VedgeSat: An automated, open‐source toolkit for coastal change monitoring using satellite‐derived vegetation edges

Public satellite platforms offer regular observations for global coastal monitoring and climate change risk management strategies. Unfortunately, shoreline positions derived from satellite imagery, representing changes in intertidal topography, are noisy and subject to tidal bias that requires correction. The seaward‐most vegetation boundary reflects a change indicator which shifts on event–decadal timescales, and informs coastal practitioners of storm damage, sediment availability and coastal landform health. We present and validate a new open‐source tool VedgeSat for identifying vegetation edges (VEs) from high (3 m) and moderate (10–30 m) resolution satellite imagery. The methodology is based on the CoastSat toolkit, with streamlined image processing using cloud‐based data management via Google Earth Engine. Images are classified using a newly trained vegetation‐specific neural network, and VEs are extracted at subpixel level using dynamic Weighted Peaks thresholding. We performed validation against ground surveys and manual digitisation of aerial imagery across eroding and accreting open coasts and estuarine environments at a site in Scotland. Smaller‐than‐pixel vegetation boundary detection was achieved across 83% of Sentinel‐2 imagery (Root Mean Square Error of 9.3 m). An overall RMSE of 19.0 m was achieved across Landsat 5 & 8, Sentinel‐2 and PlanetScope images. Performance varied by coastal geomorphology, with highest accuracies across sandy open coasts owing to high spectral contrast and less false positives from intertidal vegetation. The VedgeSat tool can be readily applied in tandem with waterlines near‐globally, to support adaptation decisions with historic coastal trends across the whole shoreface, even in normally data‐scarce areas.


| INTRODUCTION
There is a fundamental need to rapidly, repeatedly and reliably quantify coastal land change.Around 1.3% of the global population are currently impacted by 100-year coastal flood events (Muis et al., 2016).These populations will be increasingly exposed to ever-more intense interacting hazards, including storm wave impact and overtopping, and coastal erosion (IPCC, 2019;Ranasinghe, 2016).Increasing exposure with no adaptation by the end of the century could lead to damage equal to 10% of global GDP (Hinkel et al., 2014).Transport and utility networks, urban and rural infrastructure, industry and communities are all at increasing risk, with levels of potential impact being socially and geographically unequal (IPCC, 2022).Areas most at risk and communities most at disadvantage must be identified to ensure support and funding for coastal change adaptation, and resilience is directed effectively (Cutter et al., 2009;Dunkley et al., 2021).Yet robust, cost-effective tools to enable accurate, repeatable monitoring of coastal change (and support adaptation to these changes) are still largely underdeveloped.Regular and repeated monitoring and analysis of coastal change is required at local up to global spatial scales that is also capable of measuring short-term event-season to annual-decadal timescales (Vitousek et al.,2023).
On soft, sandy coasts, the shoreline position (the vector boundary between land and sea) is a key coastal change indicator (Boak & Turner, 2005).Which boundary is best used to contour and measure change from is a subject of ongoing debate, with the wet-dry contour at different tidal stages (mean low/high water, mean sea level, lowest/ highest astronomic tide), cliff top/toe, seaward edge of vegetation, erosion scarp and lower beach crest all being applied in different contexts (Boak & Turner, 2005;Beuzen, 2019;McAllister et al., 2022).
However, these indicators evolve over a range of spatial and temporal scales, and their recording of other coastal processes means they also operate on varying scales of complexity (Cowell & Thom, 1994).
Ground-based and airborne techniques to measure coastal change indicators are labour-intensive, and spatial and temporal coverage is severely limited by costs, logistics and time constraints (Boak & Turner, 2005;Morton et al., 1993;Splinter et al., 2018).In the UK, these limitations have led to a fragmented picture of coastal change and an unsystematic approach on how to address environmental changes and mitigate the impacts of coastal risks (Lazarus et al., 2021).The main problems for coastal change monitoring lie therefore in scalability, repeatability and applicability.
Solutions to these scaling problems can be found in the recent proliferation of satellites for Earth Observation (EO) at improved spatial and temporal resolution and with increasing public data accessibility (Finer et al.,2018).Thousands of multispectral and hyperspectral image platforms now capture the Earth's surface at low ( ≥ 30 m), moderate (3-30 m) and high ( ≤ 3 m) spatial resolutions (McCarthy et al., 2017).Between Copernicus' Sentinel collection and US Geological Survey's Landsat collection, members of the public can currently access near-global multispectral images taken of Earth at resolutions of 10-15 m, with revisit times of 5-8 days (European Space Agency, 2022;NASA, 2022).EO satellites provide great potential for monitoring of the coastal environment, with coverage large enough to capture entire regions and nations in one image and one instant (European Space Agency, 2022).Additionally, with the introduction of cloud-based data storage, sharing and processing for satellite imagery, scientists are now using Application Programming Interfaces (APIs) like Google Earth Engine (GEE) to drive efficient and automated analyses of Earth's surface without the limitations of local data management and computational resources (Gorelick et al., 2017;Yang et al., 2020).
A number of recent studies have developed approaches to automate the analysis of coastal change from EO-derived shorelines.Luijendijk et al. derived shorelines from Landsat imagery from 1984-2015 using a Normalised Difference Water Index (NDWI) (McFeeters, 1996), with edge detection based on Otsu thresholding (Otsu, 1979), to semi-automatically define the boundary between water and land (Luijendijk et al., 2018).These data have subsequently fed near-global, decades-long assessments of shoreline position change to infer trends in shoreface dynamics (Mentaschi et al., 2018;Vousdoukas et al., 2020).The similar toolbox CoastSat (Vos et al., 2019) rivals global-scale shoreline change timeseries, by offering flexibility in user requirements and accuracy improvement in using Sentinel-2 imagery versus Landsat 5 and 7, as well as a number of related tools to populate a coastal monitoring framework (Vos et al., 2020).Semi-automatic extraction of shoreline positions has allowed coastal scientists to piece together past and current trends and use these to help predict future change (Montaño et al.,2020).While the wet-dry shoreline position (hereby referred to as the waterline) is a powerful indicator of change representing shifts in intertidal topography, it retains an inherent bias due to instantaneous waterline positions depending on the tidal condition at the time an image was acquired.The intertidal slope further influences the land-sea contour, such that even minor tide height uncertainties between satellite images may affect waterline positions more than actual coastal topography changes (Camfield & Morang, 1996).This bias was addressed and reduced by Luijendijk et al. by generating annual average waterline positions, but not entirely eliminated due to the presence of longer-term tidal cycles (Luijendijk et al., 2018).Vos et al. introduced tidal correction into the CoastSat toolbox to adjust cross-shore waterline positions using the known tidal elevation at the time of image acquisition (Vos et al., 2020).However, this algorithm requires a datum-based tidal height timeseries which may be hard to obtain in areas with observation gaps.It also requires an estimate of beach slope from which to correct for instantaneous tide elevation, an estimate which may vary depending on the length and completeness of the waterline timeseries (Vos et al., 2020).Additionally on mesotidal and macrotidal beaches with low slope, errors in satellite-derived waterline positions due to swash-related spectral noise can reach up to 30 m (Castelle et al., 2021;Vos et al., 2023).While a number of limitations have been identified in using the waterline as a coastal change indicator, the popularity of satellite-derived coastal change metrics continues to increase, highlighting the value of rapid, cheap, standardised coastal monitoring (Payo et al., 2020;Turner et al.,2021).
The current research on coastal monitoring has shown that, to gain a deeper understanding of climate impacts on coasts, alternative coastal change indicators need to be extracted and explored (McAllister et al., 2022).In a recent benchmarking exercise for satellite-derived waterline algorithms, one of the key recommendations was to develop alternative shoreline change measures, to address positional errors related to tidal bias and noisy signals on mesotidal and macrotidal shores (Vos et al., 2023).In contrast to water-based shoreline position detection, the seaward edge of stable coastal vegetation, or vegetation edge (VE), offers an alternative shoreline proxy (Keijsers et al., 2015).Coastal VEs capture destructive change via flood and erosion events, and accretional recovery periods through the progradation of dunes.They do so with no bias or high variability in position associated with tidal stage, a key source of uncertainty in waterline metrics (Toure et al., 2019).Coastal vegetation is a biogeomorphological marker of topographic and hydrodynamic changes across the shoreface, as well as exercising influence on these processes through sediment trapping and wave attenuation (Koch et al., 2009;Durán & Moore, 2013).While there has been a proliferation of studies focused on satellite-derived shorelines, studies investigating the automatic extraction of coastal VEs from satellite imagery are few.VEs are normally collected using ground surveys or manual digitisation of aerial imagery (Coyne et al., 2012;Guy, 1999;Morton & Paine, 1985;Morton et al., 1994;Priest, 1999), or more recently using semi-automated techniques such as thresholding of band indices like the normalised difference vegetation index (NDVI) in multispectral imagery (Le Berre et al., 2005;Rahman et al., 2011;Trépanier et al., 2002).Pixel-based vegetation coverage has been explored more extensively than the extraction of vector-based edge features, which are needed to generate a standardised edge position for repeated coastal change measurements.The pixel-based methods used to date have used a large variety of spectral band indices and supervised or unsupervised machine learning approaches to classify and extract vegetation presence and health from satellite imagery (Xie et al., 2008;Zeng et al., 2020).Inland vegetation studies for forested habitats and agriculture health also dominate this literature, with little in the way of coastal edge or extent delineation.in a holistically nested edge detection approach which yields errors of < 6 m.While promising, the application of this method to VEs is relatively novel and only trained using commercial, high-resolution imagery (Rogers, 2021)

