Upland forest retreat lags behind sea‐level rise in the mid‐Atlantic coast

Ghost forests consisting of dead trees adjacent to marshes are striking indicators of climate change, and marsh migration into retreating coastal forests is a primary mechanism for marsh survival in the face of global sea‐level rise. Models of coastal transgression typically assume inundation of a static topography and instantaneous conversion of forest to marsh with rising seas. In contrast, here we use four decades of satellite observations to show that many low‐elevation forests along the US mid‐Atlantic coast have survived despite undergoing relative sea‐level rise rates (RSLRR) that are among the fastest on Earth. Lateral forest retreat rates were strongly mediated by topography and seawater salinity, but not directly explained by spatial variability in RSLRR, climate, or disturbance. The elevation of coastal tree lines shifted upslope at rates correlated with, but far less than, contemporary RSLRR. Together, these findings suggest a multi‐decadal lag between RSLRR and land conversion that implies coastal ecosystem resistance. Predictions based on instantaneous conversion of uplands to wetlands may therefore overestimate future land conversion in ways that challenge the timing of greenhouse gas fluxes and marsh creation, but also imply that the full effects of historical sea‐level rise have yet to be realized.

. For example, recent studies based on modeled tidal datums predict that a 1.0-1.5 m mean global sea-level rise will translate into hundreds of thousands of hectares of upland forests replaced by salt marshes across the conterminous United States within this century (Osland et al., 2022;Warnell et al., 2022).The resulting loss of wood production and stimulation of methane emissions contribute to a predicted net increase in the global warming potential of coastal ecosystems over large regions of the US coast (Baustian et al., 2023;Warnell et al., 2022).
However, it is unclear to what extent the predicted magnitude of forest loss will be realized, as multiple lines of evidence suggest that coastal forest retreat may not be synchronized with rising seas (Chen & Kirwan, 2022b;Schieder & Kirwan, 2019) and that other factors also play a role in modulating fine-scale patterns of coastal tree line dynamics (Fagherazzi et al., 2019;Poulter et al., 2009).For example, site-specific stratigraphic reconstructions over the past 2000 years suggest periods of time where upland conversion was slower (Schieder & Kirwan, 2019) or faster (Miller et al., 2021) than concurrent RSLRR.These reconstructions are consistent with field observations of mature trees that persist for decades under chronic flooding and salt stress (Field et al., 2016;Kirwan & Gedan, 2019;Poulter, Christensen, et al., 2008;Williams et al., 1999), and the paradigm that storms are necessary to facilitate forest retreat (Fagherazzi et al., 2019).Topography, disturbance, and biotic interactions are all factors previously invoked to interpret site-scale patterns of coastal tree line dynamics in response to rising seas (Chen & Kirwan, 2022a;Field et al., 2016;McDowell et al., 2022;Molino et al., 2022;Poulter et al., 2009;Ross et al., 1994;Schieder et al., 2018;Smith, 2013;Williams et al., 1998Williams et al., , 1999)).Nonetheless, it is largely unknown how rates of coastal tree line retreat will manifest across broad spatial scales that stretch wide gradients of environmental context (e.g., salinity, disturbance, climate, and tidal regime).
Here, we leverage extensive Landsat satellite images between 1984 and 2020 to explore landscape-scale patterns, including rates and drivers of both lateral and vertical coastal tree line retreat along the US mid-Atlantic coast (Figure 1), a global hotspot for accelerated sea-level rise (Sallenger et al., 2012).In contrast to static inundation models that assume instantaneous coastal ecosystem shifts with sea-level rise, we find that only a fraction of upland forests (~40% within elevations of 0 and 2 m) retreated inland between 1984 and 2020.Moreover, the rate of vertical forest retreat is merely half of contemporary RSLRR, pointing to a pronounced lag between sealevel rise and upland conversion that suggests surprising ecosystem resistance.

