The Northwest Territories Thermokarst Mapping Collective: a northern-driven mapping collaborative toward understanding the effects of permafrost thaw

Circumpolar landscapes are among the most sensitive environments in the world due to the rapid rates of warming compounded by the effects of permafrost thaw. Climate change and natural and anthropogenic disturbances can degrade ice-rich permafrost (Smith et al. 2010; Kokelj et al. 2017; Biskaborn et al. 2019), altering the frozen foundation of northern terrain and leading to the formation of thermokarst landforms (Fig. 1) (Kokelj and Jorgenson 2013). The initiation or acceleration of thermokarst processes, mainly involving thaw subsidence, can rapidly modify landscape morphology (van der Sluijs et al. 2018), hydrology (Connon et al. 2018), biogeochemical regimes (Tank et al. 2020), climate feedbacks (Turetsky et al. 2020), and the structure and functioning of northern ecosystems (Lantz et al. 2009; Chin et al. 2016; Schuur and Mack 2018). Thermokarst accelerated by a warmer and wetter climate has emerged as a stressor to the integrity of northern infrastructure (Bush and Lemmen 2019; GRID-Arendal 2020) and traditional land use and food security of Indigenous communities (Calmels et al. 2015; Spring et al. 2018). The nature and magnitude of thaw-driven landscape change are closely related to specific terrain, ground ice, soil, ecological, and climate conditions, producing a diversity of thermokarst and periglacial landforms (Fig. 1) that are valuable indicators of terrain sensitivity (Jorgenson et al. 2015). This suite of permafrost processes can lead to the formation of discrete features, such as retrogressive thaw slumps (Kokelj et al. 2015), deep-seated translational failures (Young et al. 2022), or drained lakes (Marsh et al. 2009), or they can affect large contiguous areas where conditions give rise to distinct landscape types, such as degrading permafrost peat plateaus (Vitt et al. 1994; Quinton et al. 2011), thaw–lake complexes (Morse et al. 2017), or ice-wedge thaw ponds and polygonal-patterned ponding (Fraser et al. 2018). The terrain, environmental, carbon, and societal consequences of permafrost thaw vary with the diversity in permafrost landscapes. However, broad-scale empirical data that describe the patterns in thaw-sensitive permafrost terrain do not exist for most circumpolar regions. This significant knowledge gap constrains our understanding of the environmental effects of permafrost degradation, and the evaluation of remote sensing and spatial modelling products that predict the consequences of thaw-driven change, limiting our ability to inform climate change planning and adaptation.

An increasing amount of circumpolar-scale remote sensing and modelling products are highlighting the global significance of thermokarst and permafrost thaw (Olefeldt et al. 2016; Chadburn et al. 2017; Hjort et al. 2018; Karjalainen et al. 2020). However, fine-scale information generated through a range of methods commonly focused on development corridors or communities (Wolfe et al. 2014; Sladen et al. 2022; Allard et al. 2023) is required to support environmental and infrastructure management and climate change adaptation planning. Optical, thermal, radar, or laser scanning methods are used to map permafrost terrain conditions across fine (10¹–10³ km²) and intermediate (10³–10⁴ km²) spatial scales, and they often focus on specific thermokarst processes or thaw-sensitive landforms (Aylsworth et al. 2000; Quinton et al. 2011; Segal et al. 2016; Steedman et al. 2017; Wolfe and Morse 2017; Fraser et al. 2018; van der Sluijs et al. 2018; Zwieback et al. 2018; Lewkowicz and Way 2019; Turner et al. 2021; Wang et al. 2023). These datasets are usually geographically limited by the logistical and operational constraints associated with detailed mapping. In contrast, rule-based schemes have been implemented in GIS platforms to estimate broad-scale distributions of thaw-sensitive landscapes or ice-rich terrain. These studies can use expert knowledge in a collaborative framework to approximate the distribution of thermokarst terrain and associated carbon storage (Olefeldt et al. 2016) or portray spatial patterns of thaw-sensitive or ice-rich terrain by integrating an existing knowledge base (O'Neill et al. 2019). These modelling approaches generate spatial data rapidly and have had high uptake among users due to their broad geographical coverage. They are also used as inputs for modelling climate change effects, such as global-scale estimates of carbon release (Turetsky et al. 2020) or thaw-driven impacts on water resources (Spence et al. 2020). However, rule-based GIS modelling schemes are inherently constrained by the existing state of scientific knowledge, the variable quality and resolution of input data layers, and the paucity of robust validation datasets. Consequently, they are of limited utility where fine-scale information or quantitative representation of terrain conditions are required (Gibson et al. 2021; Allard et al. 2023).

The need for acquiring fine-scale information on terrain and permafrost conditions across broad spatial scales is being partially overcome by rapid advances in Earth observation sensors, data acquisition platforms, and processing power (Chen and Wang 2018; Giuliani et al. 2019), which have accelerated remote detection of thaw-driven permafrost landscape change at multiple spatial scales (Fraser et al. 2018; Nitze et al. 2018). Machine learning techniques hold significant promise for automating image analyses to recognize fine-scale spatial patterns of Earth surface processes and landforms, with applications to the vast, rapidly changing, and remote permafrost environments (Siewert 2018; Huang et al. 2020). Machine learning requires systematically collected fine-scale training data from a variety of locations, but robust training and validation datasets of permafrost and periglacial processes remain sparse. The automation of Earth observation and rapid progress in modelling permafrost change at broad spatial scales heightens the need for complementary efforts in advancing empirical mapping of thermokarst terrain for training and validation and to foster a field-based understanding of the thaw-driven processes that are transforming permafrost environments.