| Satellite imagery
VEs were derived from satellite imagery obtained from Landsat 5, Landsat 8, Sentinel-2 and PlanetScope platforms as detailed in Table 1.Top-Of-Atmosphere images from each platform were used to ensure comparability between the reflectance values of each platform, as opposed to using atmospherically corrected images that may differ in their correction approaches and ranges of values (Sterckx & Wolters, 2019).In order to perform the satellite validation exercise robustly, the satellite images were also transformed to the Ordnance Survey Great Britain (OSGB1936) coordinate reference system and georeferenced using GNSS ground control points gathered at visual markers with fixed locations across the test site (such as road markings and building edges).This georeferencing process is related to yet different from coregistering, which involves lining up each image to one another, to reduce 'wobble' from offsets in the overlap of images when extracting timeseries.In this context, coregistering can be thought of as improving the precision of satellite-derived products, and georeferencing as improving the accuracy.For this validation exercise, any coregistration errors are dealt with when each image is georeferenced to a common set of ground control points.The errors for each image associated with this process are listed in Table 1.

| Procedure for identifying VEs
The process of VE extraction is separated into four main sections: (i) identifying the user requirements; (ii) satellite image pre-processing involving cloud masking and quality assurance; (iii) supervised satellite image segmentation and classification; and (iv) thresholding and contouring of the classified features to refine the VE position extracted (Figure 1).Our procedures to identify VEs are an extension of the CoastSat infrastructure (Vos et al., 2019).We have streamlined the interaction with satellite images by using the geemap package for processing GEE imagery on cloud servers (Wu, 2020), reducing processing time and effort.Because the tool builds upon the existing CoastSat infrastructure, the capability to extract wet-dry boundaries using the original shoreline extraction functions is retained alongside the new VE routines.As VedgeSat has the capability to use PlanetScope imagery, the CoastSat.PlanetScope extension has also been incorporated to allow wet-dry boundary extraction alongside VEs at high resolution (Doherty et al., 2022).

| User requirements
The user is required to identify the desired satellite platforms, the date range & area of interest, and provide a reference shoreline vector and buffer distance.The available platforms are listed in Table 1; Landsat 7 and 9 imagery is also available within the tool, but was not included in the validation process due to a lack of overlap between satellite and validation data.The reference shoreline is a user-digitised T A B L E 1 Satellite platform information for images used in the validation exercise.vector representing the region over which to extract VEs, with an associated buffer distance to constrain a coastal swath.The reference shoreline is also used to create shore-normal transects, at an equal spacing defined by the user, and extending onshore and offshore at distances defined by the user.These transects were generated for the tool validation process at a regular alongshore spacing of 3 m (to reflect the smallest image resolution of 3 m) and cross-shore length of 100 m inland and 300 m offshore (to capture the full range of validation VE positions while minimising transect overlap).

| Image preprocessing
Each satellite image is separated out into individual spectral bands, including any quality assurance bands which hold cloud masking information.Panchromatic bands, which have a higher resolution than the visible light bands (red, green, blue; RGB), are used where available to pan-sharpen the RGB bands and enhance their resolution (Tu et al., 2001).For example, Landsat 7, 8 and 9 have RGB bands at 30 m resolution but can be pan-sharpened using their panchromatic bands to 15 m resolution.Landsat 5 has no panchromatic bands, so is downsampled using bilinear interpolation.Cloud masks exclude cloudcovered pixels from further analysis; some thinner cloud edges and shadows cast by clouds sometimes remain unmasked (Skakun et al., 2022).Previous algorithms used to denote wet-dry coastal boundaries were faced with the issue of peripheral unmasked clouds surrounding masked cloud objects being potentially mistaken for whitewater from breaking waves (Pardo-Pascual et al., 2018).An additional buffer of 30 m (representing one Landsat pixel prior to downsampling) is therefore also masked around any pixels classed as cloud, to reduce the potential for false identification of cloud edges as VEs or waterlines.

