Geographic object‐based image analysis (GEOBIA) of the distribution and characteristics of aeolian sand dunes in Arctic Sweden

Current climate change in the Arctic is unprecedented in the instrumental record, with profound consequences for the environment and landscape. In Arctic Sweden, aeolian sand dunes have been impacted by climatic changes since their initial formation after the retreat of the last glacial ice sheet. Dune type, location and orientation can therefore be used to explore past wind patterns and landscape destabilisation in this sensitive area. However, knowledge of the full spatial extent and characteristics of these dunes is limited by their inaccessibility and dense vegetation cover. Geographic object‐based image analysis (GEOBIA) permits the semi‐automatic creation of reproducible parameter‐based objects and can be an appropriate means to systematically and spatially map these dunes remotely. Here, a digital elevation model (DEM) and its derivatives, such as slope and curvature, were segmented in a GEOBIA context, enabling the identification and mapping of aeolian sand dunes in Arctic Sweden. Analysis of the GEOBIA‐derived and expert‐accepted polygons affirms the prevalence of parabolic dune type and reveals the coexistence of simple dunes with large coalesced systems. Furthermore, mapped dune orientations and relationships to other geomorphological features were used to explore past wind directions and to identify sediment sources as well as the reasons for sand availability. The results indicate that most dune systems in Arctic Sweden were initially supplied by glaciofluvial and fluvial disturbances of sandy esker systems. Topographic control of wind direction is the dominant influence on dune orientation. Further, our approach shows that analysing the GEOBIA‐derived dune objects in their geomorphological context paves the way for successfully investigating aeolian sand dune location, type and orientation in Arctic Sweden, thereby facilitating the understanding of post‐glacial landscape (in)stability and evolution in the area.

