Our study focused on an 8 km stretch of coast to the east of Drew Point on the North Slope of Alaska (Fig. ). This study area largely consists of exposed ice-rich bluffs with narrow beaches along the coast and thermokarst lakes and drained lake basins onshore . Most of the shoreline in this region is less than 3 m high, although the bluffs near Drew Point reach as high as 6 m . Erosion is largely driven by thermal mechanical notching of bluffs, followed by bluff collapse . The area around Drew Point is among the most rapidly retreating locations in the Arctic, with decadal-scale shoreline change rates in excess of -10 m a⁻¹ and evidence that retreat rates are increasing in recent years . found highly variable change in this study area, with mean shoreline change averaged over a 9 km stretch of shoreline varying from -6.7 to -22.0 m per open-water season between 2007 and 2016. Although they found that the greatest retreat occurred in the year with the most storms and the warmest air temperatures, they did not find a robust relationship between spatially averaged retreat rates and open-water days, storm activity, air temperature, permafrost temperature, or sea surface temperature over the entirety of the 10-year study period. Maximum local retreat was 2–3 times higher than spatially averaged retreat, pointing to the potential influence of local controls (such as morphology) on shoreline change.

We divide our study area into three regions (Fig. b.) with repeated cross-shore ICESat-2 profiles. We delineate these regions primarily based on visual analyses of 2018 optical imagery collected by CNES Airbus that was accessed through Google Earth and draw on , , and the ShoreZone database to describe the morphological setting of each region. ShoreZone includes a classification of variable-length segments of the shoreline based on aerial photography taken in 2007. These classifications are based on the substrate and morphology of the shoreline as well as the dominant erosion and accretion processes thought to be present. Given the high amount of shoreline retreat in the last decade , specific local features identified by ShoreZone in 2007 may have been removed or modified, but we expect the general morphologic setting to be consistent with current conditions. Region 1, the westernmost portion of the study area, primarily consists of steep, ice-rich coastal bluffs . ShoreZone indicates the sporadic presence of low, vegetated peat-rich sediment, which is likely the remnants of old drained lake basins that have since been removed by erosion. Region 2 consists of the headlands and is characterized as having the same shoreline types as Region 1 (coastal bluffs and drained lake basins). Region 3 is a large breached thermokarst lake that is classified in ShoreZone as a large lagoon behind a narrow strip of low vegetated peat. The easternmost portion of this shoreline is classified as an inundated tundra environment, where the nearshore elevation is below sea level and where there has been significant thaw subsidence and flooding.

To investigate the drivers of year-to-year variations in shoreline positions, we considered ocean wave conditions, air temperature, and sea ice concentration from the European Centre for Medium-range Weather Forecasts Reanalysis v5 (ERA5) dataset . ERA5 provides hourly estimates of atmospheric and sea ice conditions at a 0.25° resolution and wave conditions at a 0.5° resolution.

We estimated annual shoreline positions using 3 m multispectral (red, green, blue, and near-infrared) images from Planet Labs' Super Dove, Dove R, and Dove Classic satellite constellations . We used four images, each collected near the beginning of each open-water season from 2019 to 2022 (Table , Fig. ). Frequent cloud cover in 2019 necessitated the use of imagery from 25 June, when a small amount of snow or ice remained visible near the shoreline.

Historically, shoreline change has been estimated from satellites via manual delineation of the shoreline . Recent workflows such as CoastSat have been developed to automatically detect the shorelines at a subpixel resolution, but they have focused on lower-latitude beaches and may not perform well in Arctic regions where sea ice is present. Here, we implement our own shoreline detection method, following some of the same steps as . For each image, we calculated the normalized differenced water index (NDWI) from the green (G) and near-infrared (NIR) bands: 1NDWI=G-NIRG+NIR. We identified the NDWI threshold corresponding to the land–water boundary using Otsu's method, which determines the threshold that divides a set of pixels into two classes such that the interclass variance is maximized (see the histograms in Fig. ). We identified the subpixel land–water boundary from our NDWI images using a marching squares algorithm implemented in matplotlib contour in Python . We found that calculating our threshold using all image pixels resulted in an adequate shoreline estimate for all four of our images, such that an initial identification of land and water pixels (as is done in ) is not necessary. We visually identified and masked out regions where the shoreline appeared to be misidentified due to an ambiguous land–water boundary (shaded regions in Fig. ). In order to improve the visual agreement of the derived shorelines with the visible shoreline in imagery and to ensure regular along-shore sampling intervals, we smoothed each shoreline using a 30 m along-shore running mean and sampled every 10 m alongshore to produce our final shoreline segments (Fig. ). Finally, we estimated shoreline change using an approach similar to that of the USGS Digital Shoreline Analysis Software . In this workflow, a reference shoreline (hereafter referred to as the baseline) is selected. Cross-shore transects are generated at regular intervals along-shore that are perpendicular to this baseline. The change between consecutive shorelines is then calculated along these transects. We created a baseline by smoothing the raw 2020 shoreline with a 60 m along-shore running mean in order to make sure that adjacent cross-shore transects did not intersect on the time interval of our study site. Transects were generated every 10 m along the baseline using a modified version of the SDS_transects routine from the CoastSat Toolbox , which generates transects that bisect the shoreline. The change along each transect was calculated between each successive shoreline.

