Published August 25, 2014
Pedology
Active-Layer Thickness across Alaska:
Comparing Observation-Based Estimates with
CMIP5 Earth System Model Predictions
Umakant Mishra*
Environmental Science Division
Argonne National Lab.
9700 S. Cass Ave.
240-6143
Argonne, IL 60439
William J. Riley
Earth Sciences Division
Lawrence Berkeley National Lab.
One Cyclotron Rd.
50-4037
Berkeley, CA 94720
Predicted active-layer (AL) thicknesses of permafrost-affected soils influence
earth system model predictions of C-climate feedbacks; yet, only a few observation-based studies have estimated AL thicknesses across large regions and at
the spatial scale at which they vary. We used spatially referenced soil profile
description data (n = 153) and environmental variables (topography, climate,
and land cover) in a geographically weighted regression approach to predict
the spatial variability of AL thickness across Alaska at a 60-m spatial resolution.
The predicted AL thickness across Alaska ranged from 0.14 to 0.93 m, with a
spatial average of 0.46 m and a coefficient of variation of 30%. The average
prediction error and ratio of performance to deviation were 0.11 m and 1.8,
respectively. Our study showed mean annual surface air temperature, land
cover type, and slope gradient were primary controllers of AL thickness spatial
variability. We compared our estimates with Coupled Model Intercomparison
Project Phase 5 (CMIP5) earth system model predictions; those predictions
showed large interquartile ranges in predicted AL thicknesses (0.35–4.4 m)
indicating substantial overestimate of current AL thickness in Alaska, which
might result in higher positive permafrost C feedback under future warming
scenarios. The CMIP5 predictions of AL thicknesses spatial heterogeneity were
unrealistic when compared with observations, and prediction errors were several times larger in comparison to errors from our observation-based approach.
The coefficient of variability of AL thickness was substantially lower in CMIP5
predictions compared to our estimates when gridded at similar spatial resolutions. These results indicate the need for better process representations
and representation of natural spatial heterogeneity due to local environment
(topography, vegetation, and soil properties) in earth system models to generate a realistic variation of regional scale AL thickness, which could reduce the
existing uncertainty in predicting permafrost C-climate feedbacks.
Abbreviations: AL, active-layer; CALM, Circumpolar Active Layer Monitoring Network;
CMIP5, Coupled Model Intercomparison Project Phase 5; ESM, earth system models;
GWR, geographically weighted regression; IPCC, Intergovernmental Panel on Climate
Change; MEE, mean estimation error; NLCD, National Land Cover Database; RPD, ratio
of performance to deviation.
P
ermafrost-affected soils contain about twice the amount of C currently in the
atmosphere (Schuur et al., 2009; Tarnocai et al., 2009). These large C stocks
are stabilized primarily because of low temperatures, cryoturbation, and peat
formation (Gorham 1991; Bockheim 2007). Over the next century, high-latitude regions, where permafrost-affected soils are an important characteristic, are predicted
to experience much higher temperature increases than temperate or tropical regions
because of polar amplification processes (Holland and Bitz, 2003; IPCC, 2007). The
Soil Sci. Soc. Am. J. 78:894–902
doi:10.2136/sssaj2013.11.0484
Received 15 Nov. 2013.
*Corresponding author ([email protected]).
© Soil Science Society of America, 5585 Guilford Rd., Madison WI 53711 USA
All rights reserved. No part of this periodical may be reproduced or transmitted in any form or by
any means, electronic or mechanical, including photocopying, recording, or any information storage
and retrieval system, without permission in writing from the publisher. Permission for printing and for
reprinting the material contained herein has been obtained by the publisher.
Soil Science Society of America Journal
potential thawing of permafrost C makes these regions a vulnerable component of the global C cycle (McGuire et al., 2009; Schuur
et al., 2008; Schuur and Abbott, 2011), with the potential for large
positive feedbacks with the atmosphere.
Within a soil profile under laid by permafrost (the portion
of the soil profile where temperatures have remained at or below
0°C for at least two consecutive years), the AL is the top portion
of the soil column that freezes in winter and thaws in summer
(Zhang et al., 2005). The AL thickness or depth is measured as
the maximum depth of thaw in any particular year. The spatial
variability of current AL thickness is a critical component in prediction of the C-climate feedbacks associated with permafrostaffected soils (Nelson et al., 1998; Zhang et al., 2012).
Ground temperature observations from Alaska over the
last 30 yr have indicated an increase in permafrost temperatures
of 0.5 to 3°C (Osterkamp, 2005). These temperature increases
have been linked to increased AL thickness and decreased permafrost extent (Zhang et al., 2005; Jorgenson et al., 2010). The
large increases in temperature (3.2–3.5°C) predicted by 2100
in high-latitude regions (IPCC, 2007) could deepen AL thickness, thereby decreasing permafrost thickness and moving soil
organic C into the AL, making it more susceptible to microbial
decomposition and erosion. Despite recent advances in landsurface models, accurate predictions of AL thicknesses at regional and circumpolar scales remain a major scientific challenge
(Riseborough et al., 2008).
Previous observation-based studies have reported AL thickness at several sites in Alaska. At 21 sites in the Arctic Coastal
Plain and Arctic Foothills of Alaska, AL thickness ranged from
30 to 90 cm (Bockheim, 2007). In an observation-based spatial
modeling study, Nelson et al. (1997) showed variability of AL
thickness to range from 0 to more than 70 cm in a heterogeneous landscape in the Kuparuk region of north central Alaska.
Hinkel and Nelson (2003) showed spatial patterns of average
(1995–2000) AL thickness ranging from 20 to 120 cm at seven
circumpolar monitoring sites in northern Alaska. These studies
reported varying environmental controls of AL thickness depending on spatial scale. For example, at the landscape scale, AL
thickness was strongly correlated with mean annual air temperature. At the local scale, topographic attributes, vegetation types,
and soil properties were primary controllers of AL thickness. We
note that different controls across scales may emerge when considering processes important to climate change prediction, e.g.,
C and nutrient cycling, which will also be impacted by other processes (e.g., moisture, vegetation) with their own spatial scaling
properties (Crow et al., 2012).
