The CB region was surveyed as part of an aircraft based campaign in 2021 that delivered high-resolution multi-spectral imagery of the permafrost landscapes using the Modular Aerial Camera System (MACS) . The entire survey, conducted on 25 June 2021, covers approximately 25.22 km2. The data collection involved multiple flight lines, with the aircraft maintaining an altitude between approximately 1490 and 1510 m above ground level. The imagery captured by the MACS system was processed into four-band orthophotos (blue, green, red, and near-infrared) and DSMs both featuring a spatial resolution of 20 cm. Data post-processing was performed using photogrammetric software to produce high-resolution orthomosaics . In this study, a selected part from tile 11-2 of the CB sub-project 1 RGB orthophoto was used as the study region. The area of the study region corresponds to 0.714 km2 see.

The BH region was surveyed using the DJI Mini 2 drone as a Citizen Science activity with students from the Moose Kerr School in Aklavik. The survey was conducted on 24 September 2022, with the aim of creating a drone image based orthomosaic. The dataset consists of 3557 individual Digital Negative (DNG) images with RGB color channels captured at 120 m flight height resulting in an average resolution of 4.97 cm per pixel. Later resampling in the course of photogrammetric post-processing resulted in a resolution of about 10 cm per pixel, covering an area of 1.58 km2. The Unoccupied Aerial Vehicle (UAV) system was equipped with a camera using a 12 megapixel CMOS sensor. To enhance the accuracy of the DSMs, we employed a novel spiral flight pattern developed within the UndercoverEisAgenten project and presented in . The images were post-processed using Agisoft Metashape 1.8.4, generating sparse and dense point clouds, DSMs, and orthomosaics, with altitude data adjusted using the local 2 m resolution ArcticDEM for improved vertical accuracy. The dataset is openly available here .

The remote sensing data were analyzed by volunteers, primarily as part of workshops at schools and universities organized as mapping events (“mapathons”). The participants consisted of university level Geography students in Germany (two mapping events) and the Netherlands (one mapping event) and school students in grades 7–12 (approximate age range 12–18 years, seven mapping events) in German Higher Secondary Education (Gymnasium). Contributions to the mapping projects were made from a total number of 105 distinct user accounts.

All mapping events were preceded by an introduction adapted to the level of expected prior knowledge of the target group that comprised information on the topic of permafrost thawing, the formation of ice-wedge polygons, and the consequences of permafrost thawing for the environment and infrastructure. Subsequently, the students participated in mapping activities. The mapping projects differed between the mapping events in regard to area of interest (subsets of the two study regions) and task format. Following the micro-mapping approach, the larger areas of interest were split into smaller regularly shaped and equally sized micro-tasks. Participants mapped groups of spatially adjacent tasks one after another until either the area of interest was completely mapped or the mapping event ended.

While several task designs were tested with different audiences, this study is based on the output of point digitization tasks (see Appendix ). In the context of this task design, participants were instructed to mark the approximate centroids of ice-wedge polygons within the respective boundaries of each task area. Such digitization tasks aim to produce digital geographic objects (vector features) using given georeferenced image data . In the crowd-sourced mapping application, the participants could interact with a web map by placing, modifying and deleting point markers within the task area boundaries.

Classification tasks of geographic crowd-sourcing – usually consisting in enriching georeferenced images with labels – are considered the task type with the lowest level of spatial cognitive complexity . However, classifying ice-wedge polygons on aerial imagery of the Arctic surface has proven to be a considerably more challenging task for volunteer contributors than e.g. the detection of buildings . Our further experimentation demonstrated that point digitization tasks allowed volunteers to detect ice-wedge polygons with higher agreement (as an intrinsic indicator for accuracy) and at the same time in a more time-efficient manner. Beyond information on the presence of ice-wedge polygons (as the output of classification tasks), point digitization of polygon centers allows to quantify and locate individual polygon structures. While line digitization of ice-wedge polygons (i.e. the tracing of polygon outlines), on the other hand, would provide additional information on the polygon sizes and shapes, initial trials have shown that this task type overwhelms volunteer contributors and reduces the area to be mapped compared to point digitization. Assigning volunteer contributors with less difficult point digitization tasks while deriving further geomorphological and hydrological information through the network reconstruction method described in Sect.  was thus identified as a promising task design to monitoring Arctic permafrost with VGI.