| Regional context
We studied coastal forest migration in response to sea-level rise across the US mid-Atlantic coast (~12,000 km 2 ; Virginia, Maryland, Delaware and New Jersey).This geophysically variable region encompasses the largest US estuary, the Chesapeake Bay, and the adjacent Delaware Bay (Figure 1).Soil texture is relatively homogenous in the region, largely characterized as silt and silt loam (Walkinshaw et al., 2022).The region was selected because it is a known global sea-level rise hotspot (Sallenger et al., 2012), and spans strong gradients in salinity, topography, and rates of relative sea-level rise rate (RSLRR) (Table 1).Moreover, the relatively rural coast of the US mid-Atlantic represents a great opportunity to observe how sea-level-driven landscape reorganization proceeds across broad scales with minimal anthropogenic obstacles (Molino et al., 2022).Indeed, massive marsh encroachment and forest mortality have been documented across the region over recent decades concurrent with increasing flooding and saltwater intrusion towards uplands (Schieder et al., 2018;Smith, 2013;White et al., 2022).
To capture the full spatial extent of sea-level rise impact (Chen & Kirwan, 2022a), we included areas between 0 and 5 m above sea level (relative to NAVD88, the mean sea level in the region) (Figure 1).
The elevation range extends from permanently flooded lowlands to coastal uplands free from seawater flooding (Pekel et al., 2016).All elevation data refer to the high precision Coastal National Elevation Database (CoNED) (Danielson et al., 2018)

| Landcover mapping
We mapped regional landcover using Landsat satellite images acquired around 1984 and 2020, and estimated lateral and vertical patterns of coastal forest retreat between 1984 and 2020 (Tables S1 and S2).We did not include an intermediate time-step after taking into account the relatively slow processes of coastal forest retreat (Chen & Kirwan, 2022b;Schieder & Kirwan, 2019) combined with comparatively coarse spatial resolution of Landsat images.The extended 36-year (1984-2020) time span allowed us improved confidence in change detection (Chen & Kirwan, 2022a).We generated two landcover maps (one in 1984 and one in 2020 that include each of six classes: Marsh, Forest, Farmland, Urban area, Water and Sandbar, Table S1) with special focus on the marsh-forest boundary using the classification algorithm we developed earlier for accurate mapping of retreating forest in coastal landscape (Chen & Kirwan, 2022b).It is worth mentioning that "Forest" studied here refers specifically to upland forest (Table S1), and it does not include forested wetlands (i.e., freshwater swamps).We mapped all upland forests across our study region, which stretch from higher elevations entirely devoid of seawater inundation to low-lying, salt-intruded forests (i.e., transition forest) (Chen & Kirwan, 2022a, 2023a) at the coastal transgression front where forest species are dominated by relatively salt-tolerant evergreen trees like Loblolly pine (Pinus taeda) and red cedar (Juniperus virginiana) (Brinson et al., 1995;Kirwan et al., 2007).Both maps were created at 30 m resolution using random forest classifier in R (v. 4.1.1,packages of "caret" and "randomForest").A detailed description of our coastal mapping approach can be found in Chen and Kirwan (2022b).
Briefly, we complemented the multispectral Landsat satellite images acquired from contrasting seasons in the year of mapping with a set of phenology metrics derived from the annual Landsat NDVI time-series for optimal differentiation between encroaching marsh and retreating forest at the upland-wetland transition (Chen & Kirwan, 2022b).For each mapping, we trained the classifier with 50% of reference sites collected earlier for different landcover types across the mid-Atlantic region (Chen & Kirwan, 2022a), and used the remaining sites for validation.All reference sites (~30,000) were identified according to field campaign, drone images, or highresolution aerial images acquired in 1982-1986 (for mapping in 1984) and in 2018-2020 (for mapping in 2020) (Chen & Kirwan, 2022a), and the sites were divided randomly by landcover type in the ratio of 1:1 for training and validation.
The resulting maps were processed further for enhanced accuracy following similar post-processing steps as addressed in Chen and Kirwan (2022a).First, we assigned all areas where flooding frequency is identified by Global Surface Water dataset (1984-2020) (Pekel et al., 2016) as greater than 95% to water.Next, areas of potential misclassification of marshes were identified and removed according to the rules of flooding frequency <5% and elevation greater than 2.5 m (upper tidal range of mid-Atlantic (Danielson et al., 2018)).
Finally, we manually digitized all areas (~5% of the study region) precluded from auto-classification due to contamination by cloud/ cloud-shadow in the input Landsat images using high-resolution aerial images following the approach by Chen, Lara, et al. (2021).The final landcover maps (Chen & Kirwan, 2023a) were validated extensively across the region, which achieved an overall classification accuracy of 92.4% (Kappa coefficient = 0.91) and 94.5% (Kappa coefficient = 0.93) for the map in 1984 and 2020, respectively (Table S2).