Recent process-based and earth system modeling studies
(Lawrence and Slater, 2005; Marchenko et al., 2008; Schaefer et
al., 2011; Koven et al., 2011; Koven et al., 2013) have predicted
spatial variability of AL thickness across Alaska and the Northern
Hemisphere at coarse spatial resolutions (>100 km). The modeled AL thicknesses from earth system models (ESMs) influence
the predicted C-climate feedbacks at a global scale. Among different ESMs, the Coupled (i.e., interactive two-way coupling bewww.soils.org/publications/sssaj
tween the land and atmosphere) Model Intercomparison Project
phase 5 (CMIP5) model simulation experiments used state-ofthe-art ESMs that incorporate mechanistic descriptions of the
soil and surface energy balance and C cycle dynamics to study
global C-climate feedbacks (Taylor et al., 2009). The purpose of
these experiments was to address scientific questions that arose as
part of the Intergovernmental Panel on Climate Change (IPCC)
fourth assessment report, improve understanding of climate, and
provide estimates of future climate change that could be used by
climate change adaptation policies (Taylor et al., 2009). The results from these CMIP5 models are being used in the current
IPCC fifth assessment report. Though these models predict
general patterns of permafrost distribution and their responses
to climate change, they are difficult to use for land-use planning
and management and for ecological monitoring and assessment (Zhang et al., 2012) because they are unable to represent
the fine-scale spatial heterogeneity in climate (e.g., temperature,
moisture) and environmental parameters (e.g., topography, vegetation, soil, and hydrologic conditions) reported to affect AL
thickness (Nelson et al., 1998; Riseborough et al., 2008).
Thus, tracking subgrid heterogeneity (e.g., in topography)
over regional scales, such that the predicted AL depths reflect
the heterogeneous controls of environmental variables, remains a
critical scientific challenge (Riseborough et al., 2008). To evaluate land model predictions of current and future high-latitude
C exchanges with the atmosphere, we believe there is a need for
observationally constrained characterizations of AL at spatial
scales consistent with existing heterogeneity. To that end, we developed finely resolved observation-based estimates of the spatial
variability of AL thicknesses across Alaska and linked these values to climate and environmental controllers. Finally, we compared our results with other observation-based studies, several
uncoupled (i.e., not prognostically coupled to an atmospheric
model) land model predictions, and eight CMIP5 ESM predictions (Table 1).
MATERIALS AND METHODS
Study Area and Soil Profile Observations
This study was conducted in Alaska, which covers a land
area (excluding ice, water, and bare rocks) of 1,221,272 km2. A
total of 153 georeferenced permafrost soil profile observations
were collected from a recently published database (Michaelson
et al., 2013). This database updated all the existing pedon observations from Alaska and contains observations collected from
the University of Alaska Fairbanks northern soils research program and USDA-NRCS. In this database, the AL thickness of
59 samples were directly measured by digging soil profiles and
measuring the depth to the permanently frozen portion of the
soil profile, and in the remaining 94 samples AL thickness was
calculated as depth to soil horizons with permanently frozen layers as indicated by the “f ” suffix, such as Cf, Cgf, and Bgf (Soil
Survey Staff, 2010; Mishra and Riley, 2012; Ping et al., 2013) of
the soil profile. Both model calibration and validation data are
sparsely and unequally distributed across Alaska (Fig. 1a) pri895
marily due to logistical difficulties and Table 1. Eight Coupled Model Intercomparison Project Phase 5 (CMIP5) model outputs
used for active-layer thickness comparison across Alaska.
extreme working conditions of permaModels
Modeling groups
References
frost terrain.
Environmental Datasets
BCC-CSM1-1
CanESM2
GFDL-ESM2G
GISS-E2-R
HadCM3
HadGEM2-CC
HadGEM2-ES
MPI-ESM-LR
Beijing Climate Center
Canadian Earth System Model
Geophysical Fluid Dynamics Laboratory
Goddard Institute for Space Studies
Hadley Center Coupled Model
Hadley Center Global Environmental model
Hadley Center Global Environmental model
Max Plank Institute Earth System Model
We used a digital elevation model
of 60-m spatial resolution from the
USGS database (Multi-Resolution
Land Characteristics Consortium,
2011). The elevation of the study area
ranged from sea level to 6188 m. Highelevation areas are located in the southeastern part of Alaska, and
low-elevation areas are located in the western and northern parts
of Alaska. We derived several primary and secondary topographic attributes to evaluate their use in predicting AL thickness.
These topographic attributes included elevation, slope, aspect,
curvature (plan, profile, and total), upslope contributing area,
flow length, soil wetness index, sediment transport index, stream
power index, terrain characterization index, and slope aspect index (Thompson et al., 2006). Land cover data of 60-m spatial resolution were extracted for Alaska from the National Land Cover
Database (NLCD; Multi-Resolution Land Characteristics
Consortium, 2011). We reclassified the NLCD land
cover types into nine major categories (Table 2). The
largest land area was under the scrub (shrub less than 5
m tall) category (43%), followed by forest (25%), barren (8.5%), herbaceous (7%), and wetlands (7%). The
remaining surface area (9.5%) was under open water, perennial ice, barren, cultivated, and developed categories.
For climate data, we used the 30-yr (1961–1990) mean
annual air temperature and annual precipitation from the
PRISM database (Daly et al., 2000). Across Alaska, the
mean annual air temperature and mean annual precipitation ranged from −18 to 6°C and 150 to 8500 mm, respectively. Indicator variables for the presence or absence
of land cover types were created and used in the model
selection process. Climate datasets that were available
at 2 km spatial resolution were resampled to a common
spatial resolution of 60 m by using the resample function
of ArcGIS (ArcGIS version 10, Environmental Systems
Research Institute, Inc., Redlands, CA).
