comparison and validation of four arctic sea ice thickness products

COMPARISON AND VALIDATION OF FOUR ARCTIC SEA ICE THICKNESS
PRODUCTS OF THE EC POLAR ICE PROJECT
C. Melsheimer(1) , M. Mäkynen(2) , T.S. Rasmussen(3) , Ø. Rudjord(4) , M. Similä(2) , R. Solberg(4) , and N.P. Walker(5)
(1)
Institute of Environmental Physics, University of Bremen, Otto-Hahn-Alleee 1, 28359 Bremen, Germany,
Email: [email protected]
(2)
Finnish Meteorological Institute, Marine Research, Erik Palmenin aukio 1, 00101 Helsinki, Finland,
Email: [email protected], [email protected]
(3)
Danish Meteorological Institute, Lyngbyvej 100, 2100 København Ø, Denmark, Email: [email protected]
(4)
Norwegian Computing Centre, P.O. Box 114 Blindern NO-0314 Oslo, Norway, Email: [email protected]
(5)
eOsphere Limited, Satellite Applications Catapult (R103) Harwell Space Cluster, Fermi Avenue OX11 0QR, United
Kingdom, Email: [email protected]
ABSTRACT
Sea ice thickness (SIT) is an important parameter for
monitoring Arctic change, modelling and predicting
weather and climate, and for navigation and offshore operations. However, SIT is still not very well monitored
operationally. In the European Commission (EC) FP7
project “POLAR ICE”, three novel SIT products based on
different satellite data as well as SIT from a state-of-theart ocean and sea ice model are fed into a common data
handling and distribution system for end users. Each SIT
product has different scopes and limitations as to, e.g.,
spatial and temporal resolution, ice thickness range and
geographical domain. The aim of this study is to compare the four different SIT products with each other and
with SIT in-situ measurements in order to better understand the differences and limitations, and possibly give
recommendations on how to best profit from the synergy
of the different data.
1.
INTRODUCTION
Sea ice thickness (SIT) is a key parameter for understanding changes taking place in polar regions, and for
ship navigation and planning of off-shore operations in
ice-infested waters. In addition, in polar seas, the oceanatmosphere heat, momentum and gas exchanges are controlled by the SIT distribution, which is therefore of interest in climate studies. As SIT in-situ measurements are
sparse and only represent points in time and space with
limited spatial and temporal connections, it is important
to have more frequent SIT data sets that can be derived
from various satellite sensors and ocean-sea ice models.
An additional advantage of having multiple data sets is
the ability of estimating uncertainties through intercomparison and comparison with in-situ SIT data.
Currently daily Arctic-wide SIT information is operationally available mainly as WMO ice type classification in ice charts provided by national Ice Services and
as estimates by various ocean-sea ice models. The EC
FP7 POLAR ICE project [1] distributes in 2014-2016
three new SIT products for the Arctic based on satellite data for operational usage: (1) daily thickness chart
of thin sea ice (up to about 0.5 m) in the entire Arctic
retrieved from brightness temperature measurements of
the L-band microwave radiometer on the ESA satellite
SMOS (Soil Moisture and Ocean Salinity), with a resolution of 30 to 40 km; (2) daily thickness chart of thin
sea ice (up to about 0.5 m) in the European Arctic derived from VIIRS (Visible Infrared Imaging Radiometer Suite, on NOAA satellite Suomi-NPP) thermal imagery – subject to cloud cover – with resolution of 750 m,
and (3) daily SIT chart (detected thin ice, and ice thickness 30 to 255 cm) of the Barents and Kara Seas derived
from multisensor (AMSR2, SENTINEL-1) data and the
TOPAZ-4 (Towards an Operational Prediction system for
the North Atlantic European coastal Zones) ocean-sea ice
model at resolution of 1 km. All three thickness products
are restricted to cold winter conditions (dry snow cover),
roughly October to May in the Arctic. Within POLAR
ICE, a state-of-the-art coupled ocean and sea ice model
(HYCOM/CICE) runs twice a day in an operational mode
and produces a 5-day forecast with a resolution of 10 km.
Among other parameters, the model also predicts SIT.
POLAR ICE integrates these SIT products with other sea
ice information products derived from satellite and sea
ice model data, so that end-users can readily visualise the
products, received in near real time, in synergy with one
another.
Here we intercompare the three satellite-based SIT charts
and the output of the coupled ocean-sea-ice model, and
compare them with available in-situ ice thickness data.
