Soil moisture, stressed vegetation and the spatial structure of soil erosion in a high latitude rangeland

Soil erosion has been a persistent problem in high‐latitude regions and may worsen as climate change unfolds and encourages increased anthropogenic exploitation. We propose that soil moisture is likely to shape future erosion trends, as moisture stress reduces the capacity of vegetation cover to retard erosive processes. However, the spatial variability of soil moisture in high‐latitude soils—and the ways in which this variability drives the spatial distribution of erosion features—is poorly understood. We addressed this knowledge gap with a study of andosol erosion in southern Iceland. Our study used a combination of high‐resolution (<3 cm) remote sensing data (using normalised difference vegetation index (NDVI) and normalised difference red edge as metrics of plant vitality) and long‐term, in situ measurements of soil moisture to unpick the relationship between moisture stress, vegetation vitality and patchy soil erosion. Mean NDVI increased with distance from eroded areas, varying from ~0.6 in vegetated areas on the margins of erosion patches to ~0.8 in areas >10 m from eroded terrain. We found lower moisture availability close to existing erosion features: mean volumetric soil moisture content varied from 17% (proximal to erosion patch) to 36% (distal to erosion patch). We also found that variability in soil moisture decreased with distance from eroded areas: the coefficient of variation (CV) in soil moisture varied from 0.33 (proximal to erosion patch) to 0.13 (distal to erosion). Our findings indicate that the margins of erosion patches have a stressful soil environment due to exposure to the atmosphere. The vegetation in these locations grows less vigorously, and the exposed soil becomes more vulnerable to erosion, leading to erosion patch expansion and coalescence. If these conditions hold more generally, they may represent a feedback mechanism that facilitates the lateral propagation of soil erosion in high‐latitude regions.


| INTRODUCTION
Soil erosion is a growing global crisis (Olsson et al., 2019). It is often conceptualised as a problem of low-and mid-latitude locations. However, soil erosion also occurs in high-latitude regions and may worsen as these regions warm and are subjected to increased anthropogenic exploitation (Grosse et al., 2011;Heindel et al., 2015;Hilton et al., 2015;Owczarek et al., 2022). Many factors combine to influence the extent and severity of high-latitude soil erosion, including climate, vegetation cover, parent material, soil type and land management (Arnalds, 2000;Grosse et al., 2011;Heindel et al., 2015). Soil moisture availability is likely to be particularly important, as water affects the physical, chemical and biological characteristics of soils and, by extension, their capacity to resist erosion. We explored the biogeomorphological dimensions of this phenomenon, focussing on the interaction between moisture availability, plant vitality and soil erosion. Water is a limiting resource for plant growth in many high-latitude terrestrial ecosystems (Kemppinen et al., 2019); the marginal conditions for plant growth in these habitats mean that even small variations in moisture availability can have a profound impact on ecosystem function (Kemppinen et al., 2018;Myers-Smith et al., 2015;Tyystjarvi et al., 2022). In the context of soil erosion, vegetation stressed by soil moisture conditions has a reduced capacity to retard soil loss. However, the spatial variability of soil moisture in high-latitude soils-and the ways in which this variability drives the spatial distribution of soil erosion-is poorly understood. Our research addresses this knowledge gap by using a combination of high-resolution (drone-acquired) remote sensing data and long-term, in situ measurements of soil moisture to unpick the relationship between moisture stress, vegetation vitality and patchy soil erosion in a sub-Arctic habitat.
High-latitude soil erosion is particularly prevalent in areas that support pastoral farming on marginal soils, e.g., southern Greenland (Jacobsen, 1987;Massa et al., 2012) and Iceland (Arnalds, 2015;Barrio et al., 2018;Dugmore et al., 2009Dugmore et al., , 2020Gíslad ottir et al., 2010;Ólafsd ottir et al., 2001;Streeter et al., 2015). These regions have experienced rapid environmental change in the last few decades, warming at 2-4 times the global average (Post et al., 2019;Rantanen et al., 2022). Climatic amelioration is likely to increase the exposure of high-latitude regions to human exploitation in the coming decades. Thus, soils here could face increased stress from climate change and increased anthropogenic impacts, both of which are likely to accelerate existing erosion or initiate erosion where it does not currently occur (e.g., Gíslad ottir, 2006).
