We searched for GT and GST data from scientific, northern- and permafrost-specific data repositories and databases including Nordicana D, the Yukon Geological Survey's Permafrost Database, the Northwest Territories Geological Survey's (NTGS) Geoscience Publications and Reports database, Arctic Data Center, Borealis, Mendeley Data, and Natural Resources Canada's (NRCan) Open S&T Repository. In addition to published datasets, we also sent personal requests and received contributions of unpublished ground temperature data from individuals.

We use the tempcf tool, a program with a graphical user interface for inspecting and cleaning ground temperature data . It allows the read-in of a variety of different file formats supported by TSP, a python library for working with permafrost ground temperature data . Data cleaning operations performed in tempcf are automatically documented in a log file, making the resulting data easily traceable to the original observations, which is foundational for reproducibility. For final data output, this tool is also used to create both text and NetCDF files. The former files were exported using TSP into a standardized ground temperature data format (the definition of which is included as File S2 in the Supplement), and the latter are based on the NSERC PermafrostNet ground temperature NetCDF template.

Data processing consisted of three main components: (1) compiling metadata sheets for unpublished datasets and extracting baseline metadata for all datasets according to CF conventions, (2) quality checking GT and GST data, and (3) generating the final data outputs. These steps and their relationships are summarized in Fig. .

The standardization of this dataset using the CF and NetCDF ecosystem allows us to also make it available on an ERDDAP endpoint (https://data.permafrostnet.ca/erddap). Although a permafrost-specific data distribution software does not yet exist and there are several choices of existing platforms for this purpose, ERDDAP offers several features that make it valuable for distribution of permafrost data . It is already well-established and supports common standards (WMS, ISO-19115, DAP) for data transport and metadata description. This makes the data more FAIR – for example, by making it discoverable on data aggregators such as Google dataset search – and also serves to integrate it within the broader earth science data ecosystem. Because it is general-purpose, it is possible to accommodate both discrete observations like ground temperature alongside model output and gridded products. The interface provides both interactive and programmatic ways to access, query, and download the data. Finally, ERDDAP servers from different organizations can be federated in such a way as to make data collections searchable across organizational boundaries.

The dataset combines 13 unpublished datasets and 29 published, including six with updates to the original releases (Table ).

Unpublished datasets were supplied by university and government researchers. One dataset was provided directly by ECCC and is distributed under their license agreement (see File S1). Published datasets were sourced from online generalist and northern-focused repositories and databases.

The datasets by and that we retrieved from Mendeley Data are distributed under a Creative Commons Attribution 4.0 International (CC BY 4.0) licence (https://creativecommons.org/licenses/by/4.0/, last access: 21 July 2026) without additional endorsement from the authors for this publication. The original datasets were modified for use in this study and all modifications are documented in the log files included with the final compiled dataset. Full creator names, dataset titles, and DOIs are provided in the references.

Data reproduced from the Report on 2010–2015 Permafrost Thermal Investigations in the Yellowknife Area, Northwest Territories and the Report on 2012–2017 Shallow Ground Thermal Investigations on the Tibbitt to Contwoyto Winter Road Portages, Northwest Territories, both published by the Geological Survey of Canada (NRCan), are included in this compilation. This reproduction is a copy of official NRCan works and has not been produced in affiliation with, or with the endorsement of, NRCan.

The dataset includes 427 GT and 491 GST sites located across British Columbia, Northwest Territories, Nunavut, Québec, and Yukon, with 60 % of the GT sites recording permafrost (Fig. ). Approximately 90 % of the recording depths are shallower than 10 m and have measurements that occur after 2008 (Fig. ).

Temperatures in the dataset range from -49.8 to 33.6 °C. The distribution of temperatures is negatively skewed for both ground surface (μ = -2.16 °C, median = -0.84 °C) and deeper ground temperatures (μ = -2.10 °C, median = -0.75 °C). Both distributions are strongly leptokurtic and this is more pronounced for observations below 20 cm than for observations at the surface (Fig. a). Few boreholes extend past 20 m and there are clear preferential measurement depths (e.g., at 10, 12, 15, and 18 m) corresponding to greater observed ranges (Fig. b).

The complete dataset is freely available for download at 10.5281/zenodo.18022924 and via an ERDDAP server (https://data.permafrostnet.ca/erddap).

Data are provided in several formats to ensure the broadest possible user base (Table ). Individual sites are exported as NetCDF datasets using the “Time Series of Profiles” discrete feature geometry available in the CF conventions . These site-level datasets are self-describing and include relevant information on provenance and, where available, links to the original publication.