The aim is to see how well the four SIT products agree,
and to find out where and under which conditions mismatches typically occur. The intercomparison of the
SMOS and VIIRS SIT charts also allows comparing thin
ice thickness estimates based on two different physical
relationships; ice thickness vs. microwave emission for
SMOS, and ice thickness vs. surface temperature for VIIRS. Comparison to in-situ data gives estimates of the absolute accuracy of the POLAR ICE SIT charts which is
the main information required by end-users.
The intercomparison is conducted over the Barents and
Kara Seas where all four products overlap. In addition,
in the Barents Sea, East of Svalbard, a sea ice campaign
was conducted in March 2014 (joint ESA SMOSIce campaign and test cruise of German IRO2 project) where,
among other parameters, sea ice thickness was measured
with ship- and helicopter-borne electromagnetic induction (EM) instruments (we did not use the helicopter EM
data here).
The next section briefly describes the four SIT products,
Section 3 reports on their comparison with the in-situ
data. Section 4 shows and discusses results of intercomparing the four SIT products, and Section 5 finally gives
conclusions and an outlook.
2.
SEA ICE THICKNESS PRODUCTS IN “POLAR ICE”
This section outlines the SIT products from the three different satellite data based retrieval schemes and from the
HYCOM/CICE ocean and sea ice model. Note that all
three SIT products that use remote sensing data can only
be retrieved in freezing conditions, i.e., from the autumn
freeze-up until the spring melt.
2.1. Sea ice thickness from L-band radiometer data
The University of Bremen (UB) retrieves SIT from microwave brightness temperatures (TB ) at 1.4 GHz (L
band) measured by the ESA satellite SMOS (Soil Moisture and Ocean Salinity). SIT up to 0.5 m (average over
one antenna footprint) can be retrieved. If sea ice is
thicker than this, it will be recorded as 0.51 m. The coverage is daily during the winter season, i.e., October to
May in the Arctic and March to September in the Antarctic. SMOS satellite data availability starts in mid 2010.
The gridding interval is 12.5 km, while the resolution
(antenna footprint size) is 30 to 50 km. The relative error of the retrieved SIT is about 30%. Details about the
retrieval and validation can be found in [2]. Here is a
brief outline: The L-band radiometer of SMOS is fully
polarimetric, i.e., it measures all four components of the
Stokes vector of the emitted radiation. The algorithm developed at UB uses the first two Stokes components, I
and Q, i.e., the average of the vertically and horizontally
polarised TB (usually called intensity) and their difference (the polarisation difference), respectively. Data at
incident angle between 40◦ and 50◦ are used. Comparison with a training data set of SIT shows that during the
winter period, the intensity is highly correlated with the
SIT whilst the polarisation difference is anti-correlated,
and non-linear functions can be fitted that express intensity I and polarisation difference Q as functions of SIT
(see [2]). From these two functions, a retrieval curve can
be constructed in the I-Q-plane that indicates which pair
of (I,Q) corresponds to which SIT. This way, SIT of up to
0.5 m can be retrieved: for a given pair of (I,Q), the point
on the retrieval curve of minimum Euclidean distance is
determined. The retrieved SIT is then the thickness at
that point of the curve. If the ice is thicker than that, the
SMOS signal is saturated, i.e., it does not depend much
Figure 1. Sea ice thickness chart in the entire Arctic on
12 April, 2016, based on SMOS data. Sea ice thickness
above 50 cm is shown in pale orange.
on the SIT any more. In this case, the retrieval yields a
flag “thickness above 0.5 m”. An example chart is shown
it Fig. 1. The data are projected into polar stereographic
coordinates with a reference longitude of 45◦ W and are
delivered as GeoTIFF file.
The retrieved SIT agreed reasonably well with measurements of helicopter-borne electromagnetic sounding (the
EM-bird instrument [3]) and MODIS thermal imagery
SIT retrievals [4].
2.2. Sea ice thickness from thermal infrared data
The Norwegian Computing Center (NR) has developed
an ice thickness retrieval algorithm for thin sea ice, based
on thermal satellite imagery. A prototype algorithm was
developed in the ESA PRODEX funded ThinIce project
[5]. The algorithm and processing chain was further developed within POLAR ICE.
The algorithm uses a thermal model of the ice surface [6]
to estimate the ice thickness. The required input is surface temperature and atmospheric forcing data (air temperature and wind speed). Surface temperature is obtained from the thermal VIIRS M15 (11 µm) and M16
(12 µm) bands using the split-window technique by [7],
while the atmospheric data are provided by the Norwegian Meteorological Institute (MET Norway).