The estimated shorelines are subject to uncertainty from image geolocation errors, tidal- and storm-driven wave runup influences on the land–water boundary, and errors in the threshold determined with Otsu’s method. The tides in this region tend to be less than 0.2 m, but storm surges can result in temporary relative sea level increases of up to 1.4 m . However, given that much of this region consists of steep bluffs with narrow or no beaches , changes in the local relative sea level are not expected to have a large impact on the observed land–water boundary. To estimate the uncertainty in our shoreline estimates, we identified six clusters of three to six images (25 images in total) taken within a week of one another in the early summer (mid June through late July). Modeling of coastal bluff retreat in this region by suggests that large (> 2.5 m) retreat events are concentrated in late summer (late July–late September) and driven by storms. Therefore, we do not expect large retreat events during early summer, such that the dominant source of change over short periods will be transient signals, such as relative sea level changes and geolocation offsets. We calculated the cross-shore difference between each shoreline position and the mean position of its cluster and pooled the residuals across all shorelines (Table ). We defined our shoreline estimation uncertainty to be the standard deviation of the pooled residuals.

Outside of the poles, the majority of ICESat-2 revisits between 2019 and 2022 were off-pointed from their nominal ground track location to increase aerial coverage, such that subsequent revisits did not cover the same ground location. However, “target of opportunity” requests over the North Slope of Alaska resulted in exact repeats of every fifth descending reference ground track starting in April 2019. We analyzed repeat ICESat-2 elevation profiles from three ground tracks (ground tracks 3r, 2r, and 1r, labeled in Fig. a) from a single reference ground track crossing our study area. Due to frequent cloud cover in the summer and fall, as well as the relatively low surface reflectivity of the snow-free tundra, the majority of observation dates with sufficient surface-reflected photons to accurately identify the surface occurred in the winter (January and April), with 2021 being the only year with snow-free profiles available. We selected four observation dates spanning the 2019–2021 open-water seasons (Table ), resulting in 12 individual profiles.

For each profile, we used 200 m of cross-shore ATL03 photons for our analysis. Each photon from ATL03 is assigned a confidence code ranging from 0 (noise) to 4 (high confidence) to indicate how likely it is to be a transmitted photon that was reflected from the surface (i.e., a signal photon). From each ATL03 profile, we selected all high-confidence (confidence score = 4) and medium-confidence (confidence score = 3) photons for further processing. These photon data were plotted and inspected to provide a qualitative analysis of small-scale features. We used the SlideRule Python client to run a customized version of the ATLAS/ICESat-2 L3A Land Ice Height (ATL06) algorithm , which estimates surface elevations using an iterative process of filtering and linear fitting of individual photon elevations. In this process, a linear fit is performed over fixed-length along-track segments of signal photons, outliers are filtered out based on a specified window above and below the linear surface, and the remaining photons are refit. This process is iterated using a successively narrower outlier window until either it converges or the outlier window reaches a user-specified minimum value. The reported elevation for each segment is the height of the midpoint of the final linear interpolation. SlideRule provides the root-mean-square (RMS) error between the photons used in the final fitting and the final linear fit, as well as the photon-level elevation error that is propagated through the linear fit. The photon elevation error is assumed to be uniform for a given segment and is estimated as the maximum of the segment's RMS error and the robust spread of photons as defined by . This robust spread is based on the vertical distribution of signal photon heights and the estimated rate of background photons (i.e., photons that are not surface signals).