tion after the retreat of the last glacial ice sheet. Dune type, location and orientation can therefore be used to explore past wind patterns and landscape destabilisation in this sensitive area. However, knowledge of the full spatial extent and characteristics of these dunes is limited by their inaccessibility and dense vegetation cover. Geographic object-based image analysis (GEOBIA) permits the semi-automatic creation of reproducible parameter-based objects and can be an appropriate means to systematically and spatially map these dunes remotely. Here, a digital elevation model (DEM) and its derivatives, such as slope and curvature, were segmented in a GEOBIA context, enabling the identification and mapping of aeolian sand dunes in Arctic Sweden. Analysis of the GEOBIA-derived and expert-accepted polygons affirms the prevalence of parabolic dune type and reveals the coexistence of simple dunes with large coalesced systems. Furthermore, mapped dune orientations and relationships to other geomorphological features were used to explore past wind directions and to identify sediment sources as well as the reasons for sand availability. The results indicate that most dune systems in Arctic Sweden were initially supplied by glaciofluvial and fluvial disturbances of sandy esker systems. Topographic control of wind direction is the dominant influence on dune orientation. Further, our approach shows that analysing the GEOBIA-derived dune objects in their geomorphological context paves the way for successfully investigating aeolian sand dune location, type and orientation in Arctic Sweden, thereby facilitating the understanding of post-glacial landscape (in)stability and evolution in the area. Climate change has been framed as the defining issue of our time. 1 With the Arctic changing at an unpreceded speed and with knock-on effects for global climate, 2,3 there is a pressing need to improve the understanding of past Arctic environments and landscapes to enhance the robustness of future projections. One archive of past Arctic environment and surface processes is aeolian sand dunes, which are common in Arctic Fennoscandia, especially in Sweden and Finland. 4 As periglacial landforms, dunes in Arctic Sweden tend to be parabolic in type, are part of complex sedimentary associations and coexist with other periglacial landforms, such as eskers and drumlins. Cold-climate dune formation and movement is forced by agents such as wind, temperature and moisture, often via disturbance or changes to vegetation.
Variability in dune morphology and reworking is thus a result of a combination of these factors, [5][6][7] and cold-climate dunes are therefore sensitive indicators of environmental changes. 5,6,8 Aeolian sand dunes in Arctic Sweden are believed to have initially formed within a few hundred to a thousand years following the last deglaciation, 9 which occurred in this area at c. 10.5-9.9 cal kyr BP. 10 However, radiocarbon and luminescence dating studies conducted on dune sediments in Arctic Finland have revealed multiple episodes of reactivation during the Holocene, as well as inconsistencies over the ages of dune cores. 5,[11][12][13][14] The reactivation of Fennoscandian Arctic dunes may be related to climate or human-induced fires, or both, which destabilise anchoring vegetation and facilitate dune movement.
Analysis of dune stratigraphy in Finland has shown numerous charcoal horizons and buried podsols, indicative of a complex history of Holocene movement that most recently has been linked to abrupt climate changes in the North Atlantic. 12 Today, most dunes in the region are stabilised by forest vegetation, especially Scots pine (Pinus sylvestrus) and mountain birch (Betula pubescens subsp. tortuosa), as well as Betula nana, Calluna vulgaris, Empetrum nigrum and Vaccinium vitisidaea dwarf shrubs above the tree line. Pine invasion into Arctic Finland occurred approximately 9,000 years ago, 15 with pine considered as initial colonising species in Fennoscandia (see 12,16 ). Dunes in some areas of Arctic Fennoscandia often exhibit currently active blowouts, indicating periodic, ongoing dune disturbances. 12,17 In addition to recording these periodic disturbances, parabolic dunes can be used as indicators of prevalent sand-transporting wind directions during dune formation of the parabolic form. 18,19 Previous work on Arctic Fennoscandian dunes has tended to focus on dunes in Finland, and although some links to climate have been proposed, the details of this relationship are debated. Recent research on dunes in Arctic Sweden is lacking. In Sweden, aeolian research was strong in the 1920s (e.g. [20][21][22] ) when a school with a focus on aeolian geomorphology formed in Uppsala. 19 A pause in publications occurred afterwards, with a few exceptions, 4,6,23,24 primarily focused on mapping. Recently, interest in investigating Swedish dunes has grown (e.g. 8,25,26 ), although with a primary focus on aeolian sand dunes in central and southern Sweden. A first step in utilising Arctic Swedish aeolian sand dunes to infer climate and landscape development is accurate, consistent, systematic, and has detailed mapping and characterisation. 27 Gaining information about sand dune location and form allows the investigation of wind patterns dominating during dune formation and modification, 18,19 as well as the origins of sand supply and overall driving mechanisms of formation. 28 To date, Arctic Swedish sand dune locations are solely available as point features based on a combination of field data, aerial imagery interpretation 29 and hand-drawn geomorphological maps. 4,23 However, given the large, often inaccessible areas these dunes are found in, and the at-times dense vegetation cover, such point mapping is likely to be incomplete and yields limited spatial information.
While satellite missions such as Landsat have rendered mapping aeolian sand dunes based on remotely sensed data possible, the increased availability of and accessibility to spaceborne data has led to a large variety of quantitative and qualitative analyses of aeolian landforms ( 30,31 ), featuring manual mapping approaches (e.g. 32,33 ) as well as semi-automatic and automatic ones (e.g. [34][35][36]. Geographic object-based image analysis (GEOBIA), a semiautomatic mapping approach, permits the creation of reproducible parameter-based polygons, referred to as objects, through segmentation and classification of spatial data. [37][38][39][40][41] In a geomorphological sense, GEOBIA allows for the delimitation of landforms as objects, e.g., alpine landforms, 42 synthetic drumlins 43 and aeolian landforms. 34 Focusing on dunes as objects that represent landforms rather than points, which cannot entail geomorphic information such as shape and orientation, allows for a better interpretation of the spatial differentiation pattern of, e.g., their occurrence or form. 44 We apply GEOBIA as a means of semi-automatic mapping of Arctic Swedish aeolian sand dunes using digital elevation data to provide insight into the type, location and orientation of the characteristics of aeolian landforms as displayed today, in an area where GEOBIA has not previously been applied. The data set resulting from GEOBIA analysis and expert-acceptance allows for the interpretation of wind directions during parabolic dune formation, geomorphic associations, sand sources and the reasons for sand availability, in short, the causes of dune formation. We thus aim to explore the applicability of objectbased mapping for aeolian sand dunes in Arctic Sweden to assess wind dynamics, topographic control, sediment routing and the timing of dune formation, with the overall goal of understanding post-glacial landscape (in)stability and evolution in the area.