The product is a daily SIT chart, based on a mosaic of
VIIRS data from the previous day. The chart has a spatial
resolution of 750 m and covers most of the Arctic (except
the Pacific side). The relatively high spatial resolution
makes it possible to distinguish finer structures in the sea
ice, such as frozen leads. The chart shows ice thicknesses
up to 50 cm (see example in Fig. 2).
Figure 2. Sea ice thickness chart in the European Arctic on 17 March, 2016, based on VIIRS data. Sea ice is in various
shades of blue, clouds are yellow, open water is purple.
The product is however, limited due to the presence of
clouds. In order to limit the possibility of cloud contamination in the product, it is necessary to use a conservative
cloud mask. As masking out clouds is very challenging
in the Arctic, this leaves relatively small areas visible.
2.3. Sea ice thickness with multisensor approach
The Finnish Meteorological Institute (FMI) has developed a sea ice thickness chart (ITC) for the Barents and
Kara Seas based on multi-sensor satellite data and modelled ice thickness [8]. The first operational version of
the ITC was demonstrated in the ESA funded ANISTIAMO project (Addressing new challenges in satellite
based maritime surveillance and Arctic monitoring) running from 2012 to 2014. The development of demonstration activities of the ITC have been continued in POLAR
ICE.
Multi-sensor sensor data now used in the ITC production include SENTINEL-1 EW mode dual-polarised images and AMSR2 radiometer 36 and 89 GHz brightness temperature (TB ) data. The modelled ice thickness
field comes from the CMEMS TOPAZ-4 coupled oceansea ice data assimilation system [9]. Thin ice areas are
first detected based on 36 GHz polarisation ratio and Hpolarisation gradient ratio between 36 and 89 GHz calculated from the AMSR2 data, see details in [10]. The
maximum thickness of detected thin ice is estimated to
be 30 cm corresponding to the maximum thickness of the
WMO young ice type. It is noted that AMSR2 data may
assign the thin ice class to areas which according to the
TOPAZ-4 model are thicker ice. We consider identification of thin ice areas with the AMSR2 data (empirical
method) more reliable than using the TOPAZ-4, or any
other model data. For thicker ice areas SAR backscatter data are used to modulate locally the TOPAZ-4 ice
thickness field. The exclusion of thin ice areas reduces
significantly the ambiguity of the radar backscatter in the
SAR images. The average ice thickness corresponding to
a SAR backscatter level is derived using the TOPAZ-4 ice
thickness field and the equivalent deformed ice thickness.
Open water and areas with sea ice concentration (SIC)
less than 70% are identified using the AMSR2-based SIC
data provided by JAXA and are excluded from the ice
thickness estimation.
The ITC is a daily product based on AMSR2 and
SENTINEL-1 data mosaics from swath data acquired
during the previous 24 and 48 hours, respectively. The
spatial resolution of ITC is 1 km, and it shows a thin ice
class (< 30 cm), ice thickness in 31-255 cm range, and
three WMO SIC classes for SIC< 70%. The spatial coverage is 1850 by 2200 km. The ITC is in polar stereographic (PS) coordinate system with reference longitude
of 55◦ E, and is delivered as 8-bit GeoTIFF-file. An example ITC is shown in Fig. 3.
Figure 3. Sea ice thickness chart over the Barents and
Kara Seas on 12 Apr. 2016. The chart is based on
SENTINEL-1 EW images, AMSR2 radiometer data, and
TOPAZ-4 modelled sea ice thickness. Thin ice with maximum thickness of 30 cm is shown with purple colour.
Light blue show and light purple colours shows areas
with ice concentration less than 70%.
Figure 4. Sea ice thickness chart of the Arctic Ocean on
12 Apr, 2016, calculated with the HYCOM/CICE model.
2.4.
Sea ice thickness from the HYCOM/CICE
model
The Danish Meteorological Institute (DMI) runs an operational ocean (HYCOM v2.2.58) and sea ice model
(CICE v4.1) coupled using the Earth System Modeling
Framework (ESMF). The physical model system is based
on a system mainly developed by the Naval Research Lab
at Stennis Space Center (USA) and is all open source
and state of the art. HYCOM [11] is different from
other ocean models in that its vertical coordinate is hybrid: the main coordinate is isopycnal but in shallow and
well mixed areas it reverts to either sigma coordinate (terrain following) or z-coordinates (fixed depth intervals).
CICE is a dynamic and thermodynamic multi-category
sea ice model that calculates the ice growth and the ice
drift [12, 13]. The dynamics includes forcing from ocean
and atmosphere. The internal strength is based on the
elastic-viscous-plastic model described by [13, 14].
