The study area in the north of West Siberia is 445 226 km² and includes the Yamal, Gydan, and Tazovsky peninsulas (Fig. 3). To ensure the completeness of the RTS dataset we reviewed previously published RTS datasets for the region, all of which were mapped using automated methods (Fig. 3). We manually filtered RTS datasets from Nitze et al. (2018, 2025), Runge et al. (2022), Bernhard et al. (2022), and Huang et al. (2023) to verify the presence of RTS and ensure that only true positives were included. This verification was conducted using the same available datasets that we later used for manual point collection, as described further below.

Manually collected RTS dataset published in 2021 (Nesterova et al., 2021) was also integrated: points were revised, classified, and renamed.

For our visual identification and manual collection of RTS points, we created a regular grid of 3.9×3.9 km cells covering the entire study area (Fig. 4a). This cell size was chosen as the optimal for visual inspection of the area and progress tracking, balancing detail and generalization. The ESRI satellite basemap was used as the primary source of information for RTS point collection due to the best quality of its recent very high-resolution imagery. This included high-resolution imagery (up to 0.31 m) largely with low cloudiness and an almost complete absence of visual artifacts. In rare cases, when the ESRI basemap did not fulfill visual quality criteria, such as no clouds, summer time of the image acquisition, and no artifacts, we used the Yandex Maps satellite basemap instead. In exceptional cases when neither the ESRI nor the Yandex basemaps fulfilled the visual quality criteria, we additionally checked the Google satellite basemap.

The majority of the high-resolution satellite images used in the ESRI basemap mosaic are recent Maxar images obtained after 2015 (Fig. 4b). Over a third of the study area is covered by satellite images from 2023 (Fig. 4b). Since the ESRI basemap was utilized as the primary source, all metadata related to the satellite images (image date acquisition, image resolution, image accuracy, min and max map level, satellite description, ESRI release name) in the mosaic for identifying the RTS is stored within the inventory dataset's metadata. Yandex Maps basemap presents a mosaic of various satellite imagery taken in 2016–2018 with spatial resolutions ranging from 0.4 up to 15 m. The majority of images are dated July 2017 (Nesterova et al., 2021). For the Google satellite image layer, no individual image metadata was provided.

RTSs were identified at a 1:1000 mapping scale in the satellite imagery based on visual indicators such as a clear outline of the headwall, the presence of a mudflow, and the sharp contrast in colors between the disturbed slump floor with bare ground and the adjacent intact tundra vegetation (Fig. 4c). Thus, stabilized RTSs were also identified when the indicators were still visible. For each identified feature, we created a point in the location of the RTS within the visible outlines of the RTS with the best possible approximation to the visual center of the landform.

Each digitized point represented one feature that would be classified (see Sect. 2.2). Due to the complex nature of coastal RTSs sometimes stretching along coastal segments (Fig. 4d), we decided to identify each elongated contour with visible semicircles embedded inland as one feature. Such contours were often separated from each other by little streams or watercourses. This approach allowed us to utilize a single technique for all coastal RTSs, regardless of their size and shape.

The RTS points underwent two visual corrections by the first author. To differentiate the process of coastal erosion from thermodenudation (Günther et al., 2012; Nesterova et al., 2024) and thereby distinguish other coastal landforms from RTSs, a special correction was applied to all coastal RTSs and thermoterrace RTSs (see Sect. 2.2). This involved verifying the headwall retreat of the RTS outline using the ESRI Wayback Machine - a digital archive of the World Imagery basemap of different versions providing multi-temporal imagery (ESRI World Imagery Wayback, 2024). The same verification procedure was applied for the identification of RTS in the southernmost part of our West Siberian study area, where no reliable data on massive ground ice distribution is available and thus permafrost landforms can have different origins. The literature specifies the limits of massive ground ice extent in the north of West Siberia only very approximately (Baulin and Danilova, 1998).

We classified each RTS point based on terrain position, morphology, spatial organization, and concurrent cryogenic processes (Fig. 5). The four main criteria had a total of 15 parameters.