For quality assurance of crowd-sourced data, it is common practice that multiple contributors map the same area. The individual contributions are then compared and aggregated to a collective result . Here, each micro-task was assigned to more than one contributor, each of them producing an individual point data set for the respective task area. These individual task results were merged into comprehensive vector point data sets covering the two study regions. To derive an “aggregated” result data set from the individual task results, it was necessary to cluster the point markings (see Sect. ) to match the identified polygon centers before further analysis.

Along with the VGI data, an additional set of ice-wedge polygon centroids in the same two study regions was contributed by experts for the purpose of quality assessment. The study regions were analyzed by three authors of this paper analogously to the procedure of the VGI data, but in this case using the GIS software QGIS. It is referred to as the expert data set. For further validation of the Voronoi polygons in the reconstructed networks, a set of “ground truth” reference polygons were created through manual digitization by experts for subsets of the two study regions.

To derive the target variables describing the geomorphological and hydrological characteristics of polygonal tundra, a specific multi-stage workflow was applied on both the volunteer-contributed and the expert-contributed ice-wedge polygon centroids (Fig. ). In contrast to the expert contributions, multiple volunteer contributions were collected for the same micro-tasks in order to ensure the quality of the results. It was thus necessary to aggregate the individual volunteer contributions of ice-wedge polygon centroids into average results, i.e. the mean locations of ice-wedge polygon centroids. To aggregate the individual results, clusters of the contributed ice-wedge polygon centroids were identified using DBSCAN (Density-Based Spatial Clustering of Applications with Noise), a density-based clustering algorithm originally developed by and the mean of each cluster was used for further analysis. One-point clusters were excluded, as the aim here was to ensure that an ice-wedge polygon was recognized by at least two volunteers. The centroids of the resulting clusters were considered as the crowd-contributed locations of ice-wedge polygon centroids.

To establish the number of volunteer contributors needed to map the same region, we compared the positional accuracy of volunteer-contributed ice-wedge polygon centroids by cluster size (i.e. number of volunteer contributions establishing the centroid location) using a pairwise Tukey Honestly Significant Difference (HSD) test as implemented by the Python library statsmodel .

A second step of clustering needed to be applied on both the expert-contributed and the aggregated volunteer-contributed centroids to delineate ice-wedge polygon sub-networks, as it cannot be assumed that a study region is characterized by only a single connected ice-wedge polygon network. A combination of Delauny triangulation and alpha shape was used to identify the sub-networks in both study regions. An alpha value of 0.067 was chosen for this application. These steps were applied consistently across both research regions, and thus have the potential for replication in other contexts.

Voronoi diagrams were generated from both the expert-contributed and the aggregated crowd-sourced ice-wedge polygon centroids to reconstruct the polygonal ice-wedge networks. and have demonstrated that automatically derived Thiessen polygons can represent ice-wedge networks with sufficient accuracy i.e. a goodness of fit of R2=0.84 for the linear regression values of manually mapped polygon sizes against Thiessen polygon sizes according to. In our workflow we made use of the relationship between Delauny triangulation and Voronoi diagrams to derive polygon networks from the individual ice-wedge polygon centroids within each sub-network. Ice-wedge polygon centroids at the edge of each (sub-) network were omitted as they would generate incorrect Thiessen polygons due to the lack of neighbors.

To assess the accuracy of the polygons generated from points digitized both by volunteers and experts through application of the Voronoi method, we compared them against a set of reference polygons whose outlines (not: centroids) were manually digitized by experts (see Appendix ), considered as ground truth. These reference polygons were created by experts through visual interpretation of the same high-resolution aerial imagery, ensuring a high level of accuracy in delineating individual ice-wedge polygons. The validation process involved the calculation of four key metrics being precision, recall, F1-score, and Intersection over Union (IoU). Precision measures the proportion of correctly identified polygons among all polygons generated by the Voronoi method. Recall quantifies the proportion of correctly identified polygons out of all true polygons in the reference data. The F1-score provides a harmonic mean of precision and recall, offering a balanced measure of overall accuracy. IoU calculates the ratio of the intersection area to the union area between the predicted Voronoi polygon and its corresponding reference polygon reflecting the degree of overlap.