| Supervised object classification
The supervised classification of 'vegetation' and 'non-vegetation' objects is performed using a pretrained artificial neural network (ANN) model generated using a multilayer perceptron (MLP) classifier.
The correction factor L sits between 0 for environments with dense vegetation (and an output equal to the NDVI, Equation 1) and 1 for sparse vegetation, with 0.5 being the default value (Huete, 1988).
While higher L values can improve the sensitivity of the classifier to diffuse vegetation (Rhyma et al., 2020), this led to increased false positives of vegetation classification in model training tests, especially in areas with wet marsh sediment.The default L value of 0.5 was therefore used to maintain applicability across varied environments (Qi et al., 1994).

| Thresholding and contouring
Once the image has been segmented into two classes, the boundary between them is extracted using a hybrid ANN-threshold approach.
This is done in order to avoid oversimplification of the boundary between classified pixels and achieve subpixel VE extraction.The CoastSat algorithm used two-dimensional Otsu thresholding to statistically define an isoline (Vos et al., 2019).Otsu thresholding maximises the variance between probability density functions (PDFs) of pixel values for two classes.This method makes use of classes identified by the ANN but increases the accuracy of the output boundary by reducing spectral noise from nearby features.This hybrid ANN-thresholding approach has been applied to the extraction of boundaries in medical images with background noise, varying quality and the potential for undersegmentation or oversegmentation using an ANN (Gao et al., 2012;Wang et al., 2021).It was originally applied to satellite-derived shorelines to reduce the error introduced to the boundary by breaking wave whitewater (Hagenaars et al., 2018;Vos et al., 2019).
A recent study by Doherty et al. applied CoastSat to highresolution PlanetScope imagery, using a variation of band indices and thresholding techniques.When the outputs of different configurations of the shoreline extraction tool were compared with in-situ survey data, it was concluded that the thresholding technique that yield the highest accuracy output was a NIR minus Blue band index with Weighted Peaks thresholding (Doherty et al., 2022).The Weighted Peaks method was originally derived for extracting the wet-dry boundary in oblique-angle RGB imagery of coasts (Harley et al., 2019).The Weighted Peaks approach minimises the tendency of Otsu thresholding to bias towards the class with the larger variance, by implementing a bias towards the opposite class: In Equation ( 6 2022), with the former listing ω wet and ω dry as 0.3 and 0.7, respectively, but the latter identifying weights of 1 3 and 2 3 , respectively.In this application, Equation ( 6) is adapted to better capture sparse seaward vegetation with lower NDVI values: where I is NDVI (Equation 1).We tested the use of Weighted Peaks thresholding for the identification of VE by comparing a mapped VE to classified Landsat 5, Landsat 8 and Sentinel-2 imagery.The weights were incrementally adjusted from 0 to 1 as shown in Figure 2 and a threshold calculated according to Equation ( 7) that was used to contour NDVI.The root mean square error (RMSE) was calculated for each weighting combination and for each satellite image platform, based on cross-shore distances between these threshold contours and ground-truth VE surveys: In Equation ( 8), X and x refer to cross-shore positions of groundtruth and satellite-derived VEs extracted using the shore-normal transects introduced in Section 2.3, i is the transect index and n is the total number of transects.The lowest RMSE value for each platform corresponds to ω veg and ω nonveg values of 0.2 and 0.8, respectively (Figure 2a).The threshold calculated with these weights is then used with the Marching Squares algorithm (Cipolletti et al., 2012;Liu et al., 2017).The linear interpolation between pixels that the Marching Squares algorithm performs creates smoothed vector outputs at a subpixel level.et al., 2018;Vos et al., 2019).With coastal vegetation, the transition zone may instead reflect sparse, discontinuous or juvenile vegetation at the seaward edge, which are all relevant to coastal geomorphology processes at play.This transition zone is therefore mapped as pixels with NDVI values within this range of overlap (Figure 2b).The range is defined using the NDVI PDF of the 'vegetation' class, with the minimum value representing the first point on the PDF where the 'vegetation' class is statistically significant, and the maximum value representing NDVI the point above which pixels are more likely to be classed as 'vegetation' than 'non-vegetation'.These correspond, respectively, to the 0.5 th and 10 th percentiles of the 'vegetation' class PDF.

| Validation sites
In order to test our ability to map VEs from satellite imagery, we ide-  explored more thoroughly in each of these distinct geomorphic environments (see Figure 3).This subclassification also serves to provide a more focused examination of the satellite-derived vegetation edges, but it should be noted that not all validation surveys cover the whole test site.The full range of validation dates is therefore not reflected in each subsite.

| Outer estuarine vegetation edges with armoured and mixed coast
Greater uncertainty is seen along the west side of the peninsula, where the VE at high water is marked by agriculturally managed fields and rock armour, and pockets of intertidal saltmarsh and seaweeds at low tide which makes automatic identification of VEs from satellite imagery more challenging (Figure 7).The overall RMSE is 25.5 m which is skewed seaward, particularly by the errors of the Landsat 5 2007-04-17 and 2011-04-28 satellite-derived edges (Figure 7b).T A B L E 2 Metadata of the vegetation edges for ground-truthing, with corresponding satellite image metadata.

| Inner estuary results
Within the more varied vegetation setting of the Eden Estuary, the errors are high compared to the open coast.On both the south and north sides of the estuary, grass-and shrub-covered promontories sit surrounded by saltmarsh vegetation that fluctuates seasonally in coverage, and is only visible at lower tides (Figure 8a).As seen by the peaks in the positional error PDFs (Figure 8b), the offset of the earli-