| STUDY AREA AND REGIONAL SETTING
The main dune fields of Arctic Sweden focused on here lie between 67.7 and 68.5 N in Norrbotten County in Swedish Lapland (part of Sápmi). The area is characterised by higher elevations and steeper slopes in close proximity to the Norwegian border, with the east and south-east of the area characterised by flatter wetlands (Figure 1).
Today's landscape of Arctic Sweden shows the extensive impact of previous glaciation (e.g. 10 ). Dead ice topography, eskers and drumlin orientations indicate dominant Quaternary ice flow directions towards the south-east. 10,45 Non-flow-parallel drumlins in parts of the area are interpreted as part of a relict, older pre-late last glacial landscape. 45 Assemblages of eskers in north-east Sweden highlight the position of the last remnants of the former ice cover in the northern Scandinavian mountains, 10 and the former presence of ice still impacts the region through a modern isostatic rebound $10 mm yr À1 . 46 Today, Arctic Sweden is characterised by a mean geostrophic wind speed of 8 to 9 m/s and a predominant north-westerly wind direction. 47 Wind speed and direction data from weather stations reveal a prevalence of southerly, northerly and north-westerly winds for Muodoslompolo, as well as strong westerly, south-westerly and southerly winds for Karesuando (see online appendix 1 and 2).

| Regional dune morphology and dune type
Dunes in Arctic Sweden are predominantly characterised by parabolic shapes, 4,6,19,23 which are typical for their location in a periglacial area. 19,48 Characteristics are U-or V-forms, arms pointing upwind and arms being lower than the main crest 19,28,48 (Figure 2). U-or V-forms are well represented by digital elevation models (DEMs) based on highly resolved light detection and ranging (LiDAR) point data from which the impacts of vegetation cover have been removed 50 (Figure 2a). The windward sides of the dunes are more gentle and contrast with a more pronounced, fine material-dominated lee side. 4,28 Although today's predominant dune shape is parabolic, this may not always have been the case during the early phases of post-glacial dune activity; dominant form today is a function of the emerging form during initial stabilisation and the subsequent effects of reworking events. 28 The parabolic aeolian sand dunes of Arctic Sweden tend to be located close to sediment sources, such as eskers, glaciofluvial deltas, outwash plains and glacial drainage channels, and are often located in valleys. 4 The presence of Arctic Swedish dunes on former glaciofluvial deltas is less prevalent than further south in Sweden. 23 as well as blowouts similar to those described for Arctic Finnish dunes above, indicating dune reactivation and reworking. 23,52 In neighbouring Arctic Finland, stratigraphic evidence of fires and dune reworking suggests a long history of blowout activity during the Holocene. 5,12,14 The precise cause of this reworking is unclear, with some authors evoking natural climate changes, 12 while others suggest that more recent events are anthropogenically forced. 23,53 F I G U R E 1 Map of the study area located in Arctic Sweden (part of Sápmi) at the borders of Finland and Norway. (a) shows published point data for sand dune locations along with surrounding quaternary geology and a digital elevation model (DEM). The DEM with a spatial resolution of 5 m is draped over a hillshade layer with an azimuth of 180 and an altitude of 45 . Point locations were georeferenced from Bergqvist 23 and Seppälä 4 as well as extracted from SGU. 29 Sedimentary deposits were extracted from SGU. 29 Figure 7d, location 1). 3D mesh was created in CloudCompare based on LiDAR point data by Lantmäteriet. 49 Vegetation was removed using the cloth simulation filter developed by Zhang et al. 50 Note that z-scale is 1.