As the last step, geomorphological and hydrological parameters were derived from the network graphs of the two entire study regions and of subsets for the two regions used in the comparison with reference polygons. Geomorphological parameters include polygon area, perimeter, and distance to nearest neighbor. The latter describes the Euclidean distance from the center of each polygon to the centers of directly neighboring polygons per sub-network. The results were averaged for each polygon center. The hydrological properties were derived using the Python package NetworkX which has already been successfully used for the delineation of ice-wedge networks by . In this study, NetworkX was used to calculate betweenness centrality, which quantifies the centrality of troughs within the ice-wedge network. The betweenness centrality is a measure of the importance of individual troughs for maintaining the flow of water within the network . Thus, a high betweenness centrality indicates troughs that are likely to feature an increased water discharge and could therefore be more susceptible to erosion.

For the CB data set, 21 volunteers mapped 10 337 points over five calendar days. 88 % of the points were mapped during two school visits (Fig. ). The remaining 12 % were mapped individually outside of any organized event. The aerial image illustrates that the ice-wedge polygons are evenly distributed over the study region, as are the contributed ice-wedge polygon centroids.

For the BH data set, 86 volunteers participated in the mapping process. A total number of 7878 points (see Fig. ) were mapped over 29 calendar days. 76.61 % of the points were contributed during one of the mapping events at three different schools and one university. Most points were contributed by students from Higher Secondary Education (Gymnasium). The individual events resulted in different numbers of digitized points. The events were organized independent from each other at different locations. The exact format of each mapping event was slightly different, as was the number of participants. About one quarter of the points (23.39 %) were not mapped during any mapping event. In contrast to the CB study region, the aerial imagery clearly shows that ice-wedge polygons are distributed in spatial clusters over the BH study region. Accordingly, the contributed points form visually noticeable clusters within the study region. Furthermore, in both regions, the points contributed by multiple volunteers visibly form smaller clusters within the observable boundaries of ice-wedge polygons (see inset maps of Fig. ). This indicates that different volunteers contributing to the same mapping micro-tasks effectively identified the same ice-wedge polygons and approximately the same centroid locations.

One concern regarding VGI is the data quality and its fitness for a specific purpose . The purpose of generating the dataset of volunteer-contributed ice-wedge polygon centroids is to derive geomorphological and hydrological properties of the polygon networks in the two study regions. The dataset's fitness for purpose depends on (i) the feature completeness of the ice-wedge polygon centroids and (ii) their positional accuracy. Errors such as commission, omission, as well as misplacement of the centroids will influence the accuracy of the reconstructed networks and, consequently, of the derived properties.

In absence of ground truth data captured in situ, these two relevant dimensions of data quality are compared to a dataset of expert-contributed ice-wedge polygon centroids (see Sect. ). While the expert-contributed data cannot be considered ground truth in the narrower sense, the extent of the deviation between volunteer- and expert-contributed data is expected to provide an indication of the data quality.

The feature completeness can be assessed by comparing the total number of ice-wedge polygon centroids in the volunteer and expert datasets. For the CB region, the number of clustered volunteer-contributed centroids (1710) amounts to 88.74 % of the number of expert-contributed centroids (1927). In the BH region, however, clustered volunteer-contributed centroids (769) only amount to 70.81 % of the number of expert-contributed data (1086). The lower completeness of the volunteer-contributed dataset in BH, assessed against expert data, can be explained by the difference in the configuration of the polygon networks. Volunteers particularly often omitted ice-wedge polygons in areas at the borders of networks, and in smaller sub-networks (that characterize the BH region, see Sect. ), whereas the single large network of CB as well as the larger sub-networks in BH are better represented.

