of Barter Island and its neighboring barrier islands (Gibbs and others, 2019a, 202022). For this study, 19 in situ shorelines were available for comparison with the SDS.

3.2. Cape Cod

Cape Cod (fig. 3), located in southeast Massachusetts, has a relatively recent geologic history; the retreat of the Laurentide ice sheet and subsequent rise in sea level began roughly 18,000 years before present (U.S. Geological Survey, 1976). The open-ocean Atlantic coast of Cape Cod (the “outer Cape”) includes Cape Cod National Seashore and Monomoy National Wildlife Refuge, where a variety of ecosystems can be found (Schwarzman, 2002).

The approximately 65 km of shoreline in Cape Cod National Seashore is composed of dune-backed beaches in the north (where the dunes can be more than 25 m tall), high bluffs (as high as 30 m) fronted by narrow beaches along the central section, and barrier islands and spits in the south. There are no rivers supplying sediment to the beaches of the outer Cape; all beach sediment is derived from the eroding bluffs. There are also no coastal engineering structures along the shores of the National Seashore, making it an entirely natural littoral system. Grain size ranges from medium sand in the south to very coarse sand at the northern spit (Fisher and Leatherman, 1987). Longshore sediment transport is bi-directional, and a nodal point is near the Wellfleet-Eastham border, south of the study area (Berman, 2011; Giese and others, 2011). North of this nodal point, longshore sediment transport flows from south to north. South of the point, it flows in from north to south.

Outer Cape Cod (the outer Cape) has a semidiurnal tidal regime with a range of 2 to 3 m and a dominant east-northeast wave direction. The outer Cape is a storm-dominated coast affected by extratropical cyclones (nor’easters) and, less frequently, hurricanes. Long-term (about 100 years) shoreline-change rates, which were derived from historical maps and imagery and modern lidar data, show an overall erosional trend for most of the outer Cape, whereas short-term (about 30 years) trends show more variable patterns (Hapke and others, 2011).

Cape Cod National Seashore was selected for this study because of the large in situ dataset available for comparison with the SDS. Between 1998 and 2005, an intensive USGS field program was conducted along 45 km of the outer Cape, from Nauset Inlet in the south to Race Point in the north (Farris and others, 2020). A ground-based all-terrain vehicle system called SWASH (Surveying Wide-Area Shorelines) was equipped with GPS sensors and was used to traverse a single, shore-parallel line along the foreshore to collect horizontal and vertical position and beach slope. The shoreline position was determined by extrapolating from the driven track to the mean high water (MHW) elevation. Over the 7 years the field program was active, 111 shoreline surveys were conducted. Most of these were biweekly, but several pre-storm, post-storm, and recovery triplet surveys were also performed. A 9.5 km alongshore subset of this dataset is used for comparison with the SDS.

3.3. Elwha River Delta

The Elwha River delta (fig. 4) and its corresponding littoral system lie within the Strait of Juan de Fuca in the northern part of the State of Washington. The littoral cell is roughly 20 km long and ends at a terminal spit. The Elwha River encompasses an 831-square-kilometer (km²) watershed that drains the interior of the Olympic Mountains (Warrick and others, 2009). Its beaches are characterized by a steep, reflective foreshore backed by a berm with a foreshore slope between 0.1 and 0.2. East of the river mouth, the Elwha River delta beach includes a flat low-tide terrace that generally has a grain size of gravel to cobble with slopes of about 0.02. Sediment in the region is categorized as mixed sand and gravel with a mean grain size of roughly 40 millimeters. Significant wave heights within the region are usually less than 1 m, and tides are diurnal with a range of 2.22 m between the mean lower low water and mean higher high water (Warrick and others, 2019). Surveys in the area between 2004 and 2007 have shown that most of the erosion occurs in the winter (November through March) with lower rates of change in the summer (June through August) (Warrick and others, 2009).

Beaches near the delta are placed into two morphological units defined by their location east and west of the river mouth. East of the river mouth, beaches are characterized by a gravel low-tide terrace, mixed sand and gravel foreshore, and a low berm. This area has also seen periods of high erosion, especially from 1939 to 2006. Beaches west of the river mouth have been much more stable and have a high berm and no subaerial low-tide terrace like the eastern portion (Warrick and others, 2009; Miller and others, 2011). Because of the relatively steep beaches around the delta, waves break directly onto the beach, which limits the width of the surf zone. Alongshore transport is greater along the middle and eastern sides of the delta because waves approach at an angle of roughly 265 degrees (Miller and others, 2011; Gelfenbaum and others, 2015).