For coastal applications, a high along-track resolution is preferable to capture abrupt elevation changes at the shoreline. However, shorter segments may not provide enough photons for a robust linear fit and may result in height estimates that are subject to along-track variations in photon density. A longer segment length results in a smoother profile at the cost of not capturing small-scale features. In order to strike a balance between these two considerations, we implemented the SlideRule ATL06 algorithm for 10 m long segments spaced every 2 m along track, such that consecutive segments overlapped by 80 %. A 10 m section was only considered valid if it contained at least five signal photons and if those photons were distributed over at least 1 m along track. Bright surfaces such as surface ponds can lead to after pulses in the ATL03 data, which appear as secondary surfaces starting ∼ 0.45 m below the true surface . The original ATL06 algorithm sets the minimum height of the outlier window to 3 m, which leads to the inclusion of these after pulses in the final surface elevation estimate. To avoid including after pulses in our analysis, our custom ATL06 processing allows for an outlier window height as narrow as 0.80 m (i.e., 0.4 m above and below the identified surface).

To estimate shoreline change from our custom ICESat-2 elevation profiles, we needed to reliably identify shoreline boundaries from the derived elevation profiles. The presence of sea ice and snow in three of the ICESat-2 tracks prevents the accurate identification of a land–water boundary. Instead, we identified the boundaries of the backshore, defined here as the relatively steep region between the beach or ocean and the onshore region. We manually identified the point corresponding to the backshore–onshore boundary (henceforth referred to as the “upper shoreline”) and backshore–offshore boundary (the “lower shoreline”) based on the visual breaks in the along-track slope (see Fig. ). Since beaches in this region are very narrow when present , we expect the lower shoreline from ICESat-2 to be similar to the land–water boundary. We identified the intersection between each ICESat-2 track and the corresponding imagery-derived shoreline and compared the shoreline positions and cross-shore retreat estimates derived from Planet and the two ICESat-2 boundaries. Under the assumption that the majority of the observed position change in our ICESat-2 boundaries is due to cross-shore change, we projected our ICESat-2-derived change estimate in the local cross-shore direction. The local cross-shore vector was based off of the baseline transect (as defined in Sect. 3.2) that was located closest to the geographic midpoint of the four ICESat-2 transects associated with each ground track.

In order to characterize the morphology of each profile, we calculated both the backshore elevation and the backshore slope. The backshore elevation was defined as the elevation difference between the backshore–onshore boundary and the mean offshore (north of the backshore–offshore) elevation from the 2 July 2021 elevation profiles, which we used as a proxy for local sea level (illustrated in Fig. ). The backshore slope was estimated using a linear fit of all points between the backshore–onshore and backshore–offshore boundaries, projected into the local cross-shore direction.

To estimate the uncertainty in our ICESat-2-derived boundaries, we use the geolocation uncertainties estimated by , who coregistered ATL03 data with ArcticDEM and reported the distribution of offsets between the original and shifted locations for each beam. We consider the total uncertainty, defined as the mean offset plus 1 standard deviation for each beam. This total uncertainty ranges from 2.8 to 4.8 m for individual beams. We consider the worst-case scenario where this error is entirely in the cross-shore direction. The satellite performs a “yaw flip” twice every 502 d where it is reoriented by 180° . This means that the specific beam sampling a given ground track, and therefore the associated geolocation uncertainty, alternates between repeats. We take the uncertainty in the shoreline change between two observation dates as the sum in quadrature of the uncertainties in the individual measurements and find that they range from 4.0 to 5.9 m.

We observed the most extreme environmental conditions as measured by all measured metrics in 2019 (Table ). The 2019 erosion year had the longest open-water season, with sea ice breaking up sooner (late June) and re-forming later (mid-November) than in the other 2 years. It also had the highest wave energy, the most storms, highest air temperatures, and over twice as many ocean thawing degree days as the following 2 years. By contrast, the 2021 open-water season was the shortest, with sea ice breakup occurring in late July (although ERA5 suggests a 1 d ice breakup on 27 June) and freeze-up occurring in late October. There were slightly warmer ocean and air temperatures but lower wave activity (in terms of both wave energy and hours of high waves) in 2021 compared to 2020.

