Introduction
Since the end of the 1990s, the Greenland ice sheet (GrIS) has been losing
mass as a result of both surface meltwater run-off increase (representing
∼ 60 % of the mass loss) and iceberg discharge increase
. This recent acceleration of the ice dynamics is
likely a consequence of the increase of meltwater availability and ocean
warming, although the role of meltwater remains unclear
. The recent surface melt increase
likely results from global warming, enhanced by the Arctic amplification
and general circulation-observed changes in summer
favouring advection of warm air masses over GrIS . In
view of the impact of GrIS on the observed global sea level rise
, it is important to consider this recent surface mass
loss in a longer perspective.
The warming observed in the 1930s is, for example, often
mentioned as equivalent to the warming observed since the end of the 1990s,
suggesting that the recent surface mass loss may not be unprecedented in the
100-yr Greenland climate history. A first estimation of the surface mass
balance (SMB) through the entire previous century has been made by
, and more recently by , based
mainly on observations and statistical regressions and corrections. However,
the number of in situ observations is too sparse over Greenland before 1950
to build reconstructions at the daily time scale as in , and
many uncertainties remain in the available reconstructions, whose time scales
are monthly at best. The recent development of a new reanalysis dataset
covering the entire last century and constrained by observations from both
Greenland and outside Greenland offers a new opportunity to evaluate SMB over
GrIS through the last century . Regional climate models (RCMs)
especially developed for polar regions and coupled with complex
snow models are powerful tools to physically downscale the 6-hourly
reanalyses and estimate SMB. Both the spatial resolution and the snow model
used in reanalyses are not yet adequate to directly derive SMB from
reanalyses . Among the “polar” RCMs, there is the
MAR (Modèle Atmosphérique Régional) model, fully coupled with
a snow energy balance model and developed and extensively evaluated to study
the present Greenland climate as well as to perform future projections of
GrIS SMB for the last IPCC report .
In the present paper, we force the MAR model with eight reanalyses over
Greenland at a resolution of 20 km for the period 1900–2015 to (i) evaluate
the uncertainties coming from reanalyses in RCM-based reconstructions (while
reanalyses are assumed to represent exactly the same climate) and
(ii) estimate SMB before 1950 with two new reanalyses covering the entire
century. All previous RCM-based SMB estimations use the ECMWF (European
Centre for Medium-Range Weather Forecasts) reanalysis as forcing until now
and cover only the second half of the last century (e.g.
).
After a brief description of the MAR model and reanalyses used as forcing in
Sect. 2, Sect. 3 compares the eight reanalyses used for MAR forcing over
Greenland as well as the reanalysis-forced MAR results over our reference
period (1980–1999). In Sect. 4, we present a validation over 1958–2010 of
the eight MAR-based series vs. in situ observations from ablation stakes, ice
cores and satellite-derived melt extent. Finally, Sect. 5 discusses the time
evolution of the GrIS SMB since 1900 as well as the uncertainties coming from
the reanalyses.
Data
The MAR model
The version of MAR used here is 3.5.2 (called MAR hereafter) and has been
used by , , , and . See
, and for a detailed
description of the MAR model and its surface scheme SISVAT (soil ice snow
vegetation atmosphere transfer) dealing with the energy and mass exchanges
between surface, snow and atmosphere. Compared to version 2 of MAR (hereafter
MARv2) and the set-ups used in , the following changes were
implemented.
A resolution of 20 km (instead of 25 km) as well as the GrIS topography
from (instead of ) were used here. In addition,
to each atmospheric MAR 20 × 20 km2 grid cell, we associated two
sub-grid cells covered by tundra and permanent ice according to the
ice mask. This fractional ice sheet mask in MARv3.x allows the
computation of SMB outside the original MAR ice sheet mask (with the aim of
forcing ice sheet models at higher resolutions), and a grid cell will be
considered hereafter as an ice sheet grid cell if its permanent ice cover is
higher than 50 %. In addition, when integrated over the whole ice sheet,
the surface mass values will be weighted by the permanent ice cover of each
grid cell (i.e. for cells at least 50 % covered by permanent ice).
According to the MAR bare ice albedo overestimation found by
using MARv3.2, the bare ice albedo has been improved in MARv3.5.2 by
exponentially varying between 0.4 (dirty ice) and 0.55 (clean ice) as a
function of the accumulated surface water height and slope. For densities
lower than 550 kg m-3, the CROCUS snow model albedo is
used with a minimum albedo value set to 0.7. Concerning snowpack with surface
density higher than 550 kg m-3 (representing the maximum density of
pure snow), the minimum allowed albedo is a linear function with a smooth
transition between the minimum pure snow albedo (0.7) and clean ice albedo
(0.55).
emphasized an overestimation of accumulation simulated
by MARv2 in the interior of the ice sheet. This bias was in part corrected in
MARv3.5.2 by slightly increasing the snowfall velocity, which enabled more
precipitation along the ice sheet margin and less inland.
Finally, MARv3.5.2 is now parallelized with OpenMP, its outputs are
CORDEX compliant and some usual bug corrections have been made since MARv2.
Reanalyses
In this study, we use the reanalyses listed below to force MAR every 6 h
at its lateral atmospheric boundaries (temperature, humidity, wind and
pressure at each vertical MAR level) and over oceanic grid cells (sea ice
cover, SIC, and sea surface temperature, SST).
ERA-Interim (ECMWF Interim Re-Analysis) over 1979–2015, available at a resolution of ∼0.75∘
from ECMWF. As in , this state-of-the-art third generation reanalysis
is used as reference over our chosen reference period (1980–1999) and assimilates
the greatest fraction of the in situ and remote observations available .
ERA-40 over 1958–2001 (resolution: ∼1.125∘), the second generation
reanalysis from ECMWF . One of main differences between ERA-40 and
ERA-Interim is a fully revised humidity scheme in ERA-Interim ,
which impacts the snowfall amount simulated by MAR as shown by .
