We compiled the database from a combination of ground-truthed field observations and remotely sensed data across Alaska. Table 1 provides a summary of the sources used to extract thaw occurrence locations. We sourced these locations from a variety of formats including geospatial databases, coordinates reported in published journal articles, field campaigns, and photo-interpreted sites of landscape change. Given that spatial datasets were often in different formats such as points, polylines, and polygons, we standardized the database by converting all locations to point features based on the centroid of the feature (polygons) or the midpoint of the line (polylines).

Each of these sources employed different data collection methods resulting in varying levels of accuracy in their final outputs. In particular, the spatial and temporal resolution of remote sensing data has improved significantly over time. Older studies often relied on moderate-resolution imagery such as Landsat (30 m), while more recent measurements frequently utilize high-resolution datasets like ArcticDEM (2 m), enabling more precise detection and characterization of thaw features. We did not manually verify each individual feature, and the features in the database reflect the accuracy of their source dataset. While a comprehensive and quantitative statistical uncertainty analysis is not possible due to the heterogeneity and lack of validation data for our input sources, we have provided metadata on source type, satellite(s) or sensor(s) used (if remotely sensed data), data year(s), and spatial resolution. This level of documentation allows the reliability and limitations of the dataset to be evaluated transparently by users. In addition, the open-access and collaborative nature of the database enables community feedback including the identification and correction of errors, duplicates, omissions, or regional gaps, and provides clear opportunities for continued refinement and expansion as new data become available. To increase consistency across sources, we applied several post-processing techniques to improve data quality (Table 2). These methods included filtering out points located outside the zone of mapped permafrost in Alaska, using only the most recent thaw data when multiple years of thaw data was presented from each source, and removing duplicate locations. Duplicate entries of thaw features with identical names were removed, keeping only the first occurrence of each named feature. For example, both Wang et al. (2018a, b) and the Global Terrestrial Network for Permafrost included sites from the Circumpolar Active Layer Monitoring Network (CALM).

For each point feature, we recorded the identifying source information (authors, DOI), the type of source (field or remote sensing) and its publication status, feature name (if applicable), latitude and longitude (in WGS 84), feature type as reported by the source (i.e., retrogressive thaw slump, thermokarst fen), feature category (feature type simplified into a broader category), type of thaw (abrupt or non-abrupt), and the imagery used along with relevant mapping information such as imagery dates and resolution. Table 3 details these attributes and Table 4 defines the thaw feature categories used in our database. We classified features as abrupt or non-abrupt thaw according to the framework outlined in Webb et al. (2025a), specifically using the decision tree in Fig. 5. Under this scheme, a thaw event is considered abrupt if it develops within 30 years and meets at least one of the following criteria: it involves substrate with high ground ice content (>20 %) or it results in a major ecosystem impact, even in the absence of high ground ice. All other features were classified as non-abrupt.

We divided Alaska into ecoregions based on the EPA's Level III Ecoregions map (Gallant et al., 1995), using Level II classifications with some adjustments (Fig. 1). We distinguished the Brooks Range from the Arctic Tundra, grouped the southern mountain ranges together under a single “Southern Mountains” ecoregion, and classified all areas along the southern coast as “Maritime”. Then, we spatially joined the thaw database with the ecoregion map and quantified the occurrence of various abrupt thaw features within each unique ecoregion.

To characterize terrain conditions at observed thaw features, we extracted high-resolution topographic variables from the ArcticDEM (V.4.1) (Porter et al., 2023) using Google Earth Engine (GEE) (Gorelick et al., 2017). We calculated elevation (m), slope (degrees), and aspect (degrees) using the terrain function in GEE. The terrain function derives slope in degrees using a 3 × 3 pixel window around each point. Since the ArcticDEM has a spatial resolution of 2 m, slope was computed from a 36 m² neighborhood centered on each point. To quantify a thaw feature's topographic position on the landscape, we computed relative elevation by subtracting the mean elevation within a 100 m circular neighborhood from the elevation at each thaw point (Eq. 1). This first-order approximation yields values less than 0 for depressions and values greater than 0 for elevated terrain. 1hreli=hi-hSi‾ hreli = Relative elevation; hi = Elevation at point i; hSi‾ = Mean elevation within a 100 m circular neighborhood around i.

We calculated a solar radiation index (SRI) following the methods outlined in Fu and Rich (2002) (Eq. 2). The SRI estimates the potential incoming solar radiation under clear-sky conditions at solar noon on the summer solstice (21 June). To define a representative solar azimuth angle for Alaska, we used the NOAA Solar Calculator and selected the geographic center of the state (64.73° N, 152.47° W) since solar azimuth varies minimally across Alaska that time of year. We approximated the solar zenith angle, which varies with latitude, by calculating the absolute difference between each thaw location's latitude and solar declination on the solstice. We acknowledge that the SRI is a simplified proxy for solar energy input, and that local factors such as canopy cover, microclimate, and seasonal daylength variation also influence site-specific solar exposure. However, the SRI provides a useful estimate of potential solar energy input that may affect permafrost thaw vulnerability and is more informative than aspect alone. Because Alaska is at high latitude, the solar radiation index values are centered around ∼ 0.75 for flat terrain, with higher values indicating steeper, south-facing slopes and lower values indicating north-facing slopes or less-exposed terrain. 2SRI=cos⁡θZcos⁡β+sin⁡θZsin⁡βcos(ϕ-α) θZ = solar zenith angle; β = slope; ϕ = aspect; α = solar azimuth angle.