Climate change has major environmental and societal implications for northwestern Canada's rapidly warming and ice-rich permafrost regions (Kokelj et al. 2021; Smith et al. 2022). The Northwest Territories (NT) covers an area of 1.346 million km² and comprises much of Canada's most ice-rich permafrost terrain (Fig. 2) (O'Neill et al. 2019). The study region is also comprised of diverse ecoregions that host a significant portion of northern Canada's community, transportation, mining, and oil and gas infrastructure. The NT comprises the traditional territories of several Indigenous groups who are reclaiming decision-making authority through the establishment of seven regional and three community-based governments that have been negotiated, or are in the process of negotiating self-government agreements with the Government of the NT (GNWT) and the Government of Canada (https://www.eia.gov.nt.ca/en/indigenous-governments). Northern regulators, planners, policy-makers, and communities require better information to understand the environmental and societal implications of permafrost thaw and expressly recognize the need for improved permafrost mapping to support decision-making (GNWT 2018). The Arctic science community has mobilized to examine the physical, chemical, biological, and carbon consequences of permafrost thaw and has attempted to upscale results to broad regional and circumpolar scales (Spence et al. 2020; Turetsky et al. 2020), but many potential analyses require information on the spatial distribution and configuration of thaw-sensitive permafrost terrain. Despite these growing societal and scientific needs, no systematic, empirical inventories of thaw-sensitive permafrost terrain exist for most of the circumpolar world.

To address this significant knowledge gap, the NT Geological Survey (NTGS) initiated the NT Thermokarst Mapping Collective (TMC) in collaboration with Wilfrid Laurier University (WLU), the Geological Survey of Canada (GSC), and other government, academic, and Indigenous partners. The overarching aim of the TMC is to advance knowledge of permafrost terrain conditions and thermokarst distribution across NT and its permafrost-affected transboundary watershed areas (Fig. 2), within a framework where partners co-develop scientific knowledge and its dissemination. While this has become best practise in northern ecology (Johnson et al. 2015; Cold et al. 2020; Thompson et al. 2020; Chila et al. 2022), the general approach is still emerging in the field of permafrost science. To focus the project and address northerners’ needs, the TMC fostered collaborations across disciplines and among organizations to develop a systematic approach to map and inventory thermokarst terrain under the groupings of slopes and mass wasting, periglacial, and hydrological themes, establish a spatial data infrastructure (SDI) and management framework, and train mappers from across Canada (Fig. 3). Following methods developed by the TMC, and using ArcMap 10.8., and now ArcGIS Pro, the mappers populated a gridded vector geodatabase of the study area by (1) systematically inspecting true colour and near-infrared Sentinel-2 satellite imagery at a 1:20 000 scale for the presence and relative abundance of indicators of sensitive permafrost terrain according to criteria defined by expert teams and (2) recording the observations within each 7.5 × 7.5 km² grid cell by populating its attributes in the geodatabase according to lists of descriptors for each feature or terrain type (Figs. S1 and S2) (e.g., Gibson et al. 2020; Wolfe et al. 2022). The resulting empirical data will provide detailed information on the patterns and characteristics of sensitive permafrost terrain throughout the study area (Fig. 2).

The objectives of this paper are to (1) describe the organizational framework and general methodologies developed to implement the TMC project; (2) explain the SDI to support data sharing needs; (3) explore preliminary thermokarst mapping outputs through comparisons with published fine-scale datasets (10³ km²) as a validation of the approach, and with broad-scale national and circumpolar datasets (10⁵ km²) to explore the patterns of variation and the value of the TMC data; and (4) highlight the diversity of thermokarst-affected landscapes being quantified by these methods. This project overview and preliminary results provide a basis for discussing some limitations of the approach, and the challenges and opportunities of developing a broad-based, cross-disciplinary permafrost research collaboration, the need for linking field-scale knowledge with mapping and modelling outputs, and the benefits for advancing permafrost science and its application. In addition, we address the need for more holistic, geomorphically based frameworks to advance knowledge of the thaw sensitivity of permafrost terrain. Finally, we highlight opportunities for enhancing collaboration and knowledge sharing between northerners, scientists, environmental managers, and Indigenous partners through experiences gained in this project.

An extensive set of landforms indicative of thermokarst or permafrost sensitivity related to mass wasting, hydrological, or periglacial processes occur within the study area. These landforms are related to the suite of geological, ground ice, climate, and ecological factors that reflect variation in landscape history (e.g., Wolfe and Morse 2017; Kokelj et al. 2021). Because of the diversity of thermokarst and periglacial landforms, and the distribution of collaborators situated across Canada, a wide range of expertise was leveraged to develop the methods and tools needed to support the project. The NTGS with support from the GSC and the GNWT–WLU partnership, worked with academic and Indigenous partners to initiate the TMC project. The NTGS presented a general approach to rapidly inventory landscapes affected by different permafrost features, which was discussed at a technical workshop held in Yellowknife, NT in March 2018. Decision-makers and northern partners’ needs were determined through continued discussions with stakeholders, which together established the goal of developing territory-wide mapping of thaw-sensitive landforms. The refinement of methods for producing empirically based permafrost maps was advanced following input from this workshop, as well as activities focused on the pursuit of project funding. Participants formed expert teams to develop mapping methodologies within themes of thaw-driven landforms. The NTGS provided project oversight, and the primary theme leads from the Territorial and Federal Geological Surveys, and university researchers co-developed the scientific approach (Fig. 3). The GNWT–WLU partnership enabled the staffing of a Yellowknife-based project coordinator who reported to the TMC project. This northern-based position has worked with NTGS to advance the project by coordinating participants, developing the SDI, managing data, and supporting project reporting functions. Several funding partnerships were secured with organizations across the NT and Canada to train mappers and support mapping (Fig. 3). Linkages with community partners related to this project have been initiated through workshops and field programs but were severely hampered by COVID-19 restrictions. The northern research partners' abilities to sustain field programs through the pandemic provided opportunities to work with Indigenous partners for training in permafrost observational methods and for aerial validation of the mapping during the 2020–2022 field seasons.