ERA-20C over 1900–2010 (resolution: ∼1.25∘), the latest
generation of ECMWF reanalysis products assimilating only surface pressures
and near-surface winds over the ocean surface but starting in 1900 .
As this reanalysis assimilates much less data than ERA-40 and ERA-Interim, it is generally
less reliable than the other ECMWF reanalyses, but its reliability increases
with time with the increasing amount of assimilated observations.
NCEP–NCARv1
(National Centers for Environmental Prediction–National Center for
Atmospheric Research reanalysis version 1; referred to as NCEPv1 here) over
1948–2015 (resolution: 2.5∘), First generation reanalysis from the
NCEP–NCAR covering the second half of the last century at low spatial
resolution .
NCEP–DOE (NCEP–Department of Energy; referred to as NCEPv2 here) over 1979–2015 (resolution: 2.5∘),
second generation reanalysis using an improved version of the NCEP–NCARv1 global model
and assimilating additional satellite data with respect to NCEP–NCARv1 .
20CRv2 over 1871–2012 (resolution: 2.0∘), experimental
reanalysis based on an ensemble mean of 56 members assimilating only surface
pressure, monthly sea surface temperature and sea ice cover
. Only outputs after 1900 were used here. As for
ERA-20C, its reliability increases with time (i.e. with the amount of
assimilated data).
20CRv2c over 1851–2014 (resolution: 2.0∘), same as 20CRv2
but correcting a bias found in the sea ice distribution by assimilating new
SST and SIC data.
JRA-55 (Japanese 55-year Reanalysis) over 1958–2014 (resolution: 1.25∘), second generation
reanalysis from the Japan Meteorological Agency, described in
.
(a) In the background, the interannual variability (i.e.
standard deviation) of the JJA mean 700 hPa temperature (T700) is
simulated by ERA-Interim over 1980–1999. Units are ∘C. The contours
of the mean JJA T700 are plotted in dashed blue. (b)–(h)
Mean anomalies of the JJA 700 hPa temperature simulated by the
different reanalyses used in this study with respect to ERA-Interim over
1980–1999 (in ∘C). No comparison is shown above 2000 m a.s.l. due
to the aim of only showing comparisons in the free atmosphere (700 hPa), and
the datasets are shown here by using their native lat–long projection.
Similar figures over different reference periods and at other vertical levels
(850 and 500 hPa) can be found in the Supplement.
Evaluation over 1980–1999
MAR forcings
In , summer temperatures at 600 hPa, geopotential height at
500 hPa and
wind speed at 500 hPa from the different MAR forcing fields were
compared over 1980–1999 to explain the discrepancies between the MAR
simulations using different forcings.
According to , the JJA (June–July–August) mean 700 hPa (T700)
or 600 hPa temperatures (T600) in the free atmosphere over Greenland are a good
predictor of the melt variability in MAR. Therefore, temperature biases at
the MAR boundaries will directly impact the amount of melt simulated by MAR
. Since free atmosphere temperatures are not assimilated
either in 20CRv2c or in ERA-20C reanalysis, a comparison of this field is
presented in Fig. over our reference period (1980–1999), which is
covered by all datasets used here and during which SMB has been relatively
stable. While SMB has already started to decrease at the end of the 1990s, a
comparison over a period longer than 20 years does not change the conclusions
of this comparison as justified in .
Idem as Fig. , but for the mean annual geopotential height
(Z500) at 500 hPa over 1980–1999. Units are metres. The wind vectors
represent anomalies of wind field.
The mean 1980–1999 free atmosphere JJA temperature from ERA-40 and NCEPv1
compares very well with ERA-Interim over Greenland (see Fig. b
and e). Surprisingly, the second generation of the NCEP reanalysis does not
compare as well as with ERA-Interim and NCEPv1 because NCEP2 is too warm in
summer, except in the vicinity of Iceland (see Fig. f). As specific
humidity (used as forcing at the MAR lateral boundaries) needs to be derived
from relative humidity in NCEPv2, these temperature biases impact the
precipitation amount simulated by MAR forced by NCEPv2.
The reanalyses covering the entire 20th century are significantly
(>1 ∘C) warmer (20CRv2, Fig. g) and colder (ERA-20c,
Fig. c) than ERA-Interim in summer. Similar anomalies also occur in
winter and at other vertical levels (see Figs. S1–S4 of the Supplement). In view
of these biases in 20CRv2 and ERA-20C, a correction of -1 ∘C
(+1 ∘C) was applied to the temperature fields from these two
reanalyses at each vertical level of the MAR lateral boundaries while keeping
the relative humidity constant. These corrected reanalyses are called
CORR-20CRv2 and CORR-ERA-20c hereafter. These corrections aim at having a
good agreement with the ERA-Interim-forced MAR melt rate over the last
decades for a better comparison between the recent melt increase and past
conditions. Indeed, as the melt response to a temperature anomaly is not
linear , inaccurate current melt rates bias melt anomalies in
the past. It should be noted that no change was applied to the MAR oceanic
boundaries (SST and SIC) and that the temperatures corrections were
homogeneously applied through the whole year and over the entire period
covered by these two reanalyses, as these biases are constant in time over
1948–2010 with respect to NCEPv1 (see Figs. S1–S4 of the Supplement). As we
will see in Sect. 4.1, these temperature corrections enable a better
comparison of MAR with in situ temperature measurements than with unmodified
20CRv2 and ERA-20C-based fields as lateral boundaries. Finally, as the warm
bias from 20CRv2 is partly corrected in 20CRv2c and is now centred around
zero on average with a too-warm atmosphere at the north of Greenland and too
cold at the south-west, unmodified 20CRv2c temperatures are then used to
force MAR at its lateral boundaries.
Since the surface pressure has been assimilated in all reanalyses used here,
the general circulation (gauged here by the 500 hPa geopotential height in
Fig. ) including the North Atlantic Oscillation (NAO) compares well
over the recent decades , except for 20CRv2(c), which
underestimates wind speed at 500 hPa, inducing anti-clockwise circulation
anomalies over Greenland (see Fig. g and h). Moreover, as a
consequence of the lack of (or less reliable) assimilated data before 1940,
the general circulation variability from 20CRv2 and ERA-20C diverges
according to and explains the discrepancies between MAR
forced by 20CRv2(c) and ERA-20C before 1940 (see Sect. 6).