To examine topographic variation in slope, relative elevation, and SRI between abrupt and non-abrupt thaw locations, we performed non-parametric Wilcoxon rank-sum tests (McKnight and Najab, 2010). Because the database is heavily skewed towards abrupt thaw features, we used a bootstrapping approach to balance sample sizes (Efron and Tibshirani, 1993). We conducted 1000 iterations of stratified random sampling, selecting 500 abrupt thaw points and 500 non-abrupt thaw points with replacement in each iteration. We calculated 95 % confidence intervals for the mean values of each variable within abrupt and non-abrupt thaw groups based on the bootstrapped samples. All statistical analyses were conducted in R (version 4.5.0) and the associated code and results can be found at our GitHub repository: https://github.com/ArcticWebb/Alaska_Permafrost_Thaw_Database (last access: 9 April 2026).

Areas with high ground ice content are especially prone to abrupt permafrost thaw due to substantial volume loss when ice melts. Ground ice is widely recognized as one of the most influential factors driving abrupt thaw (Jorgenson et al., 2006; Teufel and Sushama, 2019; Turetsky et al., 2020) and many permafrost thaw prediction models use it as a key variable for identifying areas at risk. We acknowledge that precipitation, vegetation, snow cover, and other factors can also influence abrupt thaw, however, many of these drivers are indirectly represented in ground ice maps because those maps are developed using environmental indicators as model inputs. Existing ground ice maps for Alaska were produced at coarse spatial resolutions (statewide or circumpolar) which limits their utility for fine-scale mapping (Heginbottom et al., 2002; Karjalainen et al., 2022). For this study, we compared two of the most widely used datasets: one developed specifically for Alaska and another designed for pan-Arctic ground ice conditions. The Permafrost Characteristics of Alaska map by Jorgenson et al. (2008) estimates ground ice content within the upper 20 m of permafrost based on terrain originally described in Kreig and Reger (1982) and is supplemented with field observations. Inconsistent or patchy permafrost distribution is classified as variable, <10 % as low, 10 %–40 % as moderate, and >40 % as high. In contrast, the Circum-Arctic Map of Permafrost and Ground-Ice Conditions, Version 2 by Heginbottom et al. (2002) summarizes permafrost conditions and ground ice distribution across the Northern Hemisphere (20 to 90° N). Ground ice classification is also based on the upper 20 m of permafrost, with <10 % defined as low, 10 %–20 % as moderate, and >20 % as high. Because the two datasets use different percentage thresholds to define moderate and high ground ice content, we combined the moderate and high classes into a single category for each map while retaining low ground ice as a distinct class since both maps define it as <10 %. This approach groups the majority of landscapes where abrupt thaw is considered probable or highly likely (moderate or high ground ice) while retaining a smaller class where abrupt thaw should be unlikely (low ground ice).

We assessed the ability of these ground ice maps to capture ground ice distribution at a finer scale using our thaw database. Prior to analysis, we removed abrupt thaw processes that do not require ground ice melting, ensuring that the abrupt thaw points used for this analysis only represented ice-dependent thaw processes (Table 4). We then overlaid the maps with the updated thaw database points, extracted the ground ice classification, and calculated the proportion of abrupt and non-abrupt thaw points within each ground ice class. We expected abrupt thaw points to predominantly occur in high or moderate ground ice areas and non-abrupt thaw points to occur in low to no ground ice areas. Finally, we overlaid the ice-dependent abrupt thaw points from our database with the Database of Ice-Rich Yedoma Permafrost by Strauss et al. (2022) and calculated the proportion of these thaw processes within the Yedoma domain. Although duplicate features were removed (Sect. 2.1), some study sites include multiple thaw features which may result in spatial clustering of observations. Therefore, results in Sect. 3.3 should be interpreted as representing broad regional patterns rather than statistically independent observations.