The spatially averaged shoreline change rate in our study area was -16.5 m a⁻¹ (corresponding to retreat) between 2019 and 2021, with notable year-to-year and local variability (Table , Fig. ). The year 2019 had the most shoreline loss, with a mean shoreline change of -23.7 m and single-segment (i.e., 10 m scale) shoreline change values as extreme as -66.7 m. The year 2020 experienced more moderate shoreline change, with a mean shoreline change of -15.4 m (ranging from -41.0 to +6.0 m). The year 2021 experienced similar but slightly lower rates, with a mean shoreline change of -10.5 m (-30.9 to +5.6 m).

Based on the uncertainty estimation described in Sect. 2.3, we estimated the precision of our shoreline position estimates to be 2.2 m. Assuming that the error in each shoreline is independent of the others, this corresponds to a position change uncertainty of 3.1 m. We define substantial shoreline change as any value that exceeds this threshold.

Our shoreline change estimates for each region (Table , Fig. b) indicate that there was spatial variability between regions. Region 1 showed high retreat, with a mean change rate of -12.9 m a⁻¹ over 3 years and single-segment year-to-year shoreline change estimates ranging from -29.4 to +1.6 m. Shoreline change consisted almost exclusively of retreat, with the maximum observed advancement (+1.6 m) falling below our estimated uncertainty threshold (3.1 m). We observed the greatest spatially averaged retreat in 2019 (-18.5 m) and the lowest in 2021 (-7.1 m). While every shoreline segment in Region 1 underwent retreat in 2019 and 2020, 16 % of valid shoreline segments (all in the western half of Region 1) did not exhibit substantial (> 3.1 m) shoreline change in 2021.

Region 2 exhibited the largest overall retreat, with a 3-year spatially averaged mean change rate of -24.3 m a⁻¹ and year-to-year single-segment change estimates ranging from -58.2 to -3.0 m. As with Region 1, we observed the largest amount of shoreline change in 2019 (-38.4 m), with similar and smaller change in 2020 (-16.8 m) and in 2021 (-17.7 m).

Region 3 exhibited the smallest amount of shoreline change, with a 3-year spatially averaged rate of -6.0 m a⁻¹. Region 3 was the only region in our study area where substantial shoreline advance (up to +18.6 m locally in 2019) occurred. Although 2019 was not the year with the highest spatially averaged shoreline change (-0.6 m) in Region 3, it was the most dynamic year, with both the maximum observed local retreat (-66.7 m of shoreline change) and advance (+18.6 m of shoreline change). We found that 24 % of valid shoreline segments, concentrated almost exclusively in the easternmost third of the region, underwent significant retreat, whereas the central 62 % of shoreline segments experienced substantial advance. The remaining 14 % did not exhibit substantial retreat or advance. In 2020, the mean shoreline change across the region was higher (-13.5 m), but the change at individual segments was lower in magnitude (with a local maximum of -41.0 m) and spread over a larger area (the eastern 66 % of the basin). Only 3 % of shoreline segments showed substantial advance, whereas the remaining 31 % experienced no significant change. Relative to 2020, retreat in 2021 was lower in magnitude (with a maximum shoreline change of -20.7 m) and present across less of the region (40 %), resulting in a lower spatially averaged shoreline change (-3.9 m), with 19 % of segments showing significant advancement (with a maximum of +5.6 m) and the remaining 41 % experiencing no significant change.

Altimetry from ICESat-2 provides an independent estimate of shoreline change at multiple positions along the shoreline profile, which can be compared to Planet-derived shoreline positions for validation. We find that overall shoreline positions estimated from both datasets are consistent. We would expect the land–water boundary from Planet to be located seaward (north) of the lower backshore boundaries identified by ICESat-2. However, we found that the Planet-derived land–water boundary consistently falls landward (south) of the ICESat-2-derived lower boundary (by 8.6 ± 4.2 to 41.1 ± 4.5 m) and either seaward (by up to 36.0 m ± 4.5 m) or slightly landward of the ICESat-2-derived upper boundary (by up to 3.3 ± 4.0 m) (Fig. ), such that it is located on the backshore or onshore section of the elevation profiles. Given the published geolocation errors from ICESat-2 (up to 4.8 m; ) and Planet (< 10 m; ), the difference between ICESat-2- and Planet-derived shorelines is not thought to be due to a consistent geolocation offset between the two satellite estimates. The difference could be explained by a consistent landward bias in our NDWI thresholding technique (Sect. 2.3) or by differences introduced by snow cover and variations in the local water level. Although observations from ICESat-2 and Planet span different time intervals (Table , Fig. ), the number of open-water days spanned by both are similar, allowing for a direct comparison between our imagery and altimetry-defined shoreline change estimates. We perform an orthogonal distance linear regression and find a high correlation between the Planet-derived change estimates and changes in the ICESat-2-derived upper boundary (r2=0.79, p=0.0008) and moderate correlation with changes in the ICESat-2-derived lower boundary (r2=0.52, p=0.007) (Fig. a).

