Monitoring the thermal state of permafrost (TSP) is
important in many environmental science and engineering applications.
However, such data are generally unavailable, mainly due to the lack of
ground observations and the uncertainty of traditional physical models. This
study produces novel permafrost datasets for the Northern Hemisphere (NH),
including predictions of the mean annual ground temperature (MAGT) at the
depth of zero annual amplitude (DZAA) (approximately 3 to 25 m) and active
layer thickness (ALT) with 1 km resolution for the period of 2000–2016, as
well as estimates of the probability of permafrost occurrence and permafrost
zonation based on hydrothermal conditions. These datasets integrate
unprecedentedly large amounts of field data (1002 boreholes for MAGT and
452 sites for ALT) and multisource geospatial data, especially remote
sensing data, using statistical learning modeling with an ensemble
strategy. Thus, the resulting data are more accurate than those of previous
circumpolar maps (bias
Permafrost is defined as ground that remains at or below 0
Over the past half century, hundreds of permafrost maps have been compiled
at local to global scales (e.g., Heginbottom, 2002; Ran et al., 2012;
Cao et al., 2019; Zhelezniak et al., 2021). On a global scale, the first
permafrost map, the
Since the release of the IPA permafrost map, great advances have been made
in monitoring and modeling permafrost. In terms of data, the number of both
ground-based observations and remote sensing data have remarkably increased
in the past 30 years. Global and regional observation networks have been
gradually established worldwide, strengthening permafrost monitoring in many
regions, including Russia (e.g., Dvornikov et al., 2016), North America
(e.g., Romanovsky and Osterkamp, 2001; Smith et al., 2005, 2010), Central
Asia and China (e.g., Zhao et al., 2010b, 2021), and Europe (e.g., Harris et
al., 2001; Mair et al., 2011; Kellerer-Pirklbauer et al., 2016). As a result of these regional networks, the Global Terrestrial Network for Permafrost
(GTN-P) was established by the IPA in 1999 (Brown et al., 2000). The
GTN-P monitors the TSP and active layer thickness (ALT) through the TSP and
the Circumpolar Active Layer Monitoring (CALM) programs (Romanovsky et al., 2010; Biskaborn et al., 2015, 2019). At present, at a
global scale, data on ground temperature measured at approximately 700
boreholes and ALT data from more than 200 sites can be freely downloaded
from the GTN-P website (
With data accumulation and technical advances, several new hemispheric-scale
permafrost maps have been compiled. Gruber (2012) proposed a simple
semiempirical function relationship between the mean annual ground
temperature (MAGT) and mean annual air temperature (MAAT) for estimating the
global permafrost zonation at a 1 km scale using downscaled MAAT data. The
map indicates an areal extent of regions actually underlain by permafrost in
the NH of approximately
On the other hand, the maps of both the IPA and Obu et al. (2019) present permafrost zonation based on areal continuity. Areal-continuity-based map systems effectively reflect the permafrost distribution characteristics in high-latitude areas, but such systems are not suitable for describing high-altitude permafrost because the areal continuity of permafrost distribution is relative and scale-dependent (Nelson, 1989; Ran et al., 2012, 2021b). With a sufficiently high spatial resolution, all permafrost can be considered continuous or absent. Therefore, how to define the areal continuity of permafrost and the spatial resolution per se remains controversial and practically daunting. Additionally, continuity-based systems ignore the vertical distribution and longitudinal zonation of permafrost and thus cannot effectively reflect the hydrothermal conditions of permafrost, which are important for comprehensively understanding the characteristics and vulnerability of permafrost, especially at the regional scale (Cheng, 1984; Jin et al., 2014; Ran et al., 2021b). From a thermal stability perspective, for a given thermal condition and temperature increase in the air, permafrost temperature often responds more quickly in arid regions than in humid regions because of the much greater thermal inertia of wetter soils (Abu-Hamdeh, 2003). Dry soil can reduce evaporation heat consumption and increase the incident radiation on the soil (Pan and Mahrt, 1987). In addition, under different soil hydrological conditions, the response of permafrost to precipitation of various types may differ substantially (Trenberth and Shea, 2005). In humid regions, increased precipitation may heat the air and subsequently the active layer and permafrost, in contrast to that in arid regions. The difference in hydrothermal conditions is also reflected in the responses of ecosystems to precipitation because precipitation may alter or modify the effects of temperature on ecosystems (Zhao et al., 2018). Furthermore, ecosystems are very important for the thermal stability of permafrost under a changing climate (Shur and Jorgenson, 2007), although the interactions among climate, permafrost, and ecosystems are complex.