The final database contains 19 540 permafrost thaw locations spanning all ecoregions of Alaska (Fig. 1). Spatial coverage is statewide and the temporal resolution varies because the sources used to map thaw features are based on imagery spanning the past ∼ 70 years. Of these, 18 213 points represent abrupt thaw including 10 625 thermokarst lakes, 5463 active layer detachments, 1450 retrogressive thaw slumps, 280 generic thermokarst processes, 209 wildfire-induced abrupt thaw features, 134 thermokarst wetlands, 47 thermoerosional gullies, and 5 thaw ponds (Fig. 1). An additional 1327 points represent non-abrupt thaw. Abrupt thaw locations were concentrated in northern Alaska (Fig. 1), reflecting the higher abundance of continuous permafrost with near-surface ground ice in this region. The classification of thaw features follows the terminology used in the source data. If the thermokarst type was not specified (i.e., fen, water track), the observation was recorded as “thermokarst” and is hereafter referred to as generic thermokarst processes. Thaw features influenced by wildfire were designated as a distinct class called “wildfire-induced thaw” and were therefore not double-counted with other categories.

There were distinct regional differences in types of abrupt thaw across Alaska (Fig. 2). Active layer detachments and retrogressive thaw slumps dominated in mountainous areas like the Brooks Range, while lowland regions like the Arctic and Bering Tundra primarily experienced thermokarst lake expansion and general thermokarst activity. Abrupt thaw accounted for more than 97 % of features in the Arctic Tundra, Bering Tundra, and Brooks Range. In contrast, only 36 % of features in Interior Alaska and 17 % in the Bering Taiga were abrupt thaw while the Southern Mountains and Maritime ecoregions had virtually none. Interior Alaska exhibited the most diverse array of abrupt thaw features, with thermokarst wetlands and wildfire-driven thaw being the most common. We tested whether the distribution of abrupt thaw features differed significantly between Alaska's ecoregions using a chi-square test of independence with Monte Carlo simulation (10 000 replicates) to account for sparse data. The test revealed a highly significant association between thaw feature type and ecoregion (χ2 = 27 509, p<0.001), indicating that abrupt thaw features are not evenly distributed across regions.

Two out of three topographic variables (relative elevation and SRI) showed statistically significant differences between abrupt and non-abrupt thaw locations. Slope was not significantly different between abrupt (mean = 5.36°, 95 % CI: [4.70, 6.07]) and non-abrupt thaw locations (mean = 4.01°, 95 % CI: [3.50, 4.60]; p = 0.51). Relative elevation was significantly lower at abrupt thaw locations (mean = -0.66 m, 95 % CI: [-0.78, -0.56]) compared to non-abrupt thaw locations (mean = 0.04 m, 95 % CI: [-0.13, 0.21]; p<0.001). Less intuitively, solar radiation index was significantly lower at abrupt thaw locations (mean = 0.71, 95 % CI: [0.70, 0.71]) compared to non-abrupt thaw locations (mean = 0.75, 95 % CI: [0.74, 0.76]; p < 0.001). Most solar radiation index values fall between 0.6 and 1 and are clustered around 0.75 because the majority of thaw features occur on flat terrain with slopes near 0°. These findings suggest that abrupt thaw is more prevalent in lowlands or depressions and in areas with lower potential solar radiation (Fig. 3), while slope has no significant effect.

When we compared our thaw database with the Jorgenson et al. (2008) ground ice map, only 65 % of ice-dependent abrupt thaw locations fell within areas classified as high or moderate ground ice while 35 % occurred in areas labeled low, variable, or unfrozen (Fig. 4). This indicates that this ground ice map only captures about two-thirds of locations where ice-dependent abrupt thaw processes actually occur, revealing a substantial mismatch between the map and observed thaw features. Spatial overlays with the Heginbottom et al. (2002) map showed that 60 % of ice-dependent abrupt thaw locations occurred in areas classified as high or moderate ground ice while 40 % were in low-ice regions (Fig. 4). The spatial distribution of ice-dependent abrupt thaw events in areas classified as having low ground ice content is not uniform across Alaska, with most of the discrepancies being in the Seward peninsula, northern tundra, and parts of interior Alaska. Because the abrupt thaw processes used in this analysis require substantial ice content to develop, these results highlight the potential limitations of existing ground ice datasets for accurately representing permafrost vulnerability at the local scale.

We also examined the level of agreement between the two ground ice maps across Alaska (Fig. 5). Although Jorgenson et al. (2008) and Heginbottom et al. (2002) show comparable accuracy in identifying ice-dependent abrupt thaw features within areas classified as having high or moderate ground ice, substantial discrepancies remain in their spatial distribution of ground ice content (Fig. 5). Excluding regions without permafrost, the two maps assign the same ground ice class to only 73 % of the permafrost zone, while 27 % is classified differently, meaning that there is conflicting representation on more than one quarter of ground ice distribution in Alaska's permafrost zone. These findings suggest the maps are not only inconsistent with our fine-scale thaw database but also show considerable disagreement with each other.

In addition to evaluating ground ice maps, we examined the spatial relationship between ice-dependent abrupt thaw features (i.e., thaw processes that require high ground ice content to form) and Yedoma. The Yedoma domain covers 38 % of Alaska, yet it includes nearly 60 % of ice-dependent abrupt thaw features from our database, indicating that a substantial proportion of these thaw processes are located in regions underlain by extremely ice- and carbon-rich permafrost (Fig. 6).