Many factors combine to promote high-latitude soil erosion, including vegetation type and cover, soil type, spatiotemporal patterns of snow lie, the magnitude and frequency of freeze-thaw cycles, degree of exposure to winds and grazing by large herbivores. Soil moisture is a critical but frequently overlooked factor, as it has been shown to impact plant growth and-by extension-the ability of vegetation cover to retard erosion (Berdanier & Klein, 2011;Kemppinen et al., 2022). Drought is obviously stressful to plants (Grace, 1997), but variability in soil moisture levels (e.g., rapid cycling between soaked and parched conditions) can also lead to reduced growth and sparser cover (Kemppinen et al., 2019). Spatial variation in soil moisture conditions may be a factor in determining which areas are vulnerable to erosion and which are resilient. However, little is known about how patchiness in moisture stress drives the distribution of soil erosion in high-latitude habitats.
Our study aimed to unpick the spatial relationships among soil moisture, vegetation vitality and soil erosion on an eroded rangeland site in Iceland. The site is characterised by patches of eroded terrain surrounding vegetated 'islands' with thick soil cover. The boundaries between these two patches are low (a few metres high) scarps of exposed soil, named rofabards in Icelandic (Arnalds, 2000) (Figure 1a). Previous work in Iceland has argued for the importance of small erosion patches in triggering rofabard erosion (Dugmore et al., 2009;Gíslad ottir, 2001). A recent study of erosion patch formation and growth in Iceland by Streeter and Cutler (2020) demonstrated that small openings in vegetation coverincipient erosion patches indicative of stressed plantscluster around large erosion patches. Based on this observation, we hypothesised that moisture stress decreases with increasing distance from the edge of the erosion patch, i.e., plants closer to the rofabard scarps Highlights • Connections among soil moisture, vegetation vitality and high-latitude soil erosion were investigated. • A novel combination of remote sensing data and in situ measurements of soil moisture was used. • A consistent relationship among soil moisture, vegetation cover and erosion was found. • We propose a feedback mechanism that may shape the progression of high-latitude soil erosion.
will be more stressed than those further away. This is because areas close to the rofabard scarp (peripheral areas) are likely to wet and dry more quickly than those further away (interior areas), due to their exposure to wind and driving rain. They are also likely to remain drier for longer due to evaporation from the exposed soils of the rofabard scarp ( Figure 1b). We propose that the vitality of plants close to the rofabard scarp will be lower because of these soil moisture conditions (Figure 1c), although other factors, e.g., physical disturbance by sheltering sheep, may also contribute to stressful growing conditions. We used two approaches, applied at different spatial scales, to investigate relationships among soil moisture, vegetation vitality and soil erosion: 1. At a landscape scale (10s-100 s m), we used normalised difference vegetation index (NDVI) and normalised difference red edge (NDRE) data, derived from multispectral uncrewed aerial vehicle (UAV) imagery, as a proxy for vegetation vitality.
2. At a fine scale ($10 m), we supplemented our UAV imagery with measurements of soil moisture, collected over 12 months.