Our SlideRule-derived elevation profiles (Fig. b, d and f) fit the ATL03 photon data well, with RMS errors ranging from 0.04 m to 1.3 m and a mean RMS error of 0.15 m (Figs. , , ). The propagated vertical uncertainty ranges from 0.0024 to 0.27 m, with a mean uncertainty of 0.029 m. We now discuss the time evolution of the shoreline horizontal position (Table ), backshore elevation, and backshore slope (Table ) at each ICESat-2 ground track.

Ground track 3r in Region 1 (Figs. a, b, a) shows a coastal bluff (such as the one shown in Fig. a) that underwent retreat with little change in morphology. The upper- and lower-shoreline boundaries showed consistent retreat in 2019 and 2020, with cross-shore change at each boundary ranging from -16.9 to -28.4 m. During these 2 years, we observed a steepening of the backshore slope from 15 % in 2019 to 28 % in 2020 and 2021. In July 2021, we note a cluster of photons that is ∼ 1 m high and ∼ 5 m across at the base of the bluff (Fig. a) that may correspond to toppled bluff material. Between 2021 and 2022, we observed slight retreat of the upper boundary (-6.6 m of shoreline change) and advance of the lower boundary (+20.3 m), resulting in a reduction in the backshore slope to 11 %. In December 2021, we again note a ∼ 1 m high cluster of photons at the base of the bluff that gently slopes down towards the mean offshore height over ∼ 20 m (Fig. a). The backshore elevation remained stable, ranging from 4.89 to 5.02 m over the 3-year study period.

Ground track 2r (Figs. c, d, b) in Region 2 passes over the remnant basin of a ∼ 150 m diameter lake that was breached and drained as a result of shoreline retreat prior to our study. This profile underwent large changes in both the shoreline position and morphology. Aerial imagery from ShoreZone (Fig. b) showed the lake pre-drainage, and the lake basin was clearly visible in Airbus imagery from Google Earth (Fig. c) and in our 2019 ICESat-2 profile (Fig. d). Between 2019 and 2020, -45.3 m of upper-shoreline change and -67.2 m of lower-shoreline change occurred, corresponding to the loss of about half of the basin. This resulted in a 0.40 m drop in the backshore height, from 1.33 to 0.93 m, and an increase in the backshore slope from 2.4 % to 3.9 %. By contrast, between early 2020 and early 2021 there was almost no change in the upper-shoreline position (-2.5 m) and only moderate change in the lower-shoreline position (-12.5 m), resulting in a further increase in the backshore slope to 6.7 %. We also observed a slight lowering of 0.23 m of the backshore height. The photon distribution over the lake basin in July 2021 is concentrated in a narrow height band, with a secondary reflection below the surface (Fig. b). This feature is indicative of an after pulse due to surface ponding , suggesting the presence of shallow water. The rest of the lake basin was eroded between July and December 2021, resulting in -62.1 m of change in the upper shoreline and -20.3 m of change in the lower shoreline. This necessitated the definition of a new, substantially higher (2.04 m) backshore boundary and resulted in a shallowing of the backshore slope from 6.7 % to 3.5 %.

Ground track 1r (Figs. e, f, c) in Region 3 passes over a 1.71 m high dune in front of a ∼ 2.6 km wide breached thermokarst lake (shown in more detail in Fig. c) that showed little shoreline change relative to other regions. We observed between +1.7 and +10.2 m of advance in the upper boundary each year, while the lower boundary retreated in 2019 (-7.7 m) and 2020 (-10.0 m) and then advanced in 2021 (+15.6 m). We note a 0.22 m drop in the backshore elevation (from 1.89 to 1.67 m) between 2019 and 2020, after which the elevation remained stable from 2020 to late 2021 (1.64 to 1.67 m). The backshore slope steepened between 2019 and 2021, from 3.6 % to 7.0 %, and then relaxed to 4.3 % in late 2021.