The terrain position of an RTS is defined based on the location of the object to either some hydrological feature (sea coast, river bank, lakeshore, and gully) or just slope when there was no visible hydrological feature. The location lake was selected for RTSs even on the former shores of drained lakes.

We further defined three types of RTS morphologies: thermocirque, thermoterrace, or a combination of these two (Nesterova et al., 2024). Thermocirque generally presents a horseshoe-like RTS shape (Fig. 6a), while thermoterrace is applied to an elongated RTS with mostly straight headwall outlines parallel to a coastline or riverbank (Fig. 6b). The combination of these two morphologies sometimes occurs when the elongated RTS landform also contains circular isometric curves of headwall outlines (Fig. 6c). It is usually formed when a thermocirque merges with a thermoterrace or when multiple thermocirques merge in one elongated landform. The complicated shapes of these combined RTS features make it highly challenging to distinguish between individual elongated and horseshoe-like RTSs (Fig. 6c). Our decision tree to define the morphology of RTS is shown in Supplement.

Due to the polycyclic nature of RTS development, these landforms can exhibit a very complex spatial organization of nested and amalgamated RTSs (Nesterova et al., 2024). We identified two types of RTS spatial organization: single landforms and complex landforms. RTS can be classified as a single landform when its outline is distinct and clearly defined and there is no more than one actively thawing zone within this outline (Fig. 7a). RTS can be classified as a complex landform when its boundary is difficult to define and/or there are two or more actively thawing zones (Fig. 7b). All the RTSs with combined morphologies were marked as complex landforms.

The influence of concurrent (happening in parallel to RTS development) processes on RTS development is described in Nesterova et al. (2024). For each mapped RTS, we noted the possible presence of 5 concurrent processes: lateral thermo-erosion, coastal thermo-erosion, ice wedge erosion, nivation, and thermokarst subsidence. Lateral thermo-erosion was identified by the rugged outline of the RTS and visible traces of erosive channels (Fig. 8a). The Coastal thermo-erosion classifier includes not only the sea coast erosion but also river and lakeshore erosion. It was determined by a sharp dark outline of the RTS base along the coastline of a waterbody and the absence of sediment accumulation in the water (Fig. 8a). We have noted ice wedge erosion when an RTS headwall had a jagged outline resembling the adjacent polygonal surface of undisturbed tundra (Fig. 8b). Nivation in the context of this study is considered as persistent snow cover. It was detected as white patches of snowpacks that stayed over the summer within RTS (Fig. 8a). Thermokarst subsidence appears as small thermokarst ponds filled with water. It is noticeable as black patches within the RTS outline (Fig. 8b).

The dataset is presented in a GeoPackage vector file of point geometry with 6168 RTS point locations. Mapped RTSs were distributed unevenly, covering Tazovsky Peninsula where no RTS were found, the Yamal Peninsula except its northern part, and covering the Gydan Peninsula except its southern part (Fig. 10a). RTSs were significantly clustered according to Ripley's K function on a wide range of distances (p value = 0.001). The majority of areas of both peninsulas had less than 20 RTSs per 30×30 km hexagon grid cell, indicating distinct hotspots of RTS occurrence with more than 100 RTSs per grid cell. The main areas with high RTS density were the western part of central Yamal and the area between the southern-western and north-eastern parts of central Gydan. On Gydan, they clustered along a distinct linear feature on its southern edge, south of which RTSs abruptly become almost absent (Fig. 10b).

More than 75 % of all RTSs were found at lakeshores (Fig. 11a). The high-density areas of lakeshore RTSs correspond to RTS occurrence hotspots in the western part of central Yamal and the area between the south-western and north-eastern parts of central Gydan (Fig. 11c).

The density of RTSs at the sea coasts was mostly less than 10 RTSs per grid cell. The highest density of coastal RTSs was found along the northern shores of Yuribei Bay in south-western Yamal (Fig. 11b). For RTSs along river banks, gullies, and slopes, the predominating values of density were less than 10 RTSs per grid cell, not showing any spatial clustering (Appendix A).