Ji, 1995
Verseghy, 1991
Dunne et al., 2012
Rosenzweig and Abramopoulos, 1997
Cox et al., 1999
Essery et al., 2003
Essery et al., 2003
Raddatz et al., 2007
equal to the number of predictors (small bias of the regression
model). Among the environmental variables considered in this
study, only slope, temperature, and land cover type were significant predictors of AL thickness across Alaska (Mallow’s Cp
= 3.4). Selected independent variables were tested for unequal
error variance, multicollinearity of variables, normality, and randomness of residuals. The SAS statistical software (SAS Institute,
2004) was used for this purpose. Selected independent variables
were then used in a GWR approach, and the model parameters
were derived at a 1000-m regular interval throughout the study
area. The calibration R2 ranged from 0.15 to 0.65 in the study
Spatial Modeling, Environmental Controls,
and Comparison with Coupled Model
Intercomparison Project Phase 5 Predictions
We used a geographically weighted regression
(GWR) approach (Mishra and Riley 2012) to predict
the AL thickness across Alaska. First, the best subset
regression was used to identify the environmental variables that could be used in further modeling steps (Table
3 and 4). We used a Mallow’s Cp criterion for model selection (Kutner et al., 2004). In Mallow’s Cp criterion, a
subset of predictors were identified for which Cp value
was small (small total mean squared error) and nearly
896
Fig. 1. Distribution of observations (a) and predicted active-layer thickness using
geographically weighted regression across Alaska (b). Black color indicates surface
area without permafrost based on Brown et al. (1998). CALM, Circumpolar Active
Layer Monitoring Network.
Soil Science Society of America Journal
Table 2. Reclassification of USGS land cover types for this study.
=
RMSE
National Land Cover Database land cover type Reclassified land cover type
Developed open space, low intensity, medium
intensity, and high intensity
Developed
Deciduous, evergreen, and mixed forest
Dwarf scrub and shrub scrub
Shrub, sedge, and moss
Pasture and cultivated lands
Woody and herbaceous wetlands
Barren
Open water
Perennial ice
Forest
Scrub
Herbaceous
Cultivated
Wetland
Barren
Open water
Perennial ice
^
1 n
( D( xi ) − D( xi )) 2
∑
n i =1
^
where D( xi ) is the CALM-measured AL thickness, D( xi )
is the predicted AL thickness (both from our approach
and ESMs), and n is the number of validated observations
(Mishra et al., 2012). Both MEE and RMSE values should
approach zero for an optimal prediction. We calculated the
ratio of performance to deviation (RPD, defined as the ratio between the standard deviation and the RMSE), which
indicates the overall prediction ability of the selected aparea. In GWR, the weight function was chosen as an adaptive
proach (Chang and Laird, 2002). A higher RPD value indicates
spatial kernel type so that the spatial extent for included samples
greater accuracy in prediction.
varied based on sample density (Fotheringham et al., 2002). The
Environmental controls on observation-based AL thickness
GWR approach used in this study can be represented as
predictions were examined by converting temperature and slope
gradient data into different zones (creating zones of 4°C, and 10°
^
^
^
^
^
slope) and then calculating the statistical parameters, such as avD i = β0 (ui , vi ) + β1 (ui , vi ) X i1 + β 2 (ui , vi ) X i2 + ⋅⋅⋅⋅⋅ + β k (ui , vi ) X ik
erage and standard errors of AL thickness in each zone. Similar
^
where D i is the predicted AL thickness (m) at location i; (ui, vi)
calculations were performed for impacts of land cover types on
^
^
are the coordinates for location i; β0 to β k are regression coeffiAL thickness variability.
cients; Xi1 to Xik are environmental variables at location i; and k
We also compared our models’ representation of current
is the number of environmental variables. The GWR model paAL thickness distribution across Alaska with predictions from
rameters used in this study are provided in Table 4.
offline models and eight coupled ESMs that contributed to the
The prediction accuracy of the resulting AL thickness maps
CMIP5 database (Taylor et al., 2009). Active-layer thicknesses
(both from our approach and ESMs) was evaluated using data
(thaw depth) in CMIP5 models were calculated using monthly
from 34 locations across Alaska from the circumpolar activemean soil temperatures from the depth resolved ESM grid cells.
layer monitoring network (CALM, 2013) (Brown et al., 2000);
Monthly mean thaw depth was calculated as the deepest point in
these data were not included in the development of our GWR
the soil column of a given grid cell at a given month by defining the
model and so allow for an independent check on our predictions.
freezing point as 0°C. Annual AL thickness was then calculated
GWR model predicted values of AL thicknesses were extracted
as the maximum monthly thaw depth for a given year. Detailed
for these locations. The observed values from the Circumpolar
description of this calculation is provided in Koven et al. (2013).
Active Layer Monitoring Network (CALM) dataset and preFrom the eighteen CMIP5 models that reported sufficient
soil temperature predictions to infer AL thickness, we selected
dicted values from our approach and ESMs (CMIP5 coupled
simulations) were interpreted by calculating mean estimation erthe top eight of the models based on the closest median AL thickness predictions compared to our GWR prediction estimates in
ror (MEE) and RMSE:
predicting AL thickness across Alaska: Beijing Climate Center
^
1 n
=
MEE
( D( xi ) − D( xi ))
Climate System Model version 1-1 (BCC-CSM1-1), Canadian
n i =1
Earth System Model version 2 (CanESM2), Geophysical Fluid
Dynamics Laboratory Earth System Model with GOLD Ocean
Table 3. Linear regression model predictors selected from best
Component (GFDL-ESM2G), Goddard Institute for Space
subset regression approach for predicting log-transformed
active-layer thicknesses.
Studies model E coupled with Rusell Ocean Model (GISSE2-R), Hadley Center coupled model version 3 (HadCM3),
Predictor
Model coefficient
t value
P value
Intercept
4.1
56.75
<0.001
Hadley Center Global Environmental Model version 2 C Cycle
Slope
−0.013
−1.81
0.023
(HadGEM2-CC), Hadley Center Global Environmental Model
Temperature 0.04
4.8
<0.001
version 2 Earth System (HadGEM2-ES), and Max Plank Institute
Barren
0.33
2.8
0.004
Earth System Model Low Resolution (MPI-ESM-LR). For comparison between our observation-based and the CMIP5
Table 4. Geographically weighted regression model summary for logpredictions, we used median and interquartile range of AL
transformed active-layer thicknesses.
thicknesses predicted in different models and calculated
Predictor Minimum Lower quartile Median Upper quartile Maximum
MEE, RMSE, and RPD using the CALM validation dataIntercept
3.7
4.01
4.02
4.04
4.6
set. We also compared the CVs of AL thickness in selected
Slope
−0.06
−0.016
−0.008
−0.007
0.003
models with CVs of our predictions by gridding them at the
Temperature
0.01
0.017
0.027
0.038
0.09
same spatial resolutions.