In short, the current statement of permafrost maps requires addressing the need for more comprehensive and integrated efforts to map the thermal state and hydrothermal zonation of permafrost. Thus, the objectives of this paper are as follows: (1) to develop and release the MAGT and ALT datasets as a baseline at 1 km scale in the NH, (2) to provide revised zonal statistics of the thermal state and distribution of permafrost in the NH, and (3) to present a permafrost zonation map that features the detailed hydrothermal conditions of permafrost.
We first compiled a ground measurement database for MAGT and ALT in the NH
(see the workflow in Fig. 1). Then, the MAGT and ALT predictions with a
1 km resolution in the NH were produced using ensemble statistical
forecasting by integrating the remotely sensed freezing degree days (FDD)
and thawing degree days (TDD) (i.e., the total annual degree days below and
above 0
The data processing workflow used to compile the permafrost datasets (FDD, freezing degree days; TDD, thawing degree days; LAI, leaf area index; SCD, snow cover duration; MAGT, mean annual ground temperature; ALT, active layer thickness; GAM, generalized additive model; SVR, support vector regression; RF, random forest; and XGB, extreme gradient boosting).
The standardized measurements of MAGT at or near DZAA (approximately 3 to 25 m) from 1002 boreholes and those of ALT from 452 sites were compiled mainly based on the ground measurement data used in Aalto et al. (2018), but the density of data points on the Qinghai–Tibet Plateau, in the Tian Shan, and in Northeast China was greatly increased (Fig. 2). Additional MAGT measurements from 253 boreholes are used mainly from the sources in Ran et al. (2021b), which were compiled mostly from existing literature on the Qinghai–Tibet Plateau (Wu et al., 2007; Yu et al., 2008; Sheng et al., 2010; Li et al., 2011, 2016; Zhang et al., 2011; Sun et al., 2013; Wang et al., 2013; Liu et al., 2015; Qiao et al., 2015; Wu et al., 2015; Qin et al., 2017; Cao et al., 2017; Luo, 2012; Luo et al., 2018a; Wani et al., 2020; Zhao et al., 2021) and the Tian Shan (Yu et al., 2013; Liu et al., 2015). Other additional MAGT measurements from 19 boreholes in Northeast China are from Li et al. (2019) and Chang (2011). Additional ALT measurements from 149 sites were mainly compiled from permafrost studies on the Qinghai–Tibet Plateau and the Tian Shan (Zhao et al., 2010a; Luo, 2012, 2018b; Yu et al., 2013; Wu et al., 2015; Cao et al., 2017, 2018; Ali et al., 2018; Wani et al., 2020) and Northeast China (Chang, 2011; He et al., 2018; Li et al., 2020b). Only the measurement data from the undisturbed (natural) sites were used. Incomplete or inaccurate location information in some studies was corrected and checked carefully via communications with the authors. The elevations of these boreholes range from 0 to 5428 m above sea level (a.s.l.), and 99 % of these measurements were made during 2000–2016.
The distribution of the boreholes (1002) for monitoring the mean
annual ground temperature (MAGT) and the sites (452) for monitoring the
active layer thickness (ALT) used in this study. FDD refers to the freezing
degree days (
For MAGT, to reduce the potential overrepresentation of MAGT observations
around 0
Overall, the proportion of MAGTs at or near DZAA was quite large when we considered only permafrost sites (84 %). However, for all 1002 sites, only 79 % of MAGTs were at or near the DZAA (Table A1 in Appendix A). This was due to the inclusion of non-permafrost sites for which there often was no information to determine the DZAA.
To reduce the potential overfitting due to residual autocorrelation, we resampled the training data 1000 times by excluding the sampling points within a distance of less than 3 km following Ran et al. (2021b). This resulted in using an average of 776 MAGT measurements and 276 ALT measurements for model training, and an average of 76 MAGT measurements and 30 ALT measurements were used for model evaluation per cross-validation run.