The majority (72 %) of RTSs were classified as thermocirques, one-quarter of all RTSs are combined landforms, and less than 3 % were classified as thermoterraces (Fig. 12a). The majority of RTSs in all categories have a spatial density of less than 15 RTSs per grid cell.

Thermocirques were highly concentrated in hotspot areas of general RTS abundance (Fig. 10b). Combination landforms followed the high RTS abundance pattern mostly in the Gydan Peninsula but less so on the Yamal Peninsula. In contrast, thermoterraces lacked distinct high-density hotspots.

More than half of all RTSs (64 %) were classified as complex landforms and slightly more than one-third (36 %) as single landforms (Fig. 13). Both complex and single landforms followed the general spatial distribution patterns, with high-density areas being located in the western part of the central Yamal Peninsula and the southern-western and north-eastern parts of the central Gydan Peninsula. The most frequent density range for both classes was less than 10 RTSs per grid cell.

More than half (53.8 %) of all RTSs were found to have at least one concurrent process detected, more than a third (33.4 %) of all RTSs showed only one process detected, while much fewer RTSs demonstrated two or more processes detected at the same time (Fig. 14a). Lateral thermo-erosion and thermokarst were two very abundant RTS-concurrent processes (Fig. 14b). For the cases where only one process was detected per RTS, there was a predominance of thermokarst (38 %) followed by lateral thermo-erosion processes (30 %) (Fig. 14c).

Using chord diagrams (Fig. 14d–f) allowed a depiction of the co-occurrence of concurrent processes estimated for the cases when two, three, or four processes were detected for RTS. In general, the co-occurrence of the concurrent processes shows different results depending on the cases of the amount of the processes detected. There was a clear trend of the co-occurrence of nivation and lateral thermo-erosion among all 3 cases (Fig. 14d–f). The co-occurrence of lateral thermo-erosion and ice-wedge erosion gradually increased with more processes detected. The co-occurrence of the nivation and the coastal thermo-erosion, when only 2 processes are detected, was relatively low but increased significantly with more processes detected. The presence of thermokarst processes, in general, decreased with more processes detected.

RTSs attributed with concurrent processes exhibit low densities, with fewer than 5 RTSs per grid cell, regardless of the type of concurrent process (Fig. 15). RTSs with lateral thermo-erosion detected had higher densities in the western part of the central Yamal Peninsula and the central and northern Gydan Peninsula, with a hotspot in central Gydan Peninsula (Fig. 15a). RTSs with concurrent coastal thermo-erosion had higher densities in the western part of central Yamal and the north-western Gydan peninsulas, with three hotspots located at south-western part of central Yamal Peninsula, and northern and north-western Gydan Peninsula (Fig. 15b). In general, the spatial distribution of RTSs with coastal thermo-erosion did not follow the main spatial patterns detected in the Fig. 10b. RTSs with ice wedge erosion had higher densities on the northern Gydan Peninsula and rather lower densities on the Yamal Peninsula, with one hotspot located on central Yamal Peninsula (Fig. 15c). The spatial distribution of RTSs with concurrent ice wedge erosion also did not follow the main spatial patterns detected in Fig. 10b. RTSs with nivation had higher densities in central and northern Gydan Peninsula (more than 30 RTSs per grid cell) and rather lower (less than 15 RTSs per grid cell) densities on Yamal Peninsula (Fig. 15d). There were four hotspots: one on central Yamal Peninsula and three on central Gydan Peninsula. The spatial distribution of RTSs with nivation also did not follow the main spatial patterns detected in Fig. 10b. RTSs with concurrent thermokarst did follow the main spatial patterns detected in Fig. 10b and thus had higher densities and some hotspots in the western part of central Yamal Peninsula and the area between the southern-western and north-eastern parts of central Gydan Peninsula (Fig. 15e).