Barren
0.14
0.20
0.37
0.78
0.92
∑
www.soils.org/publications/sssaj
897
RESULTS AND DISCUSSION
Descriptive Statistics of Observed Data
Across the observations from the 153 permafrost pedons,
observed AL thicknesses showed a large range with positively
skewed distribution (coefficient of skewness = 1.2). The average
AL thickness for all of Alaska was 0.49 m, ranging from 0.16 to
1.27 m, with a high spatial variability (CV = 62%). In cases where
the study area is large and observations are distantly located, such
as ours, the mean value of the same variable could be different in
different locations. In such cases CV is useful for assessing the spatial variability of investigated variable (Webster and Oliver, 2007).
The median AL thickness and the standard deviation were 0.45
m and 0.20 m, respectively. Consistent with expectations, the observations indicated deeper AL thickness in the southern part of
Alaska and shallower AL thickness in the northern parts of Alaska.
Spatial Variability of Observation-Based ActiveLayer Thickness Predictions
Using the GWR spatial prediction approach described above,
we predicted the average AL thickness for Alaska to be 0.46 m
(range of 0.14–0.93 m). The average AL thickness increased
from north to south; the lowest values were in the arctic coastal
plains, and the highest values were in the subarctic coastal plains
and lowlands of the Bering Taiga ecoregion (Fig. 1b). On average
our approach under predicts the AL thickness by 3.6 cm (MEE =
−3.6 cm). The average error (RMSE) of AL thickness prediction
compared to the CALM observations was 0.11 m (25% of mean
value), and the observed RPD was 1.8. These global validation
indices showed good prediction accuracy for AL thickness across
the state (Chang and Laird, 2002). However, lower predicted CVs
(30%) in comparison to observed values (62%) likely indicate (i)
data smoothing in predicted values by GWR and (ii) the impact
of other environmental factors not included in our analysis, such
as cryogenic features (high and low centered polygons, pingos,
frost boils), organic layer thickness, and fire intensity. Geospatial
datasets of these potentially important factors do not exist now, so
we could not incorporate them in our spatial prediction approach.
The use of these datasets should be evaluated once they become
readily available. Because of the modest predictive capability of
our method (RPD = 1.8) against the CALM observations (which
were not used in the model development), we consider the GWR
predicted AL thicknesses presented here to be a first step in the
development of our method and one that will benefit from more
observations spread across the topographic, edaphic, and climatological gradients known to impact AL thickness.
Current sampling density in this study was ~7900 km2 per
sample (for model calibration data). As suggested by Fig. 1a, these
samples were primarily distributed across the road networks (accessible areas), and large surface area of Alaska remained unsampled because of logistical difficulties (both economic and human)
limiting widespread sampling efforts in the remote and extreme
environment (Kuhry et al., 2010; Mishra et al., 2013). This also
shows potential opportunity for improvement in our spatial modeling approach as additional samples become available.
898
We compared our predictions with other previously reported
observation-based values. Bockheim (2007) analyzed 21 pedons
across Alaska that were not included in the dataset applied here
and reported an average AL thickness of 0.47 cm (0.30–0.90
m). In a spatial modeling study in the Kuparuk region of Alaska,
Nelson et al. (1997) predicted AL thickness from 0 to more than
0.7 m. At seven sites in Alaska, Hinkel and Nelson (2003) predicted AL thickness from 0.20 to 1.2 m using field observations and a
linear interpolation algorithm. Our predicted mean and range of
AL thicknesses across Alaska (0.14–0.93 m) are within the ranges
of these observation-based studies.
Predicted Environmental Controls of
Active-Layer Thickness
Across the land cover categories, the largest predicted mean
AL thickness (0.54–0.56 m) was found under cultivated and developed land cover types (based on n = 7 and 14 observations,
respectively) (Fig. 2a). These categories represent disturbed land
Fig. 2. Distribution of active-layer thickness across different land
cover types (a), temperature zones (b), and slope gradient (c). Error
bar represents the standard error.
Soil Science Society of America Journal
cover types due to human activities and have reduced natural
vegetation cover in comparison to other land cover types. As a
result, the annual temperature wave causes soils to thaw at greater depths in these land cover types. Smallest mean AL thickness
(0.27 m) was found under barren land cover types (based on n =
9 points). Barren lands in Alaska are associated with mountainous terrain and higher sloped positions. Mean AL thicknesses of
forest (n = 29), grassland (n = 47), scrub (n = 36), and wetlands
(n = 11) were intermediate between the above-mentioned land
cover categories (Fig. 2a). Type and coverage of vegetation influence the AL thermal regime and its thickness primarily through
altering the surface energy balance by acting as a buffering layer
between the atmosphere and the ground (Williams and Smith,
1989). Without separating impacts of different vegetation types,
however, Nelson et al. (1997) also reported strong influence
of vegetation on mean AL depth in the Kuparuk Basin area of
north central Alaska. Consistent with our results, vegetation
types were reported as an important component to model the
spatial variation of AL thickness (Walker et al., 2003). Increased
plant biomass was reported to have a negative relationship with
AL thickness because of the insulating effects of vegetation and
development of high organic soil horizons (Walker et al., 2003).
Also, land cover types were reported to be important controllers
of fine resolution spatial variability of AL thickness (Zhang et
al., 2012).
The mean predicted AL thickness increased with increasing
mean annual air temperature (Fig. 2b). This result is consistent
with the results of several former studies and analytical estimates
(Stefan and Kudryavtsev equations; Riseborough et al., 2008).
