Since 19 October 2016, and in the framework of
Copernicus Marine Environment Monitoring Service (CMEMS), Mercator Ocean
has delivered real-time daily services (weekly analyses and daily 10-day
forecasts) with a new global
This paper describes the recent updates applied to the system and discusses the importance of fine tuning an ocean monitoring and forecasting system. It details more particularly the impact of the initialization, the correction of precipitation, the assimilation of climatological temperature and salinity in the deep ocean, the construction of the background error covariance and the adaptive tuning of observation error on increasing the realism of the analysis and forecasts.
The scientific assessment of the ocean estimations are illustrated with diagnostics over some particular years, assorted with time series over the time period 2007–2016. The overall impact of the integration of all updates on the product quality is also discussed, highlighting a gain in performance and reliability of the current global monitoring and forecasting system compared to its previous version.
Timeline of the Mercator Ocean global analysis and forecasting
systems for the various milestones (from V0 to V4) of the past MyOcean project
and for milestones V1, V2 and V3 of the current CMEMS. Real-time productions
are in yellow with the reference of the Mercator Ocean system. Available
Mercator Ocean simulations are in green, including the catchup to real time.
Global intermediate-resolution (high-resolution) systems at
Specifics of the Mercator Ocean IRG systems. In italic are the major upgrades with respect to the previous version. Available and operational production periods are described in Fig. 1.
Specifics of the Mercator Ocean HRG systems. In italic are the major upgrades with respect to the previous version. Available and operational production periods are described in Fig. 1.
Mercator Ocean monitoring and forecasting systems have been routinely operated in real time since early 2001. They have been regularly upgraded by increasing complexity, expanding the geographical coverage from regional to global, and improving models and assimilation schemes (Brasseur et al., 2006; Lellouche et al., 2013).
Mercator Ocean, which had primary responsibility for the global ocean
forecasts of the MyOcean and MyOcean2 projects since January 2009, developed
several versions of its monitoring and forecasting systems for the various
milestones (from V0 to V4) of the MyOcean project and, more recently, for
milestones V1, V2 and V3 of the Copernicus Marine Environment Monitoring
Service (CMEMS) as part of the European Earth observation program
Copernicus (
These systems are intensively used in four main areas of application: (i) maritime safety, (ii) marine resources management, (iii) coastal and marine
environment, and (iv) weather, climate and seasonal forecasting
(
In the system PSY4V3, the ocean–sea ice model and the assimilation scheme benefit from the following main updates: atmospheric forcing fields are corrected at large scale with satellite data; freshwater runoff from ice sheet melting is added to river runoffs; a time-varying global average steric effect is added to the model sea level; the last version of GOCE geoid observations are taken into account in the mean dynamic topography used for sea level anomaly assimilation; adaptive tuning is used on some of the observational errors; a dynamic height criteria is added to the quality control of the assimilated temperature and salinity vertical profiles; satellite sea ice concentrations are assimilated; and climatological temperature and salinity in the deep ocean are assimilated below 2000 m to prevent drifts in those very sparsely observed depths.
The impact of all these updates can be evaluated separately thanks to an incremental implementation taking advantage of Mercator Ocean's specific hierarchy of system configurations running with an identical setup. To this aim, short simulations (from 1 year to a few years) were performed by adding from one simulation to another one upgrade at a time using the IRG configuration or some high-resolution regional configuration.
The system PSY4V3 was run over the October 2006–October 2016 period to catch up to real time, assimilating the “reprocessed” observations (along-track altimeter, satellite sea surface temperature, sea ice concentration, and in situ temperature and salinity vertical profiles) available at that time and the so-called “near-real-time” observations otherwise. Moreover, in the development phase of the operational system PSY4V3, it was decided to systematically perform two other twin numerical simulations over the same time period, maintaining the same ocean model tunings but varying the complexity and the level of data assimilation. The first one is a free simulation (without any data assimilation) and the second one only benefits from temperature and salinity large-scale bias correction using in situ observed temperature and salinity vertical profiles. Intercomparisons between the three simulations were then conducted in order to better analyze and try to quantify the impact of some component of the assimilation system. These three versions of the system have been used to quantify the impact of some updates.
In a previous paper (Lellouche et al., 2013), the main results of the scientific evaluation of MyOcean global monitoring and forecasting systems at Mercator Ocean showed how refinements or adjustments to the system impacted the quality of ocean analyses and forecasts. The primary objective of this paper is to describe the recent updates applied to the system PSY4V3 and show the highest impact on the product quality. Updates resulting from routine system improvements are not separately illustrated and discussed (bathymetry, runoffs, assimilated databases, mean dynamic topography, etc.). So, particular focus was given to the initialization, correction of precipitation, assimilation of climatological temperature and salinity in the deep ocean, construction of background error covariance and the adaptive tuning of observation error. Another objective of this paper is to present a first-level evaluation of the system. The purpose here is not to perform an exhaustive validation but only to check the global behavior of the system compared to assimilated quantities or independent observations. Thus, an assessment of the hindcast (2007–2016) quality is conducted and improvements with respect to the previous system are highlighted in order to show the level of performance and the reliability of the system PSY4V3. A complementary study is aimed at demonstrating the scientific value of PSY4V3 for resolving oceanic variability at regional and global scale (Gasparin et al., 2018). Lastly, several scientific studies have investigated local ocean processes by comparing the PSY4V3 system with independent observation campaigns (Koenig et al., 2017; Artana et al., 2018). This reinforces the system PSY4V3 evaluation effort.
This paper is organized as follows. The main characteristics of the system PSY4V3 and details concerning the updates are described in Sect. 2. The impact of some sensitive upgrades is shown in Sect. 3. Results of the scientific evaluation, including some comparisons with independent observations, are given in Sect. 4. Section 5 contains a summary of the scientific assessment and a discussion of future improvements for the next version of the global high-resolution system.
