Accurately forecasting the sea-ice thickness (SIT) in the Arctic is a major challenge. The new SIT product (referred to as CS2SMOS) merges measurements from the CryoSat-2 and SMOS satellites on a weekly basis during the winter. The impact of assimilating CS2SMOS data is tested for the TOPAZ4 system – the Arctic component of the Copernicus Marine Environment Monitoring Services (CMEMS). TOPAZ4 currently assimilates a large set of ocean and sea-ice observations with the Deterministic Ensemble Kalman Filter (DEnKF).
Two parallel reanalyses are conducted without (Official run) and with (Test run) assimilation of CS2SMOS data from 19 March 2014 to 31 March 2015. Since only mapping errors were provided in the CS2SMOS observation, an arbitrary term was added to compensate for the missing errors, but was found a posteriori too large. The SIT bias (too thin) is reduced from 16 to 5 cm and the standard errors decrease from 53 to 38 cm (by 28 %) when compared to the assimilated SIT. When compared to independent SIT observations, the error reduction is 24 % against the ice mass balance (IMB) buoy 2013F and by 12.5 % against SIT data from the IceBridge campaigns. The improvement of sea-ice volume persists through the summer months in the absence of CS2SMOS data. Comparisons to sea-ice drift from the satellites show that dynamical adjustments reduce the drift errors around the North Pole by about 8 %–9 % in December 2014 and February 2015. Finally, using the degrees of freedom for signal (DFS), we find that CS2SMOS makes the prime source of information in the central Arctic and in the Kara Sea. We therefore recommend the assimilation of C2SMOS for Arctic reanalyses in order to improve the ice thickness and the ice drift.
Sea ice plays an important role in the Arctic climate system because it prevents the rapid exchange of heat flux between the ocean and atmosphere. A decline and a thinning of the sea-ice cover has occurred in the past decades (e.g. Johannessen et al., 1999; Comiso et al., 2008; Stroeve et al., 2012) as well as an increase of deformation rates and drift speed (Rampal et al., 2009). It is expected that these changes will have significant impacts on the Arctic Ocean Circulation (e.g. Levermann et al., 2007; Budikova, 2009; Kinnard et al., 2011) and on the future human living environment (Schofield et al., 2011; Bathiany et al., 2016). The interpretation of such changes is severely hampered by the sparseness of observations, therefore the reanalyses that can provide continuous spatiotemporal reconstructions by assimilating existing observations into dynamical models have become increasingly popular tools. In addition, recent studies (Day et al., 2014; Guemas et al., 2014; Melia et al., 2015) have shown that sea-ice thickness (SIT) anomalies play an important role for the Arctic predictability up to seasonal time scale.
Satellite observations of sea-ice concentration (SIC) have been available since 1979 and have allowed an accurate monitoring of sea-ice extent (SIE) during that period. Data assimilation of SIC has constrained the position of the sea-ice edge (Lisæter et al., 2003; Stark et al., 2008; Posey et al., 2015), but large disagreements (e.g. Uotila et al., 2018) remain in the estimation of sea-ice volume because observations of SIT are very incomplete.
Until the 1990s, the only SIT measurements were sparse in situ measurements and submarine data. With new satellites, continuous estimates of SIT on basin scale have been achieved using satellite radar and laser altimeters: European Remote Sensing (ERS), Envisat and the NASA Ice, Cloud and land Elevation Satellite (ICESat). These were used to document the rapid thinning of sea-ice in the Arctic (Laxon et al., 2003; Kwok and Rothrock, 2009).