This model system at DMI [15] runs in a regional configuration that covers the Arctic and the Atlantic oceans
to 20 degrees south of the equator and it is forced by
ERA-INTERIM meteorological reanalysis data in hindcast mode and the operational ECMWF meteorological
analysis in forecast mode. It assimilates sea ice concentration (SIC) and sea surface temperatures with a Newtonian nudging scheme. When assimilating SIC, SIT is
estimated based on the present ice category distribution,
thus when sea ice is added or removed it is adjusted according to the distribution of the current ice categories.
For instance if the SIC of category 3 is 50% and the observation shows that the total SIC should be increased by
10%,then the SIC of this category should be increased towards 55%. In operational mode it runs twice a day producing a 1 day hindcast and a 5 day forecast. An example
ITC produced by the model is shown in Fig. 4.
3. COMPARISON WITH IN-SITU DATA
The joint ESA SMOSice campaign and IRO-2 test cruise
took place in March 2014 in the Barents sea, East of Svalbard [16]. The SIT data used in this study were measured by the electromagnetic induction instrument EM31
Figure 5. Track of the research vessel Lance doing the
EM31 sea ice thickness measurements during the ESA
SMOSice campaign/IR0-2 test cruise in March 2016 in
the East of Svalbard (the island of Spitsbergen in upper
left corner, Edgeøya in the centre).
mounted at the bow of the ship. Measurements were
made twice per second, and the footprint size of EM31
was about 5 m. The data used here were acquired between 20 and 27 March, 2014.
First we have a brief look at the overall statistics of the
EM31 data. The measurement track and area are shown
in Fig. 5. The main measurement area is roughly 110 by
240 km. Note that most of the area had been free of sea
ice some weeks before the cruise, so most of the ice during the cruise was new thin ice. The ice thickness data
contain around 12.4% of negative thickness values which
were excluded from further analysis. This leaves in total
1.20×106 ice thickness data points. Mean ice thickness
is only 25 cm, cumulative 10th and 90th percentiles are 2
and 64 cm, and 75% and 87% of the ice thickness values
are below 30 and 50 cm, respectively. Thus, the EM31
data are in principle well suited for validation of SMOS
and VIIRS based thin ice charts. The problem in the comparison lies in the very different sampling pattern: 5 m
footprints every 0.5 second (EM31) versus daily retrieval
in grid cells of between about 1 km and 12.5 km. In addition, the time span of the cruise data is only 8 days.
3.1. Multisensor ITC
FMI’s multisensor ITC for March 2014 was calculated
from RADARSAT-2 ScanSAR imagery. We have compared EM31 data acquired during the previous 48 hours
to the ITC of the respective day.
EM31 ice thickness statistics for a 1 km FMI ITC pixel
were calculated from EM31 data points which were
within a 500 m radius from the ITC pixel centre. If there
were at least 100 data points then various statistical parameters were calculated, e.g., mean, standard deviation,
modal value from a histogram with 5 cm bin width, percentage of thickness values below 30 cm (thin ice class).
The number of EM31 data points varied from 101 to
2137, and the average was 576. All ITC pixel values
vs. EM31 statistics from the eight ITCs were collected
together for further analysis.
The total number of 1 km comparison pixels is 2758
Table 1. Comparison of ice thickness estimates from thermal VIIRS imagery with EM31 measurements (RMSD:
root mean square difference).
RMSD Bias
Estimate 0-30 cm
16.5
-0.05
Estimate 30-50 cm
28.8
25.2
19.2
4.78
Total
(SIC> 70%). Nearly all ITC pixel values, in total 2716,
represent the thin ice class (< 30 cm). Thus, we can only
study how accurate is thin ice identification in the ITC.
Various EM31 data statistics suggest that thin ice class
was identified in the FMI ITC with good accuracy. For
example, only in around 10% of the pixels both mean and
modal EM31 ice thickness was over 30 cm, but FMI ITC
showed thin ice class. Mismatch cases may be partly due
to ice drift and lower SICs, 70-90%, when the measured
AMSR2 TB is heavily influenced by the open water TB .
3.2. VIIRS ITC
As the EM31 measures the combined thickness of snow
and ice, while the VIIRS ITC provides only the ice thickness, an estimated snow depth was added to the SIT before the analysis. For each of the ITC grid cells in the
region, all the EM31 data points within a 375 m radius
were assigned to the grid cell. For each of these grid
cells, the mean and standard deviation of the measured
ice thickness was calculated. Grid cells with fewer than
100 EM31 data points were excluded from the analysis.
Based on the mean value of the EM31 measurements and
the ITC estimate in the corresponding grid cell, the root
mean square difference (RMSD) and bias (mean difference) were calculated. Furthermore, the samples were
split in two groups, with estimated ice thickness in the
range of 0-30cm and 30-50cm. The RMSD and bias were
found for each of the groups. The results are shown in Table 1.
