| 20 May 2026
An Ensemble Dataset of Permafrost Thaw Conditions for Northern High Latitudes from Open Satellite Data
Abstract. Permafrost in the northern high latitudes is experiencing accelerated warming, casting important implications for carbon–climate feedback and ecosystem stability. Integrating three open-access permafrost products, this study generates an ensemble permafrost dataset of permafrost percent (PP) and mean annual ground temperature (MAGT), and proposes a Permafrost Thaw Index (PTI) to measure thaw vulnerability in a two-tier ranking system. Sixteen open-access datasets derived primarily from satellite observations are ensembled to represent environmental conditions, including land surface temperature (LST), vegetation dynamics, snow cover, freeze–thaw state, soil properties, and topography. An ensemble machine learning approach, XGBoost, is employed to predict PTI from these datasets with an overall accuracy of 91.8%. Thermal variables, particularly LST, LST trends, and frozen days, contribute most strongly to PTI prediction. The PTI map reveals clear latitudinal gradients and regional variations in alignment with eco-climatic transitions. The total permafrost coverage in the northern high latitudes (> 45°N) is approximately 18.7 million km2. About 41.4% remains highly stable, while nearly half is subject to high thaw pressure. The most stable conditions persist in the High Arctic tundra of North America, the Arctic East Siberia, and the high-elevation alpine tundra of North Asia. The highest vulnerability occurs along the southern margins dominated by boreal and montane forests. A comparison of the predicted PTI with borehole records from 26 established stations reveals strong agreement (Spearman's r = 0.69). This study demonstrates the value of multi-source open satellite data in permafrost research. The proposed PTI framework provides a scalable approach for monitoring permafrost dynamics and supporting climate impact assessments over the northern high latitudes. The ensemble products of this study, including the PTI map and the ensemble land cover and permafrost distributions (PP and MAGT), can be freely downloaded at https://doi.org/10.5281/zenodo.19148960.
To the authors of Duo et al.
Certainly there is value in collating together globally datasets of permafrost conditions for use by the broader community so I applaud the authors in attempting to do so. However, there are also challenges that emerge when averaging or presenting an ensemble of datasets that suffer from the same flaws in similar geographic areas. In these cases, it is best to hatch out or not present these regions where there was no source data or where there are known problems rather than to have these issues extend throughout the analysis.
I'll call attention to one particular region of my research interest, Labrador, northeast Canada, where we've noted significant problems with representing coastal permafrost in previous works (Way and Lewkowicz, 2016; Way et al. 2018; Wang et al, 2023). Of particular relevance was the work by Wang et al (2023) which directly compared against Obu et al (2021) and others, finding a complete misrepresentation of the distribution of permafrost in southeastern Labrador. Part of this was because of poor land cover products but other parts of it relate to not representing coastal tundra snow redistribution well. This current work seems to reproduce this challenge across its maps. This issue is less prominent on the IPA map which does consider geomorphological observations in its derivation. Although lower resolution, the IPA map is still the gold standard for most areas in the discontinuous zone in Canada.
It is also notable when the datasets do not necessarily use all available public borehole sources. For example, Nordicana D includes a wide variety of boreholes from locations in northeastern Canada not well-represented by the data used by the authors.
I do understand the need for global efforts, but it's important to remember that these products get used by others and in some cases this can result in real harm during environmental assessment processes if whole geographic regions have had their permafrost or ground ice conditions mischaracterized. This is all to say that far more effort needs to be given to improving models/maps by examining where they have failed and not simply averaging out or reproducing the same issues identified with prior efforts. At the very least, a more expansive literature search is needed to identify the areas where these existing models/maps fail to accurately characterize permafrost at broad regional scales.
Way, R.G., and Lewkowicz, A.G. 2016. Modelling the spatial distribution of permafrost in Labrador–Ungava using the temperature at the top of permafrost. Canadian Journal of Earth Sciences, 53(10): 1010–1028. doi:10.1139/cjes-2016-0034.
Way, R.G., Lewkowicz, A.G., and Zhang, Y. 2018. Characteristics and fate of isolated permafrost patches in coastal Labrador, Canada. The Cryosphere, 12(8): 2667–2688. doi:10.5194/tc-12-2667-2018.