CryoSat-2, launched in April 2010, has been the first satellite dedicated to
measure with high accuracy the sea-ice freeboard, from which SIT can be
derived (Ricker et al., 2014; Tilling et al., 2016). However, the resulting
SIT estimates are still very uncertain because of uncertainties in the snow
depth (using climatology), snow penetration and sea-ice density (Kern et al.,
2015; Khvorostovsky and Rampal, 2016). Those uncertainties are large for thin
ice (
In this study, the CS2SMOS will be assimilated into the TOPAZ4 forecast
system, which is a coupled ocean–sea-ice data assimilation system using the
Deterministic Ensemble Kalman Filter (DEnKF; Sakov and Oke, 2008). The
Ensemble Kalman Filter has previously been demonstrated for assimilation of
SIT data (Lisæter et al., 2007) of freeboard data (Mathiot et al., 2012)
and of the CS2SMOS data (Mu et al., 2018) as well. TOPAZ4 is the Arctic
Marine Forecasting system in the Copernicus Marine Environment Monitoring
Services (CMEMS,
Section 2 describes the TOPAZ4 system: namely the coupled ocean and sea-ice model, the implementation of the EnKF and the observations used for data assimilation and validation. In Sect. 3, we carry out an observing system experiment (OSE) comparing the two reanalyses: one using the standard observation types used in operational setting and another assimilating the CS2SMOS in addition. Then the performance of the two runs is presented against both assimilated and non-assimilated measurements. Section 4 presents the impacts of assimilating the CS2SMOS on sea-ice drift and the integrated quantities for sea ice, and measures its relative impact compared to other assimilated observations. A summary is provided in the last section.
TOPAZ4 is a forecasting ocean and sea-ice system developed for the Arctic, having been operational since the early 2000s (Bertino and Lisæter, 2008). It uses the Hybrid Coordinate Ocean model (HYCOM: version 2.2) developed at University of Miami, which has been successfully applied in global and regional oceans (Chassignet et al., 2003; Counillon and Bertino, 2009; Metzger et al., 2014; Xie et al., 2018). The model grid is constructed using conformal mapping (Bentsen et al., 1999) with a 12–16 km resolution shown in Fig. 1a. The model uses 28 hybrid layers with reference potential densities selected specifically for the North Atlantic and the Arctic regions (Sakov et al., 2012). The model is forced by atmospheric forcing from ERA-Interim. A barotropic inflow of Pacific water is imposed through the Bering Strait, which is balanced by an outgoing flow through the southern model boundary. It has an averaged transport of 0.8 Sv, and varies seasonally with a minimum (0.4 Sv) in January and a maximum (1.3 Sv) in June consistently with observations (Woodgate et al., 2005). The model accounts for river discharge, for which a seasonal climatology is estimated by feeding the run off from ERA-interim (Dee et al., 2011) into the Total Runoff Integrating Pathways model (TRIP, Oki and Sud, 1998) over the period 1989–2009.
A simple sea-ice model using a one-thickness category has been coupled to HYCOM. The sea ice and the ocean are thus coupled every 3 h and exchange momentum, salt and heat on the ocean model's Arakawa C-grid. The sea-ice thermodynamics treat precipitations on ice as snow whenever surface air temperature is below zero (Drange and Simonsen, 1996). The ice dynamics uses the elastic–viscous–plastic rheology (Hunke and Dukowicz, 1997) with the modification suggested by Bouillon et al. (2013). There is a 0.1 m limit in the model for the minimum thickness of both new ice and melting ice.
The TOPAZ4 system uses a deterministic Ensemble Kalman Filter (DEnKF, Sakov
and Oke, 2008), which solves the analysis without the need to perturb the
observations and is therefore a square-root filter implementation of the
EnKF. In the DEnKF, if the model state is represented by
The practical implementation of the model and its perturbations follow Sakov et al. (2012): the model errors include joint perturbations of winds and heat fluxes as originally recommended by Lisæter et al. (2007). The precipitation perturbation has however been increased from 30 % to 100 %, following a log-normal probability distribution of errors (Finck et al., 2013), which also increased the spread of ice thickness.
The following observations are assimilated sequentially every week in the
TOPAZ4 system (Xie et al., 2017): along-track sea level anomaly; in situ
profiles of temperature and salinity; gridded Operational Sea Surface
Temperature and Sea Ice Analysis (OSTIA) SST; Ocean and Sea Ice Satellite
Application Facility (OSI-SAF) sea-ice concentration and sea-ice drift from
satellite observation (Lavergne et al., 2010). All measurements are retrieved
from CMEMS,
The weekly SITs of CS2SMOS were retrieved from
In order to estimate the representation error for the SIT observation, we
have performed a preliminary sensitivity assimilation experiment for
November 2014. We used the diagnostics by Desroziers et al. (2005) as an
indicative lower limit for the observation error in the TOPAZ4 system based
on the misfits to the CS2SMOS data. Desroziers et al. (2005) estimate the
optimal observation error as the following matrix:
Observation error uncertainties as a function of sea-ice thickness for the original CS2SMOS data set (black line), the estimated observation error using the Desroziers diagnostics with red-triangle line (see Eq. 3) and the one used in the TOPAZ Test run with blue-square, with an additional error term as Eq. (4) to the original uncertainty.
