The mass balance of the Greenland Ice Sheet (GrIS) in a warming climate is of
critical interest to scientists and the general public in the context of
future sea-level rise. An improved understanding of temporal and spatial
variability of snow accumulation will reduce uncertainties in GrIS mass
balance models and improve projections of Greenland's contribution to
sea-level rise, currently estimated at 0.089
Assessing the stability of the Greenland Ice Sheet (GrIS) in a warming world
is crucial for predicating future global sea-level rise and its societal and
economic impacts (Dumont et al., 2014; IPCC, 2014). The mass balance of the
GrIS decreased over the 1988–2016 period, with a conservative estimate of
ice sheet mass loss of 272
Here we develop a record of GrIS snow accumulation over a large portion of the GrIS interior from AD 1712 to 2014 using the airborne NASA Operation IceBridge accumulation radar (Leuschen et al., 2011). Airborne and ground-based radars have been used to map spatial patterns of accumulation in Greenland over decadal (Hawley et al., 2014; Miège et al., 2013) and annual resolutions (Koenig et al., 2016; Medley et al., 2013). Operation IceBridge collected accumulation radar data from 2009 to 2014, and it has been used in several studies (Karlsson et al., 2016; Forster et al., 2014; Leuschen et al., 2011; Medley et al., 2013) to calculate local accumulation. We examine accumulation radar data from every IceBridge flight across the Greenland interior during the 2013 and 2014 seasons to measure accumulation rates over the majority of the dry and upper percolation zones.
Regional climate models and reanalysis products provide spatially and
temporally comprehensive estimates of accumulation at ice-sheet scales. The
magnitude of mesoscale model uncertainty can be as large as the natural
variability, or larger in areas with sparse in situ measurements like ice
cores, potentially obscuring climate fluctuations with random error (Burgess
et al., 2010; Box et al., 2006). A 2013 study (Vernon et al., 2013)
determined that 1960–2008 climate model SMBs differ by as much as
130 Gt a
We calculate a spatially continuous record of accumulation along 17 730 km of NASA Operation IceBridge accumulation radar flights (hereafter “IceBridge accumulation”). Operation IceBridge was designed to bridge the gap in polar observations between the Ice, Cloud, and Land Elevation Satellite (ICESat; 2003–2009) and ICESat-2, which is scheduled to launch in 2017. Laser altimeters, four to five different frequency radars, a gravimeter, and a magnetometer are mounted on NASA's P-3B and DC-8 airplanes, which conduct airborne surveys in both the Arctic and Antarctic each spring.
The IceBridge accumulation radar captures a continuous electromagnetic
profile of the top few hundred meters of the ice sheet, displaying distinct
internal reflecting horizons (IRHs) that can be traced for hundreds of
kilometers (Leuschen et al., 2011). The accumulation radar operates in the
600–900 MHz range and has an average vertical resolution of 0.28 m in
snow/firn, which is fine enough to resolve IRHs that have been shown to
represent isochrones (Medley et al., 2013; Rodriguez-Morales et al., 2014;
Spikes et al., 2004; Hawley et al., 2014). The average distance between radar
traces is 16 m, which we then average over 10 adjacent traces to increase
the signal-to-noise ratio. The position of each trace is known from
differential GPS receivers mounted on the aircraft. We do not perform any
time variable gain or additional filtering on the IceBridge accumulation
data. Depending on signal attenuation within the snowpack, IRHs can be traced
to a depth of 50–150 m and provide accumulation records over the past
100–300 years (Fig. 1). For areas with high attenuation (i.e., shallow
penetration of the radar signal), such as those at relatively lower
elevations (e.g., below
To calculate accumulation rates using ice penetrating radar, one must know the amount of snow mass between IRHs and their relative ages. The mass between IRHs is a function of the depth–age scale, travel time–depth conversion rate, and firn or ice density. We obtain both the density profile and depth–age scale from two dated ice cores collected at Summit Station (Mary Albert, personal communication, 2015; Cole-Dai et al., 2009). These ice core sites are 3 and 7 km from the closest IceBridge radar trace, and we assume similar accumulation rates across this small distance. We correct for the 7-year difference between ice core collection and IceBridge radar flights by extrapolating the depth–age curve.