| Study area
The study was conducted at Hamragarðaheiði (63.61393 N, À19.96424 E, approx. 200 m above sea level), a site on the western flank of Eyjafjallajökull, Iceland ( Figure 2). The site was selected because (1) it has suffered soil erosion in the past; (2) the vegetated surfaces are relatively flat (so the impact of slope on soil hydrology in these areas is minimised) and (3) the vegetation cover is homogeneous (in terms of both structure and composition) at a landscape scale. Hamragarðaheiði experiences cold winters and cool summers and receives around 1380 mm of precipitation each year (Figure 2d,e). Prevailing winds are from the F I G U R E 1 Soil moisture and erosion (a) diagram illustrating Icelandic erosion features known as rofabards with indicative soil moisture sensor locations marked as red and blue circles (b) hypothesised changes in soil moisture following a rainfall event, according to distance from the edge of an erosion patch (delineated by the top edge of the rofabard scarp); peripheral locations close to the rofabard scarp are shown with a red line; interior locations are indicated with a blue line; (c) hypothesised relationship between plant vitality/moisture and distance from a rofabard scarp. east (Einarsson, 1984). Climate models suggest mean annual temperatures in Iceland will increase by 1.3-2.1 C by the mid-twenty-first century, coupled with small (1.2%-4.3%) increases in mean annual precipitation, probably concentrated in the late summer and early autumn (Icelandic Meteorological Office, 2023). In the future, one would therefore expect higher temperatures to promote evapotranspiration in summer, leading to soil moisture deficits during the growing season.
The Hamragarðaheiði site is grazed by sheep at low stocking densities from May to September. The soil cover in this region is Brown Andosol up to around 2 m thick. Brown Andosol is an Icelandic classification that corresponds to the World Reference Base for Soil Resources (WRB) category of Haplic/Mollic Andosol (Arnalds, 2015). Icelandic Brown Andosols are derived primarily from tephra parent material and aeolian deposition rather than bedrock weathering, and they are correspondingly fine (the largest fractions are silt and sand-sized) and free-draining (Arnalds, 2004). Soil organic matter content is highly variable but is typically in the range of 5%-7% (Arnalds, 2005). The high proportion of Al and Si in the glassy parent material promotes the formation of allophane, a clay mineral with a high capacity for water retention. However, the allophane tends to form stable, silt-sized aggregates and Brown Andosols lack phyllosilicates that provide cohesion in other soils (Arnalds, 2005). Hence, Brown Andosols are friable when dry, rendering them vulnerable to erosion. The Hamragarðaheiði site has experienced severe soil erosion in the past, with many large (10s m) erosion patches. The patches expand at the margins by a few cm per annum, a process largely driven by aeolian erosion (Arnalds, 2000;Dugmore et al., 2020). Water retention in Brown Andosols is relatively high, but so is hydraulic conductivity (Arnalds, 2005); thus, a period of just a few weeks of low rainfall can be enough to trigger severe moisture stress (Arnalds, 2015).
The geomorphology of the Hamragarðaheiði site is characterised by 'islands' of vegetation surrounded by erosion patches (Figure 2b,c). The islands of vegetation are bounded by scarps and have flat tops covered by a thick, grassy sward. The adjacent eroded areas consist mostly of bare lava or sand and are generally not suitable habitats for vascular plants (Arnalds, 2015). In this study, we use the term 'rofabard island' to encompass both the erosion scarps and the flat-topped, vegetated areas that they surround.

| Remote sensing
We deployed a multispectral drone survey across the whole site (approx. 64,000 m 2 ) to assess spatial variation in NDVI. The survey was conducted in the latter part of the growing season (16 August 2021) and was carried out using a DJI Phantom 4 multispectral (P4m) quadcopter (refer to Supplementary Information for camera settings). Flight planning and execution followed the recommendations of the HiLDEN drone network protocol (Assmann et al., 2019). Seven ground control points (GCP) were deployed and geolocated with a Spectra Precision Pro-Mark 120 GPS system (Spectra Geospatial, Westminster, CO). Additionally, photographs of the reflectance targets (Mapir Inc., San Diego) were taken during the survey for later radiometric calibration.
The survey was flown twice along parallel flight lines (the second set of survey lines was perpendicular to the first). The surveys were flown at an altitude of 48 m, resulting in an average ground sampling distance of $2.6 cm. Images were acquired with 80% front and side overlap and close to solar noon with a mean absolute difference to solar noon of 1.5 h (max. 2 h). The weather conditions during the flight were overcast with continuous cloud cover.