| Diffuse and dynamic vegetation edges
Where rapid coastal change has taken place, for example at Tentsmuir (Figure 10a), validation is more challenging.Figure 10b   the average RMSE value was 41.9 m and highest R 2 was 0.21 across the five algorithms tested at the meso-macrotidal site (Vos et al., 2023).
As a tool, VedgeSat can and should be used in a nested approach with smaller resolution satellite imagery, after an initial coarser analysis of significant coastal change and risk to infrastructure.
Comparing the VE with the waterline as a coastal change indicator, the waterline can change position substantially within the intertidal zone even in the course of a single day (Cowell & Thom, 1994;Stockdon et al., 2006).To highlight this, waterlines were extracted alongside the VE shoreline for each image in the validation dataset using the original CoastSat model and the CoastSat.PlanetScope update.These cross-shore waterline positions were tidally corrected using data from the 2014-2016 Finite Element Solution (FES2014) global tide model (Lyard et al., 2020).At an eroding transect across Tentsmuir (Figure 3c), the satellite-derived VEs show an erosional long-term trend of À7.2 m/year across the validation period, when a linear regression is fitted to the full cross-shore VE timeseries.For the same transect, the satellite-derived waterline change rate is À16.1 m/year.For this steadily eroding section, the rate of change is over twice as fast across the intertidal zone as it is across the dune front.
Investigating these rates further reveals the R 2 of the regression line used to quantify an average rate of waterline change is 0.15, whereas R 2 is 0.58 when measuring change in VE.Therefore, while both coastal change metrics show an erosional trend, the confidence in the calculated rate of change is much lower for the waterline than for the VE, owing to the much higher variability in waterline positions over this time period (even after tidal correction).This intertidal variability and uncertainty is also seen in the eroding transect across West Sands (Figure 3d), where the VE change rate is À1.2 m/year with an R 2 of 0.59, but the waterline change rate is 13.0 m/year with an R 2 of 0.11.Assessing change in VE does not require tidal correction, and the potential for misclassification due to cloud cover is lower due to the spectral properties of vegetation being quite different to that of clouds, whitewater and wet sediment (Pardo-Pascual et al., 2018).
While deriving shoreline change from waterlines will remain incredibly useful for assessing beach processes, the waterline remains highly variable in time and space, particularly so on all but steep, microtidal beaches (Castelle et al., 2021;Vos et al., 2023).The reduced variability of results using satellite-derived VEs compared to waterlines demonstrates the potential of VE as a robust coastal change indicator.That being said, along coasts where vegetation is sparse or nonexistent, waterline positions will still offer an acute understanding of coastal change.This comparison demonstrates the value of using VE and waterline metrics in tandem, rather than as standalone indicators.
Combined indicator analyses can illuminate beach migration patterns across a greater breadth of timescales.Long-term, predictive changes in VE could provide an informative and complementary metric for assessing coastal change rates, and thus, the future impacts of coastal erosion on coastal assets, infrastructure and communities (Rennie et al., 2021).Conversely, mapping VE recovery can then be useful for assessing the resilience of vegetated biogeomorphic coasts to stormdriven disturbance (Balke et al.,2014).This is interpreted to be related to how much hydrodynamics directly influence the VE in different geomorphic regimes.In contrast to the inner estuary, the VE sits above the tidal range and is free from the influence of hydrodynamics, resulting in a single, clearly defined edge.Differences in hydrodynamic impact, in turn, control the types of vegetation that dominate a coastal section.For example, the reflectance values between marram grass and saltmarsh vegetation may differ due to their water content (Zhang et al., 1997).The non- est.Nevertheless, VedgeSat offers a more objective approach that, while subject to its own interpretive error and lacking in contextual information available to humans, will at least be more systematic in its delineation of the VE.The addition of the transition zone then offers a quantification of the level of interpretive uncertainty at pixel-scale.

| Areas for potential improvement
Our focus in this contribution has been to develop approaches to extracting and analysing coastal vegetation that are broadly applicable, and have the potential to be applied at regional scales (similarly to CoastSat (Vitousek et al., 2023)).There exist several avenues of exploration to potentially improve the accuracy of the VedgeSat tool, especially for site-specific applications.One is the choice of machine learning algorithm and combination of spectral bands used to classify vegetation and non-vegetation pixels.Of the similar black-box machine learning classifiers available, support vector machines have comparable computational efficiency to ANNs, but often outperform them in terms of accuracy (Salah, 2017).Based on the higher error values in estuarine environments (Section 3.4), the accuracy of the tool might further be improved with the inclusion of a band index created to maximise the difference between water, wet sediment and vegetation.The Superfine Water Index is an example of a band index designed to better delineate between water and dense vegetation that may have similar spectral properties in multispectral satellite imagery (Sharma et al., 2015).Furthermore, the inclusion of other contextual spatial information (such as topographic data) into the ANN classifier may improve its accuracy, especially along coastal cliffs whose shadows may complicate the extraction of a VE position (Chen et al., 2015).
The accuracy of satellite-derived VEs is not just dependent on the correct classification of vegetation pixels, but also on the calculated In relation to improvements in the VedgeSat thresholding process, there may be benefit in increasing the computational complexity of the tool.In its current form, the complex nuances of different for the NDVI, and then a statistical model could be built in to adaptively select an optimal threshold value from these thresholds (Huo et al., 2010).This approach may increase computational expense, but also increase accuracy of the final threshold value used to extract the VE.
Considering the accuracy of the tool, not just in extracting a single vegetation edge but in analysing a timeseries, the potential for coregistration errors should be considered.Section 2.1 explains the assumption that georeferencing each satellite image validated to a common set of ground control points means that all images are coregistered to one another.For future applications in locations with no ground control information, there is the potential for satellite images per platform to be slightly misaligned with respect to one another.
This may impact the position of vegetation edges extracted in of change (Doherty et al., 2022;Rufin et al., 2021;Scheffler et al., 2017).
While the St Andrews site offers a variety of vegetation types and coastal geomorphologies on which to assess the tool, and while the tool itself has been trained on a variety of vegetation types and settings, it is limited to a marine temperate coast in the Northern Hemisphere.A blind application on more diverse international sites will increase the robustness and applicability of the tool.To demonstrate broader applicability, a tropical deltaic setting is a desirable next step for automated VE exploration.Tropical mangroves offer natural coastal protection from storm-driven impacts, with multiple ecosystem benefits that outweigh traditional man-made coastal protection (Barbier, 2016;Dasgupta et al., 2019).Changes in the coverage of these forests, particularly driven by flooding and sea level rise, may affect their protective function for the farmland and residential areas northwest behind the delta (Morris et al., 2023).Coupling a lack of seasonal observations with their socioeconomic importance and diversity in tropical vegetation species and tidally dominated morphodynamics (Wang et al., 2019), mangrove settings offer a fascinating future site for blind-testing the efficacy of satellite driven VE extraction.