The methodology implemented by the TMC builds upon an approach described in detail by Segal et al. (2016) and employed by Kokelj et al. (2017) who used 10 m resolution SPOT 4/5 satellite imagery (from 2005 to 2010) to identify thaw-driven mass wasting features within a standardized grid of 15 × 15 km² cells throughout northwestern Canada. The uniquely identified grid cells are used to partition and attribute observations of terrain features into a set of discrete mapping units, which can then be used to quantitatively evaluate spatial variation. After some testing, the TMC used uniquely identified quartiles (i.e., 7.5 × 7.5 km²) of the grid system from Segal et al. 2016, significantly improving resolution and resulting in mapping units deemed to be of sufficient granularity to produce a valuable broad-scale inventory of permafrost landscape characteristics over the entire study area in a timely manner (Fig. 4). Grid cells entirely over water, such as in the ocean or in the middle of Great Bear or Great Slave lakes, were removed from the database using the Canada Coastline and inland lake boundaries (Natural Resources Canada 2016). The TMC developed guidance documents to describe methods for inventorying mass wasting, hydrological, and periglacial features, supporting the design of a geodatabase in ArcMap 10.8 to enable mappers to identify and attribute the landforms or terrain features that characterize each grid cell. Procedural manuals for inventorying hydrological features and peatlands, including intermapper comparisons for permafrost peatland identification are published as NTGS Open Reports (e.g., Gibson et al. 2020; Wolfe et al. 2022). The procedural manual for inventorying mass wasting features builds on Segal et al. (2016) and along with the manual for periglacial features are in the process of publication as NTGS Open reports. The raster datasets from which the inventory data were derived, and the supporting data layers, are summarized in Table S1. Thematic or regionally focused spatial datasets will accompany scientific publications.

The purpose of the grid-based approach is to build an initial inventory of feature presence and relative abundance over large areas (Segal et al. 2016; Ramsdale et al. 2017; Chiasson and Allard 2022). Here, we summarize how the rubrics to identify and interpret thermokarst and periglacial features on the satellite imagery were developed, and how inventory data were recorded and integrated into an SDI, and we touch on some limitations of the mapping approach. More detailed explanations of the methods and feature interpretation criteria upon which the mapping was based are provided in NTGS Open Reports that were developed (Gibson et al. 2020; Wolfe et al. 2022) or are being refined and published with anticipated, theme-specific research papers. For the retrogressive thaw slump (RTS), polygonal terrain, and ramparted lake–lithalsa (RL–L) terrain maps presented in this paper, Appendix A provides several examples of each of these landforms in an oblique aerial image and its appearance on SPOT 6/7 imagery (2018) and Sentinel-2 imagery (2016–2017). The general approach and geodatabase structure ensured that the data types are correct, data entry is logical and consistent, and database entries are unique.

Permafrost experts initially advanced the mapping workflow by creating a comprehensive list of thermokarst landforms and indicators of permafrost thaw sensitivity and their morphological attributes (Fig. 5). Features were categorized into themes according to geomorphic processes and environmental setting: (1) mass wasting (e.g., Segal et al. 2016), (2) hydrological features (Wolfe et al. 2022), and (3) periglacial features. An organic theme (Gibson et al. 2020) focused on identifying the presence and abundance of peat plateaus as indicators of thaw-sensitive permafrost and at a higher spatial resolution, so it was merged with the periglacial theme where attributes describing peatland morphology and the nature of degradation patterns had been developed. The list of features and attributes included in each theme was refined as leads and mappers learned by locating, identifying, and attributing landforms across well-studied environments using the Sentinel-2 satellite optical imagery (e.g., Appendix A). Following Tobler's (1987) rule for depicting land use and land cover change using satellite imagery, an optimal viewing scale for the Sentinel-2 imagery was set at 1:20 000. Consequently, the 10 m resolution of Sentinel-2 imagery imposes limitations that necessitated experts to group thermokarst landforms or landscape indicators of permafrost thaw sensitivity and simplify rules for identification and attribution (Fig. S2 and Appendix A). For example, landslides were grouped generally as either “thaw slumps” or “deep”, or “shallow” slides based on morphology and were not further differentiated by type. Polygonal (ice-wedge network) terrain was attributed as upland or lowland, but high- or low-centred polygons were not differentiated. Given the mapping constraints, methods were developed to identify and inventory 25 distinct thermokarst landforms or landscape indicators of permafrost thaw sensitivity according to the three primary themes, which are organized according to the spatial data model implemented in the geodatabase with corresponding primary tables (Figs. S1 and S2).

The TMC theme leads and mappers reached consensus on how to estimate counts or % cover of each feature according to bins of counts (e.g., 0; 1; >1; and >10) or approximate % cover during preliminary trials (e.g., 0; <10%; 10%–33%; 33%–66%; 66%–100%) (Figs. 4 and 5, and Fig. S2). These bins and the final set of landforms and indicators of thaw-sensitive permafrost terrain, and their attributes were informed by results of preliminary trials and mapper feedback, which followed regular online sessions with theme experts and mappers held to discuss landscape interpretations, refine rubrics, and promote iterative learning. The bins can be reduced to three categories (i.e., [absent (0); few (1); and many (2)] or [absent (0); trace cover (<10% cover); and abundant cover (>10% cover)] or further reduced to binary presence or absence, giving several categorical data formats for quantitative comparisons among landforms or with other datasets. With the feature sets, their attributes, and bins of count and cover defined, we linked the geodatabase to a GIS where mappers inspected imagery and populated the specified grid cell attributes according to theme-specific dropdown menus that reference look-up tables specifying the features and attributes (Fig. 4). The look-up tables ensured the format of the information entered is consistent and replicated by all mappers.