Regarding the positional accuracy of the ice-wedge polygon centroid, the clustered volunteer-contributed centroids can be assessed by the deviation from the nearest expert-contributed centroid in space. For both study regions, the distributions of distances of all volunteer-contributed centroids from their nearest expert-contributed neighbor are clearly right-skewed, with rather small deviations for most of the centroids (Fig. ): The median distance is 1.29 m (mean: 1.56 m) in CB and 1.38 m (mean: 2.66 m) in BH. For comparison: the approximate median distance between ice-wedge polygon centroids, based on the aggregated volunteer-contributed centroids, was determined as 19.34 m (CB) and 13.15 m (BH) respectively (see Table ).

With regard to the efficiency of the use of VGI in monitoring Arctic permafrost, it is vital to determine the optimal number of volunteers per micro-task required to map ice-wedge polygons with the sufficient quality. From the distributions of the distances between volunteer-contributed centroids and their nearest expert-contributed neighbor per cluster size (i.e., the number of volunteers that contributed to the centroid), it can be seen that, as a general trend, the higher the number of contributors per polygon, the better the positional accuracy (Fig. ). For CB, the positional accuracy of the volunteer-contributed centroids does not improve significantly after surpassing the number of five contributions per polygon (see Table ). For the BH region, the data does not support clear conclusions due to the relatively low number of volunteer-contributed points of medium cluster sizes (see Table ).

Following the generic workflow described in Sect. , ice-wedge polygon networks could be effectively reconstructed from volunteer-contributed ice-wedge polygon centroids in both of the study regions, despite their very different characteristics. For both of the study regions, the resulting networks appear plausible upon visual inspection. In the CB region with its rather evenly distributed ice-wedge polygons visible on the surface across the entire study region, the proposed approach has a single contiguous network of 1490 polygons as an output (see Fig. ).

In the BH region, the reconstructed network is in agreement with the clearly clustered occurrence of ice-wedge polygons in specific regions of the study region. The resulting network consists of 21 different sub-networks with an average number of 16 polygons, ranging from one single polygon to 133 per sub-network (see Fig. ).

Comparing the originally contributed points (see Fig. ) with the resulting network (see Fig. ), it becomes, however, evident that due to the necessary removal of centroids at the edge of networks (see Sect. ), some smaller sub-networks may be entirely omitted.

The results of the accuracy assessment of reconstructed ice-wedge polygons derived both from centroids contributed by volunteers and by experts against polygons digitized by experts (see Sect. ) indicate a reasonably good agreement between the Voronoi polygons and the reference data across the two study regions and data sources (Table ).

In general, the expert-derived polygons exhibit higher values across the accuracy metrics compared to the VGI-derived polygons. In the CB region, a difference in the F1-score is driven by a better recall in the expert set (experts: 0.88, VGI: 0.81), whereas precision and median IoU do not vary substantially between the comparison data sets. This concludes that volunteer mappers failed to detect more ice-wedge polygon centroids than experts, but did not commit more errors in the points that they did identify than the experts. However, the difference in the values is more pronounced, and extends to the precision (experts: 0.72, VGI: 0.65). Despite the overall good performance, the moderate IoU values (CB: 0.71, BH: 0.57) suggest some discrepancies in the shapes and sizes of the polygons. Visual inspection revealed that these discrepancies are more pronounced in regions with heterogeneous landscapes, where the underlying environmental factors influencing ice-wedge polygon morphology are more complex.

The workflow described in Sect.  results in CB being represented as a single network of 1490 ice-wedge polygons in the volunteer dataset and 1679 in the expert dataset (Table ). BH consists of multiple sub-networks (21 in the volunteer dataset and 25 in the dataset mapped by experts), encompassing 1–133 and 1–140 polygons respectively. Both the average number of polygons per sub-network and the number of sub-networks in this study region are lower in the volunteer dataset than in the expert dataset. On the average, ice-wedge polygons in BH are smaller than the ones in CB (Fig. ). The median polygon area is 304 m2 for CB and 147 m2 for BH, with the expert-contributed data set showing a slightly higher median area for BH (160 m2) and a similar one for CB (303 m2). The ice-wedge polygons have a median perimeter of 68 m in CB and 48 m in BH, differing by only 1–2 m from the values in the expert data set. The median distance between neighboring centroids is approximately 19 m for CB and 13 m for BH, which is consistent with the expert dataset. The relative standard deviation values for all parameters are in high agreement between the expert- and volunteer-contributed datasets.