Our estimates of spatially averaged (regional) mean annual shoreline change (-10.5 to -23.7 m a⁻¹) across our study region are higher than long-term historical estimates and are similar to recent observations (Fig. ). estimated a regional mean of -6.3 m a⁻¹ and a local maximum of -18.6 m a⁻¹ between Drew Point and Cape Halket between 1947 and 2002. estimated shoreline change across this region over multiple time intervals and found -6.8 m a⁻¹ of change between 1955 and 1979, -8.7 m a⁻¹ between 1979 and 2002, and -13.6 m a⁻¹ between 2002 and 2007. A follow-up study by estimating shoreline change over a 9 km region covering our study area found a 10-year mean shoreline change rate of -17.2 m a⁻¹ between 2007 and 2016. In addition to estimating a 10-year mean, reported regional year-to-year rates, which ranged from -6.7 to -22.6 m a⁻¹. Our mean retreat rates in 2020 and 2021 fall within the range of the year-to-year rates observed by , whereas our mean retreat in 2019 (-23.7 m) slightly exceeds that range. Our 3-year mean of observed retreat rates (-16.5 m) is similar to the decadal-scale estimate from and is consistent with an increase in local shoreline change rates compared to 2002–2007 . Our retreat estimates and the post-2002 estimates of and are all higher than the pre-2002 decadal-scale estimates of and . This increase could be reflective of a long-term increase in retreat rates, although it may also be due to differences in timescales between studies, as short-term estimates of shoreline change tend to be higher in magnitude than long-term estimates .

Previous observations suggest that the spatial distribution of change rates across our study area has been variable over time. found the highest rates of shoreline change in Region 3 (-10 to -18 m a⁻¹), intermediate rates in Region 2 (-8 to -12 m a⁻¹), and slightly lower rates in Region 1 (-6 to -12 m a⁻¹) between 1947 and 2002. found a similar pattern between 1955 and 1979, with the shoreline change in Region 3 occurring at rates in excess of -10 m a⁻¹ and shoreline change in most of Region 1 and Region 2 occurring at rates between -5 and -10 m a⁻¹. However, they found that shoreline change rates in Region 2 increased to be in excess of -10 m a⁻¹ between 1979 and 2002, and change rates in Region 1 followed suit between 2002 and 2007 such that change rates across the region were relatively uniform (averaging -18 m a⁻¹) between 2002 and 2007.

In contrast, our observed shoreline change rates between the three regions between 2019 and 2021 are not uniform. Although the relatively high rates of retreat we observed in Region 2 compared to Region 1 are consistent with and , the stability of the central portion of Region 3 that we observed with both ICESat-2 altimetry and Planet imagery appears to be a recent development. sampled a transect across the central portion of Region 3 from satellite-imagery-derived shorelines from 1974, 1985, 1992, 2001, 2009, and 2017 and estimated a mean shoreline change rate over the full study period of -55.9 m a⁻¹. When considering the rates between successive shorelines, found that 2009–2017 was the only time interval during which there was no observed shoreline change, suggesting that stabilization occurred in this time frame after over 4 decades of retreat.

found that the environmental drivers of year-to-year variability on the Beaufort Sea coast are not well-defined, but they did observe high retreat in years with extreme weather, which is consistent with our findings of both high retreat and extreme environmental conditions in 2019. Compared to 2020 and 2021, 2019 had a long open-water season and elevated air and ocean temperatures (Table ), all of which likely contributed to the elevated spatially averaged retreat across the study region.

Although there is correspondence between mean year-to-year shoreline change and year-to-year variations in wave and temperature conditions, the response of the shoreline to these time-varying conditions is not spatially uniform (Fig. b). In 2019, Region 1 and Region 2 both experienced high retreat, while Region 3 underwent high retreat across parts of the shorelines and moderate advancement across others. While we do not expect air and ocean temperatures to vary greatly spatially over the study region, local variations in wave energy due to shoreline orientation, position, and morphology may contribute to the observed spatial variability. For indented shorelines such as our study area, wave energy is expected to be concentrated towards the headlands due to wave refraction, such that shorelines with uniform composition will straighten over time . This provides a potential explanation for increased retreat in Region 2, particularly in 2019 when wave action was high. The year 2021 saw the lowest amount of wave energy and fewest storms (Table 2), and the concentration of wave energy in the headlands of Region 2 would result in particularly low wave energy in the hinterlands of Region 1 and Region 3. We note that while Region 1 and Region 2 both consist primarily of ice-rich coastal bluffs, the shoreline in Region 3 is characterized by ShoreZone primarily as low-lying peat (Fig. b.) that is expected to be subject to low incident wave energy . The eastern edge of Region 3 is characterized as inundated tundra, which refers to areas where thaw subsidence and surface ponding are present . Based on these characterizations and on our observed patterns of shoreline change, we posit that shoreline change in Region 1 and Region 2 is sensitive to year-to-year variations in wave energy and ocean and air temperatures, while the western and central portion of Region 3 are not. The eastern portion of Region 3 (inundated tundra) is likely subject to ground thaw, high wave energy, and flooding, resulting in high retreat in response to high wave activity and high temperatures. The advancement observed in Region 3 may be driven by along-shore transport of material lost from the surrounding regions.