We calibrate a Herron and Langway (1980) depth–density model at Summit using data
from both ice cores, then use the calibrated model parameters to estimate
density profiles elsewhere in our study region. Input parameters for this
model include satellite-derived mean annual temperature (Hall et al., 2012),
modeled accumulation (Burgess et al., 2010), and an estimate of surface snow
density from field measurements along ground traverses, shallow firn cores,
and MAR model output. Since we are using the density profile to calculate
accumulation based, in part, on modeled accumulation, the results could be
seen to be circular. However, our results are largely insensitive to changes
in this modeled accumulation input because accumulation estimates are
minimally affected by input variations to the Herron–Langway model. For
example, adjusting input accumulation and surface density by
We convert the radar travel time to depth by iteratively multiplying the
velocity of the electromagnetic wave by the signal's travel time to each IRH.
The electromagnetic speed of the radar wave,
In turn, the dielectric permittivity is calculated from the density,
The snow surface reflection is readily identified in each radar profile from the large signal amplitude. We then calculate the depth for each subsequent radar sample in the profile using the radar travel time and velocity profile from Eqs. (1) and (2), following Hawley et al. (2014).
We manually select 19 clear, strong IRHs to consistently trace from Summit Station towards the NNW and SW along two main flight paths (5 April and 2 May, 2014, respectively; see Fig. 1). When a layer appears to bifurcate due to changes in accumulation, we continue to trace the layer based on the trajectory of surrounding IRHs. Horizons are not traced in areas where the signal-to-noise ratio made them too difficult to discern.
Internal reflecting horizons for the other 23 flights in this study are traced from crossover locations with the two main flight paths. Wherever possible, we trace IRHs outwards from crossover locations along the two main flight paths to locations where those traced layers cross another flight path. Whenever we have accumulation differences at crossover locations larger than our accepted error, we review IRHs to determine which layers are incorrectly traced.
Date of oldest resolvable internal reflecting
horizon (IRH) along 25 IceBridge
accumulation radar flights totaling 17 730 km. Locations are shown for
A–A
Finally, we calculate snow accumulation using the ice core depth–age scales,
modeled depth–density profiles, and traced IRHs. We calculate accumulation
between each pair of adjacent IRHs for every radar trace along the flight
lines. Spatial changes in accumulation are evident from varying vertical
distances between IRHs along each flight line. Temporal changes in
accumulation are evident from examining accumulation during different epochs
at one radar trace. We calculate the water-equivalent accumulation,
We do not correct for ice flow due to advection of the ice sheet since
nearly all of the radar traces occur in areas with surface velocities
Uncertainty in accumulation can arise from independent errors in tracing IRHs, errors from incorrectly dating the ice core, and/or errors in the densities used for converting from separation distance to water-equivalent accumulation.
To reduce tracing errors, two authors separately retraced each IRH along the
two main flights paths four times each. Close inspection of the IRHs reveals
that the peaks defining IRHs are within
We take uncertainty in dating the Summit ice cores to be
The error associated with measuring density using similar techniques has been
estimated to be 1.4 % (Karlöf et al., 2005). However, following
Hawley et al. (2014) we conservatively assume that our measurements have an
error of up to twice this large, corresponding to a maximum accumulation
error of
The three error sources are all random, non-systematic, and thus can be
assumed to be non-additive (following Hawley et al., 2014). Over the extent
of the data set we can assume that the errors are not correlated, and thus we
estimate accumulation uncertainty from all sources at
We compare our IceBridge accumulation results with annual outputs from Polar MM5 (1958–2008; Burgess et al., 2010), MARv3.5.2 (1948–2015; Fettweis et al., 2016), RACMO2.3 (1958–2015; Noël et al., 2016), and Box13 (1840–1999; Box et al., 2013). Grid cell sizes for these model outputs are 24, 5, 1, and 5 km, respectively. Since accumulation can be bilinearly interpolated over the distance of these grid cells without significant loss of detail (Box and Rinke, 2003), we choose to compare IceBridge accumulation with bilinearly interpolated model grid output to compare accumulation at corresponding spatial locations.
The Box13 data set is corrected using a correction multiplier grid, which is estimated using a triangular irregular network interpolation of the ratio between 1961–1990 average Box13 ice core accumulation rates and RACMO2.1 output. The multipliers have respective minimum and maximum values of 0.605 and 1.891. We assume that the calibration coefficients are stationary in both time and space, since Fettweis et al. (2016) show that MAR accumulation reconstructions are similar to those from Box13 after 1930.