The remote sensing data were processed using the Pix4D Mapper software (Pix4D SA, Switzerland). Photogrammetric procedures were applied along with georeferencing based on the GCPs and radiometric calibration based on photographs of the calibration targets (refer to Supplementary Information for details). The processed output comprised a digital surface model (DSM) and orthomosaics for each individual band. NDVI was calculated as the normalised difference between the nearinfrared (NIR) and red bands: We additionally calculated the NDRE index: We calculated NDRE because it is more robust than NDVI against oversaturation (Nguy-Robertson et al., 2012). This is due to the lower absorption sensitivity of chlorophyll in the red edge region and its deeper penetration into vegetation canopies.
Analysis of the processed images was carried out with QGIS v 3.10. First, surface in the survey area was divided into rofabard islands (following our definition of scarps + flat, vegetated tops) and eroded areas. The rofabard islands had thick soils and continuous vegetation cover; the eroded areas lacked these characteristics. We used the following criteria to classify the rofabard islands: • The area had an NDVI value >0.4. • The area had to be connected to a multi-metre-scale vegetated area, not a small (less than metre-scale), isolated fragment of vegetation in an otherwise eroded area. • The boundary of the rofabard island was defined by a break in slope, identified from a digital elevation map (DEM).
In addition, we cropped the vegetation cover in the northwest corner of the image, as visual inspection indicated it was recent, secondary regrowth on an eroded area, and thus qualitatively different from the vegetation cover on the top of the rofabard islands.
The area classified as rofabard islands was clipped, and the distance to the nearest rofabard scarp was calculated for each pixel. The pixels were then placed in bins at 1 m intervals, and the mean NDVI and NDRE for each bin were calculated ( Figure S1). To ensure that each bin had sufficient points for a representative estimated mean, the distance interval of the last bin was chosen so that its pixel count was at least 10% of the first bin.
F I G U R E 3 Result of drone survey (a) an overview of the site, showing rofabard islands, eroded areas, and the moisture sensor transect (red dots connected with a red line); (b) the site after being classified into rofabard islands (black) and eroded (white) areas; (c) a section showing the transect location and the profile of the rofabard island on which it was installed (vertical scale exaggerated); note the flat top of the rofabard island.

| Moisture sensors
In June 2019, we established a transect of nine SMT150 soil moisture probes (Delta-T Devices, Cambridge, UK). The probes determined volumetric soil water content. They were designated CH1-CH9 and arranged along a transect running perpendicular to the edge of a rofabard scarp. This grassy area, like many others in Iceland, is covered with low (cmscale) earth hummocks. The sensors were placed in the hollows between these hummocks. Probe CH1 was closest to the rofabard scarp (distance = 0.2 m), and Probe CH9 was furthest away (distance = 11.1 m). At each probe location, a small turf block was removed so that the sensors could be placed at a depth of 0.05 m below the surface ( Figure S2). We chose a depth of 0.05 m because this placed the sensor in the rooting zone with sufficient soil cover but not so deep that excessive excavation was required. Once the sensor was installed, the turf and soil that had been removed were carefully replaced on top of the sensor to completely cover it. Measurements of soil moisture were made every hour over a 12-month period (1 October 2019-30 September 2020) and recorded on an in-situ Delta-T GP2 data logger (Delta-T Devices, Cambridge). In addition to compiling time series data, we also calculated soil moisture summary statistics, including the coefficient of variation as a metric of variability in soil moisture (Kemppinen et al., 2019). Vegetation composition along the transect was subsequently surveyed using contiguous 0.5 m Â 0.5 m quadrats. All plant species present were recorded and quantified according to the Domin scale.