Although the specific instrument used to measure ground temperature is sometimes provided in the source datasets, this is not always the case. However, the accuracy, resolution, and drift for common ground temperature sensors are tabulated in , and these can be used alongside typical instrument precisions reported in this dataset to estimate instrumental error when it is missing. Reported accuracies from manufacturers range from 5 to 400 mK and precision ranges from <1 to 40 mK. We recommend conservative estimates of 400 and 40 mK for the instrumental accuracy and precision, respectively, for sites where these values are not provided. Instrumental noise, but not bias, will be reduced by daily averaging. Sensor drift reported by manufacturers is between 0.002 and 0.1 °C yr⁻¹, but studies suggest it is often a nonlinear phenomenon caused by degradation of equipment under field conditions . Drift is most readily detected at the ground surface via change to temperatures in successive zero-curtains (periods during thaw or freeze when temperature is steady near 0 °C). Drift at depths beneath the influence of seasonal temperature fluctuations can be suspected when a trend is present that is not visible in neighboring sensors. For this contributions, we checked for obvious signs of sensor drift during the manual inspection phase; confirmed drifting data was removed and suspected drift was flagged.

The calculation of daily means from data of differing sampling frequencies introduces frequency- and amplitude-dependent error. We can consider the possible bias by approximating the daily temperature fluctuation as: 3T(t)=Tavg+Acos⁡2πtP where Tavg is the true daily mean, A is the amplitude, and P is the period (1 d).

For data with only one observation per day, the sampling frequency is less than the Nyquist frequency for the diurnal signal and the observation can differ significantly from the true daily mean; biases as large as the amplitude A are possible.

For single observations per day, if we assume the sampling time-of-day is randomly distributed, we can use the definition of the standard error of the mean (SE) 4SE=σN along with the standard deviation for a sine wave . 5σ=A2 to quantify more precisely the standard error of the mean as: 6SE=A22N which, for n=1 yields A2.

Sampling rates between twice-daily and 2-hourly, while greater than the Nyquist frequency for the diurnal signal, are subject to bias that will depend on how much the daily temperatures deviate from a simple harmonic behaviour. These errors are difficult to quantify but will change in time and with different ground conditions. Zero-curtain behaviour in ice-rich ground will, for example, tend to worsen estimates of the daily mean.

For daily means calculated from sensors sampling at greater frequencies, we can estimate bias in a worst-case scenario where missing data is concentrated at the warmest or coldest part of the day: 7ΔTmax⁡=1αP∫P2(1-α)P2(1+α)T(t)dt-Tavg8ΔTmax⁡=-Asin⁡(πα)πα which, for completeness ratios α of 70 % , 80 % and 90 % of available daily data evaluates to maximum (worst-case scenario) biases of roughly 36 %, 23 %, and 11 % of the amplitude at any given depth, based on Eq. (). Completeness ratios can be estimated using the expected and total number of observations provided in the data (netcdf and aggregated csv file only: see Table ). A description of typical amplitudes are provided below.

In all cases, averaging error is expected to scale with amplitude. In near-surface sensors, the amplitude changes significantly with both depth and season. Empirically, we find that the distribution of amplitude values is positively skewed and when partitioned according to whether the daily minimum temperature is above or below zero, the median amplitude is 3.28 °C in warm, unfrozen conditions (IQR = 5.20 °C) and 0.17 °C in cold conditions (IQR = 0.50 °C). Importantly, for depths below 1.8 m, the error on daily means is negligible (Appendix ).

Errors associated with the estimation of daily mean will affect applications comparing daily values between locations. However, this error is less significant when comparing trends between sites. Daily mean temperature bias caused by incomplete sampling or instrument bias will persist when calculating mean annual values.

The assumption of a simple harmonic diurnal cycle is most appropriate for conditions where latent heat effects are negligible. In near-zero temperature conditions where the effects of phase change dampen or distort the diurnal signal, the true bias may be greater than the worst-case estimates provided here. , working with empirical ground temperature data to estimate MAGT in a context where latent heat effects are present, report a qualitatively similar pattern of estimation error: uncertainty increases with both temperature amplitude and sampling interval. This corroborates the behaviour predicted by our analytical approach.

At 134 sites, the UTC offset of source datasets was not known with certainty. In these cases, we treated the data as though they were in UTC time. Within Canada, UTC offset ranges from -03:30 to -08:00. This introduces up to an 8 h offset in the means.

At an additional 64 sites, data were obtained as daily averages from observations in local time. Without original sub-daily measurements it was not possible to perform a time zone correction. For internal consistency and ease of use, we present all data as UTC daily means. However, for the 200 sites mentioned above, there is an additional source of error corresponding to the shift in sampling period. This is indicated as “possible_utc_error” in the data.