The elevation profiles from ICESat-2 data provide additional information on topographic and morphological change at specific ground tracks (Table ), which can provide insight into specific erosion and accretion processes that drive local shoreline change. Specifically, we consider changes in backshore elevation, backshore slope, and relative change between the upper- and lower-shoreline boundaries.

Based on previous studies in the Beaufort Sea coast area e.g.,, we infer that the coastal bluff shown in ground track 3r in Region 1 (Fig. b) likely retreats primarily through the formation of thermal erosional niches, followed by bluff collapse. As the shoreline retreats, we observed no major morphological change; the backshore slope remained > 10 % with a stable (4.87 to 5.02 m) backshore elevation. However, there were year-to-year fluctuations in the backshore slope due to differences in change rates at the upper and lower boundaries of the backshore. In particular, the relatively low slope (11 %) in December 2021 is driven by a retreat of the upper boundary (-6.6 m of shoreline change) coupled with advance of the lower boundary (+20.3 m of shoreline change) (Table ). Based on the ICESat-2 ATL03 photon data from July and December 2021 (Fig. a.), we hypothesize that this reduction in slope is due to the accumulation of collapsed bluff material at the base of the cliff that is not removed by the end of the 2021 open-water season. The low retreat derived from altimetry in 2021 is consistent with the low retreat estimated from Planet imagery in the broader Region 1 (which has a spatially averaged mean shoreline change of -7.1 m) and is possibly due to relatively low wave energy and fewer storms in the 2021 open-water season (Table ). It may be that multiple locations along the shoreline in Region 1 observed in imagery were sheltered by un-eroded toppled bluff material similar to what we observe at ground track 3r.

The drained lake basin captured by ground track 2r in Region 2 (Fig. d) is an example of a more complex feature, which displays variable retreat as the shoreline migrates from the lake basin boundary to the bluff on the southern edge. The transition in shoreline morphology is reflected in an increase in backshore elevation from 0.70 to 2.04 m and moderate fluctuations in the backshore slope (2.4 % to 6.7 %; Table ). The changes in slope between 2019 and 2020 and between early and late 2021 were driven by high and variable change (-20.3 to -67.2 m) in both the upper and lower boundaries, while the steepening of the slope between 2020 and 2021 was driven by the retreat of the lower shoreline (-12.5 m of shoreline change) and by little change in the upper shoreline (-2.5 m).

In order to assess whether the presence of this lake basin impacts local retreat rates, we compared the mean Planet-derived shoreline change across the drained lake basin (which spans 15 segments over 150 m) to the distribution of Planet-derived shoreline change across the entirety of Region 2. In 2019, the mean shoreline change across the lake basin (-35.3 m) is similar to the retreat in Region 2 overall, as it falls in the 63rd percentile of observed change. However, in 2020, shoreline change across the basin (-10.5 m) fell into the 8th percentile of observed headlands change, and in 2021 it was in the upper 90th percentile (-23.1 m). This suggests that the presence of the drained lake basin may have contributed to differential retreat during 2020 and 2021 relative to Region 2 overall. Differential retreat rates in drained lake basins have been observed before, with observing generally higher retreat rates of recently drained lake basins in this region between 1955 and 2002 relative to other land types (although this trend did not hold between 2002 and 2007). They suggested that the low elevation and presence of thawed sediments likely make drained lake basins more susceptible to erosion than coastal bluffs are. Our observation of surface water in July 2021 (Fig. b) is consistent with this interpretation, as it would help keep the underlying sediments in this basin thawed.