For example, at the landscape scale, the AL thickness was strongly correlated with local mean annual air temperature (Hinkel
and Nelson, 2003). Mean annual air temperature is often used
as a sole predictor of permafrost occurrence for the mapping of
large geographical areas (Riseborough et al., 2008). Also, summer air temperature has been widely documented as a principal
control on AL thickness (Zhang et al., 1997, 2005). Various authors have used summer air temperature to estimate the annual
thawing index and multiplied this index with edaphic factor(s)
(e.g., soil conductivity, moisture content, bulk density, latent
heat of fusion) to estimate AL thickness over several spatial scales
(Romanovsky and Osterkamp,1995; Nelson et al., 1998; Brown
et al., 2000; Hinkel and Nelson 2003). Despite its extensive use,
use of thawing index in predicting AL thickness across regional
scales has also been associated with large errors primarily because
it doesn’t account for (i) winter and spring warming of soils
(Serreze et al., 2000), (ii) key process governing AL thickness
such as less winter cooling of the surface soil (Zhang et al., 2005),
and (iii) surface morphology and vegetation impacts (Nelson et
al., 1997).
Predicted mean AL thickness decreased with increasing
slope gradient (Fig. 2c). Topographic attributes including slope
gradient have been reported to have strong control over the spatial variability of AL thickness at several scales, as it influences
microclimate, hydrology, soil moisture, and vegetation (Hinkel
www.soils.org/publications/sssaj
and Nelson, 2003; Riseborough et al., 2008). In nonhuman altered landscapes such as much of Alaska, topographic attributes
will be more prominent in describing variability of soil properties in comparison to human altered landscapes. Higher slope
gradients indicate well drained soils with lower soil moisture and
extremely low thermal conductivity (Nelson et al., 1997, 1998)
in comparison to saturated or water-logged soils, which typically
have lower AL thicknesses. The negative relationship between
AL thickness and slope gradient can also be explained by considering the top-soil thickness as a balance between soil production and erosional processes (Heimsath et al., 1997). That means
higher sloped positions in the landscape are more erodible, indicative of shallower top-soil thickness in comparison to lower
sloped positions that are less erodible, and indicative of deeper
top-soil thickness. Further, higher sloped positions and mountainous landscapes covary in Alaska, both indicative of shallower
AL thicknesses.
Comparing Observation-Based AL Thickness
Estimates with Offline Model and Coupled Model
Intercomparison Project Phase 5 Predictions
Previously reported offline (meaning uncoupled with a
prognostic atmospheric model) land model predictions of AL
thickness were deeper than our observation-based predicted values and other reported observations discussed above. For example, Marchenko et al. (2008), using a spatially distributed numerical modeling approach, predicted AL thickness across Alaska to
be 0 to 3 m. Schaefer et al. (2011), using a climate-scale landsurface model, predicted AL thickness across the entire northern
circumpolar region; their predictions for Alaska ranged from
0.8 to 2 m. Our observation-based AL thickness predictions are
within the lower range of these large-scale modeling predictions
(0.14–0.93 m).
We also compared our results with the coupled CMIP5 earth
system model predictions. Making these comparisons was complicated by the fact that, in Alaska, many of the CMIP5 models
predicted that permafrost did not exist where the observations suggested it should exist (Koven et al., 2013). We, therefore, selected
the top eight of the models based on the closest median AL thickness predictions across Alaska compared to our observation-based
estimates (Table 1). Because of these complications, we report the
comparisons here to argue that substantial work is required in the
land models before they produce believable permafrost temperature
predictions, and that there is value to generating observation-based
high-resolution estimates for this type of comparison.
Among the eight ESMs analyzed here, GFDL-ESM2G and
HadCM3 showed the most realistic representation of AL thickness across Alaska (spatial CVs of 30.5 and 31%, respectively).
Although these models did not predict accurate current permafrost extent over the circumpolar region, they reported the lowest bias compared to inferred current climatic controls and circumpolar median AL thickness (Koven et al., 2013). The eight
CMIP5 model predictions of median AL thickness across Alaska
ranged from 0.35 to 3.4 m, with a spatial CV ranging from 11 to
899
Fig. 3. Predicted active-layer thickness from eight Coupled Model Intercomparison Project Phase 5 (CMIP5) earth system models. The dots and
error bars indicate the predicted median thicknesses and inter-quartile ranges across Alaska. Our observation-based approach is labeled GWR for
geographically weighted regression.
47% (Fig. 3). The average of the median prediction of AL thickness across these models was 1.9 m, with a spatial CV of 23%.
Our results showed substantially lower (3.5 times) AL thickness
(median = 0.54 m) and a different spatial variability in comparison to these estimates when gridded to the same spatial resolution (spatial CV of 28–32%). These comparisons suggest that
ESMs substantially overestimate current AL thickness in Alaska,
which might result in higher positive permafrost C feedback under future warming scenarios. Without an ability to query the
individual ESM models directly, it is not possible to tell what
causes the discrepancy between our observation-based results
and the model predictions nor is it possible to know whether an
ESM with a good comparison to the observation-based estimates
has low bias for the correct reasons.
Our geospatial interpolation approach considers the impacts of surface mean annual air temperature, land cover types,
and topographic attributes. In a previous analysis of circumpolar
AL thickness predictions in CMIP5 models, Koven et al. (2013)
focused on surface mean annual air temperature as the primary
controller of AL thickness, although other ESM variables were
either not represented (topography) or not reported (land cover
type). That analysis demonstrated a wide range of AL thickness
predictions between models across the circumpolar Arctic, with
the differences being attributed to differences in the modeled
coupling between (i) near-surface air and surface soil temperatures or (ii) surface and deeper soil temperatures.
When compared with CALM validation datasets, the
CMIP5 predictions of AL thickness showed large inaccuracies.