| MATERIAL AND METHODS
GEOBIA allows meaningful investigation of context within a raster data set while focusing on objects instead of pixels. The approach is rendered possible by high-resolution data where objects of interest are larger than pixels. 39 Software such as eCognition Developer (Trimble Geospatial) can be used to partition remotely sensed data into segmentation-derived image objects. Each object can be semiautomatically classified considering a range of characteristics such as shape and size, texture, information about the context (neighbourhood) of an object, as well as spectral properties. Multiresolution segmentation, which was used in this study, starts with a seed pixel and creates objects that accentuate intra-segment homogeneity and intersegment heterogeneity, as long as a set of shape and compactness parameters is not exceeded. [37][38][39]41 This study makes use of the similarities of geomorphic mapping theory and segment-based image classification regarding parameter set-up, 41 with the classification of the segments being rule-based, meaning that a set of characteristicstargeting rules is followed to determine potential aeolian dune sites.
Knowledge of and experience in the targeted terrain is essential to this knowledge-based, supervised classification of geomorphological features. 54,55 Working with objects allows for morphology-oriented analysis, whereas pixel-based approaches are limited to spectral information and lead to salt-and-pepper effects, that is, noisy classification results. Object-based results are often considered to be easier to interpret, potentially because GEOBIA mimics human perception. 38,39,56 However, the distinguishability of object-candidates is scale-dependent, a reason why GEOBIA is mainly applied to very high-resolution images. 39

| Deriving suitable input data for the GEOBIA analysis
A LiDAR-derived DEM was used as it simulates the surface morphology of the landforms. 44 The 2 m resolution DEM is part of the Ny Nationell Höjdmodell, which is more precisely referred to as 'GSD-Höjddata, grid 2+', and is georeferenced to SWEREF 99 TM and RH 2000, 57 see Table 1. The high-resolution DEM was resampled to 5 m resolution with the aim of decreasing its size and reducing artefacts.
This resampled DEM was used to create a residual relief separated DEM following the first stage of residual relief separation, as outlined by Hiller and Smith 27 and originally introduced by Wessel. 58 Returning the median height to a central point was favoured to reduce the strong impact of outliers, which would likely occur when returning a mean value. Furthermore, standard curvature and planar slope data sets were derived from the original, non-resampled DEM. The curvature was unilluminated and, despite amplifying noise, 27 is free from bias. 59 The use of slope values allows for an investigation of the land surface that is not prone to azimuth biasing [59][60][61] and is independent of other DEM derivatives. 61 The processed DEM and the curvature and slope data sets served as inputs to the geographic object-based image analysis.

| GEOBIA workflow
The GEOBIA workflow can be divided into segmentation, classification and export stages ( Figure 3). First, the multiresolution segmentation algorithm in eCognition Developer 9.5 was used. A weighting of 3 for the standard curvature data set, 2 for planar slope data and 1 for   23 Seppälä, 4 and extracted from SGU. 29 All classified image objects are visualised in green; dark-blue polygons remain after expert decision. Numbering refers to locations described in the text, as well as shown in Figure 7. White space in the upper-right corner of (d) is due to lack of data for Finland (river constitutes the Swedish-Finnish border). Point and polygon data are visualised on two hillshades (azimuth 135 and 315 , altitude 45 ) to minimise azimuth biasing. All data are projected in SWEREF 99 TM [Colour figure can be viewed at wileyonlinelibrary.com] predominantly c-shaped (e.g. Figure 6b, location 2) and that single polygons occur next to coalesced polygons (e.g. Figure 6a, locations 1 and 2). Displaying the results in a map further shows that polygons that are highly pronounced and well developed in their shape occur next to less-pronounced polygons (Figure 6a, locations 1 and 2).