| Coastal tree line and coastal forest retreat
Using the landcover maps generated above, we then extracted coastal tree lines following the approach of Chen and Kirwan (2022b), and computed lateral/vertical rates of coastal forest retreat between 1984 and 2020 (Chen & Kirwan, 2023b).
Coastal tree lines in this study refer specifically to the marsh-forest boundary (or in less frequent occasions where coastal forests meet F I G U R E 1 Sea-level rise along the mid-Atlantic coast of North America.(a) Yellow circles indicate the locations of all tide gauges (n = 19) across the study region, where long-term information in sea level is available (1984-present).The size of the circles is proportional to the relative sea-level rise rate between 1984 and 2020.Elevation is relative to NAVD88 (mean sea level in the region).(b) Regional sea-level rise trend averaged across all tide gauges in the region.Data shown as mean ± 1 standard deviation.The mean linear regression trendline is bounded by the 95% confidence interval.(c) Drone image showing retreating forest in the Blackwater National Wildlife Refuge taken in 2020 (Image credit: Tyler Messerschmidt).Map lines delineate study areas and do not necessarily depict accepted national boundaries.
seawater or sandy shores as commonly seen on barrier islands) (Chen & Kirwan, 2022a, 2022b;Schieder et al., 2018;Schieder & Kirwan, 2019), and they do not include tree lines where forests border human land use like farmland or urban areas, which were removed prior to analysis.To understand the spatial distribution and temporal changes in coastal tree lines along topography, we systematically sampled the elevation and slope data along all coastal tree lines every 100 m ('Generate Points Along Line' tool in ArcGIS v10.7) from the CoNED DEM (Danielson et al., 2018) (Figure 2).In general, forest losses in P1 and P2 are usually small in size, collectively accounting for less than 10% of regional forest loss, with the remaining 90% areas of forest loss roughly equally represented by P3 and P4.
Next, we generated transects running through the polygons to represent paths of forest retreat (Figure 2).For forest loss in patterns of P3 and P4, tree lines are present in both years to indicate directional retreat from 1984 to 2020.For these areas, we placed points along all polygon boundaries at regular distance (100 m), from where we created perpendicular lines ('Create Perpendicular Lines', ArcGIS v10.7) to intersect the opposite tree line (Figure 2).Only those connecting paired tree lines were selected as a retreat path, the intersection with the tree line in 1984 was determined as the start of the path, and the intersection with the tree line in 2020 was the end of the path.Unlike P3 or P4 polygons of paired tree lines, the P1 and P2 polygons have a single tree line, present either in 1984 or in 2020.For each of these polygons, we generated a theoretical start (P1) or end (P2) point according to the CoNED DEM to direct the path of forest retreat.To be specific, the start point of the P1 polygon was identified as the location that has the lowest elevation within the polygon, whereas the end point of the P2 polygon referred to the location of the highest elevation within the polygon.In the same way, we generated points along boundaries of P1 and P2 polygons every 100 m, and connected these points with the start or the end point to represent directional change of forest from 1984 to 2020 (Figure 2).Finally, we computed the length of each path to represent the magnitude of lateral forest retreat, and estimated the elevation difference between the start and the end of the path to represent the TA B L E 1 Sea-level rise in the US mid-Atlantic region.Tessellation," ArcGIS v10.7) to generate regional forest retreat maps at a spatial resolution of 0.075 km 2 (Hexagon grid, side length of 170 m).The value of each grid is calculated as the mean of all rate samples inside the grid, and grids outside polygons are assigned to a value of 0 as they correspond to areas of no forest change (Figure 2).