Average and standard deviation (gauging the interannual
variability) of the annual SMB components simulated by MAR over 1980–1999
and from Box's reconstruction (; interpolated to the MAR 20 km grid).
Units are GTyr-1 and the acronym of each simulation
(RCMforcings) is given in the first column. The surface mass
balance (SMB) equation is
SMB = snowfall + rainfall - run-off - water fluxes. The run-off
is the fraction of water from both surface melt and rainfall that is not
refrozen before reaching the ocean.
Simulation acronym
SMB
Snowfall
Rainfall
Run-off
Water
Meltwater
fluxes
MARERA-Interim
480±87
683±56
28±5
220±52
12±4
427±82
MARERA-40
529±89
716±57
31±6
210±54
9±3
418±86
MARERA-20C
500±71
624±76
18±4
126±35
15±3
296±59
MARCORR-ERA-20c
491±84
665±59
26±6
190±48
10±3
399±77
MARNCEPv1
467±88
675±59
28±6
228±53
8±4
440±82
MARNCEPv2
486±86
672±60
22±5
200±46
8±4
409±74
MAR20CRv2
420±102
703±60
31±6
221±52
12±2
432±82
MARCORR-20CRv2
459±88
670±57
22±4
309±67
5±5
559±102
MAR20CRv2c
456±92
680±59
25±6
241±63
8±4
462±97
MARJRA-55
482±88
670±57
29±5
209±52
9±4
412±83
Box (2013)
502±74
735±62
229±47
424±71
(a) Difference between the mean annual SMB (in
mm.w.e. yr-1) simulated by MARv2 forced by ERA-Interim and MARv3.5.2
forced by ERA-Interim over 1980–1999. (b) Difference between the
mean 1980–1999 annual SMB simulated by MARv3.5.2 forced by ERA-40 and
simulated by MARv3.5.2 forced by ERA-Interim. (c) Idem
as (b) but for ERA-20C. (c) Idem as (b) but for
CORR-ERA-20c. (e) Idem as (b) but for JRA-55.
(f) Idem as (b) but for NCEPv1. (g) Idem
as (b) but for NCEPv2. (h) Idem as (b) but for
20CRv2. (i) Idem as (b) but for CORR-20CRv2.
(j) Idem as (b) but for 20CRv2c. Finally, the areas where
the differences are lower than the interannual variability of MARv3.5.2
forced by ERA-Interim over 1980–1999 are hatched.
Same as Fig. but for snowfall (in mm.w.e. yr-1).
MAR results
Considering their interannual variability, the mean SMB components from the
different MAR simulations compare very well with each other when they are
integrated over the entire ice sheet, except for the non-corrected ERA-20C and
20CRv2-forced MAR simulations (see Table ). However,
when looking at spatial differences (see Figs. and ),
the comparison with the MARERA-Interim simulation over 1980–1999
shows the following.
MARERA-40 slightly overestimates precipitation because the
ERA-40 high atmosphere is wetter than ERA-Interim, as a result of biases in the
ERA-40 humidity scheme that were later corrected in ERA-Interim .
However, this wet anomaly is homogeneous over the whole integration domain and
explains why there are no locally significant discrepancies between MARERA-Interim
and MARERA-40.
Although the temperature corrections of +1∘ degree at the MAR
lateral boundaries reduce the underestimation of melt by MARERA-20c,
MARCORR-ERA-20c is still too cold in summer. Both ERA-20C-forced
simulations also significantly underestimate precipitation along the south-western coast
(60–70∘ N) because not enough humidity is advected at the south-west
lateral boundaries of our integration domain, from where the prevailing flow over
south Greenland comes. This too-dry and cold main flux is a consequence of the
ERA-20 underestimation of the free atmosphere temperature and wind speed in this
area (see Figs. c and c).
Most of the differences between MARERA-Interim and
MARNCEPv1 (MARJRA-55) are within the interannual
variability of MARERA-Interim over 1980–1999 and are therefore
insignificant. We can see an underestimation of precipitation along the south-east
coast with respect to MARERA-Interim, but it is not significant.
MARNCEPv2 is too wet (too dry) in the south-west
(south-east) of the ice sheet despite the fact that the general circulation
(here Z500) from NCEPv2 compares very well with ERA-Interim. However, NCEPv2 is
too warm (too cold) in the south-west (south-east) of Greenland,
which impacts the amount of humidity advected by MAR from its lateral boundaries.
This is because the specific humidity is derived from the NCEPv2 relative humidity
and is then affected by the temperature biases found in NCEPv2 with respect to ERA-Interim.
The patterns of anomalies of MARCORR-20CRv2 and MAR20CRv2c
with respect to MARERA-Interim are similar and mainly result from
anomalies in precipitation. We can see that the temperature correction in
CORR-20CRv2 reduces the MAR20CRv2 run-off overestimation vs.
MARCORR-20CRv2, but this correction does not impact the simulated
MAR precipitation: MAR20CRv2(c) is too wet (dry) along the
north-eastern (north-western) coast, as a result of the anti-clockwise
circulation anomalies simulated by 20CRv2(c) with respect to ERA-Interim (see
Fig. ). Finally, except along the south-western margin where
CORR-20CRv2 and 20CRv2c are too cold in summer, MARCORR-20CRv2
and MAR20CRv2c weakly overestimate run-off with respect to
MARERA-Interim.
Validation
Near-surface climate
As validation of the near-surface conditions simulated by MAR, a comparison
with daily measurements from the automatic weather station (AWS) of the
PROMICE (Programme for Monitoring of the Greenland Ice Sheet) network
starting in mid-2007 is presented over the common
period covered by the forcing datasets used here: 2008–2010. The raw PROMICE
data are used here without any filtering or withdrawing of aberrant values.