The uncertainty of the VIIRS ITC estimates is still relatively high, in particularly for the thicker ice (30-50 cm)
where there is a strong bias. This indicates that the ice
chart currently is most useful for thinner ice. However,
there is also some distribution of EM31 ice thickness
measurements within each ITC grid cell. We have calculated the fraction of ice thickness estimates that were
within 1 or 2 standard deviations from the mean EM31
ice thickness estimate within the grid cell. 34% of estimates were found to be within one standard deviation,
and 58% were found to be within two standard deviations
of the mean.
3.3. SMOS ITC
Following a similar approach, for each of the SMOS ITC
grid cells, EM31 measurements within a radius of 50 km
from the cell centre were averaged, which corresponds to
the resolution (footprint size) of SMOS. For SIT below
30 cm, the RMSD is 14 cm and the bias is −2 cm. As
there was only one sample with thicker ice of 30-50 cm
(the SMOS ITC grid is much coarser than the VIIRS
SIT grid), no comparison was possible. So for thin ice
(<30 cm), the retrieval agrees reasonably well.
3.4. Modelled ITC
Model output exists with a frequency of one hour and a
spatial resolution of 10 km in each direction. Therefore
all ship measurements are binned into the nearest grid cell
and time. Based on this the average and the standard deviation for each bin has been calculated. The average
values have then been compared with the modelled point
values. As the model does not distinguish between different ice thicknesses, no differentiation has been made
between thin and thick ice categories.
The result is an RMSD of 44.6 cm and a bias of 30.9 cm,
where the model overestimates the ice thickness. When
looking at the temporal variation of the model it has large
daily variations which indicates that the ship has been
close to the edge of the thick ice in the model, and that
the thin/thick ice edge potentially has been moved back
and forth. Fig. 6 (below) yields the same picture, where
the model seems to overestimate the ice thickness in the
marginal ice zone near Svalbard.
4.
INTERCOMPARISON OF PRODUCTS
The main challenges in the intercomparison of the different ITC products are: (1) different spatial grid cell sizes,
about 1 km on the one hand (multisensor and VIIRSbased ITC), but about 10 km on the other hand (SMOSbased and model-based ITC); (2) the different ice thickness sensitivity ranges: one thin ice class (< 30 cm) and
then SIT from 30 cm to 255 cm for the multisensor ITC,
but SIT from 0 to 50 cm and a thick ice class (> 50 cm)
for SMOS and VIIRS-based ITCs.
Fig. 6 shows examples of the four POLAR ICE ITCs, all
resampled to the same polar stereographic grid with grid
cell size of 12.5 km. Resampling the 1 km resolution
multisensor ITC required particular care because of the
thin ice class – the procedure is described in Section 4.1.
Note the following: (1) in the SMOS and VIIRS ITC,
ice thicker than 50 cm is shown uniformly as dark red
(1 m SIT); (2) in the multisensor ITC, the thin ice class
(SIT< 30 cm) is represented by middle blue (15 cm SIT);
missing data areas (no satellite data, or cloud cover) are
white.
The best qualitative agreement is between SMOS (top
left) and multisensor (bottom left) ITCs, see details in
Section 4.1. The qualitative agreement between the
SMOS ITC and the model-based ITC (bottom right)
seems to be quite good, but the latter generally shows
thicker ice. There are several possible reasons which are
discussed in Section 4.2.
The ITC based on VIIRS data has a lot of gaps from cloud
masking as just the data for one day are taken. While
the ice edge also roughly agrees with the other products,
there seems to be a strong positive bias for thicker ice
here as well (East of Svalbard: dark red).