Additionally, we compare our IceBridge accumulation with an accumulation map krigged from 295 snow pits and ice cores and 20 coastal weather stations (Bales et al., 2009). While this map estimates accumulation over the time domain of the oldest ice cores, we choose to compare IceBridge accumulation with the highest accuracy accumulation estimates from 1950 to 2000, which include weather stations and recent ice cores.
Average accumulation over the temporal domain of each radar trace calculated from IceBridge accumulation radar over all 25 flights. IceBridge accumulation matches large-scale accumulation patterns from ice cores and snow pits from Bales et al. (2009).
IceBridge accumulation patterns are consistent with observed large-scale spatial patterns from ice cores and snow pits (Bales et al., 2009), with high accumulation rates in the southeast and southwest and lower accumulation rates in the northeast and at higher elevations of the ice sheet interior (Fig. 3). The number of traceable layers is highest towards the interior of the ice sheet and lowest in warmer areas towards the coast and in the south, where enhanced surface melt attenuates the radar signal and reduces the density gradients that produce IRHs (Fig. 2).
We assess the internal consistency of IceBridge accumulation by comparing the
accumulation at 87 locations where IceBridge flight paths cross one another
(hereafter “crossover points”). Differences at crossover points are most
likely due to errors in layer picking where isochrones become difficult to
detect or distinguish. There are no spatial or temporal patterns in
accumulation differences at crossover points over the data set. Moreover, the
differences are normally distributed with a mean of
0.017
Comparison of IceBridge accumulation rates determined at 87
crossover locations for each epoch, totaling 1241 measurements. There are no
temporal or spatial patterns in crossover location accumulation differences.
Shaded region is the calculated uncertainty of
Averaged ice core accumulation compared with IceBridge (IB) accumulation averaged over the overlapping time domain of each ice core. Uncertainty figures represent 1 standard deviation of ice core accumulation and average IceBridge accumulation at the closest radar trace to each core, respectively. Trends and their standard deviation are reported for both ice core accumulation and nearest IceBridge accumulation.
IceBridge accumulation (blue) with uncertainty (blue circles) compared with Camp Century, D3, and D4 (see Fig. 2 for locations) ice core annual accumulation (thin red lines) and ice core accumulation averaged over corresponding epochs (thick red lines). A red square denotes 1 standard deviation of ice core annual accumulation over each epoch. Note the longer timescale for the D4 ice core. There is no statistically significant difference between IceBridge and ice core accumulation for any of these ice cores.
Accumulation rates derived from ice cores collected at Camp Century, D3, and
D4 (see Fig. 2 for locations) correspond closely with our IceBridge
accumulation rates, matching their long-term mean and tracking their decadal
variability (Fig. 5). Additionally, we compare IceBridge accumulation rates
and trends to the NASA-U, NEEM, D5, B26, B29, NGRIP, and PARCA ice cores over
corresponding temporal domains (Table 1). IceBridge accumulation rates and
accumulation trends are statistically indistinguishable from each of these
cores at a
In Fig. 6 we compare IceBridge accumulation to snow pit measurements at
station T-31 on the Expédition Glaciologique Internationale au Groenland
(EGIG) traverse (Fischer et al., 1995; Hurbertus Fischer, personal
communication, 2015), and to accumulation rates calculated at this location
from the Airborne SAR/Interferometric Radar Altimeter System (ASIRAS; Overly
et al., 2016; see Fig. 2 for location). IceBridge accumulation rates are
statistically indistinguishable (
IceBridge accumulation results at EGIG T-31 (see Fig. 2 for location) from 1957 to 2014 are statistically indistinguishable from Airborne SAR/Interferometric Radar Altimeter System (ASIRAS) accumulation Overly et al., 2016) and field measurements (H. Fischer, personal communication, 2015). Error bars are 1 standard deviation of ASIRAS accumulation over data points from that time period.