The MAR values at each station are based on an interpolation of the four nearest
MAR grid cells weighted by the inverse distance to the station. As the
elevation difference between MAR and AWS is not corrected, the comparison is
only carried out on the 12 AWSs listed in Table S1 of the Supplement that have
an elevation difference within 100 m of the interpolated MAR 20 km
topography. Scatter plots are shown in Fig. and statistics are
listed in Table .
On average for the 12 AWSs, the comparison of MARERA-Interim with
the measured daily near-surface temperature is excellent, with a correlation
above 0.96 and a RMSE (root mean square error) of 2–3 ∘C,
representing less than 30 % of the daily variability. The improvements
with respect to MARv2 are evident. The biases with the downward shortwave
(longwave) radiation remain, however, high in both MAR versions, with the
RMSE representing 25 % (70 %) of the daily variability of these
fluxes. Due to an underestimation of the cloudiness, MAR slightly
overestimates (highly underestimates) downward shortwave (longwave)
radiation. Such biases in the short- and longwave were also found in the
regional RACMO2.3 model , suggesting that
improvements are still needed in the clouds and/or radiative schemes of
(regional) climate models. As a result, MARERA-Interim is
slightly too cold (-0.29 ∘C at the annual scale), in particular in
summer (-0.65 ∘C) when the underestimation of the downward
infrared flux is the highest (a bias of -18 W m-2 compared with a
daily variability of 43 W m-2). Finally, MAR overestimates the bare
ice albedo as it is limited to 0.40 in MARv3.5.2, while values ranging from
0.2 to 0.4 (due to the presence of impurities not taken into account into
MAR) are observed in some PROMICE AWSs . In view of
the sensitivity of the simulated SMB to the bare ice albedo formulation
(; ), improving its
representation in MAR should be a priority for future developments.
Mean correlation, bias, RMSE and correlation over the 12 AWSs
listed in Table S1 of the Supplement between MAR forced by the different
reanalyses and daily observations from the PROMICE network over 2008–2010.
Statistics are given for the surface pressure (SP), near-surface temperature
(TAS) over the entire year and for the summer months only (for JJA; Summer TAS), shortwave
downward flux (SWD) and longwave downward flux (SWD).
Simulation acronym
SP
TAS (∘C)
Summer TAS (∘C)
CORR
BIAS
RMSE
CORR
BIAS
RMSE
CORR
MARERA-Interim
0.99
-0.29
2.32
0.96
-0.65
2.38
0.95
MARERA-20C
0.99
-1.04
2.78
0.95
-1.42
2.92
0.93
MARCORR-ERA-20c
0.99
-0.26
2.56
0.95
-0.61
2.64
0.93
MARNCEPv1
0.99
-0.04
2.48
0.95
-0.26
2.47
0.93
MARNCEPv2
0.99
-0.19
2.52
0.95
-0.44
2.51
0.93
MAR20CRv2
0.98
0.30
3.16
0.92
-0.27
3.07
0.90
MARCORR-20CRv2
0.98
-0.42
3.21
0.92
-1.02
3.25
0.89
MAR20CRv2c
0.98
-0.33
3.09
0.93
-0.76
3.05
0.91
MARJRA-55
0.99
-0.56
2.51
0.96
-1.08
2.62
0.94
MARv2ERA-Interim
0.99
-0.98
2.73
0.95
-1.39
2.90
0.94
Simulation acronym
SWD (W m-2)
LWD (W m-2)
BIAS
RMSE
CORR
BIAS
RMSE
CORR
MARERA-Interim
3.42
27.07
0.96
-16.92
28.13
0.84
MARERA-20C
4.05
30.54
0.96
-19.98
32.35
0.79
MARCORR-ERA-20c
3.17
30.43
0.96
-16.33
30.29
0.79
MARNCEPv1
1.84
29.58
0.96
-14.19
29.64
0.79
MARNCEPv2
2.70
29.74
0.96
-14.64
30.10
0.79
MAR20CRv2
1.75
33.51
0.95
-14.28
32.55
0.74
MARCORR-20CRv2
0.21
32.30
0.95
-14.34
32.54
0.74
MAR20CRv2c
0.73
32.21
0.95
-14.28
32.55
0.74
MARJRA-55
3.71
26.92
0.96
-17.98
29.41
0.83
MARv2ERA-Interim
-1.8
27.64
0.95
-19.52
31.42
0.81
Using other reanalyses than ERA-Interim as boundaries forcing does not
significantly change the comparison of MAR with the PROMICE measurements.
Table shows the relevance of the ERA-20C
temperature correction while the warm bias of 20CRv2 mitigates the cold bias
found in MARERA-Interim. Statistically,
MARERA-Interim and MARJRA-55 show the best agreement
with PROMICE, whereas MAR20CRv2 shows the worst.
(a) Scatter plot of the MARERA-Interim daily
near-surface temperature vs. near-surface daily temperature recorded by 12
AWSs from the PROMICE network over 2008–2010. The number of observations
used here is listed in red and units are ∘C. (b) Same
as (a) but for the surface albedo. (c) Scatter plot of the
MARERA-Interim SMB (in m w.e.) with respect to ice core
measurements in the accumulation area (in blue) and SMB measurements (in red)
from the MACHGUTH16 dataset over 1958–2010. We refer to the text for more
details on how this comparison is performed. (d) Same
as (a) for the shortwave downward radiative flux (in W m-2).
(e) Same as (d) for the longwave downward radiative flux.
(f) Daily melt extent (in % of the ice sheet area) simulated by
MARERA-Interim over the 1979–2010 summers (May–September) vs.
the satellite-derived one. More information about the thresholds used for
retrieving the melt extent is given in the text.
Surface mass balance
As validation of the SMB simulated by MAR over 1958–2010 (the period covered
by seven of the datasets here), we use the following.
The ice core measurements in the accumulation area from
and . The MAR accumulations values (here in m.w.e. yr-1)
for each of the 246 records are averaged over the years listed in the three previous
references (the mean from 1958–2010 is used if the period is not given or before 1958)
and come from an interpolation of the four nearest inverse-distance-weighted MAR grid cells.