4.1. Multisensor ITC vs. SMOS ITC
The time period for this ITC comparison is from 15 October 2015 to 31 March 2016. FMI multisensor ITC
(MITC) was processed to the polar stereographic grid
Figure 6. Four different ITC for March 24, 2014, resampled to a common grid, top left: SMOS-based, top right: VIIRSbased, bottom left: multisensor-based, bottom right: model-based.
and pixel size (12.5 km) of the UB SMOS ITC (SITC)
in the following way: (1) MITC (1 km pixels) in its original polar stereographic grid was divided into 13×13 km
blocks; (2) blocks which have more than 75% of pixels
with ice information (thin ice flag or thickness in range
of 31–255 cm) or a SIC class assigned (0-10%, 10-40%,
40-70%) were identified; and then either (3) the weighted
average of ice thickness for a block was calculated as:
Hiw =
Ny y Nt
H +
mean(hti )
N i
N
(1)
where N is the total number of pixels with ice information in a 13×13 km block (min.: 126 pixels), Ny is the
number of thin ice pixels (< 30 cm), Nt is the number of
thick ice pixels (31–255 cm), Hiy is the thickness value
chosen for the thin ice class (here:15 cm), and mean(hti )
is the average thickness of the thick ice pixels; or (4) a
SIC class with most pixels is assigned to the block. Resulting values below 30 cm were marked as 15 cm. Finally, (5) MITC in 13 km blocks was rectified to the
SITC polar stereographic grid using nearest neighbour
sampling.
The MITC vs. SITC comparison was conducted using
statistics from the daily ITC pairs and from co-incident
pixel values collected from all ITC pairs.
4.1.1. All co-incident ice thickness data
The number of co-incident ice thickness samples is
around 625000 and those for the MITC SIC class vs.
SITC thickness over 1.7×106 . Unfortunately, the comparison of the MITC and SITC thickness data is not
straightforward due to the thin ice class in the former and
the thick ice class in the latter. Thus, it is not possible to
calculate overall mean bias and RMSD between the two
ITCs. Therefore, we have investigated statistics of MITC
data in different SITC ice thickness bins: 10 cm wide bins
within 0-30 cm thickness range, a bin from 31 to 50 cm,
and the SITC thick ice class. In each SITC thickness bin
various statistics were calculated for the MITC data, e.g.,
the relation of thin ice class (< 30 cm) pixels to total
number of pixels, and mean and mode of the MITC thick
ice data (31–255 cm).
Thin ice areas in the SITC and MITC match very well:
in the 0-10 cm SITC thickness bin nearly 100% of the
MITC pixels also show thin ice. In the 21–30 cm bin thin
ice percentage is still quite high, 83.3%, and the other
statistics show that most of the MITC thick ice values are
quite close to the 30 cm thin ice upper limit.
In the SITC thick ice class the match is poorer: around
21% of the MITC pixels show thin ice class and only 55%
of the pixels show ice thickness larger than 51 cm. In the
intermediate SITC thickness bin, 31–50 cm, only 26%
of the MITC pixels are in the same thickness range, and
half of the MITC pixels show thin ice class, but the modal
thickness is within the SITC bin.
Finally, we have investigated distribution of SITC ice
thickness values when the MITC has a WMO SIC class
(0–10%, 10–40%, 40–70%) assigned. SITC thickness
values should be small due to the large effect of low
open water TB on the measured total TB . In the smallest
WMO SIC class, 0–10%, 93% of SITC thickness values
are 0 cm. In the next SIC class, 10–40%, only 9% of the
SITC values are 0 cm, but 90% of them are below 8 cm.
In the last SIC class, 40–70%, corresponding figures are
1% and 12 cm. The percentage of thickness values over
30 cm is only 0.1% in all SIC classes combined. Thus,
when MITC shows low WMO SIC classes then SITC has
open water or thin ice pixels.
4.1.2.
Thin ice vs. thick ice classification in daily ITC
pairs
In the daily chart pairs, we have studied the match between thin and thick ice areas: type I difference – SITC
shows thick ice class (> 50 cm) but MITC show thin
ice class, and type II difference – SITC show thin ice
(< 30 cm) but MITC show thick ice (> 50 cm). The average percentage (average is over all the days with chart
pairs) for the type I difference is 32.5% (std. dev. is 27%),
and for the type II difference it is only 2.5% (std. dev. is
3%). The distribution for the type I difference is very
wide, percentages cover the full range from 0 to 100%
and one quarter of them are over 50%. The results show
good match in the detection of thin ice, but for thick ice
detection the match is rather poor.
4.1.3.
Conclusions
The SITC vs. MITC comparison showed that the charts
match very well on the average in detection of thin ice
(<30 cm), but in thick ice detection (>50 cm) the match is
rather poor. The comparison results do not indicate which
MITC is generally better. However, results here and from
other studies conducted at FMI, including visual analysis
of MITCs and comparison of the AMSR2 thin ice chart
with MODIS ice thickness charts (processed at FMI; [4]),
suggests that improvement in the AMSR2 thin ice detection algorithm is needed; too often AMSR2 detects thin
ice over areas of thick ice likely due to sensitivity of the
36 and 89 GHz TB data to various snow and ice properties (measured thin ice signatures resemble those of thick
ice). In the SITC thin vs. thick ice detection should be
more reliable due to L-band TB data which are much
less sensitive in this regard, but at the poorer resolution
of SMOS (larger spatial averaging of various ice type
TB signatures) compared to AMSR2 likely leads to some
pixel class differences.