A parallel OSE is conducted from 19 March 2014 until end of March 2015. The two assimilation runs cover two special time periods: the onset of ice melting in March–April 2014 following by a data period free of CS2SMOS, then a whole cold season from October 2014 to March 2015. The control run named the Official run uses the standard observational network in the TOPAZ4 system (Xie et al., 2017), which assimilates on a weekly cycle the SLA, SST, in situ profiles of temperature and salinity, SIC and sea-ice drift (SID) data. Another assimilation run named the Test run includes as well the SIT from CS2SMOS.
We discard the SIT closer than 30 km from the coast to account for
differences of coastlines between the model and observations. The innovation
of SIT in Eq. (1) is calculated in terms of sea-ice volume:
In the following, we investigate the misfits of the forecasted model states
by evaluating the bias and the RMSD:
Three types of independent SIT observations are used for validation. First,
the drifting ice mass balance buoys (IMB:
The first assimilation time is the 19 March 2014 and the last is the 25 March 2015. The monthly SITs from the two OSE runs are compared to CS2SMOS in Fig. 3. The SITs in April 2014 are presented for comparison in the upper row of Fig. 3. In the Official run, the thick sea ice to the north of the CAA is underestimated but thickens slightly in the Test run: the 3 m SIT isoline covers a wider area, in better agreement with the observations. The areas of thinner sea ice north of the Barents Sea, west of the Kara Sea, and the coast of the Beaufort Sea, which were too thick in the Official run, have all been improved also shown by reduced area delimited by the isolines of 1 or 2 m SIT in the Test run.
Monthly SIT from CS2SMOS, column
After summer of 2014, measurements of SIT from CS2SMOS restart at the end of
October. Results are presented for November 2014 in Fig. 3: the thick sea ice
in the Central Arctic has been further improved in the Test run. The thickest
sea ice (
In the last month of the experimental period (March 2015), the thick sea-ice pattern in the Test run, shown as the 2 m isoline, is more similar to CS2SMOS. The maximal SIT within the 4 m isoline is located north of the CAA in the Test run and in CS2SMOS, while in the Official run it spreads further out from the northern coast of Canada to north of Greenland. In addition, the SIT north of the Fram Strait is thicker than in the Official run. The SIT is similarly improved near the coast of the Beaufort Sea and to the northwest of Svalbard. As expected with data assimilation, the Test run agrees clearly better with the assimilated product. Those improvements are largest in the ice pack and in the marginal seas, where the model deviates considerably from the CS2SMOS SITs. On the contrary, the thickness near the sea-ice edge is not strongly impacted by the assimilation.
The above results are confirmed quantitatively by comparing misfits of weekly SIT from the two runs with the corresponding CS2SMOS observations. Time series of bias and RMSD calculated as in Eqs. (6) and (7) are shown in Fig. 4a. In the beginning of the period, the SIT RMSD in the Test run decreases quickly from 0.6 to 0.4 m before the observations are interrupted for the summer. The biases are reduced equally in both runs. After the observations resume in the end of October 2014, the SIT RMSD is comparable between the two runs but the bias is slightly lower in the Test run. There is large spike in the bias and RMSD for both systems that relates to an inaccuracy of the CS2SMOS observations (see Sect. 4.2). After the spike, the RMSD and bias in the Test run are lower than in the Official run. The bias in the Test run converges to 0 and fluctuates around that level but this is probably not due to the assimilation since the bias in the Official run also converges to 0 during that time. This is rather due to the compensation of seasonal and regional errors. On average, the SIT bias (too thin) is decreased from 15 to 5 cm by the assimilation of CS2SMOS. The RMSD of SIT is 38 cm in the Test run, which corresponds to a reduction of 28.3 % relative to the error in the Official run.