A
recent seminal study by Rogers et al. applied a nested convolutional neural network to a collection of spectral band indices designed to identify VEs (Rogers et al., 2021).Rogers et al. progressively decrease the resolution of PlanetScope images (3 m resolution) . However, there is freely accessible mid-resolution imagery from 1984 to present via Landsat and Sentinel.Platforms like these present strong potential to assess change in coastal VEs over the spatial and temporal scales needed to generate robust evidence required to support effective coastal adaptation planning.This study has developed a new methodological technique to automate the delineation of VEs from a range of publicly available, global EO datasets, from moderate-resolution (15 m for Landsat 5 & 8, 10 m for Sentinel-2) to high-resolution products (3 m for PlanetScope).The peninsula of St Andrews and the Eden Estuary in Scotland were used to validate VE extraction, due to the variety of coastal geomorphic settings and vegetation types across the site.By assessing the accuracy of the resultant VEs against an array of ground-truth surveys, a seasonal-annual picture of coastal change across the Eden Estuary can be built from the repeated surveys across the four-decade lifespan of Landsat, Sentinel-2 and PlanetScope imagery.
. The training of the MLP network was performed across 10 km 2 areas of Dornoch and Aberdeen in the Scottish Highlands.These areas have a similarly diverse landcover as the validation site at St Andrews, including saltmarsh, marram grass, golf links, forestry, gorse and shrubs, pebble and sand beaches and urban areas.The MLP used in the tool was trained using manual interpretation and digitisation of regions of vegetation and non-vegetation (e.g.water, bare sediment, urban surfaces), using a selection of satellite images from Landsat 5 and 8, Sentinel-2 and PlanetScope.High-resolution (25 cm) aerial imagery of the training sites was used alongside the satellite images as an interpretation aid when training.More in-depth training information can be found in the supporting information.RGB, near infrared (NIR) and shortwave infrared (SWIR) pixel values were compiled for each class in each training image (where available), as well as standard deviations on each of these bands.A separate classifier was trained for PlanetScope imagery as this platform has no SWIR band.Pixel values of suitable band indices for identification of vegetation were also included (Price, 1987; F I G U R E 1 Flowchart of vegetation edge extraction process, with colours corresponding to the four subsections of: defining user requirements; image preprocessing; supervised object classification; and thresholding and contouring.[Color figure can be viewed at wileyonlinelibrary.com] ), I is the band index being used for the PDFs, I O is the optimal threshold band index value being calculated, ζ is the peak value on each PDF, ω is the weight applied to each maximum value and the subscripts wet and dry represent the 'wet' class (with weaker NDWI signal) and 'dry' class (with stronger NDWI signal), respectively.In the original Weighted Peaks application, the values of ω were chosen to skew the threshold towards the landward water's edge or upper swash zone.The values of ω differ slightly between Harley et al. (2019) and Doherty et al. ( ζ veg and ζ nonveg correspond, respectively, to peak NDVI values of the 'vegetation' and 'non-vegetation' classes, with the larger weight placed on the class that the threshold should be pulled towards.In this case, this is the 'non-vegetation' class, to better capture sparse seaward vegetation with lower NDVI values.The PDFs are generated from NDVI values constrained to the coastal swath zone defined by a reference shoreline (if provided by the user), to better represent a coastal NDVI distribution.
NDVI values that fall either side of the contour value identified by Weighted Peaks are not disparate but represent a region of overlap, where pixels with similar NDVI values are classified as either 'vegetation' or 'non-vegetation'.This overlap can be interpreted as a zone of transition from vegetation to non-vegetation.This transition zone is explored in the CoastSat algorithm, since with water-to-land transitions the overlapping NDWI values often represent undesirable error associated with whitewater and wet sand from swash activity (Pardo-Pascual ally required a location with: (i) a diversity of coastal environments, including varied geomorphology and ecology; and (ii) existing survey data on the position of VE mapped by ground surveys and/or digitised from high-resolution aerial imagery.Sites with coastal assets of significant natural or economic value were also considered desirable for delivering useful findings to stakeholders.The coast surrounding the Eden Estuary in Scotland was chosen, which includes the world-renowned St Andrews Links golf course (Figure3).The 46 km 2 validation site consists of a mixture of open coast sand and mixed sand-gravel beaches backed by coastal dunes, grasses, shrubs and forestry; the inner Eden Estuary exhibits estuarine saltmarsh, mud and fine gravels(Joint Nature Conservation Committee, 2021; Rennie   et al., 2021).Due to the spatial limitations enforced by the GEE server to process satellite tiles no larger than 512 x 512 pixels, the site was split into two approximate halves of 23 km 2 .We characterised these halves as the 'open' and 'inner' coast, dependent on the level to which coastal processes are dominated by waves versus tidal and fluvial hydrodynamics respectively.These are informed by analysis from the Dynamic Coast project(Hurst et al., 2021).2.8 | Ground-truth vegetation edgesTo validate the VEs extracted by the VedgeSat tool, satellite-derived VE and ground-truth VE positions were compared on each 3 m-spaced shore-normal transect (similarly to Vos et al. (2023) and Pardo-Pascual et al. (2024)).Transect intersections between each pair of satellite and ground-truth VEs yields a sample size of 33,075 validation distances.The list of ground-truth datasets, their associated capture dates used to match to the closest satellite images capture dates, and each ground-truth date's geographic coverage and validation sample size can be found in Table 2. Ground-based VEs were mapped with Global Navigation Satellite Systems (GNSS) survey equipment on foot.The particular GNSS equipment used has a positional accuracy of 20 cm when capturing points.Points were captured and joined together (streamed) at an interval of 1 m to approximately match the stride of the surveyor (Eos Positioning Systems, 2023).VEs were also mapped using high-resolution (25 cm) aerial orthophotos from GetMapping and Ordnance Survey, obtained via an EDINA Digimap aerial licence (EDINA Digimap, 2023).These edges were manually digitised by tracing along the visually interpreted boundary.Due to the logistics of ground-truth data collection at the coast and the timetable of aerial imagery flyovers, the majority of the surveys are captured in Northern Hemisphere spring and summer months.Surveyors interpreted VE as the seaward boundary between the presence of F I G U R E 2 (a) Accuracy results from the Weighted Peaks weighting sensitivity analysis, for one image each from the four satellite platforms tested (L5: Landsat 5, L8: Landsat 8, PS: PlanetScope, S2: Sentinel-2).Minimum positional errors for each platform are outlined in red, corresponding to across all platforms tested.(b) Probability Density Functions (PDFs) of Normalised Difference Vegetation Index (NDVI) values across an example satellite image, classified by the vegetation detection model into vegetation (veg) and non-vegetation (nonveg).The peak NDVI values (ζ) for each class and resulting NDVI threshold (I O ) are marked with dashed lines.The vegetation transition zone (TZ) is also bounded by the 0.5 th and 10 th vegetation PDF percentiles.[Color figure can be viewed at wileyonlinelibrary.com] stable vegetation and bare sediment.Three autumn and winter surveys were also assessed.However, through the assessment of the effects of seasonality in training data on accuracy of the classifier(explored in the supporting information), we found no significant seasonal effect on accuracy.3| RESULTS3.1 | Ground-truth validationSatellite images were available on average 12 days either before or after each validation survey.Unfortunately, the 2012-03-27 survey occurred during a period of limited cloud-free satellite data coverage, resulting in the closest suitable satellite image date occurring 151 days earlier on 2011-10-28.Nevertheless, this date was retained in order to validate a suitable number of Landsat 5 images.