The manual collection of RTS points using the ESRI satellite basemap was effective across a large region but also had several limitations. One challenge was the resolution and zoom limitations, as the minimum detectable landform width was 20 m, potentially excluding smaller features. Seasonal variability of the images in the ESRI satellite basemap further complicated the process, with snowpacks identifiable only in summer images, excluding all autumn (September) imagery. On the other hand, more extensive snow cover on certain images obscured some areas, hindering the accurate inventory of RTS and their attributes in these regions. Additionally, visual artifacts (blur, glare, clouds, contrails) in some imagery led to the omission of some cells, though this accounted for less than 0.5 % of the total dataset.

Temporal constraints posed another issue, as working with a single satellite image captured at a specific time could mean that some features were not visible or detectable under those conditions, leading to potential underrepresentation of RTS features. The rapid evolution of RTS in this area (i.e., 35 % increase in RTS number in the central Yamal key site over 8 years reported by Ardelean et al., 2020) added difficulty for static inventory not only in the amount, with two RTSs of a single morphology potentially merging into a complex morphology, creating challenges in morphology classification. Similar challenges were reported in the literature (Huang et al., 2020; Rodenhizer et al., 2024). Additionally, updates to the ESRI satellite basemap during the mapping effort sometimes introduced inconsistencies across different stages of our workflow, e.g. between the initial mapping of RTS as points, the subsequent addition of attributes, and the later correction loop (Fig. 2). To alleviate some of these challenges, we effectively used the ESRI Wayback time series to verify uncertain landforms or attributes.

Visual identification also had several challenges. Stabilized RTSs were difficult to recognize. Challenges were also faced when classifying partially stabilized RTS. The limitations concerning distinguishing slowly stabilizing slumps from stabilized slumps using optical data were also reported in the literature (Bernhard et al., 2020). The sediment accumulation as a secondary indicator for coastal thermo-erosion was found to be debatable due to its temporary nature. Some landforms, such as curved riverbanks, wave-cut lakeshores, active layer detachments (ALDs), and first-stage thermokarst mound (baydzherakh) development, could have been easily misclassified as RTS, leading to false positives in the final dataset.

Atypical for this area, Yedoma RTSs found in our inventory in the northern Gydan region differed significantly in appearance from the majority of the rest mapped RTSs. Yedoma deposits in West Siberia were not included in the Circum-Arctic Map of the Yedoma Permafrost Domain (Strauss et al., 2021), yet were described in the fieldwork in northern Yamal coast and northern Gydan coast (Vasil'chuk and Vasil'chuk, 2018; Vasil'chuk et al., 2022). Since Yedoma mapping was not the aim of this inventory, we did not mark Yedoma RTSs. Moreover, Yedoma RTS's visual characteristics were not properly addressed in the initial visual identification protocol, leading to potential misidentifications.

Human subjectivity, even if mapping is conducted by experienced researchers, can influence the results and contribute to dataset uncertainties. For RTS mapping, this has been demonstrated before in a mapping exercise with multiple operators with varying degrees of experience (Nitze et al., 2024). Our subjectivity assessment using a subset of 120 RTS samples revealed that 3 %–16.6 % were classified as non-RTS, with an average false positive rate of approximately 8.5 % and a median of 4.1 %. Consequently, the accuracy of our dataset based on this experiment averages around 0.91. We acknowledge that involving additional experts in visual correction could have improved accuracy and reduced subjectivity.

The degree of classification similarity among the five co-authors, compared to the original dataset, exhibited a clear trend influenced by spatial organization, morphology, and two concurrent processes – coastal thermo-erosion and lateral thermo-erosion – which were generally the most subjective. Spatial organization emerged as the most subjective parameter, with classifications showing the alignment in only half of the 120 sample points on average (Fig. 16a).

To further quantify classification variability, we calculated Jensen-Shannon distances (Fig. 16b), a metric for measuring similarity between probability distributions. This value ranges from 0.0, indicating identical distributions, to 1.0, representing completely distinct distributions. The results confirmed the overall trend of morphology, coastal thermo-erosion, and lateral thermo-erosion being the most subjective parameters, except for spatial organization, which showed minor differences in probability distributions. Coastal thermo-erosion exhibited the highest variation in classification probability distributions, likely due to two distinct hotspots observed in the heatmap.