One of the largest dam removal projects occurred on the Elwha River between 2011 and 2014 (Warrick and others, 2009), when the Elwha and Glines Canyon Dams were removed. The dams were built in 1912 and 1927 and were 32 m and 64 m tall, respectively, which disrupted the natural flow of the Elwha River and halted the passage of fish and sediment. More than 90 percent of the watershed was captured by these dams, which altered downstream habitats and ecosystems (Warrick and others, 2009). Within 2 years of the dams’ removal, more than 2.5 million cubic meters of sediment were deposited immediately offshore of the river mouth (Gelfenbaum and others, 2015). During the first 5 years following their removal, nearly 20 million metric tons (Mt) of sediment were eroded into the river and transported downstream, and 90 percent of this sediment (18 Mt) reached the coast (Ritchie and others, 2018). During the same 5-year period, the delta area grew by approximately 60 hectares, shifting the coastal area from a region of erosion to deposition (Ritchie and others, 2018; Warrick and others, 2019).

In situ data were collected along beach transects of the Elwha River delta beach via GNSS surveys between February 3, 2009, and November 27, 2023 (Stevens and others, 2017; Miller, 2025). From these survey data, 249 shorelines were used for comparison with SDS. Details are listed in section 4.0.

3.4. Madeira Beach

Madeira Beach, located on Florida’s west coast on the 22.3 km long Sand Key barrier island in Pinellas County (fig. 5), is comprised of quartz sand beaches backed by low dunes. Structures and seawalls extend seaward of the dune in some areas. Long-term (approximately 150 years) MHW shoreline trends derived from lidar, historical maps, and imagery that incorporate the proxy-datum bias show a slight erosional trend (−0.3 meters per year [m/yr], ±0.65); whereas short-term (approximately 30 years) trends are slightly accretional (0.3 m/yr, ±0.21; Kratzmann, 2022).

Madeira Beach is an example of a highly managed beach. In 1957, the City of Madeira Beach built 37 timber and concrete groins to slow erosion. In 1961, the city built a terminal jetty and placed approximately 30,000 cubic yards of fill north of the jetty. Continued beach nourishment on a 10-year cycle since the late 1980s has further influenced and complicated the interpretation of shoreline trends.

The area experiences semidiurnal microtides with a range of 7 to 80 cm. Coastal morphodynamics are storm-dominated, mainly from tropical cyclones in the summer and fall (June through November), but also winter storms. Otherwise, the shore is dominated by small waves for most of the year, and alongshore transport is predominantly southward.

In situ data were collected for Brown and others (2018) by walking a set of beach profiles equipped with a GPS receiver and a GPS antenna affixed to a surveying backpack between 2016 and 2023. The walking surveys are used primarily to calibrate the USGS CoastCam at Madeira Beach but extend below MHW and can be used as in situ shorelines (Brown and others, 2018). There are 42 shorelines available for comparison with the SDS results from CoastSeg.

3.5. Rincón

Rincón (fig. 6) lies on the northwest coast of the Island of Puerto Rico, and the coast is comprised of reef-fronted stretches of beach backed by unconsolidated sediment and alluvial deposits. Beaches are generally sandy and narrow, showing varying levels of erosion (Thieler and others, 2007; Henderson and others, 2021). Littoral cells in this area are isolated, and there is little transport between them. New sediment primarily comes from nearshore reefs and alluvial deposits. Alongshore currents are generally to the south.

As noted by Thieler and others (1995), beach morphology at this location is influenced by seasonal storms’ increased wave energy and rising sea levels, both of which can drive short-term changes in erosion rates. Construction of a marina facility in 1983 and changes in sand-management practices are thought to have contributed to increases in erosion rates at about the same time. Prevalence of urbanized shoreline fronted by seawalls and revetments has also increased in the region since the 1990s (Thieler and others, 1995).

In situ data were collected from topographic surveys, lidar, and manual extraction from imagery between 1901 and 2023, and compiled MHW and high-water line shorelines were processed and released in cooperation with the Coastal Research and Planning Institute of Puerto Rico (Henderson and others, 2021). Timing and temporal density of the in situ data do not resolve shoreline changes associated with individual wave or storm events (Thieler and others, 2007), and recent in situ data are very sparse. Only two shorelines are used from the survey data as a comparison with the new SDS results from CoastSeg because of insufficient time overlap with the SDS (refer to section 4.0 for more details). However, the Rincón site is included because many coastal study locations and communities across the United States also lack sufficient spatial and temporal data needed to evaluate changes in shoreline positions.