Nine environmental and climate variables were selected as predictors in a statistical learning model to estimate the permafrost thermal state based on previous studies (Aalto et al., 2018; Ran et al., 2021b). These variables were derived from high-quality datasets available at present (Table 1).
Environmental and climate variable datasets used in this study to predict MAGT (mean annual ground temperature) and ALT (active layer thickness). MODIS, Moderate Resolution Imaging Spectroradiometer; LST, land surface temperature; AVHRR, Advanced Very High Resolution Radiometer; GLASS, Global Land Surface Satellite.
The multiannual average FDD and TDD with a spatial resolution of 1 km
The annual SCD data were estimated from satellite-derived bimonthly global
snow cover extent (SCE) products obtained from the Japan Aerospace
Exploration Agency (JAXA) Satellite Monitoring for Environmental Studies
website. The SCE products with 0.05
The multiannual average values of LAI from 2000 to 2016 were derived from the Global Land Surface Satellite (GLASS), an 8 d, 1 km resolution LAI product. The GLASS LAI product was proposed based on the integration of MODIS and CYCLOPES LAI products, and the remaining cloud contamination values were removed using general regression neural networks (Xiao et al., 2014). The validation results show a higher accuracy for the GLASS LAI product in comparison with the MODIS and CYCLOPES LAI products (Xiang et al., 2014).
The soil organic content, bulk density, and coarse fragment content source
from SoilGrids250 (
The WorldClim v2.1 climate variables with a 1 km resolution (
The four statistical learning modeling techniques used in this study
include the generalized additive model (GAM) (Hastie and Tibshirani, 2017),
support vector regression (SVR) (Vapnik, 1995), random forest (RF) (Breiman,
2001), and extreme gradient boosting (XGB) (Friedman, 2001; Chen and
Guestrin, 2016). The techniques were implemented based on the R packages
mgcv (Wood, 2011) for GAM, randomForest (Liaw and Wiener, 2002) for RF,
e1071 (Karatzoglou et al., 2006) for SVR, and xgboost (Chen and Guestrin,
2016) for XGB. The GAM is a semiparametric extension of the generalized
linear model in which a smooth function is specified to fit the nonlinear
function of multiple predictors to the response variable at the same time
(Aalto et al., 2018). In this study, the maximum smoothing function was set
to three, and the thin plate regression spline was used as the smoothing
function. The RF is an ensemble learning algorithm for building and
aggregating multiple decision trees. In this study, the number of trees is
set to 400. Each tree is built using a randomly selected training dataset
and three environmental variables to split each tree node. SVR features
nonlinear kernel transformation, sparse solution, and maximal margin control
(Awad and Khanna, 2015). It assumes that the maximum deviation (
To account for the uncertainty of a single run and single model, an ensemble average of the 1000 runs for the four model means was used to represent the distribution of MAGT and ALT in this study.
The probability of permafrost occurrence is obtained by calculating the
fraction of predicted MAGT
We propose a classification system based on the hydrothermal condition of
permafrost by synthesizing the thermal condition system proposed by Cheng
(1984) and later modified by Ran et al. (2018) and by taking into account
the aridity system proposed by Jin et al. (2014) and UNEP (1997). This
system divides permafrost into nine categories using a two-level
hierarchical index system, i.e., with two criteria consisting of the MAGT
and climate aridity index (CAI). At the first level, permafrost is divided
into cold (MAGT
Model performance was assessed on the basis of the root-mean-square error
(RMSE), bias, and square of the correlation coefficient (
The cross-validation indicates that the ensemble average of four statistical
techniques (GAM, SVR, RF, and XGB) achieved the highest accuracy for MAGT
(RMSE
Based on the datasets (named NIEER in reference to the Northwest Institute of Eco-Environment and Resources, Chinese Academy of Sciences), we analyzed the distribution characteristics of MAGT, ALT, and the permafrost area/region changes along latitude, elevation, and aridity index transects and the hydrothermal conditions in the permafrost regions.
The predictive performance (mean
GAM: generalized additive model; SVR: support vector regression; RF: random forest; and XGB: extreme gradient boosting.
Distribution of the mean annual ground temperature (MAGT) at the zero annual amplitude depth (approximately 3 to 25 m) average in the Northern Hemisphere for the period of 2000–2016.