We compare IceBridge accumulation to RCM accumulation results along the
length of each flight. IceBridge accumulation is averaged over 1957–2014 to
compare with averaged Polar MM5 (1958–2008), MAR (1948–2015), RACMO2
(1958–2015), and Box13 (1840–1999). An example of this comparison along a
single flight (B–B
The model output and IceBridge accumulation time domains do not match
identically, but these minor differences do not significantly affect our
results. The largest time domain discrepancy is with the Polar MM5
comparison, where model output is averaged from 1958 to 2008 and IceBridge
accumulation is averaged from 1957 to 2014. The top panel of Fig. 7 shows Polar
MM5 output averaged from 1958 to 2008 compared to IceBridge accumulation
averaged from 1957 to 2004. The difference between IceBridge averaged over
1957–2014 and IceBridge averaged over 1957–2004 along this flight is
0.00096
Next, we compute the magnitude and percent differences between RCM output and
IceBridge accumulation over the entire domain of this data set. Averaged over
all 25 flights, the RMS difference between the models and IceBridge
accumulation is 0.036
Comparison of 1957–2004 averaged IceBridge accumulation (solid
line) and uncertainty (shaded region) to averaged Polar MM5 (1958–2008;
triangles) along a 977 km flight in northern Greenland. Location of flight
shown as B–B
Percent and magnitude differences between average 1957–2014 IceBridge accumulation and average model accumulation in each of the six GrIS drainage basins. Positive numbers indicate that the model overestimates accumulation in that basin. Plus minus figures represent 1 standard deviation. Statistically significant differences are indicated in bold.
Magnitude (left) and percent (right) differences between averaged
1957–2014 IceBridge accumulation and
We divide the GrIS into six major drainage basins (see Fig. 8)
following Vernon et al. (2013) to evaluate and discuss the spatial
differences between model and IceBridge accumulation. Table 2 shows both
percent and magnitude differences between the models and 1957–2014 averaged
IceBridge accumulation in each of the six drainage basins. Statistically
significant differences (
Averaged across basin A, the northern basin with generally low accumulation
rates, there are no statistically significant differences between IceBridge
accumulation and any of the RCMs used in this study. Although the models
disagree with each other in this basin, as suggested by Vernon et al. (2013),
the differences from the IceBridge accumulation are neither large nor
statistically significant. Basin B in the northeast has some of the largest
differences between models and IceBridge accumulation. Averaged across all
815 points in basin B, MAR and Box13 underestimate by 18.68
Basin D in the southeast is poorly covered by our data, but we find that MAR
significantly overestimates accumulation by an average of
23.31
In summary, the RCMs do an excellent job of calculating accumulation averaged over basins A and E, but there are large differences between model and IceBridge accumulation in basins B and C. We note that RACMO2.3 does not significantly differ from IceBridge accumulation in any of the basins. Areas where RCM and IceBridge accumulation differ the most are concurrent with areas without many in situ measurements (e.g., in the southeast), and where ice cores were collected several decades ago (e.g., NASA-U, Camp Century). Additional field measurements would be beneficial to validate both our IceBridge accumulation and RCMs in these data-poor regions.
Averaged across all 25 flights, the Bales09 accumulation model krigged from
ice core and snow pit measurements differs from averaged 1957–2014 IceBridge
accumulation by 0.033
Basins B, E, and F have sufficient data coverage to extrapolate over these
basins' spatial domain to estimate the model uncertainty of their SMB
estimates. We obtain total model uncertainty (in Gt a
A study by Karlsson et al. (2016; hereafter Karlsson16) uses a very different
method to calculate accumulation from IceBridge accumulation radar data near
NEEM and NGRIP. We compare data from their study, representing flight lines
in 2011 and 2012, to a repeat flight during the 2014 IceBridge season
analyzed using our method. In Fig. 9, the 1921–2014 accumulation rates (this
study) are plotted against 1911–2011 Karlsson16 accumulation rates and the
RCMs used for comparison in this study. On average along the 350 km flight
line, the accumulation rates calculated in this study are
0.002
We can analyze spatiotemporal trends in snow accumulation using our IceBridge accumulation record spanning 17 700 km of flight paths over the past 300 years. We perform an empirical orthogonal function (EOF) analysis on the data set to evaluate temporal changes in accumulation and assess potential atmospheric forcing mechanisms (Fig. 10). We limit our EOF analysis to 1889–2014 to capture the maximum spatial variability since layers older than 1889 are difficult to trace in the southern region (see Fig. 2). We find that EOF1 and EOF2 represent most of the variance within the data set, explaining 33 and 19 % of the variance, respectively.