The innovation statistics taken at each assimilation time are used to
evaluate how well our data assimilation system is calibrated. In the
reliability budget of Rodwell et al. (2016), the total uncertainty of an
ensemble data assimilation system is calculated as follows:
The innovation statistics for SIC are mostly identical in the two runs (not
shown), the mean misfits for SIC vary around
Four IMB buoys are available as independent validation of the impact of the assimilation of CS2SMOS. The buoys are drifting in the Canadian Basin (Fig. 1a), and only one buoy (2013F) lasted during the whole experimental time period shown (Fig. 5a). This buoy exhibits the seasonal variability of SIT: it reaches 1.5 m in spring 2014, decreases down to 1.0 m in September and rises again to 2 m in March 2015. The seasonal SIT cycle of the Official run shows excessive seasonal variability, with a thin bias in summer 2014 and a thick bias during the two winters. In the Test run (shown as the red-dashed line) the seasonal cycle is dampened and more consistent with the observations. The bias is still quite large around March–April and remains so even at the end of the study period. It should be noted that the impact of CS2SMOS seems largest in summer, when no observations are assimilated. This illustrates the persistent effects of winter SIT improving the predictability of the summer Arctic sea ice as shown in Mathiot et al. (2012). When CS2SMOS is assimilated again in the fall 2014, the Test run initially slightly overestimates the SIT measured at the buoy compared to the Official run but is slowly improving as the data is assimilated. The time-averaged SIT RMSD for buoy 2013F is reduced from 0.33 m in the Official run down to 0.25 m in the Test run, a reduction by 24.2 %.
Two other buoys (2014B and 2014C) cover the early months of the experimental period. The two runs are initially biased with a too thick SIT by 0.5 and 0.2 m compared to 2014B and 2014C. At buoy 2014B, there is a slight error reduction during the assimilation period that continues beyond the assimilation window, similarly to buoy 2013F. At buoy 2014C however, although the error is reduced during the analysis period, the two assimilation runs converge during the summer. At these three buoys the assimilation corrects the mean SIT values and the amplitude of the seasonal cycle but has little influence on the phase of the seasonal cycle.
The buoy 2014F covers the last 6 months of the experimental period. At that buoy, the assimilation seems increase the errors. It should be noted however that the constant SIT at buoy 2014F seems unlikely or not representative of the area.
In order to convert the sea-ice draft measured by ULS from the BGEP buoys to
SIT, we used the balance equation as in Tilling et al. (2018):
Time series of SIT along the trajectories of IMB buoys (
The SIT time series of the measurement and of the two runs are shown in Fig. 6, from October 2014 onwards. The gray error bars depict the daily standard deviation. The data indicates an increasing SIT from around 0.5 m in October 2014 to nearly 2 m in March 2015. The observed SIT at mooring 14D shows a very large daily variability from end of October to November 2014, especially compared with that of moorings 14A and 14B.
The weekly SITs from CS2SMOS match well the data with RMSDs of 15, 19 and 39 cm during the 6 months, which is lower than in the two model runs. Still, the SIT from CS2SMOS overestimates SIT from October 2014 to mid-January 2015 compared to the mooring 14B, and between in October and November of 2014 for mooring 14A. The SITs in the Official run are overestimated in all three locations. The SIT RMSDs are 41, 23 and 51 cm, respectively, compared to SIT measurement from the three moorings. The SITs in the Test run are closer to observations, thanks to the data assimilation of the SIT from CS2SMOS. The SIT RMSDs in the Test run are 25, 33 and 36 cm for moorings 14A, B and D, respectively. The error is reduced for moorings 14A and 14D compared to the Official run but increases for mooring 14B, mostly due to the initial mismatch between CS2SMOS and the mooring. Similarly to the comparison with IMB buoys, moorings suggests that error of SIT in the Beaufort Sea is reduced by assimilation of CS2SMOS.