The new SMB database (hereafter MACHGUTH16) compiled under the auspice
of PROMICE and available through the PROMICE web portal
(http://www.promice.dk) containing a total of ∼ 3000 measurements
from 46 sites from 1892 to 2015 and mostly covering the ablation area of the
GrIS and local glaciers . For each site, the MAR SMB value
is corrected as a function of the elevation difference between the MACHGUTH16
database and the interpolated MAR 20 km topography using a local and time varying
SMB vs. elevation gradient as explained in . Moreover, the
MAR values (here in m.w.e.) are an integration of daily MAR outputs over
the exact period given for each record in the MACHGUTH16 database. The data
are not converted to m.w.e. yr-1, as some MACHGUTH16 records sometimes
cover only several months in the melt season. Only the records included in the
1958–2010 period with an elevation difference with the MAR topography of
less than 500 m and inside the MAR ice sheet mask are considered here.
The comparison is therefore limited to 1616 records from the MACHGUTH16 database
. Similarly, the same dataset has also been used in
for the validation of RACMO2.3.
The revised version (fully described in ) of the 5 km
reconstruction of the near-surface air temperature and the land ice SMB from
, hereafter BOX13, spanning 1840–2012 and calibrated to
outputs from RACMO2.1/GR forced by ERA-40 and ERA-Interim
. In contrast to the MAR-based reconstructions, this
reconstruction is not forced with reanalyses, except for the calibration with
RACMO2, but is based on in situ observations . Absolute
uncertainty for the revised SMB estimates from is estimated by
comparison against field data. A total of 208 in situ annual ablation rates
over 1985–1992 yield an ablation root mean square error of 35 %, similar
to the one found with RACMO2.1/GR. The comparison with ice-core-derived net
accumulation time series from 86 sites shows a 30 % accumulation RMSE. A
fundamental assumption is that the calibration regression factors, derived
over 1960–2012 vs. ice cores, from meteorological station temperatures and
with RACMO2.1/GR, are stationary in time.
Comparison with SMB from the MACHGUTH16 database over 1958–2010,
ice-core-based accumulation from and , and
satellite-derived melt extent over 1979–2010. MARERA-Interim
(MARERA-40) means that MAR was forced by ERA-40 over
1958–1978 (1958–2000) and ERA-Interim over 1979–2015 (2001–2015). Finally, MARERA-Interim* means that the
extrapolation of was not used to correct the MAR SMB with
respect to the elevation differences between MAR and the MACHGUTH16
measurement sites.
Simulation acronym
SMB–MACHGUTH16 (m.w.e.)
Accumulation (m.w.e. yr-1)
Melt extent (%)
BIAS
RMSE
CORR
BIAS
RMSE
CORR
BIAS
RMSE
CORR
MARERA-Interim
+0.14
0.46
0.93
+0.02
0.08
0.91
+0.0
2.8
0.93
MARERA-40
+0.20
0.48
0.93
+0.03
0.09
0.91
-0.1
2.9
0.92
MARERA-20C
+0.39
0.67
0.91
-0.03
0.07
0.91
-2.0
3.8
0.90
MARCORR-ERA-20c
+0.22
0.52
0.93
+0.01
0.07
0.91
-0.4
3.0
0.91
MARNCEPv1
+0.13
0.45
0.93
+0.03
0.09
0.92
+0.2
2.9
0.92
MARNCEPv2
+0.26
0.52
0.93
+0.03
0.09
0.92
-0.3
2.9
0.92
MAR20CRv2
+0.01
0.47
0.93
+0.01
0.08
0.92
+2.0
4.5
0.92
MARCORR-20CRv2
+0.18
0.50
0.92
+0.01
0.08
0.92
+0.1
3.4
0.91
MAR20CRv2c
+0.14
0.49
0.92
+0.02
0.09
0.90
+0.6
3.7
0.91
MARJRA-55
+0.18
0.48
0.93
+0.01
0.07
0.92
-0.2
2.8
0.92
BOX13
+0.16
0.68
0.84
+0.00
0.08
0.92
MARv2ERA-Interim
-0.08
0.58
0.90
+0.06
0.14
0.82
+0.1
2.9
0.91
MARERA-Interim*
+0.34
0.74
0.86
Figure c illustrates MARERA-Interim SMB validation
results. Statistics are listed in Table . Correlation
exceeds 0.9 and RMSE is ∼ 40 % for 1862 samples within the MAR ice
sheet mask over 1958–2010. With respect to MARv2, the accumulation
overestimation shown by has been partly corrected in
MARv3.5.2. However, MARv3.5.2 overestimates SMB in the ablation area, while MARv2
underestimates it, as a result of the bare ice albedo overestimation shown in
the previous section for MARv3.5.2. The bare ice albedo was fixed to 0.45 in
MARv2, while it varies between 0.4 and 0.55 in MARv3.5.2. This shows the
impact and importance of improving the bare ice albedo representation in the
models, as already stated by .
When MAR is forced by reanalyses other than ERA-40 and ERA-Interim, we find
that (i) MARNCEPv1 is the most accurate because, over 1958–1978,
NCEPv1 is not affected by the humidity bias present in ERA-40 and impacting
the MAR precipitation in the non-homogeneous ECMWF time series, (ii) the use
of CORR-ERA-20c partially corrects the SMB overestimation (due to the
underestimation of melt) obtained when MAR is forced by unadjusted ERA-20C,
(iii) MAR20CRv2 is more accurate than MARCORR-20CRv2
because the overestimation of melt in MAR20CRv2 compensates for
the SMB overestimation in the ablation area due to albedo overestimation and
(iv) the results of MARNCEPv2 are worse than MAR results using
less constrained reanalyses (e.g. 20CRv2) or first generation reanalysis
(e.g. NCEPv1), as a result of the temperature biases in NCEPv2. Moreover,
while some of these data were used in the reconstruction, the
comparison of ice core measurements with BOX13 shows the same agreement with
MAR (see Table ). Regarding the comparison with the SMB
MACHGUTH16 database, the SMB values from BOX13 were corrected as a function
of the elevation difference with the MACHGUTH16 database as done for MAR.