Overall, the probability distributions of most classified parameters were either highly or moderately similar to those in the original dataset. This suggests a generally consistent perception of RTS classification among the co-authors in the experiment.

RTS location accuracy was estimated for the area around the Vaskiny Dachi research station in central Yamal and central Gydan Peninsulas, with helicopter surveys conducted in 2020 and 2023 (see Appendix B). RTS location accuracy assessments for all areas revealed very high precision compared to the ground truth, confirming the reliability of the dataset (Table 2). A relatively low recall, even after applying mapping style adjustments, indicates an approximate 50 % underestimation of small RTSs in the study area (Table 2) primarily due to the reasons described in the Data Limitations section (see Sect. 4.1). Please, note that in this context, precision specifically refers to the metric used in the F1-score calculation and should not be mistaken for measurement precision, as no actual measurements were conducted.

The relatively low F1 scores observed in our study can be attributed primarily to high underestimation (i.e., low recall) when compared to field data. Manual mapping of RTS using remote sensing data is often regarded as the most accurate approach (Swanson and Nolan, 2018; Segal et al., 2016a, b; Young et al., 2022; Luo et al., 2022). Efforts to enhance accuracy, particularly in terms of precision, have been made by incorporating multi-year datasets (Huang et al., 2021) and conducting multiple rounds of expert review (Segal et al., 2016b; Young et al., 2022). To ensure the reliability of manual mapping, Young et al. (2022) employed aerial field survey data for visual validation; however, their study did not report the initial recall of manual RTS mapping against field observations.

To the best of our knowledge, there are no existing studies that quantitatively assess the recall uncertainty of RTS manual mapping using remote sensing compared to field data, particularly over large spatial extents. Lewkowicz and Way (2019) attempted to estimate recall accuracy for manual RTS mapping in Banks Island, Canada (70 000 km²), but their evaluation was based on a comparison with another remote sensing dataset rather than ground-based field observations. This limitation is largely due to the challenges associated with field data collection in remote study areas. Moreover, since field data provides only a single snapshot in time, some RTS classified as false positives based on remote sensing data may be true RTS that were simply not captured in the field dataset.

Despite these uncertainties, manually mapped RTS datasets serve as validation sources for automated deep-learning-based mapping algorithms (Nitze et al., 2021; Yang et al., 2023; Xia et al., 2022; Huang et al., 2021). Notably, relatively high F1 scores (F1>∼0.7) for automated RTS mapping have been reported, but these assessments were primarily conducted against internal training datasets covering limited spatial extents and derived from manual mapping rather than field data (Huang et al., 2020; Nitze et al., 2021; Witharana et al., 2022; Yang et al., 2023).

Our findings demonstrate that manual mapping using remote sensing data cannot be considered a definitive ground truth and is associated with a certain degree of inaccuracy, particularly concerning recall.

Our accuracy assessment highlights the overall subjectivity in defining RTS morphology and spatial organization. These parameters critically influence what is visually identified as RTS in satellite imagery. This subjectivity aligns with previous RTS mapping experiments, where “mapping style” and the scientific background of domain experts were found to impact RTS delineation (Nitze et al., 2024). Our results demonstrate that, despite standardized instructions, both morphology and spatial organization remain the most subjective parameters in RTS classification.

The collected data on RTSs holds significant potential for future applications and research across various disciplines. It can serve as a foundation for a more detailed characterization of the permafrost region. The spatial distribution and clustering of RTSs in West Siberia, combined with cryostratigraphic and geomorphological analyses, can help unravel driving processes and improve our understanding of these dynamic landforms.

This dataset can also guide further research efforts, such as field surveys aimed at monitoring cryogenic processes as well as studies to uncover the ground ice origin. In addition, it provides a valuable reference for ground-truthing in machine learning applications, enabling more accurate automated remote sensing classifications and predictive modeling.

The dataset is particularly relevant to ecologists, biogeochemists, geomorphologists, climatologists, permafrost scientists, hazard researchers, and remote sensing specialists. This data can also be useful in the context of managing permafrost-related risks and planning sustainable development in vulnerable regions.