The same statistics computed for subsets of each study region allow for evaluation against reference polygons (described in Sect. ). The relevant subsets covered by the reference polygon data are shown in Appendix  and include about 100 ice-wedge polygons in CB and 150 in BH. The polygon area exhibits the most discrepancies between the three datasets (Table ). Both the volunteer- and expert-contributed datasets overestimate the polygon area by an average of 10–30 m2 while the data set contributed by volunteers is somewhat closer to the reference data set for both study regions. The polygon perimeter is accurately represented in both the expert- and volunteer-contributed datasets for CB with only minor deviations from the reference dataset (mean deviations: 0–2 m). In BH, the perimeters are overestimated by about 3–4 m on average by both volunteers and experts. In both study regions, only minor deviations are observed between the reference, volunteer and expert polygons as concerns the distance between neighboring centroids.

Betweenness centrality provides a measure of the importance of individual channels for water drainage within hydrological networks . Channels with high centrality act as critical connectors, linking otherwise isolated parts of the network and thereby playing a key role in maintaining or enabling overall drainage. In the context of the hydrological function of ice-wedge polygon networks, through segments with high centrality are likely to carry disproportionately large water fluxes, as they concentrate flow. Consequently, they play an important role in the transport of dissolved nutrients and other substances, while also being more susceptible to enhanced erosion and thermokarst development .

In CB, edges with high betweenness centrality values derived from the volunteer-contributed graphs, i.e., troughs potentially more affected by erosion, are located at the center of the network (Fig. ). The polygons delimited by edges with high betweenness centrality values often have relatively large areas and perimeters. When visually comparing these troughs with remote sensing data, they often coincide with areas of surface water occurrence (dark polygon centers).

In BH, the maximum betweenness centrality value is approximately ten times lower than in CB due to the smaller size of the sub-networks (Fig. ). For the same reason, the maximum betweenness centrality values of edges differ within the same region between the sub-networks, and are generally higher in larger sub-networks. Similar to CB, edges of high betweenness centrality are located in the center of the sub-networks. In addition, visual inspection intriguingly shows visible drainage path underneath edges of relatively high betweenness centrality.

Mapping the polygon network provides a primary understanding of polygonal terrain and the results can be used for a variety of applications, such as determining where ice wedges tend to form by comparing their distribution with landscape parameters like slope and surficial deposits . Furthermore, it enables quantitative characterization of ice-wedge polygons conditions and spatial properties, which provides critical insights into past and current landscape shaping and altering processes. For instance, measuring the angles and regularity of the network based on spatial patterns of polygon intersections, helps determine the maturity of ice-wedge networks and how they evolve in different geomorphological settings . Moreover, 3D subsurface models derived from the mapped polygon networks allow for estimating wedge ice volume, a key factor in predicting thermokarst formation as permafrost degrades .

Additionally, once the polygon network has been delineated, it can be used to effectively extract different properties of the polygonal terrain, contributing to the understanding of surface and subsurface processes occurring at the local scale (i.e., intra- or inter-polygons). The microtopography of the polygons can be extracted from high resolution digital elevation models (DEMs) or ground displacement can be measured from InSAR data . These data can inform on the state of the ice wedges, as degradation typically leads to subsidence of the soil above the ice wedge, progressively forming high-centered polygons . Similarly, vegetation and wetness indices can be extracted from spectral imagery to understand surface and subsurface wetness as well as vegetation distribution patterns, which control a broad range of interactions between the ground and the atmosphere e.g.,.

The ability to derive these spatial metrics and characteristics not only improves our understanding of ice-wedge distribution and condition, but also supports broader applications such as extrapolating these findings to other regions and contributing to predictive models of ice-wedge polygon evolution .