| Data analysis
We analyzed regional forest retreat rates using multiple linear regression models to identify key environmental drivers for the dynamic patterns of coastal forest change (Table 2).To explore whether the environmental controls differ between lateral and vertical forest retreat, we generated separate models for lateral retreat rate (m year −1 ) and vertical retreat rate (mm year −1 ).We fitted each model with the same set of candidate variables that includes observed RSLRR and 24 other predictors identified from literature as influential for coastal forest retreat (

| Coastal landscape reorganization
We find that 1320.8 km 2 of the areas between 0 and 5 m NAVD88 underwent landcover change from 1984 to 2020, mostly (733 km 2 ) driven by human activity (e.g., deforestation), and to a lesser degree (587 km 2 ) by sea-level rise impacts (e.g., forest transition to marsh, Figure 3a).However, closer examination of patterns of landcover change reveals that human-induced changes largely (67.3%)occurred at elevations greater than 2 m, whereas 96% of sea-level induced changes appeared at elevations between 0 and 2 m elevations (Figure 3b).Thus, we restricted all further analysis to 0-2 m above sea level.
Within elevations between 0 and 2 m, sea-level rise impacts outpaced human activity as the predominant force responsible for over 70% of all coastal landcover change, expressed primarily as forest conversion to marsh (223.6 km 2 ) and marsh transition to water (171.9km 2 ), followed by farmland loss to marsh and water (50.1 km 2 ) (Figure 3b).In particular, sea-level-driven landward marsh migration led to the creation of 257.3 km 2 of new marsh, which overcompensated marsh loss at coastal margins and resulted in an overall increase of marsh area of 78.8 km 2 from 1984 to 2020 (Figure 3).In contrast, 235.7 km 2 of forests were deforested by rising seas from 1984 to 2020.In spite of reforestation from abandoned farmland (180.8 km 2 ), the total area of coastal forest decreased by 88.7 km 2 (Figure 3b).

| Lateral forest retreat
We detect that coastal forest retreat was widespread across the mid-Atlantic region (Figure 4; Figure S1), with an average lateral retreat rate of 0.67 ± 0.01 m year −1 (mean ± SE) between 0 and 2 m elevations (Figure 4a).However, not all forested areas retreated with rising seas, and the average forest retreat rate exhibited declining trends with elevation (r = −.70,p < .001,Figures 4 and 5).Overall, 41% of coastal forests retreated, whilst 56% of the forests remain unchanged with the remaining 3% showing tree line advance (primarily in the Virginia Coastal Reserve due to natural barrier island rollover (Deaton et al., 2017; Figure 4c).The proportion of retreating forest decreased rapidly from ~70% at elevations below 0.3 m above sea level to ~10% at elevations of 1.9-2.0m (Figure 4c).Accordingly, lateral forest retreat rates declined from a maximal rate of 1.93 ± 0.09 m year −1 at elevations of 0.2-0.3m to 0.12 ± 0.01 m year −1 at elevations of 1.9-2.0 m.To explore the linkage between spatially variable lateral forest retreat rate and relative sea-level rise rates (RSLRR, n = 19) observed in local watersheds, we averaged the spatially explicit map (Figure 4a) by watershed to generate a watershed-scale forest retreat map (Figure 4b).Interestingly, we do not detect a statistical relationship between lateral forest retreat rate and RSLRR (p = .88,Figure 4d).The lack of correlation is confirmed by our multiple linear regression model (R 2 adj = .69,p < .001, Figure 6a), suggesting that lateral forest retreat rate is strongly and positively influenced by sea surface salinity (p < .01),and negatively influenced by elevation (p < .05)and topographical slope (p < .05).Whereas salinity emerges as the most influential variable responsible for 38.5% of the variance, topography-the combination of elevation and slope-accounts for the majority of overall variance (55.2%) in lateral forest retreat (Figure 6a).

| Vertical forest retreat
Whilst lateral forest retreat is conceptually simple, the rate is heavily influenced by terrain attributes (Figures 4-6).To better isolate the influence of sea level on coastal forest retreat, we then created maps of vertical forest retreat (i.e., the upward migration of forest along elevation) between 1984 and 2020 (Figure 7).Similar to patterns of lateral forest retreat, vertical forest retreat rates varied widely across the mid-Atlantic region (Figure 7a), and declined with increasing elevation (r = −.48,p < .05).Our multiple linear regression model suggests that RSLRR is the overriding variable (R 2 adj = .58,p < .01)responsible for 43.1% of the variance in vertical forest retreat (Figure 6b).Although slope (r = −.51,p < .05)and salinity (r = .46,p < .05)remain significant controls shaping the observed patterns of vertical forest retreat, they are secondary to RSLRR, explaining 21.6% and 13.3% of the variance, respectively (Figure 6b).