This section contains the main characteristics of the CMEMS system PSY4V3 and details the last updates to the system compared to the previous system PSY4V2R2 (hereafter PSY4V2; see Fig. 1 and Table 2). A detailed description of some sensitive updates is provided in Sect. 3.
The system PSY4V3 uses version 3.1 of the NEMO ocean model (Madec et al.,
2008). This NEMO version has been available for a few years and has already
been used in the previous system PSY4V2. This was the available stable
version of the code when we started the development of the system PSY4V3 a
few years ago. Note that by using this version of the code, we do not access
the better algorithms and more sophisticated parameterizations present in the
version 3.6, which is the latest official release of NEMO. The physical
configuration is based on the tripolar ORCA12 grid type (Madec and Imbard,
1996) with a horizontal resolution of 9 km at the Equator, 7 km at Cape
Hatteras (midlatitudes) and 2 km toward the Ross and Weddell seas.
The bathymetry used in the system benefited from a specific correction in
the Indonesian seas inherited from the INDESO system (Tranchant et al.,
2016). In order to solve numerical problems induced by the use of Surface wind stress computation should in principle consider wind speed
relative to the surface ocean currents (Bidlot, 2012; Renault et al., 2016).
However, this statement applies to a fully coupled ocean–atmosphere system,
which is not the case for the present system PSY4V3. Based on sensitivity
experiments and following the results obtained by Bidlot (2012), we
pragmatically consider only 50 % of the surface model currents in the
wind stress computation. The monthly runoff climatology is built with data on coastal runoffs and 100
major rivers from the Dai et al. (2009) database (instead of Dai and
Trenberth, 2002, for the system PSY4V2). This database uses new data, mostly
from recent years, and streamflow simulated by the Community Land Model
version 3 (CLM3) to fill the gaps, in all lands areas except Antarctica and
Greenland. In addition, we built mean seasonal freshwater fluxes representing
Greenland and Antarctica ice sheet and glacier runoff melting. For this
purpose we have distributed the following mean values: 545 Gt yr As the Boussinesq approximation is applied to the model equations, thereby
conserving the ocean volume and varying its mass, the simulations do not
properly directly represent the global mean steric effect on the sea level
(Greatbatch, 1994). For improved consistency with assimilated satellite
observations of sea level anomalies, which are unfiltered from the global
mean steric component, a time-evolving global average steric effect is added
to the sea level in the simulation. This global average steric effect has
been computed as the difference between two successive daily global mean
dynamic heights (vertical integration from the surface to the bottom of the
specific volume anomaly). Due to large known biases in precipitation (Stephens et al., 2010; Kidd et
al., 2013), a satellite-based large-scale correction of precipitation has
been performed, except at high latitudes (poleward of 65 In order to avoid mean sea-surface-height drift due to the large
uncertainties in the water budget closure, the following two treatments were
applied. The surface freshwater global budget has been set to an imposed seasonal
cycle (Chen et al., 2005). Only spatial departures from the mean global
budget are kept from the forcing. A trend of 2.2 mm yr
The data are assimilated by means of a reduced-order Kalman filter derived
from a SEEK filter (Brasseur and Verron, 2006), with a three-dimensional
multivariate modal decomposition of the background error and a 7-day
assimilation cycle. It includes an adaptive-error estimate and a
localization algorithm. This data assimilation system is called SAM
(Système d'Assimilation Mercator). The background error covariance is
based on the statistics of a collection of three-dimensional ocean state
anomalies. The anomalies are computed from a long numerical experiment
(2007–2015 9-year period for PSY4V3) with respect to a running mean in order
to estimate the 7-day scale error on the ocean state at a given period of
the year. A Hanning low-pass filter is used to create the running mean with
a cutoff frequency equal to
Compared to the previous HRG system PSY4V2, the following updates were done on the data assimilation part (see Table 2).
CMEMS satellite near-real-time sea ice concentration OSI SAF
( CMEMS OSTIA SST (delayed time (reprocessed) until the end of 2006:
In addition to the quality control based on temperature and salinity
innovation statistics (detection of spikes, large biases) already present
in the previous system, a second quality control has been developed and is
based on dynamic height innovation statistics (detection of small vertically
constant biases). This is detailed in Sect. 2.3. A new hybrid MDT based on the CNES-CLS13 MDT (Rio et al., 2014) with
adjustments made using the Mercator GLORYS2V3 (GLobal Ocean ReanalYsis and
Simulation, stream 2, version 3) reanalysis and with an improved post-glacial rebound (also
called glacial isostatic adjustment) has been used.
This new hybrid MDT also takes into account the last version of the GOCE
geoid. This replaces the previous hybrid MDT used in the previous system
PSY4V2, which was based on the CNES-CLS09 MDT derived from observations
(Rio et al., 2011). The new hybrid MDT significantly reduces (not shown) sea
level bias (more than 5 cm in some areas) and consequently temperature and
salinity in regions where the topography makes the mean sea
surface estimation difficult (e.g., Indonesia, Red Sea and Mediterranean Sea). A consistent along-track SLA dataset
( The CORA 4.1 CMEMS in situ reprocessed database (Szekely et al., 2016;
As the prescription of observation errors in the assimilation systems is not
sufficiently accurate, adaptive tuning of observation errors for the SLA and
SST has been implemented. The method has been adapted from diagnostics
proposed by Desroziers et al. (2005) and is detailed in Sect. 3. New three-dimensional observation error files for the assimilation of in
situ temperature and salinity data have been recomputed from the MyOcean
IRG system PSY3V3R3 (see Fig. 1 and Table 1) using an offline version of the
adaptive tuning method mentioned above. A weak constraint towards the WOA13v2 climatology on temperature and
salinity in the deep ocean (below 2000 m) has been included in the two
components (3D-VAR and SEEK filter) of the assimilation scheme to prevent
drifts in temperature and salinity and as a consequence to obtain a better
representation of the sea level trend at global scale in the system. The
method consists of assimilating vertical climatological profiles of
temperature and salinity at large scale and below 2000 m in regions drifting
away from the climatological values using a non-Gaussian error at depth.