Figure 4
Figure4shows the relationship between ground survey and satellite-derived vegetation edge positions, colour-coded by the date of ground survey acquisition.The coefficient of determination (R 2 ) on date-wise linear regression identified the overall performance across all platforms is 0.53.The lowest error was found in the 2019-08-18Sentinel-2 VE, with an R 2 value of 0.98.The highest error VE is the

Figure 6
Figure 6 shows satellite-derived VEs and associated error along the open coast of West Sands on the seaward side of the St Andrews peninsula.West Sands displayed the best overall performance, resulting in < 15 m errors (smaller than a Landsat pixel) across 99% of transects and RMSE of 5.3 m (Figure 6b).The VE is characterised here by a sharp dune cliff and homogeneous marram dune grass.Differences between satellite and ground-truth VEs relate to resolution differences, with satellite VEs having lower resolution and therefore lower sinuosity.Errors are mainly a result of the smoothing performed during the Marching Squares routine (Figure6d) compared to the more sinuous nature of VE when mapped in the field.The positional error is higher to the north, where prograding dunes created a series of vegetated ridges (Figure6c).Along this northern section, older Investigating the NDVI used to extract the 2007-04-17 VE, the complexity of this type of coast becomes apparent where defences have become vegetated over time and loose vegetation gathers (Figure7c,f).The pixel size of Landsat 5 in this inset further illustrates how the positional error of VEs is magnified in these lower-resolution images and along these complex coastal types.However, a seaward bias also occurs in the 2022-02-22 PlanetScope NDVI (Figure7d), tracing the edge of the intertidal vegetation and deposited seaweeds at low tide (Figure7e,f), and not the terrestrial vegetation along the top of the defences.The high sinuosity of the extracted VE in highresolution images is also more apparent on this date, with the contour tracing clusters of NDVI values on the foreshore that exceed the threshold.Conversely, the Landsat 8 and Sentinel-2 VEs display a good agreement with the validation edges, with smaller-than-pixel accuracy across 50% and 96% of the transects, respectively.

est two Landsat 5
VEs and 2020-04-20 PlanetScope VE are biased seaward of the validation edges.Landsat 8 was not able to be validated along the southern estuary banks due to a lack of coinciding validation surveys.The seaward bias of these edges leads to an overall RMSE of 22.6 m for the southern estuary.Sentinel-2 still performs well in this area, with 77% of transects across the 2018-06-24 VE showing smaller-than-pixel errors.Particular areas of high error occur where there are pockets of seasonal saltmarsh that are more prominent at low tide and during growth seasons.A seaward bias following a low NDVI threshold can be seen in several VEs here; Figure 8c displays this bias in the 2020-04-20 PlanetScope edge.The median positional errors for each platform across the northern side of the estuary fall within 21.2 m, with an overall RMSE of 17.5 m (Figure 9a,b).The best performance is once again seen in the Sentinel-2 VEs, specifically in the 2018-06-24 edge with an RMSE of 5.0 m.Despite being derived from the highest image resolution, PlanetScope edges validated show a median error of 7.8 m (over double the size of a 3 m PlanetScope pixel).The largest error across the northern side of the estuary is seen in the Landsat 8 2015-09-30 edge, with an RMSE of 28.8 m.As shown in Figure 9c, both the terrestrial grasses and saltmarsh vegetation are picked up by the same extracted VE in low-tide images here.This is further displayed by the majority of PDFs in Figure 9b having a bimodal distribution, with one peak near the 0 m error mark and the other several tens of metres seaward.To further investigate the relationship between estuarine vegetation extraction and tidal heights, Figure 9 plots the relationship between positional VE error and tidal stage for each image date corresponding to those in Figure 9.The negative relationship suggests that for this environment, lower-tide images have higher positional error in the satellite-derived VE due to the algorithm identifying intertidal vegetation.By comparison, the vegetation edge on the sandy open coast to the east sits well above the tidal stage and has very little intertidal vegetation.This suggests a link between the level of influence of hydrodynamic processes on a vegetated coast, and the potential to extract multiple qualifying vegetation edges from satellite images.
shows both landward and seaward bias in the platforms validated across this area, with an overall RMSE of 21.9 m.The range of error is large, with the best performing VE extracted (the 2021-07-01 Sentinel-2 edge) having an RMSE of 4.9 m, but the worst performing VE extracted (the F I G U R E 4 Cross-shore distances of VEs along every transect for validation VEs versus satellite-derived VEs.Linear regression lines are fit through the scatter plots for each validated image date.The dashed grey line represents the overall linear regression for all dates validated, and the diagonal grey line represents an R 2 value of 1. VE, vegetation edge.[Color figure can be viewed at wileyonlinelibrary.com] F I G U R E 5 Violin plots of distances between validation vegetation edges (VEs) and satellite-derived VEs along each transect across the site, separated into satellite platforms (L5: Landsat 5, L8: Landsat 8, PS: PlanetScope, S2: Sentinel-2).Median distances (η) for each platform are represented by thick white central lines, with quartiles shown as dashed white lines.The darker and lighter grey sections in the background represent 10 and 15 m pixel widths, respectively.[Color figure can be viewed at wileyonlinelibrary.com] 2007-04-17 Landsat 5 edge) having an RMSE of 35.5 m.The large uncertainties and both landward and seaward biases are interpreted to be the result of a diffuse transition zone from unvegetated to vegetated conditions (see Section 2.6).At Tentsmuir dunes (to the north of the Eden Estuary and St Andrews dunes, Figure 10c), the transition zone for the 2020-04-20 PlanetScope image is approximately 82 m wide in a cross-shore direction.This location has been accreting over the last 16 years (as shown in the north of Figure 3b), where new vegetation growth is characterised by sparse grass coverage and embryonic dune ridges.The VE extracted along this section follows the more dense, stable vegetation as opposed to the validation edge which traces the seaward edge of new vegetation growth.In contrast, West Sands on the St Andrews peninsula (Figure 10d) consists of more mature marram grasses atop partially eroding dune cliffs, presenting a sharp, steep edge where the transition zone width is approximately 16 m.The satellite-derived VE along this section agrees well with the validation VE.The difference in widths reflects the difference in NDVI values falling within the transition zone percentiles.This itself reflects the sharpness and level of transience in these two environments' vegetation boundaries, as well as the level of uncertainty of the extracted edge.