4.1. Barter Island

Custom ROIs were created for Barter Island to better capture the complex morphology and preserve ocean and land proportions in imagery (fig. 2; Vos and others, 2019b); ROIs were sized to be the same default areas created in CoastSeg (20 km²). A custom shoreline based on the 2020 in situ shoreline was created and used as the reference shoreline. DSAS was used to cast custom transects at 50 m intervals, and transects were made to be very long (up to 1,000 m) to accommodate the large changes over the historical comparison period. Because of the challenging environment and complicating factors in imagery (dynamic change, weather, narrow beaches), a maximum buffer distance (1,000 m) and minimum length or area values (50 m, 50 square meters) were used to extract data; more restrictive buffers greatly reduced the ability of the system to identify shorelines, especially in areas with the largest shoreline change such as around bluffs, at the ends of spits, and where barrier islands narrowed. Cloud masking was not used because there was a problem with the masking algorithm conducted on the collected imagery in this location. All available imagery for the typically ice-free months of June through October was downloaded (see section 4.0) and reviewed, and “bad” imagery (degraded by the presence of clouds, ice, or affected by other sensor problems) was removed manually. The CoastSeg:CoastSat shoreline detection process was run, images with bad detections were additionally removed, and the shoreline detection process-step was repeated again. Approximately 1,400 images remained (173 from L5, 460 from L7, 351 from L8, 73 from L9, and 369 from Sentinel-2). In situ data used in this comparison were referenced to MSL, and all SDS data were tidally corrected to MSL (table 2).

Transects 154–206, which cover an extremely dynamic area at the end of Bernard Spit and eastern Barter Island, were removed from quantitative analysis because of excessive noise. This area is addressed separately in section 5.0.

4.2. Cape Cod

ROIs were generated within CoastSeg:CoastSat (at default 20 km²) for all of the Cape Cod National Seashore. Three consecutive ROIs along the central portion of the outer Cape were selected for this experiment, spanning a total of 9.5 km along the coastline (fig. 3). Satellite imagery download and shoreline extraction were performed on each ROI independently. The default CoastSeg:CoastSat reference shoreline was used, and DSAS was used to cast custom transects at 50 m intervals. After downloading all available satellite imagery from May 1984 through July 2023 for all ROIs (2,170 images total) using cloud masking, the images were manually inspected to determine their quality for shoreline extraction. Images of poor quality caused by clouds, sensor problems, and other image issues were removed (moved from usable imagery to “bad” imagery), leaving 1,735 images (80 percent) for shoreline extraction. The shoreline detection process-step was performed, results were visually inspected, and erroneous shoreline detections were removed. A total of 1,589 shorelines remained for analysis (640 from L5, 646 from L7, 248 from L8, and 35 from L9). In situ data used in this comparison were referenced to MHW, and all SDS data were tidally corrected to MHW (table 2). Unique challenges presented by bad shoreline detections are addressed separately in section 5.0.

4.3. Elwha River Delta

An ROI was generated within CoastSeg around the mouth of the Elwha River to capture shoreline evolution, but a custom shoreline was used as the reference to help capture the dynamic changes over the comparison period. Custom transects used in surveys (Miller, 2025) were adjusted for correct orientation. Removal of dams on the Elwha River by 2014 allowed for considerable growth of the river mouth delta; therefore, the reference buffer was increased to be large enough to allow for hundreds of meters of changes to the shoreline driven by the increases in sediment delivery. However, buffers set at maximum settings resulted in erroneously identified small sand bars, so the buffer was set at 320 m to capture significant shoreline changes and avoid erroneous detections. The default CoastSeg:CoastSat sand-color setting (Fitzpatrick and others, 2024a) was used at the Elwha River delta, even though the delta contains grey sand and sediment. The dark sand option in CoastSeg:CoastSat led to much poorer outputs than the default setting. Clouds, haze, and even dark-colored water were incorrectly identified by the classifier as sand with the dark-sand setting, yielding poor shoreline results.

After downloading the imagery, initial detections were reviewed, imagery with bad detections caused by clouds, weather, or other image issues was manually removed, and the shoreline detection process was rerun. Approximately 1,260 images remained (522 from L5, 451 from L7, 232 from L8, and 55 from L9). Given original survey data and complex beach profiles at the site, in situ data used in this comparison were referenced to MHW, and all SDS data were tidally corrected to MHW (table 2).

4.4. Madeira Beach

One ROI of 0.5 km² was generated for Madeira Beach within CoastSeg:CoastSat to cover the 500 m alongshore extent of the in situ shorelines. The default CoastSeg:CoastSat reference shoreline was used, and transects were cast at 50 m intervals to align with the in situ transects. Defaults for reference buffer distance, minimum shoreline length, and minimum shoreline area were used.