Daily series of SIT (black line) at the BGEP mooring (14A, 14B, and 14D) compared with the two model runs – Official (blue line) and Test (red line) – and the weekly observed by CS2SMOS (green line). The black line represents the daily average at the mooring location with the standard deviation shown as the error bar. The RMSDs of the Official run, Test run and CS2SMOS are respectively indicated on the bottom of each panels.
Another independent observation of SIT with better spatial coverage is the
SIT Quick Look data from airborne instruments during NASA's Operation
IceBridge campaign (Kurtz et al., 2013). Those are available via the National
Snow and Ice Data Center (NSIDC), albeit for the months of March and April
only. Note that the airborne SITs have been reported to be slightly
low-biased by about 5 cm compared to in situ measurements (King et al.,
2015). Figure 7a shows all observed SITs (upper-left panel) from IceBridge,
collected during March and April of 2014–2015. All observed SITs are located
in the Canadian Basin and north of Greenland and cover most of the area where
sea ice is thicker than 3 m. Thicknesses between 1 and 3 m are measured in
the Beaufort Sea. The two simulated SITs in the two model runs show
systematic differences of SIT (see Fig. 7b): the Test run SIT has been
thinned in the Beaufort Sea and thickened near the North Pole. On average,
the SIT in the Test run is increased by 0.1 m and by 0.27 m north of
80
The comparisons of the two OSE runs to the IceBridge data are presented in the bottom panels of Fig. 7. The sea ice in the Official run is too thin at the north of the CAA and north of Greenland, with a deviation larger than 1.5 m. In the Beaufort Sea on the contrary, the model is too thick by 0.5 to 1 m. This bias is consistent with that reported in Xie et al. (2017), where the TOPAZ4 reanalysis (Official run) was compared to ICESat observations in the period 2003–2008. In the Test run, the biases are slightly reduced by SIT assimilation, mainly in the Beaufort Sea and north of Greenland, but the reduction is smaller than the remaining error. On average, the SIT RMSD is 1.05 m, which corresponds to a reduction of 12.5 % compared to that in the Official run.
The regression of the SIT observations from IceBridge to the two OSE runs is
shown in Fig. 8. The Test run shows improved linear correlations to the
observation. The offset at the origin is reduced (0.52 m instead of 0.93 m)
and the slope is closer to 1 than in the Official run. The linear correlation
in the Test run is slightly increased as indicated with the square
correlation
Scatterplots of SIT daily averaged of Official (blue) and Test (red)
runs compared to IceBridge data. The dashed lines are the respective linear
regression, the coefficient
The above results and assimilation diagnostics confirm that the SIT misfits can be controlled – to some degree – by assimilation of the CS2SMOS data, without visible degradation of other assimilated variables. To better understand the advantages and the limits of assimilating the merged SIT product, we further evaluate the impact of CS2SMOS in the assimilation system: first the repercussions on other sea-ice variables and integrated quantities, and then through a quantitative impact analysis of CS2SMOS relatively to other assimilated observation types.
The EnKF implemented in TOPAZ4 updates all the variables in the model state
vector using flow-dependent multivariate covariances from the ensemble
members (Eqs. 1 and 2). The direct assimilation update of ice drift is
however short-lived: the ice drift vectors quickly readjust to wind forcing
after assimilation, so the ice drift changes are mostly caused by dynamical
readjustments, related to the updated ice thickness and ice concentrations.