While the exact magnitude of these errors will depend on daily variability in the timing and magnitude of temperature peaks we can use a simplified model to get a first order approximation. Although the simple harmonic model described in Eq. () will not be affected by the UTC shift, we can add a linear temperature trend of magnitude b °C d⁻¹. In this case, a shifted daily averaging window will introduce a bias of b24Δt where Δt is the magnitude of the offset in hours. Bias will be greater when temperatures change more quickly: at shallower depths and unfrozen conditions. For this dataset, we calculate median GST changes using a 2 d lag to be 0.23 °C d⁻¹ (IQR = 0.52 °C d⁻¹) in cryotic conditions and 0.41 °C d⁻¹ (IQR = 0.72 °C d⁻¹) in non-cryotic conditions. For GT these values are 0.01 °C d⁻¹ (IQR = 0.05 °C d⁻¹) and 0.05 °C d⁻¹ (IQR = 0.14 °C d⁻¹) for cryotic and non-cryotic conditions, respectively. For a worst-case scenario of an 8 h offset in unfrozen ground near the surface, this corresponds to a bias of roughly 0.07 °C (median) up to 0.52 °C (90th percentile).

This bias, while typically small, will have the greatest impact when comparing data from affected sites to other records such as daily model outputs or meteorological observations. Site-level trends will be unaffected. Coarser aggregation to monthly or yearly averages will render the relative magnitude of the error insignificant, as will estimation of quantities like active layer thickness or depth of zero annual amplitude.

Where available in the source publications, sampling frequencies were provided for pre-aggregated sites, whereas sampling frequencies for the subdaily sites were estimated using a binary segmentation algorithm. This was found to be reliable when data completeness was high or when measurement frequency did not change. However, at sites where data completeness is extremely sporadic, there is a greater uncertainty on estimates of data completeness. Also, during the transition period between sampling regimes, there will inherently be a day which is considered under sampled in the faster sampling regime but may be fully sampled in the slower sampling regime.

The accuracy of the site coordinates is not available for all sites. Where instrumental accuracies of coordinates are not provided, we estimate them to be on the order of 10 m for sites with data dating back to 2010 (recently established locations: modern GPS devices using multiple constellations), 20 m for sites with data dating back to 2000 (after which selective availability on GPS was turned off), and 50 m for sites with data dating back to before 2000 (based on GPS with selective availability or careful positioning on local maps).

Where elevation data were not available from primary sources, values were obtained using the Google Maps Elevation API. However, Google does not publicly disclose the underlying digital elevation model(s) used by this service, and the data sources may vary by geographic region and change over time as Google updates its mapping products. Consequently, the vertical accuracy of these elevation values cannot be independently verified or consistently reproduced in future studies. This represents a limitation of the present dataset, and users should be aware that any replication of this workflow at a later date may yield slightly different elevation values for the same locations. Error in sensor position will also increase the error of our “gap-filled” elevation estimates using the Google Maps API. Elevation uncertainty is rarely provided and should be conservatively estimated to be on the order of tens of metres.

For individual GST loggers buried in the ground, the datalogger orientation, definition of the ground surface, and differences in protocols between research groups means depth values should be treated with uncertainties of ±5 cm. For ground temperature cables in boreholes, depths should also be assumed accurate to within ±5 cm. The spacing between sensors on thermistor strings in GT sites can be measured in advance and can be considered accurate to within 1 cm.

The depth at which sensors are positioned can change over time due to ground subsidence or frost-jacking of the borehole in which thermistors are installed. The effect of this can include the appearance of spurious temperature trends and changes to the amplitude of the daily and annual fluctuations over time. Although the most obvious of these errors will have been removed during the manual inspection process, there is still the possibility that more subtle errors persist. While these errors can be confidently detected by measuring changes to stick-up height – the distance between the ground surface and the top of the borehole – these measurements are rarely provided.

If changes in sensor position remain undetected, measurements will not be comparable across sites (e.g., temperatures from a subsidence-affected site compared with a bedrock site). Additionally, observations may not be directly comparable to model results during evaluation, unless the model explicitly represents processes such as ice loss and subsidence.

Classification of sites as being disturbed or within bedrock was done based on site descriptions, photos, or existing classifications in the source datasets. Values of 1 (disturbed) should be treated with higher degree of confidence than 0 (undisturbed) because evidence of disturbance may have been missed, or not included in the source dataset. Values of -1 (unknown) were reserved for sites lacking any sort of supporting evidence but will likely tend to be undisturbed rather than disturbed because of the overall tendency of datasets to report disturbance explicitly. Therefore, applications categorizing sites into disturbance categories may find some success in treating unknown sites as undisturbed. Importantly, the categorization and significance of the disturbance is not universally standardized; natural background variation caused by proximity to a water body is treated the same as sites instrumented within a road. The group of disturbed sites should be expected to have more significant conceptual variability.