However, to match the exact period of the MACHGUTH16 database, we have simply
derived daily values from the monthly BOX13 values by dividing them by the
number of days in every month. It is clear that this approximation smoothing
the melt variability can be problematic when the period of measurements
covers only a few weeks in the melt season and very likely explains why BOX13
is less correlated with the MACHGUTH16 dataset than MAR. Finally, it is
interesting to note that the comparison with MACHGUTH16 and ice core
measurements is quite constant over the entire century (see Table S2) and not
better in the recent decades than before despite the larger amount of
assimilated data. The lowest correlations are reached in the 1950s and 1960s,
but the number of observations is too limited before 1950 to allow for the
conclusion that the reliability of the MAR reconstructions are constant in
time.
Figure illustrates how MARv3.5.2 still overestimates snow
accumulation for the southern ice sheet when compared to ice cores (see
Fig. b) and BOX13 (see Fig. c). However this bias has
been partly corrected since MARv2, which was wetter in this area than the
current MARv3.5.2 (see Fig. a). MAR also underestimates
accumulation compared to ice cores in the north-east but is better than the
BOX13 results, which are based on RACMO2 and known to underestimate accumulation
in this area . In these areas, the spread in the mean
1958–2010 SMB simulated by MAR using the different reanalyses is below
25 mm.w.e. yr-1 (see Fig. d), confirming that these biases
are independent of the used forcings and that improvements in MAR should
improve absolute accuracy. Moreover, these biases are in full agreement with
the MAR biases found by with respect to 2009–2012 airborne
snow-radar-based estimates. In the ablation area (i.e. the MACHGUTH16 sites),
the MAR biases vary regionally and no systematic bias can be highlighted.
Finally, huge differences (> 500 mm yr-1) between MAR and BOX13
occur along the coastal and mountainous regions of the south-east. MAR
underestimates accumulation relative to BOX13 (where the latter is based on
RACMO2). The Polar MM5 model (24 km) shows the same underestimation with
respect to RACMO2 . attributed the higher
accumulation rates in these topographically enhanced precipitation regions to
the higher spatial resolution used in RACMO2 (11 km). However, a MAR
simulation at a resolution of 10 km (not shown here) does not simulate such
an extremely high precipitation, and the number of observations in this very
wet area is too sparse to confirm the RACMO2-based estimations, suggesting
that further accumulation measurement campaigns should focus on this area.
(a) Mean annual SMB (in mm.w.e. yr-1) simulated by
MAR forced by NCEPv1 over 1958–2010. The ice core locations from
and used to validate MAR are quoted in
blue,
while the MACHGUTH16 SMB sites are in red.
(b) Mean biases (in mm.w.e. yr-1) of MAR forced by NCEPv1
over 1958–2010 with respect to both ice core and MACHGUTH16-based SMB
estimations. The biases lower than the interannual variability of MARv3.5.2
forced by NCEPv1 over 1958–2010 are hatched. (c) Comparison over
1958–2010 between the mean SMB (in mm.w.e. yr-1) simulated by MAR and
from the BOX13 reconstruction. Again, the biases lower than the interannual
variability of MARv3.5.2 forced by NCEPv1 over 1958–2010 are hatched.
(d) Spread (i.e. standard deviation in mm.w.e. yr-1) around 6
estimations of the mean 1958–2010 SMB as simulated by MAR forced by ERA,
NCEPv1, JRA, CORR-ERA-20c, CORR-20CRv2 and 20CRv2c.
Validation with microwave satellite-derived melt extent
As in , we use the brightness temperatures collected at K-band
horizontal polarization (T19H) to retrieve the daily melt extent from the
scanning multichannel microwave radiometer (SMMR; 1979–1987) and the special
sensor microwave/imager (SSM/I; 1988–2010) data distributed by the National
Snow and Ice Data Center (NSIDC, Boulder, Colorado; ;
). A grid cell is considered as melting in MAR (in
satellite-based datasets) if the daily meltwater production (T19H) is higher
than 8 mm.w.e. yr-1 (227.5 K). We refer to for more
details about the melt-retrieving methodology.
As already presented in , the comparison of the melt extent
simulated by MAR and retrieved from the passive microwave satellites is
encouraging (see Table and Fig. ). The RMSE
represents ∼ 30 % of the daily variability found in the remote-based
melt extent over the 1979–2010 summers, and correlations are higher
than 0.9, regardless of the forcing used. As already shown in the two
previous sections, MAR20CRv2 (MARERA-20C)
overestimates (underestimates) the melt extent, fully justifying the
corrections applied to 20CRv2 and ERA-20C to reduce these biases. Finally,
MARv3.5.2 slightly improves the comparison with respect to MARv2 used in
.
(a) Time series of the annual SW Greenland near-surface
temperature (built by merging series of the Ilulissat, Nuuk and Qaqortoq
coastal weather stations from the Danish Meteorological Institute,DMI) as
observed (in brown) according to , retrieved from the BOX13
reconstruction (in black) and as simulated by MAR with the different
forcings. Values are anomalies with respect to 1980–2010 and 10-year running
mean are shown. (b) Same as (a) for the summer (JJA) SW
Greenland near-surface temperature. (c) Mean GrIS summer (JJA)
near-surface temperature (in ∘C) as simulated by MAR using the
different forcings. The ERA-20c (without temperature correction) forced MAR
time series as well as the BOX13 reconstruction-based time series are also
shown.
Time evolution
Temperature
Figure illustrates MAR's ability to simulate a time series of
observed composite near-surface air temperature from . As
the latter is based on coastal weather station measurements of south and west
Greenland, a large part of the interannual variability comes from SST
changes, which are prescribed every 6 h into MAR. The remaining part comes
from changes in the general circulation , also prescribed at
the MAR lateral boundaries. Therefore, this section evaluates the ability of
the different MAR forcings to represent the observed temperature variability.