This is detailed in Sect. 3. The time window for the 3D-VAR bias correction was reduced from 3 months to 1 month
to obtain a correction that is more in line with the current physics, which
is made possible by the good spatial and temporal distribution of the Argo
network from 2006. In the previous system PSY4V2, the SSH increment was the sum of barotropic
and baroclinic (dynamic) height increments as in Benkiran and Greiner (2008).
Dynamic height increment was calculated from the temperature and salinity
increments, while the barotropic increment was an output of the analysis.
Barotropic height was computed without the wind effect. In the system
PSY4V3, we directly use the total SSH increment given by the analysis to
take into account, among other things, the wind effect like the hydraulic
control near the straits (Song, 2006; Menemenlis et al., 2007). The uncertainties in the MDT estimate and the sparsity of the observation
networks (both altimetry and in situ profiles) on the 7-day assimilation
window do not allow us to accurately estimate the observed global mean sea
level. Moreover, the mean sea level time evolution is the result of an
imposed trend for mass inputs (2.2 mm yr The background error covariance matrices needed for data assimilation are
defined using anomalies of the different variables coming from a simulation
in which only a 3D-VAR large-scale bias correction of
To minimize the risk of erroneous observations being assimilated in the
model, the system PSY4V3 carries out two successive quality controls (QC1
and QC2) on the assimilated temperature (
The first quality control QC1 has already been described in Lellouche et
al. (2013) and can be summarized as follows. An observation is considered
suspicious if the two following conditions are both satisfied:
The second quality control QC2 is
based on dynamic height innovation (vertical integration from the surface to
the bottom) statistics. This quality control allows for the detection of small biases that are
present in the whole water column and can thus induce large errors. It
basically stipulates that the thermal or haline component of dynamic height
innovation (hdyn
Thresholds used for QC2 for the thermal component of dynamical height
innovation (
The average and standard deviation of the thermal or haline components of
dynamical height innovation have been calculated from a global simulation at
Statistics in the Azores region:
It should also be noted that the QC2 quality control rejects the entire vertical profile, while the QC1 quality control only rejects aberrant temperature and/or salinity values at some given depths on the vertical profile.
Figure 3a shows an example of a “wrong” temperature profile detected by
the QC2 (and not by the QC1) at the end of July 2008. In this case,
Statistics of suspicious temperature (
Statistics of the QC1 and QC2 quality controls are summarized in Fig. 4, in which the percentage of suspicious temperature and salinity profiles is given as a function of the year over the 2007–2016 period. This percentage is relatively stable for both temperature and salinity profiles, with little year-to-year variability, except for the years 2012 and 2013 when more suspicious temperature and salinity profiles than usual were detected. Nevertheless, this percentage remains relatively low (less than 0.35 % for temperature and 3.5 % for salinity) given that the number of temperature profiles available each year ranges between 1.1 and 1.7 million number of salinity profiles between 150 000 and 600 000.
Most of the deficiencies in the systems can be related to these main recurring problems: initialization, atmospheric forcing biases, abyssal circulation and efficiency of the assimilation schemes. The first three problems are related to uncertainties in poorly observed areas or parameters (i.e., deep ocean, ice thickness) and to intrinsic errors of the atmospheric forcing. The last problem is related to linearity and stationarity hypotheses in the assimilation schemes. In this section, we detail some solutions adopted for the system PSY4V3, reducing uncertainties in the thermohaline component and allowing flow dependence in our assimilation scheme. These solutions correspond to a part of the updates mentioned in Sect. 2 that do not result from routine system improvements.
One way to initialize physical ocean model simulations is by using climatological values of temperature and salinity from databases and assuming the velocity field is zero at the start. The model physics then spins up a velocity field in balance with the density field. Another common way to initialize a model is with fields from a previous run of that model or with the results from another model.
Given that data assimilation of the current observation network rapidly (in
about 6 months) adjusts the model state in the first 1000 m, the first
solution has been chosen to minimize potential drifts occurring after some
years of simulation. Compared with the previous system PSY4V2 starting in
October 2012 from the WOA09 three-dimensional climatology (see Fig. 1), the
PSY4V3 system starts in October 2006 using improved initial climatological
conditions. For that, we chose to use ENACT-ENSEMBLES EN4 1
Diagnostics (time series) with respect to the vertical temperature
and salinity profiles over the October 2006–December 2007 period. Mean
misfit between observations and model for salinity (
Mean 2007–2014 IFS ECMWF atmospheric precipitation bias (units in
mm day
Two free simulations (without any data assimilation) have been performed with the system PSY4V3 using either WOA09 or robust EN4 as an initial condition in October 2006. Figure 5 shows the box-averaged innovations of temperature and salinity as a function of time and depth over the October 2006–December 2007 period. Fig. 5a reveals that, using WOA09 as an initial condition, a fresh bias appears in the first 100 m of the innovation, particularly more pronounced at the surface. It is no longer the case when using robust EN4 to initialize the model (Fig. 5b). For temperature, Fig. 5c exhibits cold biases above 100 m and below 300 m that are considerably reduced by using robust EN4 as an initial condition (Fig. 5d). The warm and salty bias between 200 and 300 m is slightly reinforced. It mostly concerns the main thermocline whose motions are well correlated with the altimetry. This bias will be corrected by the assimilation of altimetry and Argo profiles. Deeper biases are reduced with this new initialization where Argo profiles are missing.