Reconstructing the ice-wedge polygon network from volunteer-generated ice-wedge polygon centroids can be a viable alternative to automated workflows, particularly if otherwise necessary data (such as a high resolution digital elevation model) is not available. The high similarity between the reconstructed networks and the polygon networks observed in RGB aerial images, despite significant differences in the ice-wedge polygon network configurations, demonstrates the effectiveness of the presented approach. The combination of low-effort, time-efficient crowd-sourced mapping by point digitizing with the reconstruction of networks can be a suitable solution to derive geomorphological and hydrological properties of polygonal tundra under different landscape conditions. However, the spatial coverage that can be reached by such an approach is limited by (i) the number of participating volunteers, (ii) the number of tasks that each volunteer is willing to complete, (iii) the overall person-hours of volunteer contributions that can be mobilized for the the crowd-sourced mapping process, and (iv) the time needed for the completion of each task. In the case of the volunteer-contributed data used in this study, the participants needed a median time of 2.9 s for each digitized point. Observations in early trials indicated that point digitization was the most time-efficient one of the tested task designs. In a trial session to evaluate different task types, participants needed a median of 2 s per 1000 m2 of area (compared to 2.8 s in a classification task design). It is important to note that the mapping speed may vary strongly with factors such as the visibility of the ice-wedge polygons in the imagery and the mapping experience of the contributors. To make efficient use of volunteer contributions, it is important to optimize the number of participants that each task is assigned to. In Sect. , we identify a number of five contributions per task as sufficient for assuring a high quality of the aggregated crowd-sourced data. This number is in line with previous findings on the optimal number of volunteer contributors to tasks of building classification from aerial imagery .

In relation to the efficiency of the mapping method presented, the additional effort required to recruit volunteers for a crowd-sourcing activity must also be taken into account. Volunteer engagement clearly demands communication, outreach, provision of context information and motivation. In our case, this entailed among other outreach to schools and teachers, creation of teaching material, and the preparation of mapping events, with 13 person months allocated to community involvement and training in the framework of a larger citizen science project. As highlight, recruiting volunteers for crowd-sourced mapping tasks can be particularly challenging when the features of interest are highly specific and unfamiliar to most people, such as ice-wedge polygons. Sustaining volunteer contributions beyond individual mapping events across larger spatial extents is therefore likely to require tailored engagement strategies. Potential approaches include: (a) substantially broadening outreach efforts to educational institutions, while equipping teachers with ready-to-use teaching materials to facilitate in-class mapping sessions; (b) integrating ice-wedge polygon mapping projects into established crowd-sourced mapping platforms with active communities and providing interactive tutorials to support self-guided learning; and (c) incorporating gamification elements such as leaderboards, badges, and activity streaks to foster motivation and sustained participation.

The application for crowd-sourced mapping of ice-wedge polygons was designed to enable contributions by people without particular domain expertise, with some information and context provided through teaching materials. While prior knowledge allowed for providing more in-depth contextualization to geography students at the respective events, the concrete mapping task was designed to depend on general pattern recognition skills. The majority of results were generated by secondary school students. This study does not provide a comparison of quality of results for different user groups from a controlled experimental setting. Such a comparison might be useful to inform any further extension of the volunteer base.

For (volunteer) contributors to be able to accurately identify polygonal structures, the quality of imagery used as a base in crowd-sourced mapping is critical. Factors such as lighting and atmospheric conditions during image acquisition significantly influence the clarity and contrast needed for recognition . In the case of this study, some areas in the BH dataset contain small blurry sections due to the UAV flight geometry and lighting conditions. However, it is well-documented that human interpreters can often discern and infer subtle structures within images even if the image quality is strongly reduced . This ability is providing a distinct advantage over traditional automated methods, which are more dependent on high contrast and clear image conditions for effective structure recognition. With the advancement of modern machine learning techniques in image analysis, such as Convolutional Neural Networks (CNNs) the differences between human and automated structure recognition abilities are increasingly fading e.g.. There is an interest in assessing how these approaches compare to human interpretation in accuracy and reliability e.g.. The VGI method presented here could contribute valuable insights, offering a baseline dataset that could serve both training and validation, potentially further enhancing machine learning models.