| Segmentation and classification process
Applying GEOBIA reduces the effort required for time-consuming and expensive field mapping. 66 Furthermore, GEOBIA inherent algorithms produce transparent, consistent and reproducible results, thus reducing human bias. 60,65 Nonetheless, the observer's influence when accepting results needs to be acknowledged, 39,65 as well as the dilemma between ensuring quality and increasing subjectivity inherent when introducing expert decisions in the workflow. 41,54,55 While acknowledging these uncertainties, we infer that the expert-control increases the quality of results based on comparing the mapping results at the target sites (cf. Figure 6, online appendix 4). It delivers a controlled means of insight into the suitability of the segmentation and classification process based on a comparison of object characteristics (cf. Figure 5). Through the sole use of a DEM and its derivatives, planar slope and standard curvature, and by abstaining from the use of optical data to segment aeolian sand dunes in Arctic Sweden, this study eliminates landform shadowing 67  Expert-based selection of classification thresholds can be avoided by using machine learning approaches, 69,70 as exemplified for the classification of aeolian landforms by Fitzsimmons et al. 34 However, limited objectivity must also be acknowledged for machine learning due to expert-chosen training sites, potential uncertainty inherent to the training data (e.g. 71 ) and manually set parameters (e.g. 72 ). We suggest that the expert-checked data sets of the 17 target sites represent valuable validation data for future studies that apply machine learning approaches to mapping aeolian sand dunes in Arctic Sweden.  Figure 6a, location 2, south of the coalesced system), while some dune objects have not been previously mapped as point data (e.g. Figure 6c, location 1).
As the classified dune objects are based on geomorphometric characteristics of aeolian sand dunes, we infer that the semiautomatic mapping, together with the expert-decision, leads to a more

| Geomorphological interpretation of the expert-accepted dune objects
The GEOBIA-derived dune sites near Vittangi, Meraslompolo, Kurrakajärvi and Karesuando are mapped alongside selected Quaternary geology to understand their landscape associations and the reasons behind dune characteristics (Figure 7). Importantly, such interpretations would not be possible without mapping these dunes as objects, as point data cannot convey, e.g., morphology. Besides the benefits, challenges in dune object analysis occur due to natural landform variability and the impact on, e.g., mode slope and mode aspect.
We hypothesise that main dune nose slipfaces are dominant in the mapped polygons as their slope values were targeted during GEOBIA analysis and each object's most dominant value (mode) was extracted.
Despite this representation of dune nose slipfaces, the mode aspect derived from these slope values can offer partial help only in determining the true orientation of parabolic sand dunes in the study area.
In addition to dune nose slipfaces, dune arm slipfaces impact the mode aspect data set (arm strike) and, therefore, the calculated mode aspect direction. This highlights the need for additional geomorpho- forest or forest stands, and indicates that considerable numbers of dunes must also exist in the birch forest and tundra zones.

| Wind directions inferred from dune polygons
The mapped Arctic Swedish aeolian sand dunes show a tendency to a south-east orientation, partially supported by prevalent dune object mode aspect values (cf. Figures 4b, 6, 7, and online appendix 3). As classified as part of a relict glacial landscape creates valleys aligning north-west to south-east. The dunes in this area are also oriented with the parabola facing south-east, suggesting that these valleys act to funnel the wind, potentially increasing the transport capacity of the airflow. 76 The dune complex near Kurrakajärvi (Figure 7c) is located so that winds driving its formation must have travelled around the fractured bedrock west of the dune complex, resulting in esker erosion, sediment transport and sand deposition. Importantly, this dune field formed behind a local topographic obstacle while sand supply is present along the south-west to north-east expanding esker system. Due to the large size of the dune complex, it appears that the wind transport capacity must have been high and/or consistent over a longer time period during dune formation to be able to transport the volume of sand deposited. Hesp and Smyth 73 highlight that airflow around a curved object causes marginal and downwind bed erosion, with vortices originating from wind deflection by the obstacle leading to a local increase in transport capacity to an obstacle's lee. 74 With respect to the dune field west of Kurrakajärvi, we hypothesise that the existence of fractured bedrock outcrop has likely contributed to dune formation as it locally increased wind speeds, thereby enabling transportation of sediment away from the esker. This mode of dune formation contrasts with the mechanism proposed from obstacles slowing down winds, thus forcing dune formation by decreasing wind transport capacity. 19,28 Acknowledging that parabolic dune orientation is a product of climatological as well as topographic factors, we infer that topographic obstacles play a major role in controlling sediment transporting winds in Arctic Sweden and that local topography has a major impact on dune orientation and position.