| Remote sensing and vegetation survey
The results of the drone survey are shown in Figure 3a. The rofabard islands appear as bright green patches, and the intervening eroded areas as orange/grey patches of exposed sediment and lava. Figure 3b shows the sites divided into rofabard islands (black) and eroded (white) areas following classification (see also Figure S3). Rofabard islands account for 70% of the surveyed area. Visual inspection of the site indicated that the vegetation cover was homogeneous; this was supported by the results of a vegetation survey (Table S1). Briefly, vegetation cover was dominated by grasses (predominantly Anthoxanthum odoratum) and small forbs (e.g., Thymus sp., Alchemilla sp.) to a height of around 30 cm. There were some mosses in the understorey, but these were a minor component of the vegetation. There were also occasional patches of prostrate shrubs (notably Empetrum nigrum and Salix sp.) but these were small in stature and did not project above the grasses.
Analysis of our drone survey indicated that NDVI and NDRE values were generally lower close to rofabard scarps ( Figure 4). Moving towards the interior of the rofabard island (i.e., away from the scarp), mean NDVI/ NDRE increased over a distance of 4 m and then levelled out, with a slight increase from 12 m. Variability in NDVI/NDRE was also greatest closer to the rofabard edge (Tables S2 and S3).
The contrast in NDVI between vegetated and eroded areas is illustrated in Figure 5, which shows NDVI values along a transect cutting through an erosion patch. NDVI values in the vegetation surrounding the patch decrease from $0.8 to $0.6 on the patch margins, dropping to $0.4 in the patch centre.

| Moisture sensors
Data from the soil moisture probes showed that locations close to the rofabard scarp were, on average, drier and more variable in terms of soil moisture than locations further away. Mean and median soil moisture increased with distance from the rofabard scarp. Mean soil moisture for each of the nine soil probes is shown in Figure 6a. Mean annual soil moisture levels varied by a factor of 2 from 17% to 36%. A quartile regression (Figure 6b) indicated an increase in median soil moisture along our transect. Probe CH2-0.6 m from the scarp-experienced drier conditions than probe CH1 (the probe immediately adjacent to the scarp) (see also Table S4). We calculated the coefficient of variation (CV) for the nine moisture sensors and observed a decrease in variability with distance from the erosion scarp (Figure 6c) for both the whole time series and a subset of recordings from the period May to September 2021, inclusive (i.e., during the growing season). Variability was somewhat lower during the growing season, but both datasets displayed a decrease in variability with increasing distance from the rofabard scarp. Other metrics (e.g., standard deviation) also showed a decrease in soil moisture variability with increasing distance from the scarp (Table S4). The interquartile range for each probe was similar (Figure 6b), indicating that the variability was driven by extreme values.
The importance of extreme values is illustrated by the time series data for individual probes. The moisture readings for probe CH1 (closest to the erosion scarp) and probe CH9 (furthest away) are shown in Figure 7. The peripheral probe (CH1, red line) recorded a much 'peakier' signal than the interior probe (blue line) and a much greater range of values (peripheral range = 66.6%, interior range = 36.5%), both high and low. In comparison, the interior probe exhibited more consistent values.
Differences among the probe locations can be seen clearly in response to (inferred) rainfall events, i.e., the abrupt spikes in the soil moisture record. We selected five rainfall events to study in detail, again comparing the soil moisture data from the ends of the transect (Figure 8). We chose week-long sections of the time series that appeared to represent discrete wetting events, i.e., periods characterised by a steep rising limb followed by a gradual decline to drier conditions. In each case, the signal from the peripheral probe was markedly more extreme (higher initial peaks). With the exception of Event 4, the shape of the peripheral and interior curvesespecially the rising limb-was similar, although the slope of the falling limb was less steep in the interior location, suggesting slower drying. Event 4 probably resulted from a short, sharp wetting event (perhaps involving driving rain on the face of the scarp over several hours). The peripheral probe recorded a very steep but short-lived spike in moisture levels during Event 4; the response in interior locations was much more muted.