With respect to spatial resolution, this study does not experimentally compare mapping outcomes generated from 20 cm (CB) and 10 cm (BH) resolution imagery. It is to be assumed that resolution – at this level – is not a decisive factor for the results. The BH orthomosaic was indeed downsampled from 5 to 10 cm for practical reasons, as the native resolution was deemed excessive for the given mapping task on visual inspection. Nonetheless, access to sufficiently high-resolution imagery remains important: given the typical dimensions of ice-wedge polygons and the width of their delimiting troughs, crowd-sourced mapping of these features would not be feasible with coarser-resolution data.

The assessment in Sect.  demonstrates that the quality, and particularly the completeness of the crowd-sourced ice-wedge polygon centroids depends on the configuration of the network, with more centroids omitted in areas of borders of networks and in smaller sub-networks. This may be explained by the fact that the identification of smaller clusters of ice-wedge polygons within regions with only few features of interest requires a more systematic approach inspecting the imagery of a specific micro-task. In addition, our approach required to omit edge polygons from sub-networks (see Sect. ). For smaller networks, edge removal will result in a larger proportion of the polygons identified by volunteers being removed from the output data, and may even lead to entire sub-networks being omitted. Therefore, results may be more accurate for regions with large contiguous areas of ice-wedge polygons (such as CB) than for regions with dispersed clusters of smaller ice-wedge networks.

The application of the Voronoi method to reconstruct polygon networks from approximate centroids means that the resulting ice-wedge polygon boundaries are inclusive of the troughs. This may partially explain differences both in average area and perimeter, when the reconstructed crowd-sourced polygon boundaries are compared to reference polygon boundaries that were captured exclusive of troughs, as shown in Sect. . In general, our proposed method does not allow for gaining information about trough sizes.

It has been demonstrated that natural crack mosaics in drying clay represent a random tessellation that evolves with repeated wetting and drying cycles towards a Voronoi mosaic that minimizes internal energy . Although there are similarities between the repetitive cracking processes in drying and revetted clays and the formation of thermal contraction cracks in frozen soils, it is not clear whether this analogy is generally transferable to formation of Voronoi structures. So far reconstructing ice-wedge networks using Voronoi diagrams are reported to be particularly effective for orthogonal polygons featuring mainly rectangular or hexagonal shapes. These structures are often associated with relatively early stages of ice-wedge polygon development (low- and flat-centered polygons). It has been demonstrated that these types of polygons can be well-represented by Voronoi tessellations .

However, as the ice-wedge network evolves, forming secondary and higher order cracks, the ice-wedge network can become more irregular. This could limit the ability of Voronoi diagrams to represent all components of mature polygon networks. In consequence this would result in overestimation of polygon areas and underestimation of wedge-ice volume . This leads to the need to better understand the structural formation of ice-wedge networks. The presented VGI method could be used to better understand to what extent the real polygon network differs from the one reconstructed by the Voronoi diagram. This could be added to the VGI concept, which could be extended by a function to directly map trough lines in addition to ice-wedge polygon centroids. The degree of deviation could deliver important insights into the evolution stage of of the ice-wedge network. However, extending the VGI concept with polygon or line digitization tasks would require further research into the consequences of such extension on mapping efficiency, accuracy, and volunteer engagement.

While our proposed method for network reconstruction from crowd-sourced ice-wedge polygon centroids is specifically tailored to the monitoring of ice-wedge polygons, the underlying approach of micro-task-based crowd-sourced mapping holds potential for a wider range of environmental monitoring applications. In particular, it could be applied to other permafrost landforms such as pingos or thaw slumps, provided these features are sufficiently visible in available imagery to non-expert volunteers. A key limitation, however, lies in the spatial distribution of the target features: if they occur too sparsely within the designated mapping area, sustaining volunteer engagement may become increasingly difficult.