By the first order approximation of the two-dimensional momentum equation
(e.g. Hibler III, 1986; Hunke and Dukowicz, 1997), the drift velocity of sea
ice is mainly controlled by (1) the interactions of atmosphere–sea ice,
(2) the interactions of ocean–sea ice and (3) the internal sea-ice forces
which can be represented by the stress tensor
Following the EVP rheology in Hibler III (1979),
the stress tensor
The evaluation in Xie et al. (2017) shows the model drift of sea ice is
overestimated by 2 km d
Sea-ice drift misfits (model minus observation, in km per 2 days)
in the Official run
The RMSD of sea-ice drift speed in 2-day trajectories is reduced by about
0.1–0.2 km in April 2014 and February 2015 for the whole Arctic, which
corresponds to a reduction of less than 5 % of the RMSD. However, near the
North Pole (north of 80
To evaluate the potential impact of assimilating the SIT from CS2SMOS on the
sea-ice motion, we further utilize the data set from the IABP buoys which
began in 1990s to monitor ice motion throughout the Arctic Ocean. Only
trajectories longer than 30 days and reporting more than 5 times per day are
used to estimate the daily drift speed of sea ice. To avoid buoys in open
water, the observations are selected based on sea-ice concentration (
The speed distribution for daily drift of sea ice from IABP is shown by a
histogram in Fig. 10a. In the Central Arctic, the averaged drift speed is
about 10.6 km d
In Fig. 3, we show that the Arctic SIT has been improved everywhere, the assessment of the sea-ice drift is less conclusive but tends to suggest a slight improvement localized in the Central Arctic. However, improving the quantitative match with available observations does necessarily warrant the physical consistency of basin-scale integrated quantities. The impact of CS2SMOS on the Arctic-wide sea-ice extent (SIE) and the sea-ice volume (SIV) are investigated for the two runs and compared with the estimates from CS2SMOS and OSI-SAF, respectively. Due to differences of resolution and land mask (especially important in the Canadian Archipelago), we focus on the Central Arctic domain shown as the red line in Fig. 1b, excluding parts of the marginal seas.
Figure 11 shows the time evolutions of SIE and SIV in the two Official and
Test runs. Both are calculated by daily averages in the two model runs. The
SIE is classically calculated in the area where the SIC is not less than
15 % in the Central Arctic. The SIE shows the expected seasonal cycle with
the minimum (close to
SIE and SIV in the official run (blue) and the test run (red) in the Central Arctic. The black stars are the corresponding weekly SIE (or SIV) estimated from CS2SMOS. The green dashed line is the daily SIE from OSI-SAF. The averaged differences of the two runs (Official–Test) are reported. The vertical cyan dashes delimits the periods when C2SMOS data is assimilated.
The time evolutions of the SIV in the two runs show larger differences in the
lower panel of Fig. 11. The maximum in the Test run is close to
The value of the degrees of freedom for signal (DFS) is commonly used to
monitor the relative impact of different observations in a data assimilation
system (ref. Cardinali et al., 2004; Rodgers 2000; Xie et al., 2018), and is
calculated as follows:
Figures 13 and 14 show the IF
Relative DFS contributions (IF) of each observation data types in
November 2014.
Same as the above but for March 2015.
In March 2015, CS2SMOS has again a large impact in the Central Arctic
relative to other assimilated observations even though previous literature
indicates a lower impact in the midst of winter than when the ice is growing
(Mathiot et al., 2012). The relative IF of SIT indeed remains high even
though the absolute DFS is decreasing, due to the lower impact of other
assimilated observations, in particular SIC (Lisæter et al., 2003). On
average, the IF value of CS2SMOS is about 40 %. The high values (
CS2SMOS is the first product to monitor the complete pan-Arctic SIT in a systematic way, although only for the winter months. It is a combination of two very different, yet very advanced, technologies onboard the SMOS and CryoSat-2 satellites, calibrated against very few in situ observations of SIT, freeboard and snow depths. Altogether, the issue of measurements uncertainties is particularly delicate for the assimilation of CS2SMOS data. On the other hand, defining proper model background errors for SIT is just as delicate, when considering that the simulated SIT accumulates errors both in the sea-ice dynamics (in particular the rheological model) and in the thermodynamics. The Bayesian approach to confront these two uncertainties is by Monte Carlo propagation of uncertainties, which is what is practiced in the present study for the model background error, although not for the observation error.
This study assesses the impact of assimilating the new SIT product from 19 March 2014 to 31 March 2015. Compared to the assimilated SIT CS2SMOS, the thin bias is reduced from 15 to 5 cm, and the RMSD also decreased from 58 to 38 cm, a reduction by 28.3 %. Other innovation diagnostics show no degradation towards other assimilated variables – namely SIC, SSH, SST and TS profiles.