In order to produce a preliminary dataset for analysis and support northern needs, we prioritized 75 × 75 km² areas surrounding each of the 33 communities in NT (Figs. 2 and 4). We supported this initial landform interpretation and mapper training with knowledge from thematic experts and higher-resolution imagery (SPOT 6/7 1.5 m). To test the validity of the mappers' data, the experts performed QC checks on all data as mappers were being trained that focused discussions on refining interpretation of features. Thereafter, the experts performed QC checks at intervals guided by mapping progression. In quartiles where the lead researcher determined that the mapper incorrectly identified or attributed a landform, the researcher would manually change the recording and add their initials to a “QAQC” field in the dataset. Administrator updates to the master database involved replacing quartiles with an edited QAQC field. Custom python tools (Arcpy module) were developed and implemented to ensure that the logic model of the SDI was respected before uploading data to the master database. For example, the scripts used relationships established in the logic model (Fig. S2) to flag fields that were incorrectly left blank. These grid cells would then be checked manually by the mapper or coordinator and filled in appropriately.

The mapping approach we have implemented is inherently challenged by human error during feature identification, and by interpreter variation (Pasher et al. 2013; Brardinoni et al. 2019). Clear guidance documents that included landform examples and how they manifest on Sentinel-2 imagery was provided to mappers (Fig. S1 and Appendix A) (e.g., Wolfe et al. 2022). The TMC employed conservative identification protocols to limit false positives (commission errors). For example, only active landslides features were counted by mappers. Moreover, active retrogressive thaw slumps are typically associated with areas of past disturbance (van der Sluijs et al. 2022), so excluding stable landslides was assumed to have little effect on final results depicting broad-scale spatial patterns. We identify that cloud cover or shadow in some instances could affect feature identification but this was minimized by obtaining images with less than 10% cloud cover. If cloud cover, snow, or other factors decreased the mappers ability to confidently assess terrain conditions of a grid cell, this was indicated in the comments box. We note that the 75 × 75 km² areas around the 33 communities, spanning the diversity of terrain across the study area underwent a QC check by a second independent mapper, but a QC check will not be possible for the entire study domain consisted of over 30 000 grid cells. Interpreter variation assessments comparing permafrost peatland coverage estimates in Gibson et al. (2020) reported that overall agreement was 84% better than chance, and comparisons will be reported for other landforms with the publication of theme-specific datasets.

Upon completion of the initial testing, clarifying landform interpretations, correcting interpreter biases, and with mappers trained, the inventory effort expanded and was divided according to the NT Ecosystem Classification Group Level IV ecoregion framework. In total, NT contains 162 Level IV ecoregions characterized by distinct regional ecological factors, including climate, physiography, vegetation, soil, water, and fauna (Ecosystem Classification Group 2008, 2009, 2010, 2012, 2013). Mappers were assigned a suite of Level IV ecoregions based on their expertise, and research interests and mapping is ongoing, with 76% of the mass wasting, 40% of the periglacial, and 37% of the hydrological theme mapping completed for NT at the time of paper submission.

A flexible and robust SDI was needed to accommodate the development of a data model, support the collaboration of a decentralized user base, and manage the generation of empirical permafrost data that covered a broad spatial domain. SDIs are geospatial data portals (Fig. 6A) that allow users to access remote or internal servers to connect to databases and interact with spatial data through a desktop GIS or web mapping service (Wang 2010; Maquil et al. 2018). Enterprise GIS systems allow multiple users to edit databases simultaneously (Maguire and Longley 2005; Kong 2015), common to large institutions with a centralized or internal user base and dedicated resources to manage networks (Fig. 6A) (Keating et al. 2003; Esri 2022). The broad-based collaborative nature of the TMC and its modest resources required that the project coordinator with geomatics expertise developed and maintained a low-cost and easy-to-implement SDI while enabling multiple novice GIS users to edit data (Fig. 6B). The project SDI that emerged allowed mappers to view, update, edit, and attribute project data easily, and accommodated the addition of new mappers from different institutions while maintaining mapping progress.

The project applied a file-based geodatabase architecture used in desktop-based GIS that was adapted to a multi-user environment (Fig. 6B). Each TMC mapper obtained the required vector and raster datasets through file-sharing networks (i.e., cloud storage) or with physical hardware (i.e., external drives), and incrementally updated personal copies of the database. In this method, the project coordinator received the updated data and ran custom-built python scripts to perform data quality assessments. The coordinator then merged the updates with a master database through additional rule-based scripts to overcome issues with conflicted copies that can otherwise lead to loss of data and confusion (MacEachren 2001; Yovcheva et al. 2013; García-Chapeton et al. 2018). The synchronous–decentralized collaborative framework consists of users across Canada with access to the tools and training to generate systematic inventory data on thermokarst and periglacial features.