Concurrent with rising sea level, the average elevation of the coastal tree line shifted upslope from 0.60 ± 0.01 m (n = 443,145) above sea level in 1984 to 0.69 ± 0.01 m (n = 468,502) above sea level in 2020 (Figure 7c).Notably, the estimated regional mean vertical forest retreat rate of 2.71 ± 0.003 mm year −1 (averaged across all forested areas between 0 and 2 m elevations, Figure 7a) is less than the regional RSLRR of 5.48 ± 0.17 mm year −1 (n = 19, Table 1).The deficit between forest retreat and sea-level rise is reaffirmed by the watershed-scale results (Figure 7b).We find that although the rate of vertical forest retreat is strongly and positively correlated with RSLRR (r = .55,p < .05), the vertical forest retreat rate is merely 48.5% ± 2.6% (n = 19, range of 34.6%-76.4%) of RSLRR (Figure 7d).For instance, as RSLRR increased from 4.2 mm year −1 in New Castle, Delaware to 6.9 mm year −1 in Yorktown, Virginia, the corresponding vertical forest retreat rate increased from only 2.2 to 3.8 mm year −1 (Figure 7b).

| Patterns and drivers of coastal forest retreat
Sea-level rise caused massive forest loss along the mid-Atlantic coast from 1984 to 2020.Notably, landward forest retreat appeared up to 10 km away from the coastline, facilitated by interconnected drainage networks.This finding complements earlier observations in coastal North Carolina (Poulter, Goodall, et al., 2008;Smart et al.,  suggesting that legacy wetland management practices may serve as effective corridors for interior salinization.Nonetheless, the very condition detrimental to forest survival is conducive to inland marsh migration, which outpaced seaward marsh loss and led to an expansion of regional marsh area by 2%.Topographic and anthropogenic barriers are well known to limit marsh migration (Enwright et al., 2016;Molino et al., 2022).Interestingly, we found that with sea-level rise from 1984 to 2020, the slope at the marsh-forest boundary increased from 0.8 to 1.1, indicating that forests are retreating into progressively higher topographic slopes, which may slow marsh transgression in the future.
In spite of widespread forest loss over past decades, not all forests retreated with rising seas.In fact, only ~40% of coastal forests migrated inland between elevations of 0-2 m.Notably, stable tree lines commonly occur in steeply sloped areas even at elevations in which tree lines would have otherwise retreated.While it is intuitive that a gentle slope is favorable to forest migration in the lateral dimension (Chen & Kirwan, 2022a;Kirwan et al., 2016;Schieder et al., 2018;Smith, 2013), previous site-based measurements suggest contrasting relationships between topographical slope and vertical forest retreat (Fagherazzi et al., 2019;Field et al., 2016;Wasson et al., 2013).By synthesizing data across broad spatial scales, we show that both lateral and vertical forest retreat are strongly, negatively correlated with slope, highlighting steep terrain as a key asset in mediating sea-level rise impacts on adjacent uplands.
We argue that steep slopes may favor forest persistence in several ways.Aside from posing direct physical obstacles for marsh encroachment (Kirwan et al., 2016;Smith, 2013), steeper slopes generally increase the drainage area for forests downslope (Hawthorne & Miniat, 2018).Thus, forests abutting steep slopes likely receive freshwater subsidies to temper saltwater intrusion.Moreover, steep slopes minimize the distance that tree roots must extend in the landward direction to reach freshwater (Messerschmidt et al., 2021).Finally, increasing slopes also tend to shorten the duration of tidal flooding and enhance soil drainage (Hussein & Rabenhorst, 2001a, 2001b), which lessens salinization and waterlogging conditions.
Previous work suggests that increases in salinity and/or soil saturation are the primary drivers of coastal forest mortality although their effects are difficult to distinguish (McDowell et al., 2022;Smith & Kirwan, 2021).Both hypoxia and salinity are hypothesized to drive similar mechanisms of plant mortality, resulting in hydraulic failure and carbon starvation (Krauss & Duberstein, 2010;McDowell et al., 2022).The range of lateral forest retreat rates that we observed across watersheds of the mid-Atlantic coast offers empirical support to both hypotheses (Figure 6), and it also indicates that coastal topography may interact with these processes (hypoxia, salinization) to dynamically modify the impacts of sea-level rise on coastal forest survivorship.(Danielson et al., 2018).