The average prediction errors (RMSE and MEE) in four of the
compared CMIP5 predictions (BCC-CSM1-1, CanESM2,
GISS-E2-R, and MPI-ESM-LR) were several times higher in
comparison to GWR estimates (Fig. 4). The remaining four
model predictions (FDL-ESM2G, HadCM3, HadGEM2-CC,
and HadGEM2-ES) showed larger but relatively comparable
prediction accuracies with GWR results. Similarly, the overall
predictive quality (RPD) was 1.3 to 36 times lower in CMIP5
Fig. 4. Prediction accuracies of eight Coupled Model Intercomparison Project Phase 5 (CMIP5) earth system modeled active-layer thicknesses and
our observation-based geographically weighted regression (GWR) approach across Alaska [compared to Circumpolar Active Layer Monitoring
Network (CALM) data].
900
Soil Science Society of America Journal
predictions in comparison to our estimates. These large errors
suggest the need for improved process representations and representation of spatial heterogeneity due to local environment (topography, vegetation, and soil properties) in earth system models to generate a realistic variation of regional scale AL thickness,
which potentially can reduce the existing uncertainty in predicting permafrost C-climate feedbacks in permafrost regions.
SUMMARY AND LIMITATIONS
In this study, we used soil profile observations, environmental variables, and a geospatial prediction approach to predict the
spatial distribution of AL thicknesses at a 60-m spatial resolution
across Alaska. The low mean bias of our predictions compared
to CALM observations and high RPD illustrate the useful influence of the environmental variables we considered in determining the AL thickness distribution. Further investigation will
be required to understand the reasons our predictions led to a
higher bias at larger AL thickness and vice versa when compared
to the independent CALM dataset. We compared our GWR AL
predictions to CMIP5 ESM predictions and found large differences, with the CMIP5 models generally predicting much larger
AL thicknesses. Comparisons with several offline models also indicated that our observation-based estimates were typically lower
than the model predictions.
Lower predicted spatial variability of our fine-resolution
spatially interpolated results compared to the independent
CALM observations indicate that the spatial datasets for other
soil-forming factors are important for high-latitude environments, such as cryogenic features, time since pedogenesis, fire
frequency, and fire intensity. Use of spatial datasets of these variables could potentially increase the accuracy of AL thickness
predictions using the geospatial methods applied here. Similarly,
the prediction accuracy of our results could also be increased by
including additional soil profile samples across the largely unsampled regions of Alaska. Because of these limitations, we consider the observation-based AL thicknesses presented here to be
initial estimates that will require future refinement. With these
limitations in mind, we believe our fine-resolution predictions of
AL thicknesses across Alaska can be a useful resource to evaluate
and improve earth system models.
Acknowledgments
This study was jointly supported by two grants from the Director, Office
of Science, Office of Biological and Environmental Research, U.S.
Department of Energy under Lawrence Berkeley National Laboratory
contract no. DE-AC02-05CH11231 and under Argonne National
Laboratory contract no. DE-AC02-06CH11357. We appreciate
help from Charlie Koven who provided the calculation of active-layer
thickness from the CMIP5 models across Alaska.
References
Bockheim, J.G. 2007. Importance of cryoturbation in redistributing organic
carbon in permafrost-affected soils. Soil Sci. Soc. Am. J. 71:1335–1342.
doi:10.2136/sssaj2006.0414N
Brown, J., O. Ferrians, Jr., J. Heginbottom, and E. Melnikov. 1998. Circum-arctic
map of permafrost and ground-ice conditions. National Snow and Ice Data
Center/World Data Center for Glaciology. http://nsidc.org/data/docs/
www.soils.org/publications/sssaj
fgdc/ggd318_map_circumarctic/index.html (accessed 7 Feb. 2013).
Brown, J., K. Hinkel, and F. Nelson. 2000. The Circumpolar Active Layer
Monitoring (CALM): Research designs and initial results. Polar Geogr.
24:165–258. doi:10.1080/10889370009377697
CALM. 2013. The circumpolar active layer thickness data. Circumpolar Active
Layer Monitoring Network. http://www.gwu.edu/~calm/data/north.
html (accessed 24 Jan. 2013).
Chang, C.W., and D.A. Laird. 2002. Near-infrared reflectance spectroscopic
analysis of soil C and N. Soil Sci. 167:110–116. doi:10.1097/00010694200202000-00003
Cox, P., R. Betts, C. Bunton, R. Essery, P. Rowntree, and J. Smith. 1999. The
impact of new land surface physics on the GCM simulation of climate and
climate sensitivity. Clim. Dyn. 15:183–203. doi:10.1007/s003820050276
Crow, W.T., A.A. Berg, M.H. Cosh, A. Loew, B.P. Mohanty, R. Panciera,
P. de Rosnay, D. Ryu, and J.P. Walker. 2012. Upscaling sparse
ground-based soil moisture observations for the validation of coarseresolution satellite soil moisture products. Rev. Geophys. 50: RG2002.
doi:10.1029/2011RG000372
Daly, C., G.H. Taylor, W.P. Gibson, T.W. Parzybok, G.L. Johnson, and P.
Pasteris. 2000. High quality spatial climate data sets for the United States
and beyond. Trans. ASAE 43:1957–1962. doi:10.13031/2013.3101
Dunne, J.P., J.G. John, A.J. Adcroft, S.M. Griffies, R.W. Hallberg, E. Shevliakova,
R.J. Stouffer, W. Cooke, K.A. Dunne, M.J. Harrison, J.P. Krasting, S.L.
Malyshev, P.C.D. Milly, P.J. Phillips, L.T. Sentman, B.L. Samuels, M.J.
Spelman, M. Winton, A.T. Wittenberg, and N. Zadeh. 2012. GFDL’s
ESM2 global coupled climate–carbon earth system models. Part I:
Physical formulation and baseline simulation characteristics. J. Clim.
25:6646–6665. doi:10.1175/JCLI-D-11-00560.1
Essery, R., M. Best, R. Betts, P. Cox, and C. Taylor. 2003. Explicit representation
of subgrid heterogeneity in a GCM land surface scheme. J. Hydrometeorol.
4:530–543. doi:10.1175/1525-7541(2003)004<0530:EROSHI>2.0.CO;2
Fotheringham, A.S., C. Brunsdon, and M. Charlton. 2002. Geographically
weighted regression: The analysis of spatially varying relationships. John
Wiley & Sons Ltd., Chichester, UK.