Although the dataset represents a variety of different climatic settings and terrain types, the spatial sampling of ground temperature monitoring is inherently biased. The drilling equipment used to establish permafrost monitoring sites is more suitable for fine-grained soils than diamicts, coarse blocky material, or bedrock. For this reason, we expect an over-representation of fine-grained soil among GT sites. Similarly, open areas are easier to drill and instrument than swamps, bogs, or dense bushes, so there will be a bias towards certain surface conditions. Nor should ground surface loggers be considered randomly sampled; in a given study region they may be intentionally placed to investigate the maximum variability within the landscape e.g.,, or to study a particular phenomena or terrain type. The sampling locations within the dataset are more numerous in the Yukon and Northwest Territories and in areas of relatively low relief. Conclusions about the general behaviour of permafrost must take into consideration this spatial bias.

For applications in which individual sensors were used to represent a larger area or grid cell, we would expect a slight cold bias in discontinuous or sporadic permafrost zones where sensors are often placed preferentially in colder ground that is more likely to be underlain by permafrost.

Individual sites have been separated by maximum measurement depth into ground surface temperature (<0.2 m) and ground temperature (>0.2 m) for ease of use. Ground temperature records may also include shallow measurements within the upper 0.2 m.

Sensor intervals for the ground temperature sites are variable and generally increase with depth. This can impact the reliability of derived information, including thaw depth estimates and talik detection. Maximum measurement depths are restricted by drilling equipment, subsurface conditions, or are determined by research specific purposes. As a result, sites with similar depths and sensor spacing often cluster according to drilling campaigns, and some boreholes may represent similar subsurface materials, as instrumentation frequently fails when encountering coarser materials.

One limitation of the dataset presented here is the lack of detailed site metadata. This is due, in part, to the absence of such information in the source material, but also to the challenge of interoperability between data sources when such information exists. Records of “disturbance” for example are common in metadata descriptions but use different terminology, which often lacks a simple one-to-one correspondence .

Several metadata templates and sampling protocols for permafrost data have been developed by research programs and institutions . These typically describe one or more of the following: what data should be collected, how it should be collected, how it should be stored (file type, structure, or schema), or how it should be shared (web protocols). Although a single, unified metadata or dataset template is unrealistic because of the differing needs of programs and organizations, it should be a priority to increase interoperability of data between existing organizations. There are a few first steps that can be taken to support this.

First, organizations with developed data or metadata templates should describe them fully to encourage adoption by others. Iterations should be versioned to allow for modification while ensuring that older data remains well-documented. Clear descriptions of file formats also make it possible to integrate them into unifying software libraries e.g.,. To this end, we have included a detailed description of the file format standard for permafrost ground temperature data used by this dataset. It is included as File S2. The format is designed for ground temperature data and can be extended to include borehole data (interval measurements) and other non-temperature sensor data.

Second, the development of shared vocabularies within and across institutions will allow better site-level description across data sources. Authoritative community resources such as the permafrost glossary developed by can be used as a rich source for permafrost-specific terms such as landforms. These also have the benefit of including an agreed-upon definition. From a technical standpoint, a good start can be made by adopting and adapting tools and practices from other scientific disciplines to turn these vocabularies into digital resources e.g.,. This will allow equivalencies or crosswalks to be made between different metadata terms to aggregate them in a repeatable, traceable way.

Finally, it is essential that the development of data and metadata protocols be done through collaboration between both permafrost modelers and field practitioners; increasingly, large monitoring installations will be made, and funded, in the context of Northern infrastructure projects where contributing to a global monitoring network is, at most, a secondary priority. Creating these interfaces will be a multi-year process requiring persistent investment and time commitment. It will also be an opportunity for better integrating field practices between differing groups and organizations, and for introducing new field practices that will be robust in a future when permafrost is increasingly dominated by thaw phenomena .

The Canadian Tri-Agency Research Data Management Policy requires that digital research data supporting published conclusions be deposited in a trusted repository, while acknowledging that what counts as relevant data remains subject to disciplinary norms. In permafrost science, norms for sharing remain inconsistent despite clear calls for open data and interoperability . While capacity constraints limit the ability to manage and publish data, fears of being scooped, losing publication opportunities, or misuse of data also continue to limit sharing behaviour, a pattern observed broadly across science and known to reduce research impact and efficiency e.g.,.

Progress toward stronger data stewardship is underway in Canada through inter-agency cooperation , investments in tools and standards , and community building through events e.g., and efforts such as this contribution. Establishing a durable disciplinary norm will require socio-technical support that lowers the burden on data providers and encourages across government, academia, and professional societies. This must also coincide with support and continued funding for field specialists and technicians without whom the instrumentation, maintenance, and data acquisition would not be possible.

As an Arctic nation, Canada has both a responsibility and an opportunity to lead in the stewardship of permafrost data see also. Building scientific sovereignty – maintaining the capacity to generate, manage, and interpret northern data – can go hand in hand with enabling collaboration and contributing high-quality, interoperable data for global benefit.