After downloading imagery, initial detections were reviewed and imagery with bad detections because of clouds, weather, or other image issues was manually removed, and the shoreline detection process was rerun. Approximately 110 images remained (0 from L5, 38 from L7, 58 from L8, and 13 from L9). In situ data used in this comparison were referenced to MHW, and all SDS data were tidally corrected to MHW (table 2).

4.5. Rincón

An ROI of 20 km² was generated for Rincón within CoastSeg:CoastSat for the area, and the default reference shoreline was used. Transects from in situ shoreline records (Henderson and others, 2021) were used in CoastSeg:CoastSat after adjusting for correct orientation. Defaults for reference buffer distance, minimum shoreline length, and minimum shoreline area were used. Cloud masking was not used because there were complications with masks in this area, and cloud cover was a significant challenge. All imagery for the comparison period was downloaded and reviewed, and bad images were removed manually. Initial shoreline detection images were reviewed, and images with bad detections were additionally removed; the shoreline detection process-step was then repeated. Approximately 150 of the images remained (5 from L5, 17 from L7, 104 from L8, and 19 from L9). In situ data used in this comparison were referenced to MSL, and all SDS data were tidally corrected to MSL (table 2).

There was a clear offset between SDS and in situ observations during the earlier portion of the comparison period (prior to 2012) because in situ shorelines prior to 2012 were derived from wet-dry line in imagery without appropriate proxy-datum bias corrections; however, correcting this offset was beyond the scope of this effort. Additionally, the beaches were narrow during this period and more difficult for manual identification in lower resolution imagery. Therefore, statistical comparisons are limited to the data after 2012, leaving only two in situ shorelines.

4.6. Techniques for Data Analysis

SDS measurements were compared with in situ shoreline data for each location (Barter Island, Cape Cod, the Elwha River delta, Madeira Beach, and Rincón) and for each experiment (no tidal corrections, low slope corrections, average slope corrections, and high slope corrections). For this analysis, all SDS observations within 10 days of an in situ measurement were considered. The resulting timeseries of comparison SDS at all transects was filtered using a Hampel filter to remove outliers. Additionally, points that were above the 75th percentile of daily shoreline change were also removed in Barter Island and the Elwha River delta. Errors between each SDS–in-situ pair across all transects were computed, and then the mean error (ME; eq. 1), mean absolute error (MAE; eq. 2), root mean squared error (RMSE; eq. 3), mean absolute percentage error (MAPE; eq. 4), and root mean squared percentage error (RMSPE; eq. 5) were computed from the distribution of errors using the following:

These results are depicted in figures for each site in section 5.0. Each figure shows the ME, MAE, MAPE, RMSE, and RMSPE, the interquartile range (IQR) of SDS observations, and the number of observation pairs at all transects noted. A variety of metrics are presented to illustrate different properties of the shoreline measurements. MAE and RMSE provide a single number for each experiment, estimating the magnitude of error (or deviation between SDS and more traditional techniques); RMSE is more sensitive to outliers, whereas ME is indicative of bias. MAPE and RMSPE further contextualize the error for each site relative to the amplitude of the signal present in each set of shoreline position timeseries. Like RMSE, RMSPE is more sensitive to outliers. MAPE and RMSPE allow for normalized comparisons between sites that show vastly different magnitudes of shoreline change. The IQR, an estimate of signal variance based on the difference between the 25th and 75th percentiles, contextualizes the MAE and RMSE estimates. For example, if the computed error metric (either RMSE or MAE) is nearly equal to the IQR for a given site, then the signal measured is likely mostly noise. However, if the computed errors have much smaller values compared to the IQR, then there is more evidence that the shoreline measurements have captured a true signal. To support the results shown in those figures, timeseries of SDS for select transects of interest (figs. 2, 3, 4, 5, 6) are also shown for each site, along with the range of SDS values and the SDS mean value from all four experiments; in situ data not used in the comparisons are also plotted for additional context. Additionally, MAE and RMSE were computed for each transect, and the per-transect MAE (from the experiment with the lowest error values) across the study area is illustrated. Linear trends of the comparison SDS over the entire period from the same experiment are also shown.