Mean surface salinity innovation (difference between the
assimilated observation and the model; units in psu) in the year 2011.
Climatological thermal expansion (
Many studies (e.g., Janowiak et al., 1998, 2010; Kidd et al., 2013) have compared reanalysis and atmospheric model precipitation fields with observation-based datasets and have shown that atmospheric model products always bring significant and systematic errors and are not able to close the global average freshwater budget. For instance, Janowiak et al. (2010) found that the IFS operational model and ERA-Interim reanalysis (Dee et al., 2011) from ECMWF perform well for temporal variability with respect to observational datasets, but they globally overestimate the daily precipitation. Although progress has been made in the ECMWF forecast model, substantial errors still occur in the tropics (Kidd et al., 2013). The correction of atmospheric forcing within ocean applications has already been successfully explored by adjusting atmospheric fluxes via observational datasets in global applications (Large and Yeager, 2009; Brodeau et al., 2010). Other studies only focused on precipitation correction (Troccoli and Kallberg, 2004; Storto et al., 2012).
Non-Gaussian error for climatology (corresponding to a weak constraint of the system in green). A cost equal to zero corresponds to an infinite observation error, namely a system operation in a free mode (without assimilation of climatology).
Mean temperature (
The proposed method in this paper consists of correcting the daily
precipitation fluxes by means of a monthly climatological coefficient
inferred from the comparison between the Remote Sensing Systems (RSS)
Passive Microwave Water Cycle (PMWC) product (Hilburn, 2009) and the IFS
ECMWF precipitation. We use the remote PMWC product because of its relatively
high
Temperature (
SLA innovation along a single assimilated track altimeter
Figure 6 represents the difference between the IFS precipitation coming
from ECMWF and the PMWC product using satellite data before and after large-scale correction. As already pointed out by Stephens et al. (2010), original
IFS forcings exhibit a systematic overestimation of precipitation within
the intertropical convergence zones (up to 3 mm day
To validate this correction, two global ocean hindcast simulations of several years, using only the 3D-VAR large-scale bias correction in temperature and salinity, have been performed, one with IFS correction and the other without. Figure 7 represents the mean surface salinity innovation (difference between the assimilated observation and the model) in the year 2011. At the global scale, the bias reduction is not very significant, but these maps demonstrate that the IFS correction is beneficial in many local areas. The strongest benefit concerns the tropics where the IFS correction allows us to reduce the magnitude of the near-surface salinity fresh mean bias down to 0.5 psu. The fresh bias reduction in the tropics reaches 0.15 psu on average.
SLA increment difference using 10 and 300 Shapiro passes as
anomaly filtering in a regional system at
2007–2015 SSH standard deviation (diagnostics made with one point
every three horizontally and 1 day every 5) of the
The model may exhibit significant drift at depth that can be related to the
misrepresentation of several processes, an exhaustive list of which would be
hard to give here. Difficulties encountered by ocean models using
For systems that assimilate observations in a multivariate way, the problem
can be more critical because of the deficiencies of the background error
covariances that may contain spurious correlations for extrapolated and/or
poorly observed variables. Unfortunately, there are very few temperature and
salinity profiles below 2000 m to constrain the model drift. Hence, the
climatology is currently the only source of information at depth to prevent
the model from drifting. Virtual vertical profiles of temperature and
salinity below 2000 m are built from the monthly WOA13v2 climatology. These
virtual observations are geographically positioned on the model horizontal
grid with a coarse resolution (1
Density difference (October–December 2008 minus October–December 2009) in the
equatorial Pacific (2
Evolution of the PDF of the ratio for the Envisat satellite from
D
As in Greiner et al. (2006), we define empirically the standard deviations
(departures from the climatology)
We then define
Envisat
If we denote
A non-Gaussian error is used to impose a weak constraint on the model at
depth (Fig. 9). That way, we correct the model drift without constraining a
slow moderate variability or trend. Basically, the hypothesis is that small
to medium departures from the climatology (
To validate this kind of assimilation, two global ocean simulations of
several years, using only the 3D-VAR large-scale bias correction in
temperature and salinity, have been performed. Due to the high computational
cost of the system PSY4V3, the assimilation of WOA13v2 below 2000 m has been
tested with a global intermediate-resolution system at
In practice, the assimilation of WOA13v2 climatological profiles below 2000 m in the system concerns mostly some regions where the steep bathymetry might be an issue for the model (Kerguelen Plateau, Zapiola Ridge, Atlantic Ridge). Figure 10 shows mean temperature (left) and salinity (right) innovations (WOA13v2 climatological profiles minus model) in 2013 at 2865 m. The assimilation of these climatological profiles occurs more or less at the same locations over the time period 2007–2016. Since the conditions of the system of Eq. (6) relate to the density innovation, we have a perfect symmetry of the temperature and salinity data that are assimilated. This has the effect of not disturbing the density gradients too much.
If we focus on latitudes between 30 and 60
The seasonally varying background error covariance is based on the statistics of a collection of three-dimensional ocean state anomalies. This approach is based on the concept of statistical ensembles in which an ensemble of anomalies is representative of the error covariance. In this way, truncation no longer occurs and all that is needed is to generate the appropriate number of anomalies. The way in which these anomalies are computed from a long numerical experiment is described in Lellouche et al. (2013).