Map of correlation between IceBridge accumulation and
The EOF1 time series has a statistically significant positive correlation
with the 1899–2014 annually averaged Atlantic Multidecadal Oscillation (AMO)
index (
The EOF2 time series is significantly correlated with the wintertime (DJF)
NAO, with
Correlation map between 1899 and 2014 IceBridge accumulation and
epoch-averaged climate indices. Statistically significant correlations (
Although it does not appear through our EOF analysis, there are significant positive correlations between the summertime GBI and IceBridge accumulation (Fig. 11b), indicating positive GrIS accumulation anomalies during summers with overall enhanced blocking. While this may seem counterintuitive, this relationship is driven by enhanced meridional flow and moisture advection into Greenland under the weak zonal flow associated with GBI positive (generally NAO negative) conditions (Hanna et al., 2016). Hanna et al. (2016), in a study based on reanalysis data, similarly find enhanced precipitation in central northern Greenland associated with positive GBI summers (their Fig. 6g). They also show negative precipitation anomalies in southeast Greenland during positive GBI summers, but our IceBridge data coverage in that region is too poor to confidently evaluate GBI relationships.
If our hypothesis is correct that a positive AMO index (anomalously warm North Atlantic sea-surface temperatures) contributes to anomalously high GrIS accumulation, then the future behavior of the AMO may have a significant impact on the rate of GrIS mass loss. Hanna et al. (2013b) found that positive AMO summers were associated with enhanced GrIS surface melting, indicating that the AMO impacts both the mass input and mass loss portions of Greenland SMB. The highest quality climate observations, reanalysis data and RCM output exist for the 1979–present interval, during which the AMO progressed from a negative phase (in the 1980s) to a positive phase (in the 2000s), with a rapid AMO warming transition in the 1990s (Fig. 10c). Paleoclimate records show evidence that the AMO was a persistent sea surface temperature (SST) mode throughout the late Holocene with a periodicity of 20–70 years (Chylek et al., 2012; Knudsen et al., 2011), and thus would be expected to continue into the future. We therefore encourage modeling efforts to evaluate the GrIS mass balance implications of a future return towards AMO negative conditions during a continued increase in radiative forcing from anthropogenic greenhouse gases.
We have developed a new data set of accumulation rates over the interior of the GrIS spanning the past 100–300 years based on 17 730 km of Operation IceBridge airborne accumulation radar data. This accumulation record is internally consistent across the data set and is validated by in situ field measurements, several ice cores, and other radar-derived accumulation measurements.
Overall, the Polar MM5, MAR, and RACMO2 Regional Climate Models, as well as
Box13 and Bales09, accurately capture large spatial patterns in accumulation
over the GrIS, but show significant differences from IceBridge accumulation
on a regional basis. For example, in the southeast, MAR overestimates
accumulation by an average of 20.89
Empirical orthogonal function analysis indicates that the first and second principal components explain 33 and 19 % of the variance and correlate with the AMO and NAO, respectively. These results are consistent with previous ice core and weather station analyses demonstrating the importance of these North Atlantic climate modes on Greenland SMB. We recommend that future modeling efforts evaluate the effects of a future return to AMO negative conditions on GrIS surface mass balance as greenhouse gas concentrations continue to rise.
Our largest accumulation uncertainties align with regions that disagree most strongly with climate models. Thus, future research should be aimed at collecting additional in situ measurements in areas with large disagreement between climate models, particularly in the southeast.
Accumulation rates for each time period are available in the Supplement for every radar trace over the domain of this study.
The authors declare that they have no conflict of interest.
This project was supported by the US National Science Foundation under grants DGE-1313911 and ARC-1417640. We would like to thank Hurbertus Fischer and Mary Albert for providing field validation measurements, as well as Xavier Fettweis and Brice Noël for providing the most recent MAR and RACMO regional climate model outputs. The authors would like to thank Xavier Fettweis and one anonymous reviewer, as well as Nanna Karlsson, whose comments greatly improved the manuscript. Finally, we would like to thank Marco Tedesco for handling the manuscript. Edited by: M. Tedesco Reviewed by: X. Fettweis and one anonymous referee