As these observations have been assimilated into BOX13, the latter
reconstruction perfectly matches the observations.
The 1900–1920 coastal temperatures were lower than the 1980–2010 average,
and their interannual variability, which is only well represented by
MARCORR-20CRv2 and MAR20CRv2c, is also lower, even
though these simulations underestimate the negative temperature anomalies
observed during this period according to BOX13. A first maximum of
temperature was reached in 1930 and is only well represented by
MAR20CRv2c. MARCORR-ERA-20c simulates this maximum
earlier while MAR20CRv2c underestimates it. This maximum is also
observed in the summer (JJA) time series but underestimated in all of the
MAR-based time series. After this optimum warm period discussed in
, there were two minor temperature maxima at the beginning of
the 1960s and at the end of the 1970s, which are overestimated by
MAR20CRv2c and MARNCEPv1 and underestimated by the
MAR time series using the other forcings. Temperature differences of several
degrees between the JRA-forced time series before and during the satellite
era (starting at the end of the 1970s) suggest biases in the JRA-based SST
before 1980. Except in MARCORR-20CRv2 (and to a lesser extent in
MARERA-Interim), where the temperature variability is very
smooth, the summer and annual warming in the 1990s and 2000s is well
represented in all of the MAR-based times series.
(a) Time series of the annual SMB (in GT yr-1)
integrated over the whole ice sheet as simulated by MAR using the different
listed forcings and coming from Box's reconstruction (BOX13).
(b) Same as (a) but for snowfall. (c) Same
as (a) but for run-off. Finally, only 10-year running means are
shown for both (b) and (c) for more readability.
When integrated over the entire ice sheet (see Fig. c), all MAR
reconstructions show a decrease of the summer mean temperature (gauging the
melt) after 1930 until the beginning of the 1990s, when an abrupt temperature
increase of ∼ 2 ∘C in 10 years is simulated. Before 1930, the
MAR reconstructions diverge even though the reanalyses are supposed to
represent the same climate variability. However, the comparison with BOX13
constrained by DMI coastal weather station measurements is the closest when
MAR is forced by (CORR-)20CRv2(c) because SST is assimilated into 20CRv2(c)
but not into ERA-20C. Finally, as absolute temperatures are shown here, we
can see that MAR is systematically 0.5–1 ∘C colder than BOX13 as a
result of the MAR cold bias discussed in Sect. 4.1.
(a) Annual snowfall trend (in mm.w.e. yr-2) over
1921–1950 as simulated by MAR forced by 20CRv2c. The observed trend (in
mm.w.e. yr-2) from some locations listed in are also
highlighted on the figure. These observed negative (positive) trends are the
values printed in blue (in red) on the map. Trends of total precipitation
(rain + snow) are also labelled in black for five coastal weather
stations from DMI. (b) Same as (a) but for
MARCORR-ERA-20c. (c) Same as (a) but for BOX13.
(d) Time series of the annual mean daily variability (i.e. standard
deviation of the daily values) of the sea level pressure around Greenland
(0∘ W ≤ longitude ≤ 80∘ W and
55∘ N ≤ latitude ≤ 85∘ N) from 20CRv2c (in
green), ERA-20C (in red) and NCEPv1 (in orange). The ensemble mean spread
(i.e. the standard deviation of the ensemble deviations at each time) from
20CRv2c over the same area is also plotted with dashes. Finally, only 10-year
running means are shown for more readability.
Surface mass balance
Time series of the SMB components integrated over the whole GrIS are
presented in Fig. . Before 1930, as for the JJA mean GrIS
near-surface temperature (see Fig. ), there are large discrepancies
between the MAR-based run-off reconstructions, suggesting that large
improvements (i.e. assimilating more data) are still needed in the reanalysis
before this period. After the warm period observed in the 1930s
, all of the MAR reconstructions suggest an increasing SMB
due to heavier snowfall and lower melt. Regarding the period 1960–1990, the
meltwater run-off amount is low and stable. The highest SMB occurred in the
1970s, but there are some discrepancies among the models. This maximum is the
highest when MAR is forced by ERA-40, which is also used to force RACMO, on
which BOX13 is based. At the beginning of this century, all the models
simulate an SMB decrease that reaches a record minimum in or after 2010,
resulting from an increasing surface melt. A second SMB minimum is simulated
around 1930 by MAR as a result of high melt and low accumulation. This
minimum is less pronounced in BOX13 because it includes considerable
smoothing by the weighted averaging of annual core and monthly station
temperature values. Therefore, BOX13 may suffer from more damping than what
MAR can produce with 6-hourly forcings. Finally, MAR suggests a significant
snowfall increase from 1900–1920 to 1950, in opposition to .
However, such an increase is also suggested in the
reconstruction and in the ice cores but is less pronounced
than in the MAR simulations (see Fig. ). Part of the MAR-simulated
snowfall increase may be caused by an artificial increase of the daily sea
level pressure variability over 1900–1950 (see Fig. d) and the
associated strengthened eddy activity. The 20CRv2c reanalysis is an ensemble
mean, suggesting that the lower the amount of assimilated data is, the higher
the spread is for a given event. This smooths the pressure fields and
therefore decreases the amount of humidity advected into the MAR free
atmosphere and subsequently the precipitation rate simulated by MAR, even if
the 20CRv2 reanalysis itself simulates higher precipitation during this
period . ERA-20C is not an ensemble mean but it is likely that
a lower amount of assimilated data also induce smoother pressure fields.
However, ERA-20C seems to suggest that the storm activity was higher at the
beginning of the last century than in the 1920–1940 period. Therefore, this
apparent significant precipitation increase from 500 to
> 600 Gt yr-1 simulated by MAR over 1900–1950 should be considered
with caution since both reanalyses-forced MAR simulations disagree on the
location where this increase takes place (western coast vs. eastern coast)
and whether a part of this increase could just be due to an artefact in the
20CRv2(c). Finally, it is interesting to note that also
showed an increase of the variability of the 20CRv2c-based Greenland blocking
index through the last century.