To illustrate shoreline behavior across each site’s entire observation period, outside of the relatively few comparison data and provide context for trends in the above analyses, all SDS data from the average-slope experiment (not only comparison SDS, but all SDS from the imagery record at each site) were plotted to show monthly change using the techniques outlined in Janda and others (2025). For sites with multiple ROIs (Cape Cod and Barter Island), the ROIs were combined to create a cohesive view of the entire site. Shoreline positions were resampled at a monthly timescale to help reduce noise, and a locally estimated scatterplot smoothing (LOESS) filter was applied to highlight monthly trends (Janda and others, 2025). Unfiltered data are also plotted, because the LOESS filter can smooth out small events. These figures are oriented in the same cardinal geographic directions as the site locations: east-west for Barter Island, the Elwha River delta, and Madeira Beach, and north-south for Cape Cod and Rincón.

5.1. Shoreline Trends

Differences in sampling and analysis can yield very different trend results in environmental data (Hess and others, 2001). Shoreline trends, both long-term (on the order of hundreds of years) and short-term (on the order of decades or less), are critical metrics for the coastal science community’s current understanding of shoreline behavior (U.S. Army Corps of Engineers, 1984); however, previously published rates for this study’s locations, as many other locations across the United States, were derived from relatively few (compared to satellite-derived data) traditionally derived shorelines. Thus, direct comparisons of trends from all SDS to existing rates are misleading. An in-depth investigation into short-term rates along select transects (figs. 2, 3, 4, 5, 6) from all study sites using only paired SDS and in situ data, showed that SDS-derived rates were comparable to traditional methods, being within the standard error of in situ rates at most transects (table 3). Across the study areas, rates from SDS were similarly comparable to rates from in situ data at 84 percent of transects at all sites except Rincón (table 4). Dynamic shoreline change, complex morphology, shadows, and excessive turbidity complicated transect-based shoreline identification and resulting comparisons at the Elwha River delta and are discussed in section 5.4. A deeper investigation into SDS–in-situ trend comparisons is presented in section 5.3.

More detailed discussions of results and comparisons follow for each study site.

5.2. Barter Island

Several challenges were encountered in extracting SDS data for Barter Island. One was the common presence of clouds and ice, and a considerable amount of time was needed to manually review imagery to filter out unusable and bad detections. The complex landscape of the barrier islands and spits and the rapidly changing shoreline also led to unique setup needs and transect orientation. For example, a particularly challenging area was at the west end of Bernard Spit, where Bernard Spit is migrating westward past the eastern Barter Island spit and Kaktovik (figs. 2, 7). The adjacency of Bernard Spit to Barter Island required special transect generation to avoid overlapping transects and spurious shoreline detection. Noisy data related to recurved finger spits associated with rapidly migrating shorelines both on Bernard Spit and at the western end of Barter Island, where an older shoreline, back side of the spit, or wetland edge would be classified as the shoreline rather than the most seaward position. Identification issues seemed to occur often because of image problems or the presence of small clouds, ice, or a gap in the L7 imagery associated with the scan line corrector. Transects 154–206 were an extreme example of this behavior and contained significantly larger RMSE than adjacent transects (figs. 7, 8); they were removed from analysis. Some outliers throughout the study area were also due to the “compute intersections” function within underlying CoastSat functions. This function, in addition to returning occasional bad solutions, was cumbersome and slow in computation. Filtering techniques (Hampel and LOESS filters) reduced a lot of the noise to get usable trends and data.

After removing numerous bad images and faulty detections, the remaining 389 SDS from the four ROIs were combined for analysis. Of the 15 in situ shorelines, 12 were within 10 days of an SDS from 1987 to 2020. Of the four slope experiments, the no-corrections experiment resulted in the highest accuracy, with an RMSE of 29.6 m and MAE of 16.1 m, although there was little difference between all four experiments (figs. 7A–D). The similarities between the experiments may have been due to the microtidal conditions at Barter Island. Most errors between in situ and SDS observations were less than 40 m, and per-transect errors were spatially variable, with higher errors occurring at the ends of the barrier islands and spits (figs. 7A–D, 7G). MAE and RMSE were a small fraction of the IQR of observed SDS positions, indicating that the change signal was captured well despite the error in the measurements. This was evident where the SDS positions follow the same trends as the in situ shorelines between 1987 and 2020: at transect 102 (near the center of Bernard Spit), the shoreline showed a relatively consistent erosional trend through time, whereas transect 342 (near the western edge of the Barter Island spit) showed gradual retreat from 1987 to 2012, then rapid progradation of the shoreline by about 400 m followed by relative stability (figs. 7E, 7F). However, this did not represent 400 m accretion; instead, it showed alongshore migration of a more seaward fingerling spit. This location exemplified why using a transect-based approach to quantify shoreline change can provide an incomplete or misleading record of shoreline change, particularly on dynamic barrier coasts where a piece of land may have completely eroded, formed, or migrated into a location where no previous landform existed (Gibbs and others, 2019b).