The SIT is also improved when compared to four independent drifting IMB buoys and three BGEP mooring buoys. The benefits persist throughout the summer although no SIT observations are available then, consistently with the experiments from Mathiot et al. (2012). This is important because it suggests that the model is not attracted to his bias solution. The assimilation reduces the low SIT biases north of the CAA and north of Greenland and the high bias in the Beaufort Sea compared to independent observations from Operation IceBridge. Both the thick pack ice in Central Arctic and the thin ice in marginal seas are corrected. On average, the SIT errors in March–April of 2014 and 2015 are reduced by 15 cm, a reduction by 12.5 % compared to the Official run.
The dynamical adjustment following the assimilation of SIT has partially
improved the sea-ice drift speeds in the Test run where the SIT has
thickened: the monthly averaged drift speed errors north of 80
In this study, the DFS information in the ensemble data assimilation system has been applied to quantitatively evaluate the relative contributions of all assimilated observation types. CS2SMOS has the highest impact near the northern coast of Canada, north of Greenland, and on the inner side of the sea-ice edge, where the contributions from CryoSat-2 and SMOS SIT were expected. The results, compared to assimilating SMOS only in Xie et al. (2016), show the importance of CryoSat-2, particularly in the winter months to constrain the SIT offsets (also shown by Mu et al., 2018, in a coupled MITgcm model system) and motivate the assimilation of CS2SMOS in the following reanalysis of TOPAZ4. However, the impact of SIT observations may vary with the evaluation of the modelling and observing system. Firstly, the SIC may have been underestimated in the Central Arctic due to the simplicity of the present sea-ice model. Further planned developments of TOPAZ include a new model rheology that is able to resolve the scaling laws of deformation of sea ice (Rampal et al., 2016) and should therefore improve the background errors of ice concentration in winter months and sea-ice drift, increase the impact of SIC and SID within the ice pack and reduce the estimated SIT impact accordingly. Other planned changes such as the simulation of melt ponds are not expected to influence these results directly since there are no melt ponds when the SIT data is available. Lastly, if a large number of in situ profiles were available below the sea ice, they would also compete with the SIT observations.
The above OSE results, like others, are necessarily contingent on adequate specifications of observation errors. Those are very much simplified in the case of CS2SMOS, which is not an uncommon case for remote sensing observations: due to the complexity of the physics involved, the specified observation errors are reflecting interpolation errors rather than a nonlinear propagation of errors from their sources (Ricker et al., 2017). In the present study, an offset has been added to account for this difference in Eq. (4), which results in a conservative error estimate with respect to the classical Desroziers optimality criterion and a suboptimal performance in the reliability budget analysis. In the one hand, reducing the observation would have accelerate the convergence to observed SIT and converge to a more accurate solution. On the other hand, this would have made the EnKF less robust to the sudden inconsistencies in the observations as seen in Fig. 11. Further versions of the CS2SMOS data will hopefully improve their temporal continuity and the impact of the data can be increased accordingly.
An alternative to using the scheme CS2SMOS data would have been to assimilate the two data sets CryoSat-2 and SMOS SIT separately and let the EnKF merge them together rather than relying on optimal interpolation, as successfully demonstrated by Mu et al. (2018). This would for instance avoid assimilating observations in places where they are the pure result of interpolation/extrapolation but would not resolve the offset between the two satellites, which is arguably the most worrying issue as of the present state of the SMOS and CryoSat-2 data. The assimilation of the separate data sets will be attempted in the future when their consistency is further improved.
The current TOPAZ reanalysis is currently reaching 2016 and extended by one year every year. The current study clearly shows the added value of assimilating SIT. In 2020, a new TOPAZ reanalysis will be provided with the upgraded version of TOPAZ5 which will include SIT assimilation from 2010 onwards.
The observations used for assimilation and validation are
available from the sources mentioned in the text. The Arctic reanalysis
after assimilation of CS2SMOS, the Test run, is available from CMEMS
(
JX designed the experiment, performed experiments and analyses. FC and LB contributed to the experimental design and interpretation. JX lead the writing phase, with FC and LB contributing to editing and review.
The authors declare that they have no conflict of interest.
Thanks to Johnny A. Johannessen for discussions and to Stefan Hendricks and
Robert Ricker for sharing the CS2SMOS data on