In this section, we detail two features of the system PSY4V3 compared to the
previous system PSY4V2 regarding the construction of the background error
covariance. First, we evaluate the impact of anomaly filtering on analysis
increment. Second, we evaluate the potential added value to the quality of
the analysis increments of the choice of the simulation from which to
calculate the anomalies. In the previous system PSY4V2, a free simulation
was used to calculate the anomalies. For the system PSY4V3, the anomalies
are computed from a simulation in which only a 3D-VAR large-scale bias
correction of
The signal at a few horizontal grid
To illustrate the impact of the anomaly filtering, we set up some
experiments with different levels of filtering. Each experiment consists of
the assimilation of a single altimeter track over one assimilation cycle.
These experiments have been performed with a Mercator Ocean regional system
at
Figure 12a represents SLA innovation along the single assimilated track.
Figure 12b, c and d represent the SLA increments obtained with 10,
100 and 300 Shapiro passes, respectively, as the anomaly filtering mentioned above
(corresponding approximately to a 3, 10 and 15 horizontal grid
Other experiments closer to real-time integration setup have been performed, assimilating all the altimeter tracks available in a 7-day assimilation window instead of one single track. Figure 13 shows the difference of SLA increments using 10 and 300 Shapiro passes as anomaly filtering (corresponding approximatively to 20 and 80 km). The conclusions are the same as those concerning the experiments with a single assimilated track. The corrections under the tracks remain almost the same for the two levels of filtering. Both analyses are close to the data under the tracks. The strongest differences occur outside the tracks where the innovation information is extrapolated to fill the gaps. Low-filtered increments (10 Shapiro passes) have small-scale structures that are statistical artifacts. Small structures can cascade in the model and stay trapped between the repetitive tracks without correction by the assimilation. This happens less when more filtering (300 Shapiro passes) is performed on the anomalies beyond the effective resolution of the model.
The system PSY4V3 was run over the October 2006–October 2016 period to catch up to real time (OPER simulation) starting from three-dimensional temperature and salinity initial conditions based on the EN4 climatology. This simulation benefited from the full data assimilation system, including the 3D-VAR bias correction and the SAM filter. Two other simulations over the same period have been performed. The first one is a FREE simulation (without any data assimilation) and the second one has exactly the same model tunings but only benefits from the temperature and salinity 3D-VAR large-scale bias correction (BIAS simulation).
Figures 14 and 15 show comparisons between this triplet of PSY4
simulations and two observational products. The first product is the
CMEMS/DUACS (Data Unification and Altimeter Combination System)
merged–gridded sea level anomaly heights in delayed time on a
Time-averaged density differences along the equatorial Pacific between two
ENSO events (October–December 2008 minus October–December 2009), computed from the PSY4
simulations and from the Roemmich–Gilson Argo monthly climatology, are shown
in Fig. 15. The SCRIPPS Argo product presents a higher density difference in
the eastern part of the equatorial Pacific. It corresponds to the change
from moderate La Niña conditions early in 2008 to moderate El Niño
conditions in 2009. The FREE simulation is not dense enough in the east
compared to observations, particularly at the pycnocline depth (1025 kg m
In summary, the BIAS simulation better represents the density fronts on the horizontal (Gulf Stream) and on the vertical (Pacific pycnocline). The covariance matrix deduced from this simulation has information on the density gradients that is well placed. This is valuable off the Equator through geostrophy and at the Equator to control the zonal pressure gradient. The variance in sea level is stronger than the DUACS one (see Fig. 14e) but the most important point for the construction of the anomalies is to have well-placed density gradients. In the OPER simulation and as mentioned in Lellouche et al. (2013) in the description of the data assimilation system SAM, an adaptive scheme will correct the variance and will give an optimal background model error variance based on a statistical test formulated by Talagrand (1998).
In order to refine the prescription of observation errors (instrumental and representativeness errors), adaptive tuning of errors for the SLA and SST has been implemented in PSY4V3. We use the Talagrand method (Talagrand, 1998) to adjust the background error. Instrumental error does not change with time. On the contrary, the representativeness error is really flow dependent. Taking into account the representativeness error is particularly important for assimilated OSTIA SST because the sky is clear only 30 % of the time on average. The method has not been used for temperature and salinity vertical profiles because of the reduced number of in situ data compared with satellite data. Three-dimensional fixed observation errors are then used for the assimilation of in situ temperature and salinity vertical profiles.
The method consists of the computation of a ratio, which is a function of
observation errors, innovations and residuals (Desroziers et al., 2005). It
helps correct inconsistencies in the specified observation errors. This
ratio can be expressed as
Figure 16 represents the temporal evolution of the ratio defined in Eq. (7) for the Envisat satellite. At the beginning of the simulation, the observation error is overestimated (ratio less than 1). The ratio tends to 1 after only a few weeks of simulation.
Evolution in time of model SST anomaly
For SLA (Fig. 17), the a priori prescribed observation error is globally significantly reduced. The median value of the error changed from 5 to 2.5 cm in a few assimilation cycles and allows for better results. This method allows us to have more realistic and evolutive observation error maps that can provide valuable information for space agencies.
First EOF
The realism of tropical oceans is crucial for seasonal forecasting applications. Tropical instability waves (TIWs) can be diagnosed from SST (Chelton et al., 2000). These Kelvin–Helmholtz waves initiate at the interface between areas of warm and cold sea surface temperatures near the Equator and form a regular pattern of westward-propagating waves. Figure 18 gives an example of the adjustment of observation error to model physics and atmospheric variability. The SST anomalies in the equatorial Pacific clearly show the propagation westwards of TIWs in the second half of the year. This is more pronounced during episodes of La Niña (mid-2007 and mid-2010). The observation error anomalies estimated by the Desroziers method show that the error increases when these TIWs are more marked. This can be explained two ways. First, the representativeness error increases because the data are not corresponding exactly at the right time and the right position to the model counterpart. In the case of clouds, SST values can result from OSTIA time or space interpolation. This would be detrimental with the fast propagation of TIWs. Second, large errors can result from a model shift of the TIW structures. The error decreases in the reverse case.