Interestingly, we find no relationship between rates of forest retreat and patterns of climate change or disturbance (i.e., storms), both of which are known to influence tree growth and mortality (Chen & Kirwan, 2022a;McDowell et al., 2022;Ury et al., 2021).Prior dendrochronological analyses on common coastal forest species (Juniperus virginiana) suggest that progressive increases in sea level suppress the impacts of climate, while strengthening the impact of tidal flooding on forest growth (Hall et al., 2022).This phenomenon potentially explains why patterns of forest retreat are not directly linked to climate, even though a warmer and wetter climate boosts forest biomass at higher elevations (Chen & Kirwan, 2022a).Similarly, although disturbance has long been regarded as important in shaping forest retreat (Fagherazzi  et al., 2019;Schieder & Kirwan, 2019;Ury et al., 2021), we find no correlations between spatially-variable forest retreat and the magnitude or duration of Hurricane Isabel, the largest storm to influence the mid-Atlantic coast since the 1950s.We suspect that stochastic processes like storms may be essential in explaining coastal forest dynamics at relatively short, local scales (Walters et al., 2021), but the impacts may average out over long, broad scales -a pattern also seen in the process of barrier island retreat (Mariotti & Hein, 2022).

| Lags with sea-level rise
Vertical forest retreat is strongly correlated with sea-level rise, yet the rate of vertical forest retreat is merely 35%-76% of RSLRR (Figure 7).This result, derived from multiple decades of modern satellite observation, is supported by paleoecological evidence from sediment cores in the region, which estimated that the magnitude of vertical forest retreat (~2 m) was approximately 60%-80% that of regional sea-level rise (~2.5-3.5 m) over past millennia (Schieder & Kirwan, 2019).Both forest retreat rates and RSLRR are accelerating in the mid-Atlantic region (Chen & Kirwan, 2022b;Ezer & Corlett, 2012;Schieder & Kirwan, 2019).However, the average vertical forest retreat rate we observed between 1984 and 2020 (2.7 mm year −1 ) most closely resembles the average RSLRR recorded between 1930 and 1950 (2.0-3.0 mm year −1 ) (Ezer & Corlett, 2012), implying that regional forest retreat lags behind sea-level rise by roughly half of a century.With the ever-growing power of Earth observation satellite, future studies that utilize higher spatial/temporal resolution images may help identify the precise lag and potential nonlinearities in the lag effects.
We hypothesize that a suite of internal and external mechanisms may be involved that buffer upland forests from the otherwise acute impacts of sea-level rise.For instance, greenhouse experiments reveal that tree species commonly found in coastal uplands (e.g., Pinus taeda, P. serotina) possess physiological traits allowing them to tolerate a range of flooding and low salinity conditions (Poulter, Christensen, et al., 2008;Williams et al., 1998).Recent study also indicates that coastal forests can actively adapt to rising seas through morphological plasticity, as reflected by the distribution of tree roots preferentially towards freshwater sources upslope (Messerschmidt et al., 2021).Moreover, forested wetlands in other regions accrete vertically through the accumulation of mineral sediment and organic matter (Craft, 2012;Noe et al., 2016), which may be amplified in our region by the expansion of Phragmites australis into transitioning forests (Langston et al., 2021).
Although forests intruded by seawater generally display reduced tree height and basal area as compared to intact forests (Krauss et al., 2009;Smith & Kirwan, 2021), remote-sensing observations and repeated field surveys suggest that many saltintruded forests did not show biomass loss over time (Chen & Kirwan, 2022a;White & Kaplan, 2021) and some even exhibited heightened growth vigor due to enhanced light availability near forest margin (Field et al., 2016).Other factors, such as biotic interactions encouraging seedling survival (Poulter et al., 2009), the effects of marsh migration on reducing saltwater intrusion landwards (Guimond & Michael, 2021), and the capacity of coastal forests to rapidly regenerate and resprout under variable salt stress (Walters et al., 2021;Williams et al., 1998) may confer additional strength for forest persistence.Thus, although upland forests may ultimately succumb to wetlands under excessive tidal flooding, the complete transition may take years to decades to fully realize.