Concurrent with the generation of TMC data, we also conducted NT-wide aerial inventories of thermokarst and periglacial landforms to document the nature of permafrost landscapes and support the validation of TMC outputs. A brief summary of the aerial survey approach is provided here. In the summers of 2020–2022, during the COVID-19 pandemic, northern project partners were supported by the Polar Continental Shelf Project to perform aerial inventories of thermokarst and periglacial landform features for over 37 000 km of flight lines across all major ecological regions of the NT. The team made over 7500 feature observations, recorded landform types and attributes, and obtained georeferenced images of each feature type inventoried. During 2020 fieldwork, the team iteratively refined the TMC database model (Fig. S2) to improve its suitability for aerial observations and implementation through a web-based GIS application that enables offline data collection (ArcGIS Survey123, ESRI™). The application allows users to collect GPS-attributed spatial data points on a field-based tablet (Apple iPad Mini with built-in GPS unit), and then upload observations to a cloud network (ArcGIS Online) when an internet connection is re-established. The aerial survey team comprises an expert observer who identifies and assigns attributes to a feature of interest, a photographer with a GPS-enabled DSLR camera who obtains georeferenced images of the feature, and a technician who enters the observation point, feature type, and its attributes into the Survey123 survey form. Survey rule sets and option pick lists enabled efficient data entry while ensuring that the data model was respected. The repetitive nature of this data collection provides training in thermokarst and periglacial landform identification and group learning, fostering a growing community of permafrost expertise. The aerial surveys also provided the first systematic, broad-scale, georeferenced inventory of thermokarst and periglacial features and their characteristics across NT and a unique dataset to confirm the nature and presence of landforms mapped by the TMC. We extracted the data for three feature types (RTS, polygonal terrain, and RL–L complexes) for comparisons with TMC data in the western Arctic and North Slave focal regions (Fig. 2 and Appendix A).

We also compared published, fine-scale empirical data products with these TMC datasets using confusion matrices and summarized prediction results. Here, we used high-resolution, regionally specific datasets on RTS (Lewkowicz and Way 2019), polygonal terrain (Lantz et al. 2017), and RL–L complexes (Morse and Rudy 2022) to evaluate the TMC approach and outputs. The TMC spatial outputs reflect the rapid assessment of features or landforms indicative of thermokarst or sensitive permafrost terrain that were quantified according to bins of counts or % area within each 7.5 × 7.5 km² quartile, whereas the independent fine-scale spatial datasets were typically polygons for spatially extensive landforms (e.g., ice-wedge polygon networks), or points for discrete features (e.g., RTS) (Appendix A). To enable comparison with the TMC grid-based inventories the validation datasets were converted to the raster framework and normalized to the TMC grid-based schema (Table S2) using a suite of geoprocessing tools.

To demonstrate the new insights that TMC datasets provide on the thaw sensitivity of permafrost terrain at broader spatial scales (10⁴–10⁵ km²), comparisons were also made with national or circumpolar scale products depicting thermokarst landscapes (Olefeldt et al. 2016), sensitive ice-wedge terrain (Karjalainen et al. 2020), or ground ice abundances (O'Neill et al. 2022). To conduct the comparisons, broadscale products were also normalized to the 7.5 × 7.5 km² grid cells, and we retained as many of their categorical attributes as possible (Table S3). Recognizing that the nature of empirically based TMC outputs differs from the modelled permafrost characteristics or terrain susceptibility, we emphasize that these comparisons aim to highlight the value of observational datasets for better understanding the spatial variation of the indicators that characterize thaw-sensitive permafrost terrain.

Spatial patterns of active or recent retrogressive thaw slumps (RTS) on Banks Island mapped by TMC highlight the thaw sensitivity of permafrost-preserved ice-cored moraine (Fig. 7A) (Kokelj et al. 2017). A point dataset of RTS derived from the Landsat time series (Lewkowicz and Way 2019) was converted to the raster framework, and the comparison with the TMC dataset shows good (84%) agreement (Figs. 7A–7C and Table 1A). There is high correspondence in unaffected terrain (95%), but lower agreement among the various disturbance classes likely reflecting differences in imagery and methodologies used for feature counts (Table 1B). The larger bottom plate of Fig. 7D shows RTS-affected terrain mapped by TMC for a 215 426 km² western Arctic region of NT, which follows general patterns described in coarser resolution published products (Kokelj et al. 2017, 2021). RTS-affected terrain comprise an area of about 60 990 km² or 28% of the quartiles, whereas shallow slides and gullied terrain affect only 16% of the quartiles. RTS-affected quartiles has an 81% correspondence with aerial survey observations (Table 2).

The broad-scale hillslope thermokarst landscape coverage from Olefeldt et al. (2016) indicates a significant departure from the TMC distribution of RTS, indicating only 12% overall correspondence (Figs. 7D and 7E, and Table S4Ai). The overall poor agreement relates mainly to the prediction of sensitive slopes or relict ice over large areas where no RTS were observed (Table S4Aii). The agreement between TMC-derived RTS-affected terrain and categories of relict ice abundance (O'Neill et al. 2022) is about 41% (Figs. 7D and 7F; Table S4Ai); however, several RTS hotspots in ice-marginal moraine are modelled as having high relict ground ice abundance (O'Neill et al. 2022). Both modelled products identify western Arctic Canada as being broadly thaw sensitive, with 83% and 41% of the TMC mapped RTS-affected grid cells occurring in moderate-to-high hillslope thermokarst sensitivity and relict ground abundance, respectively.

The mapped terrain affected by ice-wedge polygons (IWP) in uplands east of the Mackenzie Delta shows a 69% correspondence with polygon fields mapped by air photo interpretation (Lantz et al. 2017) (Figs. 8A–8C and Table 1A). These spatial datasets show an increasing northward trend in the abundance of polygonal terrain reflecting colder winter ground temperatures, frequent thermal contraction cracking, and larger, more widespread ice-wedge networks in the tundra than in the subarctic (Kokelj et al. 2014). Distinct areas of higher and lower polygonal terrain density in both datasets reflect variations in the abundance of lacustrine deposits that host polygonal peatlands (Fig. 1Biii) (Steedman et al. 2017). For the entire western Arctic region, polygonal terrain-affected grid cells have a 96% correspondence with aerial survey observations (Fig. 8D and Table 2).