| Sediment routing inferred from dune polygons
Sediment supply control is a crucial factor for cold-climate sand dune formation, specifically with regard to dune size. 77 interaction with the esker system (cf. Figure 7). Thus, mapping the dune polygons together with the surrounding Quaternary geology demonstrates the (glacio)fluvial control of geomorphic and sedimentary systems on sand availability for dune formation in Arctic Sweden.
After initial deposition, larger parabolic dunes appear to represent a sediment source to secondary dunes located downwind, e.g., in the case of multiple coalesced arcs of dunes and single dunes located east of the nose of the major Kurrakajärvi dune system (Figure 7c, location 1). Based on signs of recent erosion of the Kurrakajärvi complex perceived in the field (cf. Figure 2c), as well as reported for aeolian sand dunes in Arctic Sweden by Bergqvist, 23 we hypothesise that the depositional landforms act as sediment sources once reactivated, rather than downwind dunes being deposited earlier than the main complex.
We suggest direct dating to clarify the secondary status of the dunes by determining their age in relation to the main dune complex.

| Timing of dune stabilisation inferred from dune polygons
With dune type being the result of multiple factors, 28  We infer that reduced sand availability, wind capacity and/or depositional processes occurring over shorter timescales lead to less material being transported away from the esker system, thus, producing lesspronounced dunes for location 1. In contrast, the extensive dunes in location 2 suggest rather high sand availability, together with strong wind capacity and/or deposition over a longer timescale. Thus, the differences in the occurrence of dune shapes are inferred to be a product of process rates or timescales. Our mapping, therefore, provides the opportunity for a detailed reconstruction of post-glacial landscape evolution, in combination with direct dating of landforms.

| CONCLUSIONS
Systematic spatial analysis of Arctic Swedish aeolian sand dunes using GEOBIA allows for reproducible mapping on a larger scale, combining expert knowledge and semi-automation. This demonstrates that expert control, an essential characteristic of traditional geomorphological mapping, and the use of innovative and machine-based approaches such as GEOBIA are not mutually exclusive. As such, combining these complementary approaches contributes to a more robust outcome. Semi-automated detection as well as ruleset transferability allow for an increase of the spatial scale in a computationally nonextensive manner, while expert decision ensures the quality and suitability of the segmentation and classification process.
This large-scale analysis resulted in a higher number of mapped dunes than before, and more precise mapping and form analysis of these overwhelmingly parabolic dunes. Furthermore, the approach more reliably detects less-pronounced aeolian sand dunes in the study area and prevents misinterpretation of, e.g., coalesced systems, as erro- show that esker systems act as dominant sediment sources, with the potential for major dunes and dune complexes acting as sediment sources to secondary smaller dunes, once reactivated. Glaciofluvial and fluvial forcing of sand availability appears likely in all studied cases, either directly or indirectly via disturbance of esker systems, and implies that initial dune formation is strongly linked to landscape processes during and immediately after deglaciation. However, to what extent these dunes are reworked during the Holocene remains unclear. The detailed mapping presented here provides a framework for which this can be tested, e.g., via a coupled analysis of the spatial associations of multiple dune ridges with respect to sediment sources and information derived from direct age dating of dune movement and stability. The results of this study on the timing of dune stabilisation further highlight the need to investigate Arctic Swedish aeolian sand dunes as objects to understand local and regional landscape associations and evolution. Such large-scale analysis can help decode past landscape (in)stability and frame understanding of potential future changes in an area that already today is outpacing predictions of global climate change impact.