Linear trends of comparison SDS showed primarily erosive trends across the site, except at the western edge of Barter Island and near the center of Barter Island (fig. 7), matching in situ data. Monthly shoreline positions from all SDS showed this erosional trend and distinct spots of accretion in more detail (fig. 8). However, artifacts from ROI boundaries were visible around the 11,000 m and 16,000 m marks (fig. 8); these artifacts stemmed from transect orientation because transects were not parallel to the ROI boundaries but crossed at oblique angles and thus only had imagery for portions of the transect (fig. 8).

A comparison of the SDS, in situ shorelines, and edge of bluff positions indicated shorelines, rather than bluff edges, were classified in the imagery where both were present. However, the width of the coastal bluff shadow was similar to the width of the beach, which may have contributed to the accurate identification of the shoreline.

There were several high-error outliers skewing error distributions, but these stemmed from only a few transects. These outliers were likely a mix of classifier error (erroneous shoreline detection), the shoreline contour being detected at the backside of the barrier island, an older, abandoned fingerling on a barrier spit, or a wetland feature behind a barrier spit. These latter sources of error were common with the L7 data where the “striping” or gaps in the imagery associated with the failure of the scan line corrector fell over the seaward shoreline along a transect (fig. 9).

Despite these substantial challenges in detection, even in the most challenging and high-error areas, plenty of usable data that showed shoreline evolution and behavior in sufficient detail were still extracted (fig. 9). Selected SDS for transects 154–206 showed the complex evolution of the spit end matching in situ data including the shift from accretional to erosional trends moving westward along Barter Island. However, much more time was required to clean and filter SDS data appropriately in these areas, including identifying sources of the largest error. At this location, most of the largest errors were due to the L7 data, which included imagery gaps (“striping”) that led to the misidentification of inland features as shorelines and bad shoreline identifications (fig. 9).

5.3. Cape Cod

Unique challenges related to nearshore bars and breaking waves were encountered when extracting SDS at Cape Cod. Whereas the beach at this site is often relatively straight and uncomplicated, the occurrence of welded bars, longshore bars, and occasional breaking waves unexpectedly complicated shoreline extraction at this location. The presence of these features was periodically the cause of bad shoreline detections, manifesting as gaps in the extracted shoreline, double (or triple) shorelines at the same location, and overly sinuous sections of shoreline (fig. 10). Through iterative testing to determine the optimal CoastSeg:CoastSat settings (table 2), the occurrences of these artifacts and bad shoreline detections were reduced. However, even with optimal settings selected, 8.4 percent of the shorelines were manually removed from the final dataset because of bad detections.

The remaining SDS from the three ROIs were combined for the Cape Cod analysis. There was a total of 1,589 usable SDS, 111 in situ shorelines, and 59 in situ shorelines within 10 days of an SDS between 1998 and 2005. Of the four slope experiments, the average slope resulted in the lowest error with an RMSE of 9.9 m and MAE of 7.2 m (figs. 11A–D). Most errors between in situ and SDS observations were less than 30 m, and per-transect errors appeared to be randomly distributed along the site (figs. 11A–D, 11G). MAE and RMSE were about one-third the IQR of observed SDS positions, indicating that the SDS effectively captured shoreline change trends comparable to in situ data on the outer Cape (figs. 11E–F). The SDS generally followed the same trends as the in situ shorelines between 1999 and 2002. At transect 582, the shoreline was relatively stable during this time, whereas the shoreline at transect 649 experienced tens of meters of erosion followed by approximately the same amount of accretion. After 2002, the in situ timeseries became sparser, complicating efforts to compare the trends (figs. 11E, 11F).

Overall, from 1984 to 2023, the SDS showed a primarily erosive trend interspersed with sections or bands of accretion along this section of Cape Cod National Seashore (fig. 11H). Monthly shoreline positions from all SDS data showed these trends of broad erosion and limited bands of accretion in more detail (fig. 12). From 1990 to 2003, there were distinct bands of erosion and accretion along the 9.5 km stretch of beach. Between 2003 and 2010, the accretional bands in the northern half of the study area began to disappear, reappearing around 2012. By 2023, the entire study area had eroded (relative to the first shoreline) except for a very small section at 5,000 m alongshore.