Polygonal terrain occurs at high abundances in low- and mid-Arctic environments and decreases across the forest–tundra transition (Fig. 8D). The complex patterns of polygonal terrain distribution mapped by the TMC reflect variations in wedge ice abundance in conjunction with landscape types where polygon expression is best visible on Sentinel-2 imagery. Notably, the TMC identified polygonal terrain in the Mackenzie Delta, southwest of Inuvik, where ice-wedge networks are expressed by patterned forest (Kokelj and Burn 2004), high abundance was mapped across the ice-wedge-rich unglaciated Tuktoyaktuk Coastal lowland (Rampton 1988), and the greatest abundances were observed in association with ice-cored moraine around Paulatuk and along the northeast-trending bands of moraine on the northwest and southeast sides of Banks Island (Fraser et al. 2018). The distribution of IWP-affected terrain from the TMC is reflected by general patterns in modelled abundance of wedge ice from O'Neill et al. (2022) (Fig. 8E); however, detailed inspection reveals contrasts that result in a low agreement between the datasets (Table S4Bi). The circumpolar-scale statistical modelling of environments suitable for IWP from Karjalainen et al. (2020) was developed to assess climate change impacts on the distribution of these features. The majority of disagreement occurs because Karjalainen et al. (2020) modelled little variation in high IWP suitability across most of the terrain that the TMC inventory indicates has variable IWP abundance, producing large “errors” of commission (Figs. 8D and 8F, and Table S4Bii).

The distribution of terrain affected by thermokarst lakes and ponds that develop in ice-rich glaciolacustrine deposits of the Great Slave Lowlands was mapped by the TMC project (Fig. 9A). These thermokarst lake environments are referred to here as “RL–L complexes” following descriptions in Wolfe and Morse (2017) and Morse et al. (2017) (Fig. 1Ci). There is a 77% agreement between the TMC product and summarized data derived from high-resolution satellite imagery from Morse and Rudy (2022) (Figs. 9A–9C and Table 1A), and 74% correspondence between grid-based identification of RL–L complexes and aerial survey observations over the larger North Slave Region (Fig. 9D and Table 2). Regional patterns over an area of 51 968 km² indicate that the highest abundance of RL–L complexes occurs adjacent to Great Slave Lake, where ice-rich glaciolacustrine sediments are widespread. These thaw-sensitive thermokarst lake environments decrease in abundance with increasing elevation and distance from the lake as fine-grained deposits diminish and the landscape becomes increasingly bedrock-dominated (Fig. 9D). The patterns of modelled segregated ice abundance (Fig. 9E) (O'Neill et al. 2022) show reasonable agreement with mapped thermokarst lake landforms (55%) (Table S4Ci). The coarse–resolution surficial geology data used in the modelling indicates mainly bedrock in the region, explaining why segregated ground ice was not predicted in several areas of abundant thermokarst lake observations (Figs. 9D and 9E, and Table S4Cii). Appreciable fine-grained, frost-susceptible deposits are mapped in more detailed surficial geology products (Kerr et al. 2022), and their use would likely improve modelling outputs (O'Neill et al. 2022). Moderate-to-high density of thermokarst lake-affected landscapes were projected to occupy 77% of this study domain by a circumpolar assessment of thermokarst-sensitive terrain (Olefeldt et al. 2016) (Fig. 9F); however, the empirical mapping of this bedrock-dominated environment indicates that only about 9% of the area hosts moderate-to-high amounts of thermokarst lakes. In contrast with the possible underprediction of segregated ice abundance, the modelled thermokarst lake terrain has a poor agreement with the TMC mapping due to large errors of commission (Table S4Cii).

Here, we show that the holistic nature of the TMC inventory captures differences in thermokarst and periglacial characteristics across landscapes with contrasting permafrost, geomorphic, and Quaternary paleoenvironmental conditions (Fig. 10). We compared Eastern Banks Hills in the mid-Arctic region and the Tuktoyaktuk Coastal Plain in the southern Arctic region, both underlain by ice-rich glacial deposits and continuous permafrost, with the Great Slave Lowland within the Precambrian Shield and discontinuous permafrost zone where ice-rich terrain is limited to low-lying areas with glaciolacustrine deposits (Figs. 2 and 10).

Eastern Banks Hills and the Tuktoyaktuk Coastal Plain comprise landscape units widely affected by retrogressive thaw slumping (76% and 80%, respectively). In addition, 29% of the cells in the fluvially incised ice-cored environments of the Eastern Banks Hills are affected by larger (approximately >400 m) thaw-driven disturbances with debris tongues entering water bodies we term “mega slumps”, and 28% of the mapping units also have smaller shallow slides. Thaw-driven mass wasting features are not observed in the low relief and bedrock-dominated Great Slave Lowland. Hydrological features indicating thaw-sensitive terrains are similar between the two northern ecoregions. For example, shorelines affected by RTS occur in about half of the grid cells in each of the northern regions, with only limited evidence of recent lake expansion or lowered lake levels (recent drying or drainage).

Notable differences exist between the two northern regions regarding the nature and distribution of wedge ice. Polygonal terrain and polygonal-patterned ponding are observed across both environments; however, they are ubiquitous across the Eastern Banks Hills. On Banks Island, the top-down thaw of large ice wedges across upland terrain identified by polygonal-patterned pond expansion is also widespread, occurring at greater than trace amounts in 60% of the affected grid cells (Fraser et al. 2018). Ice-wedge thermokarst was detected at only trace amounts in 13% of the grid cells over the Tuktoyaktuk Coastal Plain, though recent top-down thaw and subsidence of hillslope ice-wedge troughs have been observed by field monitoring at Illisarvik just to the north of the mapped ecoregion (Burn et al. 2021). In contrast to the two Arctic ecoregions, the dominant indicator of thermokarst–hydrology interactions in the Great Slave Lowland manifests as RL–L complexes common throughout ice-rich glaciolacustrine deposits and observed in over 60% of the quartiles (Figs. 1Ci, 9, and 10) (Wolfe and Morse 2017; Morse and Rudy 2022). These landforms are also commonly characterized by thermokarst ponding or by dry basins, suggesting that water levels have lowered, which were observed in 31% and 34% of the grid cells, respectively.