Figure 7. Time series of SMOS SIT (black) and model SIT
(blue) at 78◦ N, 64◦ E, Oct 2015–Mar 2016.
4.2. SMOS ITC vs. model-based ITC
Fig. 7 shows the time series of SMOS SIT (black) and
model SIT (blue), at 78◦ N, 64◦ E, for the period October
2015 to March 2016. There is concurrent SIT increase
and decrease, but, as in the ITCs in Fig. 6, the modelled
SIT is mostly larger than the SMOS-based SIT (exception: about 10 to 30 of January). One reason for this
could be the assimilation scheme of the model that adds
ice to the categories where there is ice. If the model underestimates the sea ice concentration by removing too
much of the thin sea ice, new thick ice will be added according to the current distribution of sea ice between categories. This has also been reported by [17]. Secondly
the SMOS SIT is estimated for a full resolution cell. This
means that when the ice concentration is less than 100%,
the ice thickness will be underestimated. Because of the
non-linear SMOS retrieval function, a small open-water
fraction in one resolution cell reduces the ice thickness
of that cell disproportionately much. Therefore, even a
small fraction of leads causes a bias. At last the thermodynamics is influenced by the snow on the sea ice which
insulates the ice from the atmosphere. The snow in the
ocean/sea ice model is prescribed by the precipitation in
the atmospheric forcing when the air temperatures is less
than zero. The quality of this field (i.e., snow) is hard to
estimate.
5.
CONCLUSION AND OUTLOOK
POLAR ICE has distributed operationally in 2014-2016
four different SIT products for the Arctic. Here, we have
given an overview of the SIT products, and studied their
accuracies and deficiencies.
When comparing the four different SIT products with
measured thin ice thicknesses and with each other, the
first difficulty were the different sampling patterns and
grid resolutions of the products which made resampling
and interpolation necessary. Likewise, different SIT retrieval ranges make a comparison difficult. This shows
how the different data products are complementary to
each other.
Overall, the four ITCs agreed qualitatively; the thin ice
class of the multisensor ITC agreed well with SMOS SIT
below 30 cm. However, quantitatively, there were differences: (1) The VIIRS-based ITCs seem to overestimate
ice that is thicker than about 30 cm. (2) The modelled
ITCs showed generally thicker ice than the SMOS ITCs
and multisensor ITCs, but an intercomparison like this
one cannot tell if this is an overestimation by the former
or an underestimation by the latter methods. This Intercomparison might, however, indicate possible improvements, for example better thin ice detection for the multisensor ITC by using SMOS ITC. Finally, the role of the
snow layer on sea ice – which is often poorly known – in
the different retrieval schemes and models may need further investigation, which also implies further comparison
with available SIT in-situ measurements.
ACKNOWLEDGMENTS
The work described in this paper was supported by
the European Union’s FP7 project, POLAR ICE. The
SMOSice/IRO-2 measurement campaign was supported
by ESA, German Ministry of Economic Affairs and Energy (BMWi), Alfred Wegener Institute and University of
Hamburg.
REFERENCES
1. N. Walker, A. Fleming, A. Cziferszky, L. Toudal
Pedersen, T. Rasmussen, M. Mäkynen, R. Berglund,
L. Seitsonen, R. Solberg Ø. Rudjord, H. Tangen,
L. Axell, R. Saldo, H.E. Larsen C. Melsheimer,
T. Puestow, D. Arthurs, and D. Flach. POLAR
ICE: integrating, distributing and visualising ice information products for operators in polar waters. In
Proceedings of the Living Planet Symposium 2016,
Prague, Czech Republic, 9-13 May 2016. ESA, 2016.
2. M. Huntemann, G. Heygster, L. Kaleschke,
T. Krumpen, M. Mäkynen, , and M. Drusch.
Empirical sea ice thickness retrieval during the
freeze up period from SMOS high incident angle
observations. The Cryosph., 8:439–451, 2014. doi:
10.5194/tc-8-439-2014.
URL http://www.
the-cryosphere.net/8/439/2014/.