Short-term shoreline change rates at this study site highlighted the importance of using appropriately matched data for trend comparisons and how differences in sampling can affect resultant trends. In this experiment, only SDS within 10 days of in situ shorelines were used for the comparisons. If there were several SDS within the ±10-day window, only the SDS closest in time to the in situ shoreline was used. When shoreline change trends (short-term change rate) using all SDS were compared to all in situ shorelines during a common time period, the trend patterns were similar over the length of the study area, showing similar spatial relationships; however, the magnitudes of change, and in some cases, the sign of the trend on individual transects differed, as shown in fig. 13. When using all SDS and all in situ data, the mean difference in trends across the study site was 1.96 m/yr (fig. 13A). However, when limited to only those SDS–in-situ pairs within the ±10-day window, the results improved substantially (fig. 13B), with the mean difference in the trends dropping by over a meter to 0.78 m/yr.

5.4. Elwha River Delta

The Elwha River delta site provided an opportunity to evaluate the skill of CoastSeg:CoastSat for SDS estimation in a mixed sediment delta undergoing rapid coastal change. After removal of bad detections, there was a total of 858 usable SDS, 249 in situ shorelines, and 138 in situ shorelines within 10 days of an SDS between 2009 and 2023. With reference to figures 14A–D, 50 percent of SDS measurements were within about 20 m of observed values under all slope experiments. When the average slope was used, the MAE was 16.1 m. Outliers could be large (up to 300 m), and the resulting distribution skew inflated the accuracy statistics. These large outliers were caused by errors with the CoastSat ML classifier. At this site in particular, the ML classifier often failed to detect sand and water pixels, so a fallback MNDWI thresholding was employed (Fitzpatrick and others, 2024a, b); however, the classifier struggled particularly with (a) shadows on transects southwest of the river mouth; (b) turbidity around the river mouth, which could be excessive during the dam removal process; and (c) the contrast in color and texture between the sand and gravel portions of the shoreline. Per-transect errors were therefore higher on transects near the river mouth and southwest of the river mouth (fig. 14F). Despite the outlier noise in the SDS record, trends were captured well in SDS data (figs. 14E, 14G), showing greatest accretion near the river mouth. This was in part due to the very high number of SDS, approximately 4 per month since 2009. Similar to Barter Island, there were challenges related to the transect-based approach for characterization of shoreline change, given the dynamic changes at this site. Monthly SDS from all data showed this behavior in greater detail and highlighted transitions to more erosional behavior (fig. 15B).

Use of a scalar beach slope may have been problematic because of the convex-upward profile shape at this site, which meant the slope tended to steepen landwards. Therefore, a lower datum, such as mean low water (MLW), may be a better shoreline contour for future SDS work because that portion of the beach face may be more linear, and horizontal corrections would be smaller in magnitude. Time-varying and space-varying slope changes may have accounted for some error in this generally steep but variable mesotidal environment, and these changes could be addressed in a future study. We have adapted CoastSeg to be able to import a database of slope values per transect and period, so this adaptation could be used for spatiotemporal slope water level corrections in the future. The cliff shadow effect was most evident in winter (November through March) and caused shorelines to be detected more inland because the shadows were classified as water. With reference to figure 16, shorelines in the eastern portion of the study area, unaffected by shadows, showed no systematic tendency with month of year and water level (figs. 16A, 16C), whereas those in the western portion affected by shadows in winter showed a systematic landward bias in winter at all water levels (figs. 16B, 16D). The shadow effect was a classifier failure and could be potentially addressed in the future by customizing the ML classifier for use at this specific site.

5.5. Madeira Beach

No significant challenges were encountered when extracting SDS data for Madeira Beach, and as such, many CoastSeg extraction settings were left at default settings (table 2). Challenges were expected with clouds and weather, but with the large amount of available imagery, clouds did not adversely affect the amount of usable data.

There was a total of 110 usable SDS, 43 in situ shorelines, and 23 in situ shorelines within 10 days of an SDS between 2016 and 2023 (table 1). The Madeira Beach site was a stable and highly managed beach with negligible to small erosional trends in comparison SDS (fig. 17), though LOESS filtered data showed areas of accretionary behavior and an erosion event over much of the site (fig. 18B). Most errors were less than 15 m (figs. 17A–D), and per-transect error appeared to be randomly distributed; however, the MAE and RMSE were roughly commensurate with the IQR of SDS, suggesting a low signal-to-noise ratio. As a highly managed beach, Madeira Beach was nourished on a routine cycle, reducing detectable signal. Notably, the absence of tidal corrections yielded the strongest results and lowest error in this setting, likely because of the low tidal range; however, differences in error between all four experiments were small (on the order of 1 m or less).