We have also performed an empirical orthogonal function (EOF) analysis to assess the variability of the SST observation error (Fig. 19). Mode 1 is associated with the seasonal cycle and mode 2 (not shown) corresponds to the migration of the seasonal signal. Mode 3 is associated with the interannual signal with, for instance, the transition La Niña–El Niño, showing that the SST error is able to adapt to both seasonal and interannual fluctuations.
This section describes the PSY4V3 system's quality assessment with
diagnostics over particular years, together with time series over multiyear
periods. To evaluate the quality of the system, the departure from the
assimilated observations (SST, SLA,
Mean SST residuals (units in
Time series of SST (units in
The OSTIA product is assimilated in the system PSY4V3. Compared to the previous system PSY4V2, some large-scale cold biases with respect to OSTIA are reduced in the Indian, eastern South Pacific and western North Pacific oceans (not shown). On the other hand, warm biases are not reduced, especially in regions of strong interannual warm events such as the eastern tropical Pacific where strong El Niño took place in 2015/2016, but also in the ACC, the Gulf Stream and the Greenland Current (Fig. 20a). Some inconsistencies can be found between OSTIA SST and in situ near-surface temperature, particularly in the North Pacific where the system PSY4V3 presents a cold bias compared to in situ near-surface temperature but a warm bias compared to OSTIA SST (Fig. 20b). Figure 20c shows the difference between drifting buoy SST and the system PSY4V3 over the year 2015. The drifting buoy SST data are present in the CMEMS in situ database used by Mercator but they have not been assimilated in the system because the depth of these data is a nominal value and we chose to assimilate only data with a measured depth value. Although we plan to assimilate these data in the future system, we use currently these data as independent information. This allows us to see that SSTs from in situ vertical profiles and SSTs from drifting buoys are coherent with each other. We thus again find the cold bias highlighted by the comparison with SST from in situ vertical profiles in the North Pacific. It is a lack of stratification in the model that causes midlatitude cold surface biases during (boreal) summer and a warm bias between 50 and 100 m.
We also checked the time series of the mean and the RMS of the misfit
(innovation) between the observed SSTs and the model. For OSTIA SST, which
is the gridded SST assimilated into PSY4V3, we obtain a mean warm bias
of
High-resolution CATSAT SST from CLS
Time series of the 0–5000 m RMS difference between the model
analysis and the in situ observations for the previous system PSY4V2 (in blue),
the new system PSY4V3 (in black) and the WOA13v2 climatology (in red).
The CLS (Collecte Localisation Satellites) has operated a near-real-time
oceanography data service named CATSAT since 2002 for scientific, institutional or
private users (support to fishery management or to the offshore oil and gas
industry). These data include satellite observations such as chlorophyll
For the
Mean residual errors
On global average and compared to the previous system PSY4V2, the system
PSY4V3 slightly degrades the temperature statistics (
Moreover, the system PSY4V3 experiences a slight warm bias (negative observation minus forecast difference) in the subsurface (25–500 m) on global average (not shown). For the year 2015, part of this signal comes from the strong interannual ENSO signals in the tropical Pacific where the near-surface bias is also warm, as well as in the ACC and the Gulf Stream. Seasonal cold surface biases appear in the midlatitudes, linked with a lack of stratification during summer. Summer warming is injected too deep, which results in subsurface spurious warming and a mixed layer that is too shallow. However, these biases remain small on global average.
The system PSY4V3 is closer to altimetric observations than the previous one with a global forecast RMS difference of around 6 cm instead of 7 cm for the system PSY4V2 (not shown). This RMS difference is consistent with the prescribed a priori observation errors (about 2 cm for altimeter instrumental error and 4 cm for MDT error on average). The statistics come from the data assimilation innovations computed from the forecast used as the background model trajectory and give an estimate of the skill of the optimal model forecast. These scores are averaged over all 7 days of the data assimilation window, which means the results are indicative of the average performance over the 7 days, with a lead time equal to 3.5 days.
More precisely, in the year 2015, the SLA mean and RMS errors are
considerably reduced in the new system PSY4V3 compared to the previous one
(Fig. 24). The mean bias is reduced by 0.3 cm (from
Sea surface height RMS difference between tide gauge observations and the system PSY4V3 for the year 2015. Unit is centimeters.
The system PSY4V3 produces hourly outputs at the surface that can be compared with tide gauge measurements. For that, we used the BADOMAR product (Lefevre et al., 2005), which is a specific processed tide gauge database developed and maintained at CLS that consists of filtered tide gauge data from the GLOSS/CLIVAR (Global sea Level Observing System/Climate Variability and Predictability) “fast” sea level data tide gauge network (GLOSS Implementation Plan, 2012). These tide gauge data are corrected for inverse barometer effect and tides. High-frequency model SSH compares well with tide gauges in many places, with a slight improvement in PSY4V3 with respect to PSY4V2 (not shown). The best agreement between the system PSY4V3 and tide gauges is found in the tropical band, as can be seen in Fig. 25, while shelf regions and closed seas are less accurate. This confirms the latitude dependence of the correlation between tide gauges and satellite altimetry or modeled SSH discussed in Vinogradov and Ponte (2011) or Williams and Hugues (2013).
The improvements related to water masses and SLA lead to a correct global
mean sea level (GMSL) trend. We checked the system GMSL by comparing the
results with recent estimated trends from the paper of Chambers et al. (2017). We found for the model a trend of 3.2 mm yr
Time series of (observation–forecast) mean
The system PSY4V3 assimilates OSI SAF sea ice concentration in both hemispheres with a monovariate–monodata scheme. As expected, PSY4V3 is closer to the observations than the previous system PSY4V2 (not shown), in which no sea ice observations had been assimilated. As illustrated by Fig. 26, the system PSY4V3 has a slight overestimation of ice during the melting season in summer (up to 3 % on average in both hemispheres). Conversely, the mean error is stronger on average during winter (10 % to 20 % underestimation, depending on the year). RMS errors are also larger during summer (up to 20 % in the Arctic and 30 % in the Antarctic with respect to OSI SAF observations), and they drop to less than 10 % in winter. These RMS errors quantify the capacity of the system to capture weekly time changes in the ice cover.