Based on our observations and the theory presented in Jones (2009), we can infer the processes driving the 3-year evolution of the drained lake basin. In 2019, wave-driven erosion of the seaward side of the drained lake basin exposed low-lying thawed sediments. In 2020, debris from this large erosion event may have sheltered the remainder of this basin, as evidenced by the retreat at the lower boundary and stability of the upper boundary. Once this debris was removed by waves, the low backshore elevation (0.70 m) would have left the basin susceptible to overtopping by high waves, and the thawed sediments would have had low mechanical strength, making the basin susceptible to rapid erosion, as was observed between early and late 2021. The steeper backshore slope in 2019 compared to in 2021 may have also contributed to increased wave runup and therefore increased mechanical erosion.

Both the upper- and lower-shoreline positions at ground track 1r in Region 3 (Fig. f, Fig. c) are relatively stable, whereas the other ground tracks exhibit high retreat. We see moderate advance (+10.2 m of shoreline change) and a drop in elevation (-0.22 m) of the backshore boundary in 2019 but very little change in either the position (+1.7 to +1.9 m, which falls within our estimated uncertainty) or elevation (which ranges from 1.64 to 1.67 m) of the backshore over the next 2 years. Fluctuations in the slope between observation dates are driven primarily by changes in the lower boundary (which undergoes between -10.0 and +15.6 m of shoreline change per year), which is consistent with sediment deposition and removal at the beach in front of the lagoon or changes in snow cover.

In order to understand the relative rates of retreat at the upper and lower boundaries of the shoreline, we perform an orthogonal distance regression between the annual change at both boundaries (Fig. b). We exclude change for the drained lake basin between July and October 2021, where the collapse of the remaining lake basin necessitated the definition of a new upper-shoreline boundary. We find a moderate correlation between upper- and lower-shoreline boundary change (r2=0.63, p=0.009) and a slope of 0.59 in the linear fit. Thus, both boundaries exhibit comparable amounts of change, but the lower shoreline (the interface between the backshore, beach, and ocean) is more dynamic.

Analyses of both the geolocated photon data and the elevation profiles derived from ICESat-2 provide valuable insight into shoreline change. ICESat-2's geolocated photon product resolves coastal topography with a high level of detail, capturing abrupt elevation changes over horizontal distances of 5 m or less (e.g., Fig. b). We observe process-scale features such as toppled blocks and surface ponds in snow-free ICESat-2 profiles (Fig. ). We find that the simplified elevation profiles derived via SlideRule can adequately capture the shoreline for the assessment of shoreline evolution, although small-scale features such as the toppled block on ground track 3r (Fig. a) and surface ponding on ground track 2r (Fig. b) are lost. ICESat-2 profiles allow for the identification of multiple shoreline boundaries, which can be compared against imagery-derived shoreline positions (Figs. , a). Furthermore, analysis of morphologic parameters such as backshore height and backshore slope can provide insight into specific erosional and accretional processes, such as the collapse of the small drained lake and the presence of toppled blocks at steep coastal bluffs.

The highest RMS misfits between the SlideRule-derived elevation profiles and underlying photons occur in areas with steep slopes and abrupt elevation changes (Figs. , , ). This is particularly apparent for ground track 3r (Fig. ), where the upper-shoreline boundary inferred from SlideRule often appears 2 to 5 m southward of the boundary that would be visually inferred from ATL03. The 10 m section length we used in SlideRule is likely too coarse to adequately capture abrupt elevation changes, but a shorter segment length would have relied on fewer photons per segment and would therefore have been more sensitive to variations in along-track photon density. An adaptive approach based on local topography and photon density may be able to capture steep and complex features more accurately.

Given the seasonal revisit time of ICESat-2 over our study region, the high frequency of cloud cover in the summer and fall, and the low reflectivity of the snow-free tundra, the majority of usable ICESat-2 data occurs in winter and spring, when snow and sea ice are present. Snow cover will impact ICESat-2-based shoreline position estimates, particularly the location of the lower-shoreline boundary, which will in turn impact our slope estimates. Although we expect the magnitude of snow-cover-induced changes to be small relative to the large retreat events observed at ground track 3r (the coastal bluff) and ground track 2r (the small drained lake), they may be the dominant source of change in areas with lower rates of change, such as ground track 1r (the large breached lake). Future work comparing snow-on ICESat-2 tracks to snow-off ICESat-2 tracks and snow-off digital elevation models could help quantify snow distribution at the shoreline and its impact on shoreline position estimates. There is also up to 15 m of horizontal offset between repeat ICESat-2 profiles (Fig. a, c, e), meaning that differences in shoreline position may in part be due to different sampling locations.