Figure 3 shows the distribution of MAGT in the NH displaying an obvious
latitudinal gradient from the zone of extremely cold (
Distribution of the average active layer thickness (ALT) in the Northern Hemisphere for the period of 2000–2016.
The three-dimensional ground thermal regimes across the NH are investigated
on the basis of the simulated MAGT and ALT. Figure 5a illustrates the
latitudinal distributive patterns of MAGT and ALT, which show some
comparable and contrasting features between those of MAGT and ALT. Regional
average MAGT is generally stable at approximately
Mean annual ground temperature (MAGT) and active layer thickness
(ALT) along latitude
The elevational effects on MAGT and ALT are modified by latitudinal effects
at the hemispheric scale because mountains are widely distributed in the NH,
with more in the lower latitudes. The combined effect of latitude and
elevation differs among different elevation ranges. From 0 to 2500 m a.s.l.,
contrary to our understanding of the elevational effect at the regional
scale, MAGT and ALT increase with increasing elevation, with large
variations (Fig. 5b). However, the variations show opposite trends within
this elevation range: with increasing elevation, the variation in MAGT
declines, whereas that in ALT increases. In this elevation range, the
elevational effect may be completely dominated by the latitudinal effect.
The combined effect of latitude and elevation is moderate between 2500 and
5000 m a.s.l., where MAGT is stable at approximately
The trend of MAGT corresponds well to that of ALT along the rising aridity index (Fig. 5c). For CAI values from 0.2 to 0.5, the MAGT and ALT both rapidly decrease with increasing CAI (moisture). However, in the range of CAI values from 0.5 to 3.6, MAGT and ALT both increase with increasing CAI. Then, above values of 3.6 on the CAI, both MAGT and ALT are nearly stable or slightly decreasing.
The three-dimensional zonal (latitudinal, elevation, and aridity)
characteristics of MAGT and ALT discussed above probably indicate different
controlling factors for MAGT and ALT and their variations. Multilinear
regression analysis shows that the contributions of precipitation and soil
bulk density to MAGT are statistically significant (
In addition, the DZAA represented by predicted MAGT varies in the NH (approximately 3 to 25 m). In general, in the continuous permafrost region in the Arctic and Qinghai–Tibet Plateau, the range in annual air temperature is large, and the zero annual amplitude is reached only at depths of 10 to 25 m below the surface (Ran et al., 2021b). These depths are greater than those in discontinuous permafrost and midlatitude regions, where the DZAA occurs at 5–10 m or less. This pattern is notably different from those of other MAGT products for specific depths.
For permafrost distribution, two terms, i.e., permafrost area and permafrost
region, should be distinguished. In general, the permafrost area is defined
as the area with MAGT
The present-day probability of permafrost occurrence in the
Northern Hemisphere for the period of 2000–2016 (the probability is defined
by the fraction of predicted MAGT
The spatial difference of the permafrost region proposed in this study (defined as the regions where the probability of permafrost occurrence is greater than 0) with the circum-Arctic permafrost map from International Permafrost Association (Brown et al., 2002).
In general, at a global scale, latitude, elevation, and aridity/longitude mainly govern the distribution of permafrost (Cheng, 1984; Noetzli et al., 2017). The latitudinal dependence of the permafrost distribution is based on the latitudinal variation in insolation (incoming solar radiation) and surface energy balance, and thus on the subsequent latitudinal zonation of climate, soil, and vegetation. The dependence of permafrost distribution on elevation is due to the dependence of air temperature, soil, and vegetation on elevation and the strong lateral water–heat fluxes that occur at different scales from the latitudinal effect (Noetzli et al., 2017). At a global scale, climate aridity affects the distribution of permafrost primarily by climatic continentality. This reflects the vulnerability of permafrost dependence on the annual mean temperature range and variation in the net effect of precipitation and evapotranspiration. Here, we investigate the distribution of the permafrost region and permafrost area along latitude, elevation, and aridity transects in the NH (Fig. 8).