Time series over the 2007–2016 period of sea ice volume in
the Arctic for several systems: GREP composed of the four members GLORYS2V4 from
Mercator Ocean (France), ORAS5 from ECMWF, FOAM/GloSea from the Met Office (UK)
and C-GLORS from CMCC (Italy); PSY4V3 from Mercator Ocean (France); PIOMAS
product. The spread of the GREP product is represented in light red. Unit is
km
We have also checked the evolution of the sea ice volume diagnosed by the system PSY4V3. The data assimilation scheme SAM produces an increment of sea ice concentration, which is the unique sea ice correction applied in the model using the incremental analysis update (IAU) method described in Lellouche et al. (2013). The sea ice volume then adjusts to this correction considering a constant sea ice thickness. No sea ice thickness observations are assimilated in the system. The risk is therefore to obtain unrealistic drifts or trends of the unconstrained sea ice volume. Presently, sea ice volume retrievals from satellites are associated with large uncertainties (Zygmuntowska et al., 2014). Consequently, modeled sea ice volume is difficult to validate and one of the solutions is to compare modeled sea ice volume from several systems.
Figure 27 shows the 2007–2016 evolution of sea ice volume for the system
PSY4V3, the PIOMAS modeled product (Schweiger et al., 2011) and the CMEMS
GREP (Global Reanalysis Ensemble Product;
The contingency table analysis approach described in Smith et al. (2016) has
been applied to evaluate sea ice extent as compared to observation.
Satellite ice concentration coming from AMSR2 (L1B brightness with a NASA
team 2 algorithm to compute sea ice concentration) has been used as
an independent observation to provide a general assessment in the detection of
false alarms for ice coverage. Although this type of evaluation is usually
done on forecasts, we used hindcasts. For the computation of the statistics
we have used a stereo-polar grid at a 20 km resolution. In each cell of that
grid we have then computed binary values corresponding to ice–open water
conditions for the model and the sea ice observations by using a 40 %
concentration threshold. We have also restricted our study to the proportion
correct total (PCT), following the conclusion of Smith et al. (2016)
that it was more insightful to refer to the PCT rather than other
proportions. The PCT quantity is defined as PCT
Contingency table entries for sea ice verification of the PSY4V3 system compared to AMSR sea ice concentration observations.
Time series of the PCT quantity for PSY4V2 (in blue) and PSY4V3
(in black).
Figure 28 shows times series of PCT for the PSY4V2 and PSY4V3 systems. The lower PCT values are mostly due to an excessive melt in spring and summer for both the Arctic and Antarctic. However, the assimilation of sea ice concentration significantly improves the total hit rate during these periods.
Mean zonal drift innovation (m s
The aim of this section is to use velocity observations that were not assimilated into the system to assess the level of performance of PSY4V3 compared to the previous PSY4V2 system. The mean currents are checked by comparing the model to velocity observations coming from Argo floats when they drift at the surface and in situ Atlantic Oceanographic and Meteorological Laboratory (AOML) surface drifters. A paper by Grodsky et al. (2011) revealed that an anomaly in the drogue loss detection system of the Surface Velocity Program buoy had led to the presence of undetected undrogued data in the “drogued-only” dataset distributed by the Surface Drifter Data Assembly Center. Rio (2012) applied a simple procedure using altimeter and wind data to produce an updated dataset, including a drogue presence flag as well as a wind slippage correction. Therefore, we used this new drogued-only surface drifter dataset coming from CMEMS in situ TAC (Rio and Etienne, 2017) to check mean model currents.
Figure 29 represents zonal drift innovation for the PSY4V2 and PSY4V3 systems.
Although some biases persist, mostly in the western tropical basins,
significant improvements are obtained almost everywhere with the new system
PSY4V3, particularly in the equatorial Pacific. The mean bias is
reduced (from 0.1 to 0.08 m s
More locally, a comparison of the 2007–2015 averaged drifts from the system
PSY4V3 and the observations over the Indonesian region has been performed
(not shown). Currents in this region are very difficult to resolve because
of the many narrow straits and the strong tidal mixing. The retroflection of
the westward South and North Equatorial Current (along Papua and near
12
The Mercator Ocean system PSY4V3, in operational mode since 19 October 2016, benefits from many important updates. PSY4V3 has a quite good statistical behavior with an accurate representation of water masses, surface fields and mesoscale activity. Most of the components of the system PSY4V3 have been improved compared to the previous version: global mass balance, three-dimensional water masses, sea level, sea ice and currents. Major variables like sea level and surface temperature are hard to distinguish from the data.
In this paper, the updates showing the highest impact on the product quality and that do not result from routine system improvements have been illustrated and evaluated separately. Particular focus was therefore on the initialization, correction of precipitation, assimilation of climatological temperature and salinity in the deep ocean, construction of background error covariance and the adaptive tuning of observation error.
The initial climatological condition has been improved in order to be more consistent with the vertical profiles of temperature and salinity that have been assimilated thereafter. Rather than directly taking the climatological temperature and salinity of the month corresponding to the start of the simulation, we performed a point-wise linear regression, allowing us to obtain an initial condition at the appropriate time and based only on real observations; 1-year free simulations have been performed and show that biases are globally reduced.