| DISCUSSION
Our results are consistent with the hypothesis that moisture stress decreases with increasing distance from eroded areas. The remote sensing data indicated that NDVI increased with increasing distance from the rofabard scarp into the vegetated areas, up to a range of 3-4 m. We interpret this observation as indicating higher plant vitality in areas away from the rofabard scarp. The spectral reflectivity of vegetation has been shown to reflect soil moisture conditions (Engstrom et al., 2008). We therefore propose that the increase in NDVI/NDRE we observed is at least partly explained by soil moisture, which increases and becomes less variable with distance from the rofabard scarp. To show a definitive link among these variables, we would need long-term experimental studies (of 3-5 years duration, say), repeated drone surveys and coupled measurements of soil moisture and vegetation growth along multiple transects. However, our study demonstrates a correlation between soil moisture and plant growth that merits further study.

| Remote sensing data
We observed a relationship between mean NDVI/NDRE and proximity to rofabard scarps. Generally, mean NDVI/NDRE was lower close to scarps and higher further away (Figure 4). The difference was most pronounced closest to the scarps: at distances >3 m from the scarps, mean NDVI did not vary much with distance.
Previous studies have demonstrated the usefulness of NDVI for estimating biotic parameters such as plant biomass, leaf area index (LAI) and CO 2 fluxes in highlatitude biomes (Boelman et al., 2003;Walker et al., 2003). For example, NDVI has been shown to be positively correlated with aboveground biomass in northern European tundra (Aalto et al., 2021). F I G U R E 6 Variation in soil moisture with distance from the rofabard scarp (a) mean soil moisture for the nine soil probes, i.e., CH1 (left-hand side, peripheral) to CH9 (right-hand side, interior); the error bars indicate ±1 SE (b) quartile regression of the soil moisture readings; from top to bottom, the red lines indicate the 75th quartile, the median and the 25th quartile (c) coefficient of variation (cv) for the nine probes; CV for the whole dataset is indicated with filled black points; a subset of values from May-September 2021 (i.e., from the growing season) is indicated with red circles.
F I G U R E 7 Soil moisture data from the probe nearest the rofabard scarp (CH1, red line) compared with equivalent measurements from the probe furthest from the scarp (CH9, blue line). Riedel et al. (2005) studied graminoids in the Alaskan tundra and demonstrated a positive linear relationship between NDVI and biomass. Climatic amelioration in tundra biomes is frequently associated with a marked 'greening', detectable by NDVI . In contrast, stressful conditions such as drought can lead to spectral 'browning' (Bjerke et al., 2014). We therefore attribute the increase in mean NDVI with distance to increased plant vitality.
While spectral vegetation indices like NDVI can contribute to understanding of biotic parameters, they are influenced by many additional factors-notably plant community composition and structure-that could confound interpretation. Plant community composition and structure (i.e., growth form and biomass) can influence the interpretation of the NDVI signal (Goswami et al., 2015). High-latitude vegetation is characteristically patchy, and plant community composition and structure F I G U R E 8 Soil moisture data for selected 'rainfall events' between October 2020 and September 2021. The peripheral signal is from CH1, the probe nearest to the erosion scarp; the interior signal is from CH9, furthest from the erosion patch. Note the peakiness of the peripheral signal and the somewhat slower drying on the interior site.
can vary markedly over small spatial scales (e.g., the difference between lichen cover and adjacent deciduous shrubs). Moss cover-a common component of highlatitude vegetation-obscures the relationship between NDVI and biomass: in an ecosystem where bryophytes dominate, NDVI can only explain a small proportion of the variance in biomass . Furthermore, the relationship between NDVI biomass-which is linear up to an NDVI value of $0.8-becomes exponential at high NDVI readings, complicating interpretation (Aalto et al., 2021;Goswami et al., 2015).
We addressed these issues with a vegetation survey that indicated homogeneous cover of graminoids and forbs, with only sparse moss cover and no upright shrubs (Table S1). Vegetation cover was essentially homogeneous, with no major spatial variations in plant community composition and structure. NDVI values were <0.8, so saturation at high levels should not have been a problem. The rofabard tops were also more-or-less flat (Figure 3c, Figure S3), meaning that differences were not driven by microtopographic variation. We therefore believe that our NDVI values accurately reflect the vitality of the vegetation, rather than variations in vegetation cover or topography.