Our finding of a lagged response between sea-level rise and forest retreat mirrors findings in an array of terrestrial and coastal ecosystems, where sizable spatiotemporal misalignment exists between ecosystem transition and climatic forcing (Rastetter et al., 2021).For instance, the upward shifts of forest fronts in many Arctic and high-mountain regions demonstrate decadal to centennial timescale lags with climate warming (Alexander et al., 2018;Chapin & Starfield, 1997;Rastetter et al., 2021).In coastal barrier islands, the rate of barrier retreat is out of equilibrium with contemporary sea-level rise rate, but rather reflects baseline rates of past centuries (Mariotti & Hein, 2022).Similarly, marsh accretion rates lag behind accelerating sea-level rise by around 20-30 years (Kirwan & Temmerman, 2009), and marshes may persist for decades to centuries even after threshold RSLRR's are exceeded (Törnqvist et al., 2021).
Our observations of multi-decadal lags between sea-level rise and coastal forest retreat are therefore consistent with observations from a variety of earth systems responding to various facets of climate change.Numerical models of marshes, barrier islands, and terrestrial forests typically include physiological or geomorphic processes that allow ecosystems to persist under climate change until certain thresholds are surpassed (Dial et al., 2022;Kirwan & Temmerman, 2009;Mariotti & Hein, 2022).Yet, models of sea-level driven ecosystem migration are in their infancy, and typically assume that marshes migrate into adjacent uplands as soon as tidal inundation occurs (Enwright et al., 2016;Molino et al., 2022;Osland et al., 2022;Warnell et al., 2022).Incorporating newly emerging processes into numerical models are critical to predictions of coastal vulnerability and feedbacks with climate (Ward et al., 2020).In the meantime, our finding of a multi-decadal lag suggests that existing predictions based on static inundation may overestimate land conversion (Kirwan & Gedan, 2019;Osland et al., 2022), greenhouse gas emissions (Warnell et al., 2022), and marsh formation (Schuerch et al., 2018) during a given time period, but also suggests that the effects of historical sea-level rise have yet to be fully realized.
We differenced the landcover maps in 1984 and 2020 to identify areas of forest change, and then estimated rates of lateral and vertical forest retreat based on unique patterns of forest boundary change.The step-by-step methodology is illustrated in Figure 2, modified from the framework in Chen and Kirwan (2022b) to quantify both lateral and vertical forest retreat.In brief, there are four patterns of forest loss depending on coastal tree line configuration: Interior loss (P1: emerging forest loss, tree line present only in 2020), Entire loss (P2: complete patch loss, tree line present only in 1984), Linear retreat (P3: parallel retreat with conjoint tree lines in 1984 and 2020), and Radial retreat (P4: concentric retreat with disjoint tree lines in 1984 and 2020) (Figure 2).All areas of forest loss were converted to smoothed polygons ('Smooth Polygon' in ArcGIS v10.7) with the boundaries classified either as tree line in 1984 or in 2020.
magnitude of vertical forest retreat.We then divided the magnitude of lateral/vertical forest retreat by the years between 1984 and 2020 to calculate the rate of lateral/vertical forest retreat.To allow explicit representation of forest retreat pattern across the study region, we sampled forest retreat rate every 100 m along each path across all areas of forest loss, and rasterized the results ("Generate Patterns of coastal landcover change in the US mid-Atlantic region from 1984 to 2020.(a) Landcover change for all areas between 0 and 5 m above sea level.(b) Landcover change for all areas between 0 and 2 m above sea level.Alluvial plots illustrate the direction and magnitude of changes between landcover types.Numbers given indicate the percent cover of each landcover type.
2020; Ury et al., 2021) and the Gulf of Mexico (Raabe & Stumpf, 2016), F I G U R E 4 Lateral forest retreat from 1984 to 2020 across the mid-Atlantic region.(a) Spatially-explicit map of lateral forest retreat rate (resolution 0.075 km 2 ).Positive values refer to forest retreat, and negative values represent forest advance.The white box outlines the Blackwater National Wildlife Refuge, highlighted in Figure 5.(b) Watershed-scale lateral forest retreat rate (HUC10 units, NHDPlus (McKay et al., 2019)).White circles refer to relative sea-level rise rate (RSLRR) recorded by long-term tide gauges in the region.(c) Histogram showing patterns of coastal forest dynamics along elevation.(d) No statistical relationship between lateral forest retreat and RSLRR.Data shown as mean ± 1 standard deviation.Map lines delineate study areas and do not necessarily depict accepted national boundaries.