In the periglacial theme, the Eastern Banks Hills are characterized by widespread polygonal terrain. In the ice-rich Tuktoyaktuk Coastal Plain, polygonal terrain is most common throughout the lowlands, which also host abundant pingos (Wolfe et al. 2023) detected in 72% of the quartiles. Involuted topography indicating ice-cored terrain (Dallimore et al. 1996) was observed in 61% of the grid cells in the Tuktoyaktuk Coastal Plain but is absent from Eastern Banks Hills. In contrast with the two northern ecoregions, this suite of periglacial landforms was not observed in the Great Slave Lowland. All three ecoregions are characterized by terrain with thick organic soils, which occurs in most grid cells on Eastern Banks Hills and is mainly restricted to small pockets (trace = 89%). In the Great Slave Lowland, peat-rich terrain was identified in 45% of the quartiles, mostly occurring as discrete features. In the Tuktoyaktuk Coastal Plain, terrains with thick organic soils are widespread and occur over larger portions of the landscape in 45% of the quartiles where lacustrine basins host extensive polygonal peat plateaus (Steedman et al. 2017). There was little evidence of degradation in northern peatlands, whereas greater than trace amounts of degradation was observed in most of the grid cells with organic terrains in the Great Slave Lowlands.

As permafrost landscapes change with climate warming and the abundance of modelled or remotely sensed spatial products increases, a heightened need exists for empirically based permafrost data to generate new knowledge and test model outputs. Comparisons with independent data demonstrate the effectiveness of TMC outputs in depicting thermokarst landforms or thaw-sensitive terrain. For example, the inventory of landscapes affected by RTS, IWP terrain, and RL–L (thaw lake) complexes show good agreement with fine-scale spatial datasets (10²–10⁴ km²) (Table 1; Figs. 7A–7C, 8A–8C, and 9A–9C) and correspondence with aerial inventories of thermokarst and periglacial features over larger areas (10⁴–10⁵ km²) (Table 2). The spatial data products developed by the TMC are valuable indicators of thaw-sensitive terrain (Figs. 7–9), so we compared them to broad-scale spatial data products that depict thermokarst landscapes and modelled ground ice abundance. We recognize the inherent differences between the empirical TMC spatial products, which portray observed phenomena and the modelling outputs where statistical or rule-based schemes were integrated through geomatics procedures to depict landscape susceptibility, likelihood of occurrences, or relative abundances. Regardless, we show significant deviations and some notable similarities between TMC and recent broad-scale maps (Figs. 7–9 and Table S4) highlighting the importance of empirically derived data for evaluating spatial modelling and remote sensing products, and generating new information on the distribution and diversity of thaw-sensitive permafrost terrain at regional to global scales.

The low agreement between several TMC data outputs and broad-scale synthesis products depicting the sensitivity of permafrost terrain underlines the risk with their application in regional-scale land use planning and decision-making, and for extrapolating local-scale field measurements to produce regional or circumpolar-scale summaries. Often, our basic theme-specific comparisons showed that differences of omission characterized the comparison with one modelled product, and differences of commission characterized the other (Figs. 9D–9F; Table S4Cii). A more rigorous empirical evaluation of model outputs is necessary to understand their limitations, inform the improvement of rule-based assumptions or input data layers used in spatial modelling, and focus on the production of robust regional-scale validation datasets (Wang et al. 2023). Our comparisons are also useful reminders that the application of these broad-scale spatial data products should be limited to showcasing general patterns and relevance of thaw-driven processes on terrain, ecosystems, carbon release, and infrastructure (Olefeldt et al. 2016; Chadburn et al. 2017; Hjort et al. 2018; O'Neill et al. 2022). Since these rule-based spatial products insufficiently represent (i) the heterogeneity in permafrost, soil, ground thermal, and terrain conditions, (ii) variation in climate drivers that yield a diverse array of landforms, and (iii) thaw-driven landscape responses, their use in prediction or decision-making is cautioned. Conceptual models provide a useful framework to connect geological legacies and paleoenvironmental history with the nature and occurrence of ground ice or thermokarst processes and landforms (O'Neill et al. 2019), but their implementation through broad-scale maps is constrained by the quality and resolution of existing geospatial data and the current state of permafrost–landscape knowledge. The integration and syntheses of fine-scale depictions of thaw-sensitive terrain can also be challenging because most datasets are necessarily limited in spatial extent, describe a narrow range of landforms (Steedman et al. 2017; Lewkowicz and Way 2019; Wang et al. 2023), and the methods and outputs vary because they are tailored to specific project needs and questions. Our development of systematic, empirically derived spatial datasets that quantify thermokarst and periglacial landform distribution for the permafrost regions of northwestern Canada addresses a foundational data and knowledge gap through initiating a more robust characterization of thaw-sensitive terrain over broad-spatial scales.

The depth of the TMC dataset also highlights that climate-driven permafrost thaw comprises a wide range of processes and landforms that manifest in complex combinations across the landscape. Recognizing the need to develop geomorphically based conceptual frameworks of permafrost thaw supported by more holistic approaches to quantifying thaw-driven landscape change, the TMC project data model and methods describe the variable nature of thermokarst and thaw-sensitive landscapes (Figs. 1, S1, and S2). The results demonstrate that this systematic inventory captures the diversity of permafrost terrain conditions within and between ecoregions, which can be described by their contrasting permafrost characteristics or thermokarst fingerprints (Fig. 10), providing novel datasets to explore variation in thaw-driven consequences on northern environments and society.