Our drone survey was a snapshot of conditions in the latter part of the growing season. We would expect NDVI patterns to vary temporally as well as spatially. We reasoned that the impact of soil moisture conditions would be most apparent well into the growing season, hence the timing of our survey. However, it is possible that the stress response of NDVI varies with the season; for example, soil moisture availability may be more or less critical according to phenology and is also likely to vary from species to species. To fully understand these factors, we would have to run repeat drone surveys at different times of the year, coupled with detailed studies of species-level NDVI.
In addition to moisture, it is likely there are other stresses for plants growing near rofabard scarps. For example, these locations are subject to enhanced fluxes of sediment from adjacent eroded areas (Streeter & Dugmore, 2013). Plant stems on the top of rofabard scarps might also be exposed to elevated mechanical stresses and abrasion from sediment carried by high winds. Even sheep using the rofabard to shelter from poor weather might damage exposed roots (Arnalds, 2015). In our opinion, the plant species characteristic of this environment are resilient in the face of these everyday stresses, and their overall impact is likely to be minor, compared to the effect of moisture stress.

| Soil moisture
Soil moisture levels are known to be highly variable in time and space. However, we appear to have detected systematic variation in likely soil moisture stress along our transect. Locations close to the erosion patch were generally dry but subject to highly variable conditions. Locations further away experienced wetter, less variable conditions ( Figure 6). We suspect that the dryness of locations close to the rofabard scarp is due to evaporative losses to the atmosphere. Like many high-latitude locations, Iceland is extremely windy. The geometry of the rofabard scarps means that they catch the wind, and the exposed soil of the scarp lacks vegetation cover, which would shade it from the sun and wind. A previous study demonstrated that the scarp we focussed on was exposed to prevailing winds and is actively eroding (Dugmore et al., 2020). The scarp is also likely to influence the variability of soil moisture. The bare soil readily absorbs moisture, and the sparse vegetation cover does not intercept much rainfall. Hence, the soil adjacent to the scarp rapidly wets and dries, as demonstrated by our observations of inferred rainfall events (Figure 8). Sharp peaks in soil moisture in peripheral locations would be expected during rain showers, when moisture enters the soil laterally due to driving rain. Interior locations would not experience this effect.

| Implications
Our work has implications for understanding highlatitude soil erosion in Iceland and beyond. In an Icelandic context, the moisture gradients observed here provide a mechanism to explain the 'early warning signals' noted by Streeter and Dugmore (2013) in their study of changes in tephra layer thickness towards the edge of rofabard escarpments. The variations in tephra layers observed towards the edge of a rofabard are a reflection of vegetation changes (Cutler, Bailey, et al., 2016;Cutler, Shears, et al., 2016), which in turn are likely to be a reflection of moisture stress. The form and lateral erosion rate of rofabards are sensitive to the thickness of the soil mantle (Arnalds, 2015). Differences in soil thickness may affect the intensity of the soil moisture gradient at the boundary of vegetated areas and the erosion scarp. Soil thickness in Iceland is related to the frequency and magnitude of volcanic ash deposition as well as other sediment sources such as glaciers and soil eroded from elsewhere. Areas in the axial volcanic zone of Iceland, which stretches from the southwest to the northeast of the island, typically have the thickest soil mantles (up to 6 m in some locations in the south). Hamragarðaheiði is located within Iceland's volcanic zone, so it has a moderate soil mantle thickness. Thicker soil cover results in steeper and larger rofabard escarpments, which are generally more dynamic (Arnalds, 2015). Tall, steep scarps could have larger soil moisture gradients and greater moisture variability than the smaller, shallower scarps found outside the most volcanically active areas of Iceland. The lower overall height and lower prevalence of scarps in thinner mantled areas may result in more muted wetting and drying cycles than we observed. Further studies should explore how this edge effect may manifest itself in different soil thicknesses and climatic conditions within Iceland and in other similar Arctic environments.