Gorham, E. 1991. Northern peatlands: Role in the carbon cycle and
probable responses to climatic warming. Ecol. Applic. 1:182–195.
doi:10.2307/1941811
Heimsath, A.M., W.E. Dietrich, K. Nishiizumi, and R.C. Finkel. 1997. The soil
production function and landscape equilibrium. Nature 388:358–361.
doi:10.1038/41056
Hinkel, K.M., and F.E. Nelson. 2003. Spatial and temporal patterns of active
layer thickness at Circumpolar Active Layer Monitoring (CALM)
sites in northern Alaska, 1995–2000. J. Geophys. Res. 108: 8168.
doi:10.1029/2001JD000927.
Holland, M.M., and C.M. Bitz. 2003. Polar amplification of climate change in
coupled models. Clim. Dyn. 21:221–232. doi:10.1007/s00382-003-0332-6
Intergovernmental Panel on Climate Change (IPCC). 2007. Summary for policy
makers. In: S. Solomon et al., editors, Climate change 2007: The physical
science basis. Contribution of Working Group 1 to the Fourth Assessment
Report of the Intergovernmental Panel on Climate Change. Cambridge
Univ. Press, New York.
Ji, J. 1995. A climate-vegetation interaction model: Simulating physical
and biological processes at the surface. J. Biogeogr. 22:445–451.
doi:10.2307/2845941
Jorgenson, M.T., V. Romanovsky, J. Harden, Y. Shur, J. O’Donnell, E.A.G.
Schuur, M. Kanevskiy, and S. Marchenko. 2010. Resilience and
vulnerability of permafrost to climate change. Can. J. For. Res. 40:1219–
1236. doi:10.1139/X10-060
Koven, C.D., W.J. Riley, and A. Stern. 2013. Analysis of permafrost thermal
dynamics and response to climate change in the CMIP5 earth system
models. J. Climate. 26:1877–1900. doi:10.1175/JCLI-D-12–00228.1
Koven, C.D., B. Ringeval, P. Friedlingstein, P. Ciais, P. Cadule, D. Khvorostyanov,
G. Krinner, and C. Tarnocai. 2011. Permafrost carbon-climate feedbacks
accelerate global warming. Proc. Natl. Acad. Sci. USA 108:14,769–14,774.
doi:10.1073/pnas.1103910108
Kuhry, P., E. Dorrepaal, G. Hugelius, E.A.G. Schuur, and C. Tarnocai. 2010.
Potential remobilization of belowground permafrost carbon under future
global warming. Permafrost Periglac. Process. 21:208–214.
Kutner, M.H., C.J. Nachtsheim, and J. Neter. 2004. Applied linear regression
models. 4th ed. McGraw-Hill, New York.
901
Lawrence, D.M., and A.G. Slater. 2005. A projection of severe near-surface
permafrost degradation during the 21st century. Geophys. Res. Lett.
32:L24401. doi:10.1029/2005GL025080
Marchenko, S., V.E. Romanovsky, and G. Tipenko. 2008. Numerical modeling
of spatial permafrost dynamics in Alaska. In: D.L. Kane and K.M. Hinkel,
editors, Proceedings Ninth International Conference on Permafrost.
Fairbanks, AK. 29 June–3 July 2008. International Permafrost Association,
Fairbanks, AK. p. 1125–1130.
McGuire, A.D., L.G. Anderson, T.R. Christensen, S. Dallimore, L. Guo, D.J.
Hayes, M. Heimann, T.D. Lorenson, R.W. Macdonald, and N. Roulet.
2009. Sensitivity of the carbon cycle in the Arctic to climate change. Ecol.
Monogr. 79:523–555. doi:10.1890/08-2025.1
Michaelson, G.J., C.L. Ping, and M. Clark. 2013. Soil pedon carbon and nitrogen
data for Alaska: An analysis and update. Open J. Soil Sci. doi:10.4236/
ojss.2013.32015
Mishra, U., J.D. Jastrow, R. Matamala, G. Hugelius, C.D. Koven, J.W. Harden,
C.L. Ping, G.J. Michaelson, Z. Fan, R.M. Miller, A.D. McGuire, C.
Tarnocai, P. Kuhry, W.J. Riley, K. Schaefer, E.A.G. Schuur, M.T. Jorgenson,
and L.D. Hinzman. 2013. Empirical estimates to reduce modeling
uncertainties of soil organic carbon in permafrost regions: A review of
recent progress and remaining challenges. Environ. Res. Lett. 8:035020.
doi:10.1088/1748–9326/8/3/035020.
Mishra, U., and W.J. Riley. 2012. Alaskan soil carbon stocks: Spatial variability
and environmental control. Biogeosciences 9:3637–3645. doi:10.5194/
bg-9-3637-2012
Mishra U., M.S. Torn, S. Ogle, and E. Masanet. 2012. Improving regional soil
carbon
inventories: Combining IPCC carbon inventory method with
regression kriging. Geoderma 189–190:288–295.
Multi-Resolution Land Characteristics Consortium. 2011. 2001 National land
cover data (NLCD 2001). USEPA. www.epa.gov/mrlc/nlcd-2001.html
(accessed 20 Feb. 2011).
Nelson, F.E., K.M. Hinkel, N.I. Shiklomanov, G.R. Mueller, L.L. Miller, and
D.A. Walker. 1998. Active-layer thickness in north central Alaska:
Systematic sampling, scale, and spatial autocorrelation. J. Geophys. Res.
103:28,936–28,973.
Nelson, F.E., N.I. Shiklomanov, G.R. Mueller, K.M. Hinkel, D.A. Walker,
and J.G. Bockheim. 1997. Estimating active-layer thickness over a large
region: Kuparuk river basin, Alaska, U.S.A. Arct. Alp. Res. 29:367–378.
doi:10.2307/1551985
Osterkamp, T.E. 2005. The recent warming of permafrost in Alaska. Global
Planet. Change 49:187–202. doi:10.1016/j.gloplacha.2005.09.001
Ping, C.L., M.H. Clark, J.M. Kimble, G.J. Michaelson, Y. Shur, and C.A. Stiles.