Overall small mixed accretion and erosion was shown in the LOESS filtered data (fig. 18B). Despite the managed nature of the beach and low signal-to-noise ratio, certain areas displayed enduring patterns of erosion and accretion, particularly around the 50 m, 150 m, and 250 m marks (fig. 18). These observed fluctuations in shoreline dynamics may have been influenced by the positioning of transects in relation to the existing concrete groins, leading to a complex interplay between structural interventions and natural beach processes. For example, the transect at the 150 m mark showed the largest accretionary trend (figs. 17H, 18B) and was situated between two groins, whereas nearby transects showing erosion trends were located nearer adjacent groins. Erosion across much of the beach was observed at the end of 2023 (fig. 18B), which may have been due to a very active hurricane season for Florida; 20 named tropical storms formed in the Atlantic basin that year, ranking 2023 the fourth most active year for the Atlantic Ocean on record (National Oceanic and Atmospheric Administration, 2023). These events may also have driven the linear trends over this period.

5.6. Rincón

The Rincón site provided an opportunity to evaluate the skill of CoastSeg:CoastSat for SDS estimation in reef-fronted beach areas. Obtaining a complete record of imagery for this site was a challenge; 30 percent of available satellite imagery was unusable because of cloud cover, incomplete coverage, and other image quality issues, reducing the number of usable shorelines at the site compared to other locations.

There were around 150 SDS from remaining imagery, 17 in situ shorelines, and only 2 in situ shorelines within 10 days of an SDS between 2012 and 2023 (table 1). Linear trends from comparison SDS showed a more mixed pattern than broad erosional trends from in situ data (fig. 19H); this disparity may have been due to the limited comparisons available at this site. LOESS-filtered data from all SDS, however, showed more broad erosional behavior and localized bands of accretion over the study period (fig. 20B). Errors were generally less than 20 m (fig. 19), and per-transect errors were distributed randomly; however, except for the low-slope scenario, MAE and RMSE were similar to the IQR of SDS observations, suggesting a low signal-to-noise ratio. Of note, the low-slope tidal corrections experiment was most accurate and resulted in much lower error than the other slope experiments; MAE and RMSE for the low-slope scenario were approximately half the values for the no slope or other slope scenarios. Because we lacked sufficient comparison data to investigate, more in situ data or proxy-datum bias corrections would be needed to further explain these results.

The coral reefs did not affect the ability of CoastSeg to derive shorelines from satellite imagery. Although there were few observations before 1999, enough SDS were available after L7 imagery became available to support monthly observations (fig. 20). Erosional trends were broadly observed across the shoreline, but the middle of the Rincón coastline showed the greatest erosion, matching erosion hotspots from in situ data. The highest erosion was around the 2,800 m mark (fig. 20). The second-highest erosion was seen at the 600 m mark in the southern portion of the study area (fig. 20). Both erosion areas were coastal sections with relatively wider sandy beaches without coastal structures or buildings. Two areas adjacent to the central erosional hotspot (around the 2,400 m and 4,700 m marks) showed consistent accretion in monthly plots (fig. 20), which generally matched long-term trend data.

5.7. Time and File Size Usage

Details on time and data storage needed to prepare and process data at all sites were noted to understand what may be required to develop a workflow for other sites. In this section, information associated with the image extraction for Barter Island is discussed throughout because it is representative of the maximum range of time and storage needed to complete work for this study. All other sites required less time and file space, pending availability of imagery and challenges with extraction.

Several weeks were needed to fully set up ROIs, transects, and reference shorelines for Barter Island, but much of that time was spent becoming familiar with the CoastSeg software, file formatting requirements, and working with the developers on software improvements. As much as a full workday (8–10 hours) was needed just to manually screen and filter out bad imagery and initial detections, and more ROIs (greater area) required more time. Once personnel are familiar with the CoastSeg software and workflow, this time could be significantly reduced to only 1 or 2 weeks to fully set up, refine settings, and filter imagery or detections, even for the most challenging sites like Barter Island.

Work was done on a desktop computer with sufficient storage and processing capability for scientific processing. The amount of storage needed for SDS data depends on the size of the study area, the period of interest, and the satellites used. At Barter Island, sessions for raw (non-tidally corrected) and one tidal correction for all four ROIs over the entire study period (1984–2023) required 29 gigabytes (GB) to store the Landsat imagery. Because this site also required the use of Sentinel-2 imagery, the amount of storage space used by downloaded data increased to approximately 40 GB. In total, shoreline detections and associated output required about 1,200 megabytes (MB), of which the Landsat imagery accounted for around 900 MB. Thus, about 31 GB of space was needed for a Landsat workflow at Barter Island; this included downloaded imagery underpinning the final datasets, as well as bad imagery, bad detections, supporting files, and all session output. Again, this estimate represents the upper end of file size requirements. The other locations needed less storage space.