We can see in Fig. that the pattern of snowfall increase over
1921–1950 is quite different following the reconstruction and that there are
some disagreements with the ice-core-based trend listed in .
MAR20CRV2c suggests a decrease of accumulation along the west
coast and a significant increase along the eastern coast with the highest
increase at the south-east, as the other reconstructions.
MARCORR-ERA-20c suggests a decrease only at the north of the ice
sheet and a significant increase along the western coast, in disagreement
with the two other reconstructions. Finally, BOX13 suggests an increase only
at the south (south-east) of the ice sheet. The decrease seen in the ice cores in
the Humboldt–NEEM area (at the north-west) is well represented by the three
reconstructions, but they fail to simulate the decrease observed at D1 near
Tasiilaq. The other ice cores rather suggests a positive trend in agreement
with all the reconstructions, but MAR mostly overestimates the observed
trend,
while BOX13 is in better agreement with ice cores. The significant
accumulation increase simulated by MAR20CRV2c along the
north-eastern coast and simulated by MARCORR-ERA-20c along the
western coast seems to be overestimated with respect to ice core
measurements. Unfortunately, no gauge observation is available along the
south-eastern coast to confirm the significant snowfall increase simulated by
the three reconstructions in this area over 1921–1950.
Discussion and conclusions
Reconstructions of the GrIS SMB from the beginning of the last century
(1900–2015) were carried out using the regional climate MAR model forced by
eight reanalyses. Over the recent decades, all MAR time series compare very
well with in situ measurements, ice core and satellite-derived melt extent,
while temperature corrections were needed in the 20CRv2 and ERA-20C
reanalyses at the MAR boundaries. MAR forced by ERA-Interim shows the best
comparison with observations for 1979 onward, while NCEP–NCARv1 outperforms
ERA-40 and JRA-55 over 1958–1978. Among the reanalyses covering the entire
century, 20CRv2c is the only reanalysis that does not need correction at the
MAR boundaries, but its performance is not as good as the fully assimilated
reanalyses such as ERA-Interim over the recent decades.
Around 1930, all reconstructions agree on an SMB minimum concurrent with the
warm period observed in the coastal temperatures .
Afterwards, the reconstructions suggest a melt decrease until the 1970s and
an accumulation increase until the middle of the 1940s. A second minimum of
SMB occurs in the 1960s when a minimum of accumulation is reached, while the
highest SMB rates are reached over the 1970s–early 1990s, as a consequence of
lower melt and higher accumulation than before. All reconstructions then show
a significant SMB decrease resulting from a surface melt increase starting at
the end of the 1990s and lasting until the 2010s, when the SMB absolute
minimum since 1900 is reached in all time series.
Before the 1930s, there are, however, large discrepancies between the MAR
reconstructions as well as with the time series. MAR forced by
ERA-20C suggests a continuous run-off increase from the 1900s to 1930s, while
MAR forced by 20CRv2(c) and, to a lesser extent, BOX13 suggests a run-off
decrease from the 1900s until the 1910s, followed by a melt increase reaching
a first maximum at the beginning of the 1930s. Similar discrepancies can be
seen in the MAR-simulated near-surface temperatures. MAR also simulates a
significant snowfall increase from the 1910s to the 1940s. Reconstructions
from and ice cores also suggest an
accumulation increase over this period but smaller than MAR's increase, while
suggested a decrease of the accumulation. Long-term ice core
data facilitate validation of an overall ice sheet snowfall increase in the
first half of the last century, and the comparison with MAR is good where a
few ice cores are available. This increase is, however, bracketed in several
ice cores in the dry north as well simulated by MAR, but not for the only core
(see D1 in Fig. ) in the south-east showing decreasing snowfall.
Thus, the ice sheet-averaged core trend is almost insignificant, while MAR
suggests a significant increase along the south-east coastal ridge where ice
cores are missing. This suggests that new ice core drillings are needed in
this area to confirm the MAR accumulation increase. Moreover, this
accumulation increase in MAR coincides with an increase of the daily sea
level pressure variability in forcing reanalyses, which impacts the amount of
humidity advected into the MAR integration domain. The 20CRv2(c) reanalysis
is an ensemble mean of 56 members, suggesting that the lower the amount of
assimilated data is, the smoother the pressure fields are. Therefore, the
increase of the daily sea level pressure variability could just be an
artefact coming from forcing reanalysis. While ERA-20C is not an ensemble
mean, MAR forced by this reanalysis shows the same magnitude of precipitation
increase as MAR forced by 20CRv2(c), but not at the same locations. On the
other hand, the amount of data assimilated into ERA-20C is lower during this
period. Therefore, without gauge observations in the areas where the changes
are the highest, it is hard to conclude whether this MAR-based significant
accumulation increase along the south-east coastal ridge over the first half
of the last century is robust or whether it is just an artefact coming from the
forcing reanalyses (which need to be more constrained to be in agreement
before the 1930s). already showed large discrepancies
in the general circulation simulated over Greenland by these two reanalyses
before 1940, explaining the significant differences in the simulated run-off
and snowfall variability.
The period 1961–1990 has been considered as a period when the total mass
balance of the Greenland ice sheet was stable and near zero.
However, at the last century scale, all MAR reconstructions suggest that SMB
was particularly positive during this period (SMB was most positive from
the 1970s to the middle of the 1990s), suggesting that mass gain may well
have occurred during this period, in agreement with results from
.
Finally, with respect to the 1961–1990 period, the integrated contribution
of the GrIS SMB anomalies over 1900–2010 is a sea level rise of about
15 ± 5 mm, with a null contribution from the 1940s to the 2000s,
suggesting that the recent contribution of GrIS to sea level change
is unprecedented in the last century. A next step to
evaluate total mass changes should be to force ice sheet models with these
MAR reconstructions to confirm the stability of the ice dynamics over
1961–1990 and to better understand the recent acceleration of ice dynamics
. This recent acceleration of ice dynamics could partly
result from the purge of the extra mass (accumulated through the 1970–1990s)
enhanced by the recent melt increase lubricating the glaciers–bedrock
interface.