F
Regional subset highlighting dynamic patterns of coastal forest retreat in the Blackwater National Wildlife Refuge.Highresolution (~1.0 m) aerial photographs in 1984 (a) and 2020 (b) demonstrate variable patterns of landward marsh migration and coastal tree line retreat along gradients in elevation (c) and slope (d).The landcover maps in 1984 (e) and 2020 (f) were used to create the spatially-explicit maps of lateral (g) and vertical (h) forest retreat rate.The elevation and slope data refer to the CoNED DEM

F
Vertical forest retreat from 1984 to 2020 across the mid-Atlantic region.(a) Spatially-explicit map of vertical forest retreat rate (resolution 0.075 km 2 ).White box outlines the Blackwater National Wildlife Refuge, highlighted in Figure 5.(b) Watershed-scale vertical forest retreat rate (HUC10 units, NHDPlus (McKay et al., 2019)).White circles refer to relative sea-level rise rate (RSLRR) recorded by longterm tide gauges in the region.(c) Elevation of coastal tree line shifted upslope from 1984 to 2020.The inserted panel shows the boxplot of coastal tree line elevations, where the left and right edges of the box respectively correspond to the first and third quartiles, the center line refers to the median, the white point corresponds to the mean, and the whiskers represent data within 1.5× the interquartile range.(d) Strong positive correlation between vertical forest retreat rate and RSLRR.The dotted 1 to 1 line indicates where vertical forest retreat rate equals RSLRR.The mean linear regression trendline (solid line) is bounded by the 95% confidence interval.Data shown as mean ± 1 standard deviation.Map lines delineate study areas and do not necessarily depict accepted national boundaries.F I G U R E 6Multiple linear regression models for patterns of coastal forest retreat.Factors responsible for spatially-variable patterns of horizontal forest retreat rate (a) and vertical forest retreat rate (b).The mean linear regression trendline is bounded by the 95% confidence interval.The inserted pie charts present the relative contribution of each variable retained in the model to overall variance, where variables in red represent positive correlation with the response variable and variables in blue suggest negative correlation.RSLRR is short for relative sea-level rise rate, and salinity refers to sea surface salinity.
at 1 m resolution.All sea-level rise data are observed by long-term tidal gauges (Table 1), accessed from the NOAA Center for Operational Oceanographic Products and Services (Center for Operational Oceanographic Products and Services, 2023).

Table 2
Candidate predictors for modeling lateral and vertical forest retreat rate.The column "Reference" refers to prior literatures suggesting relationships between coastal forest retreat and the variables selected.
(McKay et al., 2019)2019;Ury et al., 2021;White et al., 2022)& Stumpf, 2016;Smart et al., 2020)e variables, such as RSLRR, flooding frequency(Chen & Kirwan, 2022a;Schieder & Kirwan, 2019;White & Kaplan, 2021); (4) landscape metrics, like mean forest patch size, proximity to drainage channels(Chen, Hu, & Lara, 2021;Poulter, Goodall, et al., 2008;Raabe & Stumpf, 2016;Smart et al., 2020);F I G U R E 2Flowchart for quantifying coastal forest retreat.The approach was modified from the framework developed inChen and Kirwan (2022b).See Section 2 for detailed description of the step by step procedure.and(5)disturbancevariables,includingobservedstormfrequency, and modeled inundation depth and duration of Hurricane Isabel(Fagherazzi et al., 2019;Ury et al., 2021;White et al., 2022).To be consistent with sea-level rise observation, all data (except for static variables) were processed to the same time span(1984-  2020)and watershed-scale (HUC10unit) as defined by the National Hydrography Dataset Plus(McKay et al., 2019), where the variable value of a certain watershed was computed as the mean of all forested areas within the watershed.We fitted the model with all candidate variables, and eliminated unimportant, cross-dependent/correlative variables in a stepwise manner to achieve a single reduced model that contains only significant predictors for the response variable TA B L E 2