3. C. Haas, J. Lobach, S. Hendricks, L. Rabenstein, and
A. Pfaffling. Helicopter-borne measurements of sea
ice thickness, using a small and lightweight digital
EM system. J. Appl. Geophys., 67:234–241, 2009.
doi: 10.1016/j.jappgeo.2008.05.005.
4. M. Mäkynen, B. Cheng, and M. Similä. On the accuracy of the thin ice thickness retrieval using MODIS
thermal imagery over the Arctic first year ice. Ann.
Glaciol., 54(62), 2013.
5. Ø. Rudjord, Ø. Due Trier, M. Zortea, R. Solberg,
G. Spreen, S. Gerland, A. Renner, and N. Hughes.
Thin ice thickness from MODIS: Improvement of algorithm and evaluation of product. Technical Report NR Note SAMBA/21/14, Norwegian Computing Center, 2014.
6. Y. Yu and D. A. Rothrock. Thin ice thickness from
satellite imagery. J. Geophys. Res., 101(C10):25753,
1996.
7. J. Key, J. Collins, C. Fowler, and R. Stone. Highlatitude surface temperature estimates from thermal
satellite data. Remote Sens. Environ., 61:302–309,
1997.
8. M. Similä, M. Mäkynen, B. Cheng, and E. Rinne.
Multisensor data and thermodynamic sea-ice model
based sea-ice thickness chart with application to the
Kara Sea, Arctic Russia. Ann. Glaciol., 54(62):241–
252, 2013.
9. P. Sakov, F. Counillon, L. Bertino, K. A. Lisæter,
P. R. Oke, and A. Korablev. TOPAZ4: an ocean-sea
ice data assimilation system for the North Atlantic
and Arctic. Ocean Sci., 8:633–656, 2012.
10. M. Mäkynen and M. Similä. Thin ice detection in
the Barents and Kara Seas with AMSR-E and SSMIS
radiometer data. IEEE Trans. Geosci. Remote Sens.,
53(9):5036–5053, 2015.
11. E. P. Chassignet, H. E. Hurlburt, O. M. Smedstad, G. R. Halliwell, P. J. Hogan, A. W. Wallcraft,
R. Baraille, and R. Bleck. The HYCOM (HYbrid
Coordinate Ocean Model) data assimilative system.
J. Marine Syst., 65:60–83, 2007. doi: 10.1016/j.
jmarsys.2005.09.016.
12. E.C. Hunke. Viscous-plastic sea ice dynamics with
the EVP model: linearization issues. J. Comp. Phys.,
170(1):18–38, 2001. doi: 10.1006/jcph.2001.6710.
13. E.C. Hunke and J.K. Dukowicz. An elastic-viscousplastic model for sea ice dynamics.
J. Phys.
Oceanogr., 27(9):1849–1867, 1997. doi: 10.1175/
1520-0485(1997)027<1849:AEVPMF>2.0.CO;2.
14. W. B. Hibler. A dynamic thermodynamic sea ice
model. J. Phys. Oceanogr., 9:817–846, 1979.
15. Madsen K. S, Rasmussen T. A. S., Ribergaard M.
H., and Ringgaard I. M. High resolution sea ice
modelling and validation of the Arctic with focus on
south Greenland waters. Polarforschung, 2015. in
press.
16. L. Kaleschke, X. Tian-Kunze, N. Maaß, A. Beitsch,
A. Wernecke, M. Miernecki, G. Müller, B.H.
Fock, A. Gierisch, H. Schlünzen, T. Pohlmann,
M. Dobrynina, S. Hendricks, J. Asseng, R. Gerdes,
P. Jochmann, N. Reimer, J. Holfort, C. Melsheimer,
G. Heygster, G. Spreen, S. Gerland, J. King, N. Skou,
S. Schmidl Søbjærg, and C. Haas. SMOS sea ice
product: Operational application and validation in
the Barents Sea marginal ice zone. Remote Sens.
Environ., 2016. doi: 10.1016/j.rse.2016.03.009. in
press.
17. G.C. Smith, F. Roy, M. Reszka, D. Surcel Colan,
Z. He, D. Deacu, J. Belanger, S. Skachko, Y. Liu,
F. Dupont, J. Lemieux, C. Beaudoin, B. Tranchant, M. Drévillon, G. Garric, C. Testut, J. Lellouche, P. Pellerin, H. Ritchie, Y. Lu, F. Davidson,
M. Buehner, A. Caya, and M. Lajoie. Sea ice forecast verification in the canadian global ice ocean prediction system. Q. J. Roy. Meteor. Soc., 142(695):
659–671, 2016. doi: 10.1002/qj.2555.