Soil erosion in Iceland and Greenland is characteristically patchy (Dugmore et al., 2009;Heindel et al., 2015). The initiation of erosion is usually conceptualised as a threshold-crossing event (Kéfi et al., 2007;Streeter & Dugmore, 2013), driven by a large-scale deterioration in environmental factors. Global (in the sense of whole-site) factors, e.g., climate (quantity, timing and type of precipitation, windiness, etc.), no doubt play a role in determining whether a landscape will erode or not. However, the spatial structure of the landforms that subsequently develop is probably determined by small-scale processes. Previous research on the size and spatial distribution of Icelandic erosion patches has shown that (1) once initiated, isolated 'spots' of erosion have the potential to expand and coalesce (Ólafsd ottir & Guđmundsson, 2002) and (2) small erosion patches cluster around the margins of larger features (Streeter & Cutler, 2020). The research we present here provides a biogeomorphological mechanism-localised moisture stress-that might explain the spatially clustered distribution of small erosion spots. The erodibility of the margins is likely largely determined by the degree of stress experienced by individual plants, which is highly localised, although factors, such as soil structure (texture, presence of tephra layers, etc.), will also be a factor. On a landscape scale, the area under moisture stress is determined by the spatial structure (particularly edge length) of existing patches. It therefore appears that localised processes operating at the margins of large patches have the potential to drive landscape-scale patterns of erosion.
Our study of soil moisture prioritised spatial resolution over spatial coverage. Soil moisture sensors and data loggers are expensive, so trade-offs between extent and resolution are inevitable. To fully understand soil moisture dynamics on our sites, we would need replicate transects; sadly, such a survey was beyond our resources. As well as the lateral variation that we measured, soil moisture will also vary vertically. This would be significant if plants have different rooting depths, e.g., if the vegetation included a significant component of upright shrubs. If this were true on our sites, we would have to deploy sensors through the soil profile to fully understand moisture stress. However, the vegetation we observed was dominated by short-statured forbs and graminoids; similarities in growth form, functional type and overall stature indicated that the plants on our sites have similarprobably rather shallow-root structures. We are therefore confident that our moisture probes captured conditions relevant to the growth of plants on our sites.

| CONCLUSIONS
Understanding high-latitude soil erosion is important because these regions will be subject to abrupt climate change in the coming decades. Many of these changes will impact the soil environment. For example, climate change is likely to bring increased and more variable precipitation to many high-latitude locations, along with enhanced evapotranspiration and changes to the duration and extent of snow-lie, all of which will impact the soil moisture environment (Masson-Delmotte et al., 2021). Under current predictions, high-latitude soil moisture is predicted to become more variable (Berg et al., 2017). There is already evidence that increased instances of hot, dry days have increased soil erosion in Iceland (Owczarek et al., 2022). Higher variability in soil moisture will lead to increased moisture stress, compounded by reduced organic layer depths and reduced water retention capacity . More frequent extreme weather-particularly rainstorms and periods of high winds-will increase erosive forces.
Our study provides a snapshot of the relationship between soil moisture, plant growth and erosion that should be applicable to other high-latitude regions, although the exact pattern will be sensitive to factors such as soil type. Most importantly, our results have relevance for understanding the spatial characteristics of high-latitude soil erosion. NDVI values and soil moisture are variable in time and space. However, in our study, both factors varied with distance from existing eroded areas. The margins of erosion patches have a stressful soil environment due to exposure to the atmosphere. The vegetation in these locations grows less vigorously, and the exposed soil becomes more vulnerable to erosion by wind and driving rain, leading to expansion of the erosion. If these conditions hold more generally, they may represent a feedback mechanism that facilitates the propagation of soil erosion in high-latitude regions, particularly those with similar andosols (e.g., the Kamchatka Peninsula and the Aleutian Islands).