2013. Sampling protocols for permafrost-affected soils. Soil Horiz. 54:13–
19. doi:10.2136/sh12-09-0027
Raddatz, T.J., C.H. Reick, W. Knorr, J. Kattge, E. Roeckner, R. Schnur, K.G. Schnitzler, P. Wetzel, and J. Jungclaus. 2007. Will the tropical land
biosphere dominate the climate–carbon cycle feedback during the twentyfirst century? Clim. Dyn. 29:565–574. doi:10.1007/s00382-007-0247-8
Riseborough, D., N. Shiklomanov, B. Etzelmuller, S. Gruber, and S. Marchenko.
2008. Recent advances in permafrost modelling. Permafrost Periglac.
Process. 19:137–156. doi:10.1002/ppp.615
Romanovsky, V.E., and T.E. Osterkamp. 1995. Interannual variations of
the thermal regime of the active layer and near-surface permafrost in
Northern Alaska. Permafrost Periglacial Process. 6:313–335. doi:10.1002/
ppp.3430060404
Rosenzweig, C., and F. Abramopoulos. 1997. Land-surface model development
for the GISS GCM. J. Clim. 10:2040–2054. doi:10.1175/1520-
902
0442(1997)010<2040:LSMDFT>2.0.CO;2
SAS Institute. 2004. SAS version 9.1.3 for microcomputers. SAS Inst., Cary, NC.
Schaefer, K., T. Zhang, L. Bruhwiler, and A.P. Barrett. 2011. Amount and
timing of permafrost carbon release in response to climate warming. Tellus
63:165–180. doi:10.1111/j.1600-0889.2011.00527.x
Schuur, E.A.G., and B. Abbott. 2011. High risk of permafrost thaw. Nature
480:32–33. doi:10.1038/480032a
Schuur, E.A.G., J. Bockheim, J.G. Canadell, E. Euskirchen, C.B. Field, S.V.
Goryachkin, S. Hagemann, P. Kuhry, P.M. Lafleur, H. Lee, G. Mazhitova,
F.E. Nelson, A. Rinke, V.E. Romanovsky, N. Shiklomanov, C. Tarnocai, S.
Venevsky, J.G. Vogel, and S.A. Zimov. 2008. Vulnerability of permafrost
carbon to climate change: Implications for the global carbon cycle.
Bioscience 58:701–714. doi:10.1641/B580807
Schuur, E.A.G., J.G. Vogel, K.G. Crummer, H. Lee, J.O. Sickman, and T.E.
Osterkamp. 2009. The effect of permafrost thaw on old carbon release and
net carbon exchange from tundra. Nature 459:556–559. doi:10.1038/
nature08031
Serreze, M.C., J.E. Walsh, F.S. Chapin, III, T. Osterkamp, M. Dyurgerov, V.
Romanovsky, W.C. Oechel, J. Morison, T. Zhang, and R.G. Barry. 2000.
Observational evidence of recent change in the northern high latitude
environment. Clim. Change 46:159–207. doi:10.1023/A:1005504031923
Soil Survey Staff. 2010. Keys to soil taxonomy. 11th ed. USDA-Natural
Resources Conservation Service, Washington, DC.
Tarnocai, C., J.G. Canadell, E.A.G. Schuur, P. Kuhry, G. Mazhitova, and S. Zimov.
2009. Soil organic carbon pools in the north circumpolar permafrost region.
Global Biogeochem. Cycles 23:GB203. doi:10.1029/2008GB003327
Taylor, K.E., R.J. Stouffer, and G.A. Meehl. 2009. A summary of the CMIP5
experiment design. PCMDI Tech. Rep. http://cmip-pcmdi.llnl.gov/
cmip5/docs/Taylor_CMIP5_design.pdf (accessed 15 Feb. 2013).
Thompson, J.A., E.M. Pena-Yewtukhiw, and J.H. Grove. 2006. Soil–landscape
modeling across a physiographic region: Topographic patterns
and model transportability. Geoderma 133:57–70. doi:10.1016/j.
geoderma.2006.03.037
Verseghy, D.L. 1991. CLASS—A Canadian land surface scheme for GCMS. I.
Soil model. Int. J. Climatol. 11:111–133. doi:10.1002/joc.3370110202
Walker, D.A., G.J. Jia, H.E. Epstein, M.K. Raynolds, F.S. Chapin, C. Copass,
L.D. Hinzman, J.A. Kundson, H.A. Maier, G.J. Michaelson, F. Nelson,
C.L. Ping, V.E. Romanovsky, and N. Shiklomanov. 2003. Vegetation-soilthaw-depth relationships along a low-arctic bioclimate gradient, Alaska:
Synthesis of information from the Atlas studies. Permafrost Periglac.
Process. 14:103–123. doi:10.1002/ppp.452
Webster, R., and M.A. Oliver. 2007. Geostatistics for environmental scientists. J.
Wiley & Sons, New York.
Williams, P.J., and M.W. Smith. 1989. The frozen earth: Fundamentals of
geocryology. Cambridge Univ. Press, Cambridge, UK.
Zhang, Y., W. Chen, S.L. Smith, D.W. Riseborough, and J. Cihlar. 2005. Soil
temperature in Canada during the twentieth century: Complex responses
to the changes in air temperature at high latitudes. J. Geophys. Res.
110:D03112. doi:10.1029/2004JD004910
Zhang, Y., J. Li, X. Wang, W. Chen, W. Sladen, L. Dyke, L. Dredge, J. Poitevin,
D. McLennan, H. Stewart, S. Kowalchuk, W. Wu, G.P. Kershaw, and R.K.
Brook. 2012. Modelling and mapping permafrost at high spatial resolution
in Wapusk National Park, Hudson Bay Lowlands. Can. J. Earth Sci.
49:925–937. doi:10.1139/e2012-031
Zhang, T., T.E. Osterkamp, and K. Stamnes. 1997. Effects of climate on the
active layer and permafrost on the north slope of Alaska. Permafrost
Periglac. Process. 8:45–67.
Soil Science Society of America Journal