|
Performance compared with alternative coastal change indicatorsSection 3.1 demonstrates seasonal timeseries of vegetation edge change can be generated by VedgeSat with a conservative positional accuracy within 19.0 m (1.9 times the size of a Sentinel-2 pixel and 1.3 times the size of a processed Landsat pixel).This accuracy is comparable to the VEdge_Detector algorithm that identified VE from 3 m resolution PlanetScope imagery to within two pixel widths(Rogers et al., 2021).In comparison to CoastSat which extracted waterlines, the VedgeSat tool performs more poorly, with an R 2 of 0.53 and RMSE of 19.0 m compared to CoastSat's 0.8 and 7.2 m respectively.However, the CoastSat validation was performed solely on an open-coast site and, as shown, the performance of VedgeSat differs across satellite platforms and environments.For wave-dominated open coasts, the VedgeSatR 2 value rises to 0.82 and the RMSE drops to 5.3 m.By comparison, the recent satellite-derived waterline benchmarking exercise conducted by Vos et al. produced poorer results for a mesotidal environment like the Eden Estuary validation site.The exercise found F I G U R E 6 Result of validation exercise between ground-truth vegetation edges (VEs) and satellite-derived edges, with focus on the sandy, dune-dominated open coast to the southeast.(a) Focus site location (with bounding transects of interest in grey) and resulting satellite derived VEs.(b) Probability Density Functions (PDFs) of cross-shore distance between validation and satellite-derived lines along transects in focus site.Median error distances (η) for each satellite are marked by dashed lines, and the darker grey background panels highlight distances of 10 and 15 m (Sentinel-2 and Landsat pixel sizes) respectively.(c) Zoomed inset of resulting VEs (smooth lines) compared to their validation counterparts (dashed lines) along the north West Sands dune ridges.(d) Zoomed inset of resulting VEs, with less sinuous satellite VEs leading to deviation from validation.[Color figure can be viewed at wileyonlinelibrary.com]

F
I G U R E 7 Result of validation exercise between ground-truth vegetation edges (VEs) and satellite-derived edges, with focus on the mixed VEs along the partially defended, sheltered side of the coastal peninsula.(a) Focus site location (with bounding transects of interest in grey) and resulting satellite derived VEs.(b) Probability Density Functions (PDFs) of cross-shore distance between validation and satellite-derived lines along transects in focus site.Median error distances (η) for each satellite are marked by dashed lines, and the darker grey background panels highlight distances of 10 and 15 m (Sentinel-2 and Landsat pixel sizes) respectively.(c, d) Zoomed insets of resulting VEs (smooth lines) compared to their validation counterparts (dashed lines) highlighting seaward bias, and underlying Normalised Difference Vegetation Index (NDVI) rasters used to extract VEs from.(e, f) Field photos showing the complexity of certain vegetation edges (cliffs with seaward diffuse vegetation, artificially defended sections with seaward vegetation debris).[Color figure can be viewed at wileyonlinelibrary.com] Variations in the threshold NDVI value used to extract each VE can be investigated to form a coastal typology.The threshold changes with each image, because the distribution of NDVI values within the vegetation and non-vegetation classes (and therefore the peak NDVI values used in thresholding) also changes.When the threshold values generated by the Weighted Peaks method are spatially separated into open coast and inner estuary (see Figure 3), the distribution of threshold NDVI values differs (Figure 11).For the validated images, the median NDVI threshold for the inner estuary environment along which the VE is extracted is 0.22, and the median threshold for the eastern open coast is 0.16.Additionally, the range of values is greater in the western half (À0.06-0.49versus À0.06-0.22),indicating more variation in the threshold values extracted.The ANN pixel classifications and resulting VE extracted are dependent on the type of geomorphological and ecological environment a satellite image captures.Our interpretation is that the difference in NDVI thresholds is controlled by vegetation dynamics differing between geomorphic regimes.Section 3.4 highlighted the negative relationship between tidal stage and error against validation VEs in an estuarine setting.

F
I G U R E 8 Result of validation exercise between ground-truth and satellite-derived vegetation edges (VEs), with focus on the saltmarsh, grasses, shrubs and agricultural vegetation of the southern Eden Estuary.(a) Focus site location (with bounding transects of interest in grey) and resulting satellite-derived VEs.(b) Probability Density Functions (PDFs) of cross-shore distance between validation and satellite-derived lines along all transects in focus site.Median error distances (η) for each satellite are marked by dashed lines, and the darker grey background panels highlight distances of 10 and 15 m (Sentinel-2 and Landsat pixel sizes) respectively.(c) Zoomed inset of resulting VEs (smooth lines) against validation (dashed lines), where the 2020-04-20 PlanetScope VE is extracted along lower Normalised Difference Vegetation Index (NDVI) values seaward of the estuarine promontory.[Color figure can be viewed at wileyonlinelibrary.com] vegetation class distribution may be influenced by the presence of water within the coastal buffer zone (e.g.depending on tide conditions), skewing its PDF towards much lower NDVI values (McGwire et al., 2000).The sediment type may also affect the non-vegetation class distribution, due to differences in pixel reflectance for fine muds with organic materials versus coarser, quartz-dominated sands.The original CoastSat algorithm contains different waterline extraction models trained specifically for either bright sand versus dark sand.However, the VedgeSat model takes a simplified approach, training the model with a range of different vegetation types and water characteristics, but then collating them into two overarching classes (vegetation or non-vegetation).The contrast in NDVI thresholds has shown that there may be a way to classify and monitor different ecological environments at the coast.4.3 | Humans versus machinesField surveys of VEs are costly and time-consuming.There is a large back catalogue of open-source satellite imagery, and a proliferation of commercial satellite imagery at increasing temporal and spatial resolutions.Automated approaches to feature extraction from these data are accelerating our ability to quantify coastal change at the Earth surface.It takes roughly 40 min to survey 1 km of vegetation in the field (not including time required for survey mobilisation).On a typical desktop computer, VedgeSat can process the entire back catalogue of 450 available Sentinel-2 images for a single 23 km 2 scene in under 10 h, automatically producing timeseries of waterlines and vegetation lines.Despite the overall success of VedgeSat to map timeseries of VEs from satellite images, pronounced differences exist between VEs mapped by human surveyors and VedgeSat, in locations lacking sharp contrast between vegetation and sediment.Sections 3.3 and 3.4 demonstrate seaward biases in some of the satellite-derived VEs, and Section 3.5 reveals a landward bias in several extracted VEs, relative to those mapped on the ground by a human surveyor.This is particularly present in Figures8 and 9.There may be differences in interpretation from one surveyor to the next, even when surveying standards (explored in the supporting information) are established to reduce human bias in the interpretation of the VE position(Parson, 1997;Van Coillie et al., 2014).VEs remain particularly difficult to interpret on accreting coastlines with discontinuous, patchy F I G U R E 9 Result of validation exercise between ground-truth and satellite-derived vegetation edges (VEs), with focus on the saltmarsh, grasses, and shrubs of the northern Eden Estuary.(a) Focus site location (with bounding transects of interest in grey) and resulting satellite derived VEs.(b) Probability Density Functions of cross-shore distance between validation and satellite-derived lines along all transects in focus site.Median error distances (η) for each satellite are marked by dashed lines, and the darker grey background panels highlight distances of 10 and 15 m (Sentinel-2 and Landsat pixel sizes) respectively.(c) Zoomed inset of resulting VEs (smooth lines) against validation VEs (dashed lines), with some contours tracing both the terrestrial grass and offshore saltmarsh.(d) Root Mean Square Error (RMSE) of each satellite-derived VE in Figure (b), plotted against the tidal stage when each satellite image was captured, with a negative linear regression (dashed grey line).Marker edge colours correspond to the satellite platforms in panel (b).[Color figure can be viewed at wileyonlinelibrary.com]and/or pioneering vegetation, in the presence of intertidal vegetation or vegetation debris such as detached and decaying seagrasses or seaweeds(Pollard et al., 2019).These challenges are exemplified by lowtide PlanetScope images in the estuary, where we obtained large RMSEs compared with image resolution, owning to the highresolution imagery picking up intertidal vegetation as well as terrestrial vegetation.For VedgeSat users investigating terrestrial vegetation in estuarine environments, we recommend tuning the NDVI threshold produced by Weighted Peaks to better align with their edge of inter-