Distributions of the permafrost region (permafrost
probability
The permafrost distribution depends strongly on latitude. In the NH,
permafrost occurs from 28
The permafrost distribution also clearly depends on elevation, especially in the mountainous and high-plateau regions at middle and low latitudes. Approximately 80 % of permafrost occurs below 1000 m a.s.l., and less than 10 % of permafrost occurs above 3000 m a.s.l. In the areas below 3000 m a.s.l., the area of permafrost decreases with increasing elevation, but its fraction of the land surface is approximately stable. In the areas above 3000 m a.s.l., the area of permafrost and its fraction of the exposed land surface both increase with rising elevation (Fig. 8b). This may indicate that latitude is the main controller of permafrost distribution below 3000 m a.s.l., while elevation is a more significant controller above 3000 m a.s.l.
Regarding the profile of permafrost distribution along northern longitude (Fig. 8c), our study results are generally consistent with those of existing studies (Zhang et al., 2008) but provide more spatial detail.
Hydrothermal-condition-based permafrost zonation in the Northern Hemisphere for the period of 2000–2016.
The hydrothermal conditions of permafrost differ remarkably over the NH, from the warm–arid type dominating on the Qinghai–Tibet Plateau to the cold–humid type dominating in the High Arctic. Figure 9 shows the distribution of permafrost hydrothermal conditions in the NH. In the Eurasian Arctic, permafrost occurs mainly as the cold–humid type in western Siberia and as the cold-semiarid/subhumid type in central and eastern Siberia. While cold–semiarid/subhumid permafrost prevails in the western Canadian Arctic, the eastern Canadian Arctic contains cold–humid permafrost. The cold–humid type is dominant in Greenland. In the Alaskan Arctic, there is cold–semiarid/subhumid permafrost. In the permafrost regions at middle and low latitudes, such as the Qinghai–Tibet and Mongolian plateaus, permafrost occurs mainly as the warm–arid type. Warm–humid permafrost mainly occurs to the south of the continuous permafrost zone in western Siberia and eastern Canada.
In general, regarding moisture conditions, permafrost in the NH is dominated
by the humid type, which accounts for approximately 51.56 %
(
The zonal areas of different permafrost hydrothermal conditions
(10
The list of permafrost datasets for the Northern Hemisphere produced in this study (MAGT: mean annual ground temperature; ALT: active layer thickness; GAM: generalized additive model; SVR: support vector regression; RF: random forest; and XGB: extreme gradient boosting).
The NIEER datasets generated by this study (Table 4) are publicly available and can be downloaded at the National Tibetan Plateau Data Center (TPDC)
(
This study produced MAGT at or near the DZAA and ALT datasets with 1 km
resolution for the period of 2000–2016 in the NH (named the NIEER map). The
datasets integrate unprecedentedly large amounts of ground/field measurement
data (1002 boreholes for MAGT and 452 sites for ALT) and multisource
spatial data, including remotely sensed FDD and TDD, LAI, SCD,
precipitation, solar radiation, soil organic content, soil bulk density, and
coarse fragment content, using a multiple-statistical/machine-learning model
with 1000 runs. The cross-validation shows that the accuracy of the MAGT
(bias
Data sources used to compile mean annual ground temperature (MAGT) datasets for the period 2000–2016. The number of observations, minimum, mean, and maximum depths of MAGT, and proportions of MAGT measured at or near the depth of zero annual amplitude (DZAA) are provided for each source.
YR, XL, and GC designed the study. YR prepared the datasets, wrote the manuscript, plotted the figures, and performed the analysis. YR designed the methodology, and JC implemented the statistical learning. JA, OK, JH, ML, QY, and XC contributed parts of the field data. JO provided the TDD and FDD data. XL, JA, OK, JH, ML, and HJ improved the writing and structure of the paper.
The contact author has declared that neither they nor their co-authors have any competing interests.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This article is part of the special issue “Extreme environment datasets for the three poles”. It is not associated with a conference.
This study was jointly supported by the Strategic Priority Research Program of the Chinese Academy of Sciences (grant no. XDA19070204) and the National Natural Science Foundation of China (grant no. 42071421). Olli Karjalainen and Jan Hjort acknowledge support from the Academy of Finland (grant no. 315519). FDD and TDD data were generated by the European Space Agency GlobPermafrost project (grant no. 4000116196/15/I-NB).
This research has been supported by the Strategic Priority Research Program of the Chinese Academy of Sciences (grant no. XDA19070204), the National Natural Science Foundation of China (grant no. 42071421), and the Academy of Finland (grant no. 315519).
This paper was edited by Kirsten Elger and reviewed by Stepan Varlamov and two anonymous referees.