The TMC is guided by the science priorities of a northern-based team that leveraged the expertise of Federal and academic research partners to improve knowledge of the nature and distribution of thaw-sensitive permafrost terrain in the NT. Northern-driven permafrost science initiatives are rare in Canada because the capacity to conceptualize and implement research and monitoring programs has traditionally rested in southern institutions. The evolution of Territorial and Indigenous governance has stimulated policies to strengthen northerners’ roles in Arctic science (Inuit Tapiriit Kanatami 2018; National Geological Surveys Committee 2022), and the growth of northern-based research organizations has facilitated opportunities for science collaboration and co-developing knowledge (Thompson et al. 2020). The Territories’ investment in permafrost technical capacity has actioned science strategies (GNWT 2018) and enabled research partnerships to be leveraged through more equitable collaborations that address northern priorities. A key government–university partnership, the GNWT–WLU agreement, has also provided a framework and flexibility to co-manage a Yellowknife-based TMC project coordinator position. Critical support to the development, and rigor of technical products and outputs (e.g., Gibson et al. 2021; Wolfe et al. 2022) has been the sustained commitment of Federal and academic research partners. Based on our collective experience, we emphasize that developing permafrost knowledge collaboratively with northern partners can increase project relevance and accountability, build scientific and technical capacities, strengthen relationships with Indigenous partners and practitioners, and increase the likelihood that research outputs will be understood and applied in northern planning and decision-making.

Ongoing outreach has helped to identify a diverse range of northern needs, inform our approaches, prioritize mapping, and foster new partnerships. For example, the NWT Association of Communities emphasized the need for better information on the effects of permafrost thaw, and so the TMC prioritized the mapping and synthesis of permafrost sensitivity for NT communities and has presented results at related northern workshops and conferences (Table S5). The Territorial Lands and Infrastructure Departments have worked with the TMC to adopt mapping and aerial assessment methods over liabilities and infrastructure corridors. The TMC project framework, methods, and spatial data provide an entry point to foster the discussion about permafrost terrain sensitivity, community concerns, and land management needs. Relationships with Indigenous partners facilitated through projects or programs with established community linkages have created opportunities for information sharing, and the development or renewal of partnerships (Table S5). For example, the presentation of methods and maps at workshops with the Inuvialuit Land Administration, Łı́ı́dlı̨ı̨ Kų́ę́ First Nation, and K'asho Got’ı̨nę Foundation have supported fieldwork collaboration, discussions on permafrost-driven landscape change, areas of community interest, and the development of monitoring programs led by Indigenous Guardians and Land Management organizations. The project has fostered discussions around permafrost issues to strengthen the capacity of these partners through training, implementation of methods for fine-scale mapping and monitoring, and collaboration on aerial validation activities, and in turn gained insights on observations of change and research priorities. A new linkage has been established with the Gwich'in Land Use Planning Board who have gained some capacity to manage spatial data, conduct mapping, and leverage preliminary spatial data products from the project to support land use planning. A common constraint in effective science collaboration relates to capacity limitations of many northern partners, and the time and resources required by researchers to listen and develop spatial datasets that meet partner needs.

The project adopted a low-cost, user-centered solution (Fig. 6) that facilitated the exchange of geospatial data and services between mappers and researchers situated in different organizations across Canada (Hjelmager et al. 2008; Michener 2015) and allowed for design modifications as information and performance requirements evolved. This required geomatics capacity via a project coordinator who developed and maintained an SDI that could address data management and mapping needs. Effective communication between the coordinator and researchers was critical, ensuring that the SDI accounted for the capacity and resource limitations inherent in northern organizations, and maintained the flexibility required to support an evolving database and a growing number of collaborators in a decentralized mapping framework (Figs. 3 and 5).

A file-based architecture was necessary during the project's development phase but had demonstrated inadequacies as the number of users and requirements for quality control increased. For example, managing and merging overlapping edits in the master database required automated data quality assessments and updating, which became increasingly inefficient as the number of mappers increased, (García-Chapeton et al. 2018). An enterprise GIS system, which enables simultaneous database access by multiple users and automated updates, could improve efficiency but it requires reliable internet connections to remote servers, high annual subscription fees, and resources to support database and infrastructure management (Keating et al. 2003; Kong 2015; Coetzee et al. 2017; Esri 2022). Government partners have enterprise GIS functionality, but the implementation cycles and security requirements would have hindered project development and collaboration across institutions. While these issues were resolved by implementing a user-centered design (Fig. 6B), recent advances in cloud computing and remote servers provide efficient solutions for facilitating a decentralized–synchronous collaboration (García-Chapeton et al. 2018), overcoming SDI limitations that have challenged small organizations. Looking forward, these new tools have the potential to support decentralized, northern-focused research collaborations; however, GIS capacity will still be required for their implementation.

The implementation and progress of the TMC required perseverance, adaptability, and sustained commitment from a network of experts across NT and Canada. The decentralized approach and flexible organizational structure (Figs. 3 and 6) enabled iterative design, adaptation to changes in funding, accommodation of new stakeholder interests, and progress throughout the COVID-19 pandemic and onward. This project environment permitted revisions to scope and schedules and integration of new ideas. The growing connections with stakeholders and multiyear vision facilitated a stepwise approach with initial emphasis on documenting scientific methods, recruiting and training mappers, generating data, and validating results with field/aerial observations. Gradual progress has led to some changes in participants; however, the commitment of core contributors and organizations has maintained momentum (Fig. 3). The institutional support from the Territorial Government through enhancing northern-based permafrost capacity has enabled multiyear project planning, offsetting the limitations of short-term funding cycles typical of funding in academia (Pulsifer et al. 2011). This project has matured as feasibility and value was realized and communicated to partners and stakeholders through ongoing discussions and outreach (Table S5). For very large or centrally funded projects with rigid timelines and strict definitions of collaborators and scope, or in research environments where rapid publication defines success, the TMC approach would not be feasible.