F
I G U R E 1 0 Result of validation exercise between ground-truth vegetation edges (VEs) and satellite-derived edges, with focus on diffuse edges along the accreting dynamic dunes influenced by both wave action and fluvial sediment transport in the north.(a) Focus site location (with bounding transects of interest in grey) and resulting satellite derived VEs.(b) Probability Density Functions (PDFs) of cross-shore distance between validation and satellite-derived lines along all transects in focus site.Median error distances (η) for each satellite are marked by dashed lines, and the darker grey background panels highlight distances of 10 and 15 m (Sentinel-2 and Landsat pixel sizes) respectively.(c) Satellitederived vegetation transition zone (orange pixels) extracted from the 2020-04-20 PlanetScope image, highlighting the diffuse vegetation signal along Tentsmuir.(d) Sharper dune cliff vegetation boundary along West Sands in the south, where transition zone width is much smaller.[Color figure can be viewed at wileyonlinelibrary.com] threshold NDVI value for contouring along, and thus there is potential for the spectral properties of a dominant environment to bias the threshold.This environmental difference in threshold values is explored further in Section 4.2.Refining thresholds specific to the local environment in which the routines are applied provides an opportunity to tailor the analysis and improve accuracy in site-specific applications.For example, the routine could subsequently be adapted to generate transect-specific PDFs from which to extract thresholds, enabling the VE to be more adaptive to the surrounding local environment.
spectral band and band index values are used within an ANN to make a simple final decision between vegetation and non-vegetation.That decision can introduce misclassification errors, particularly at the boundary where (particularly as seen in Section 3.5) the overlap between NDVI values for vegetation and non-vegetation create a fuzzy boundary.While useful geomorphological information has been found in the fuzziness of this boundary, it complicates the extraction of a simple vegetation boundary.To account for misclassification error, the complexity is then further distilled by contouring the VE with one threshold value, informed by the peak NDVI values of those two classes.What would happen however if the complexity was not reduced up at this step, and we were able to more dynamically make use of the sizeable amount of data fed into the ANN?For example, a threshold value between vegetation and non-vegetation classes could be generated for all feature vectors used in training, not just timeseries, and therefore introduce error into positions and thus rates of change extracted from these timeseries.While reported geolocation errors tend to be small (< 11 m or just over one pixel for Sentinel-2 Level 1C products (S2 MSI ESL Team, 2023)), they vary by location and through time.As the scope of this work is to assess the abilities of the tool in extracting single VEs, coregistering has not been incorporated into the VedgeSat pipeline as of yet.Coregistering each image to a common geolocation through phase correlation, similarly to CoastSat.PlanetScope, may improve the accuracy of calculated rates We present a new approach to assessing seasonal-decadal coastal change through objective mapping of the VE from a back catalogue of open-source satellite imagery.The VedgeSat tool makes use of cloudbased processing capabilities, machine learning for pixel classification, and Weighted Peaks thresholding to automatically extract the VE.We validated the approach across a geomorphologically diverse coastal region in Scotland, revealing that the tool can successfully extract coastal VEs from satellite images, along a variety of vegetation types, to within an average of 19.0 m of the true edge.Positional accuracy of the extracted edge increases to 5.3 m when applied across open wave-dominated coasts, but performance is less accurate along estuarine, accreting and/or artificially protected coasts.Variations in spectral contrast across the different coastal environments, and discrepancies between human and computer in the interpretation of a F I G U R E 1 1 Violin plot of Normalised Difference Vegetation Index (NDVI) values used as threshold for contouring vegetation edges along, for satellite images on the inner estuarine and open wave-dominated halves of the validation site.[Color figure can be viewed at wileyonlinelibrary.com] solid vegetation boundary, lead to larger error.By defining a novel metric for the zone of transition from vegetation to bare sediment, we have been able to quantify zones of greater uncertainty in the extracted edge position, with the potential to provide insight into geomorphological processes at play along wave-dominated vegetated sandy coasts.Vegetation is a biogeomorphological feature in these systems, both controlling and being influenced by coastal processes.Its seaward boundary should therefore be considered as a useful indicator of shoreline position at the land-sea boundary, and thus as a metric for quantifying coastal change across the upper shoreface and backshore.Our research has demostrated that satellite-derived vegetation edges can be less noisy than conventional waterlines as a coastal change indicator, showing excellent potential for widespread use by coastal researchers and practitioners.By offering automated and open source tools such as VedgeSat to analyse freely available, near-global satellite imagery, coastal managers across the globe have the potential to gather broader information to measure coastal change, and thus make better-informed decisions for their communities and assets.AUTHOR CONTRIBUTIONS Freya M. E. Muir: Conceptualisation; methodology; software; validation; formal analysis; visualisation; writing-original draft.Martin D. Hurst: Conceptualisation; funding acquisition; methodology; software; supervision; writing-review and editing.Luke Richardson-Foulger: investigation; software; writing-review and editing.Alistair F. Rennie: Funding acquisition; resources; supervision; writing-review and editing.Larissa A. Naylor: Funding acquisition; supervision; writing-review and editing