Uncertainties inherent to atmospheric analyses and forecasts can induce large errors in ocean surface fluxes. For instance, a slight shift in the position of a storm can induce local errors in salinity, temperature and currents. In the tropical band, precipitation is systematically overestimated. Moreover, large-scale salinity biases can appear because the global average freshwater budget is not closed. For this reason, IFS ECMWF atmospheric analyzed and forecasted precipitation has been corrected at large scale using the satellite-based PMWC product. This correction is beneficial in many areas, reducing the magnitude of the near-surface salinity fresh mean bias in the tropics down to 0.5 psu. This surface fresh bias reduction in the tropics reaches 0.15 psu on average.
Due to misresolved processes, the model may also drift at depth. To keep
some water mass properties, the DRAKKAR group restored modeled temperature
and salinity toward the annual climatology of Gouretski and Koltermann (2004) in
specific areas. This choice was driven by the Antarctic Bottom Water
restoring zone where this climatology is recognized as more suitable.
For Mercator systems that assimilate observations in a multivariate way,
the problem can be more critical because of the deficiencies of the
background errors for extrapolated and/or poorly observed variables. To
overcome these deficiencies, vertical climatological
We have also proposed solutions to reduce some problems related to linearity and stationarity hypotheses in the assimilation schemes. The first one concerns the construction of background error covariance. Rather than calculating the anomalies from a free simulation, we chose to calculate them from a simulation benefiting only from the 3D-VAR large-scale bias correction in temperature and salinity and better representing the density fronts on the horizontal and on the vertical. Moreover, anomalies have been filtered in order to remove the scales beyond the effective resolution of the model. The second one concerns the tuning of the observation errors. Adaptive tuning of SLA and SST errors has been successfully implemented. It allows us to have more realistic and evolutive SLA and SST error maps.
All these scientific and technical choices have been validated and
integrated into the system PSY4V3, which has been evaluated for the period
2007–2016 by means of a thorough procedure involving statistics of model
departures from observations. The system PSY4V3 is close to SLA along-track
observations with a forecast (range 1 to 7 days) RMS difference below 6 cm.
Moreover, the correlation of the system PSY4V3 with tide gauges is
significant at all frequencies; however, many high-frequency fluctuations of
the SSH might not be captured by the system because tides or pressure
effects are not yet included. The description of the ocean water masses is
very accurate on average and departures from in situ observations rarely
exceed 0.5
A global comparison with independent velocity measurements (surface drifters) shows that the location of the main currents is very well represented, as is their variability. However, surface currents of the midlatitudes are underestimated on average. The underestimation ranges from 20 % in strong currents to 60 % in weak currents. Some equatorial currents are overestimated, and the western tropical Pacific still suffers from biases in surface currents related to MDT biases. In contrast, the orientation of the current vectors is better represented.
Lastly, the system reproduces the sea ice seasonal cycle in a realistic manner. However, compared to assimilated data, sea ice concentration is slightly overestimated in winter seasons and underestimated during summer seasons. A contingency table analysis approach has also been used to evaluate sea ice extent compared to observations. This approach shows clear improvements due to the assimilation of sea ice concentration in the system PSY4V3.
Remarkable improvements have been achieved with the system PSY4V3 compared
to the previous version. However, some biases have been highlighted in
ocean surface features as well as three-dimensional ocean structure at
basin, sub-basin and local scales. The simulation biases may be due to the
initial state (especially in the deep layer for which historical observation
data are rare), the atmospheric forcing uncertainties, the river runoff
approximations, the efficiency of the assimilation scheme, and the model
errors induced by unresolved or parameterized physical processes. Numerous
projects have already been set up at Mercator Ocean to propose innovative
solutions. The integration of the ingredients from these projects into the
future CMEMS global high-resolution system is planned for 2019. The
improvement of numerical simulations could thus be carried out based on
sensitivity experiments on some model parameters (e.g., coastal runoffs,
atmospheric forcing, high-frequency phenomena including tides,
multi-category sea ice model, interaction and retroaction between ocean
currents and waves, vertical mixing and advection scheme). Better algorithms
and more sophisticated parameterizations already available in version
3.6 of the NEMO code should help in the future to resolve issues related to
important ocean processes and to reduce model biases. It is also planned to
assimilate new types of observations into the system (drifting buoys SST,
higher-resolution SST (L3 products), satellite sea surface salinity,
velocity observations from AOML surface drifters, deep-ocean
observations from Argo surface floats) to better constrain the modeled
variables and to overcome the deficiencies of the background errors, in
particular for extrapolated and/or poorly observed variables. Another
important issue is to use a shorter assimilation time window and a 4-D
analysis in the assimilation scheme to better correct the fast-evolving
processes. The next version of the global high-resolution system will also
include seasonal errors for in situ vertical profiles already used in the
CMEMS eddy-resolving 1992–2016 reanalysis GLORYS at
No datasets were used in this article.
JML designed and wrote the paper. JML and OLG performed the numerical simulations. JML and EG carried out most of the analyses, with contributions from OLG, CR, MD, FG, OH and BL. GG worked on the precipitation correction. MB and CET (RBB) contributed to the development of the data assimilation (modeling) component. Lastly, YD, ER and PYLT helped in reading and commenting on the paper.
The authors declare that they have no conflict of interest.
This article is part of the special issue “The Copernicus Marine Environment Monitoring Service (CMEMS): scientific advances”. It is not associated with a conference.
This study has been conducted using EU Copernicus Marine Service information. The authors thank Luc Vandenbulcke and the anonymous reviewer for their careful reading and for providing very constructive comments that improved the paper. Special thanks to our Mercator Ocean colleague Jerôme Chanut for his help answering questions regarding the specifics of the NEMO code. Edited by: Marilaure Grégoire Reviewed by: Luc Vandenbulcke and one anonymous referee