Observing system
experiments (OSEs) are carried out over a 1-year period to
quantify the impact of Argo observations on the Mercator Ocean 0.25
Argo is the first ever in situ ocean observing system providing in
real-time observations at global scale. The initial target of 3000 profiling
floats drifting in the ocean was reached by the international Argo program in
November 2007. Mean coverage is one float in every
3
Operational oceanography capabilities have improved dramatically since the end of the 1990s thanks to the development of real-time in situ and satellite global observing systems (in particular Argo and satellite altimetry) and the improvement of modeling and data assimilation techniques (e.g., Bell et al., 2009). Data assimilation techniques now provide efficient tools for analyzing the impact and improving the design of Global Ocean Observing Systems (GOOS) (e.g., Fujii et al., 2014; Lea et al., 2013).
The OceanObs'09 conference held in Venice in September 2009 for the international coordination of interdisciplinary ocean observation highlighted the need to consolidate and improve the design of the global ocean observing system (Lindstrom et al., 2012). To meet this requirement, it is crucial to evaluate and quantify how the existing observation system constrains ocean analysis and forecasting systems. Observing system experiments (OSEs) are a classical tool for evaluating the impact and importance of an observing system on a data assimilation system. OSEs involve the systematic withholding of a subset of observations. The evaluation of the degradation in quality of the resulting analyses and forecasts is then used to quantify the impact of the observations withheld.
In the last decade, several studies based on OSEs have analyzed the impact of
different components of the global ocean observing system for ocean analysis
and forecasting. Balmaseda et al. (2007) studied the statistical impact of
Argo on analyses of the global ocean for the period 2001–2006. Oke and
Schiller (2007) analyzed the importance of the combination of Argo, sea
surface temperature (SST) and altimeter data on a regional eddy-resolving
ocean reanalysis. Other relevant studies (e.g., Vidard et al., 2007; Tranchant et al., 2008; Guinehut et al., 2012 or
more recently Fujii et al., 2014 and Lea et al., 2014) have focused on
different observing systems and assessed their impact on analysis and
forecasting systems. The GODAE OceanView program
(
In this paper, we focus on the impact of the Argo observing system on the
analyzed temperature (
The Mercator Ocean 0.25
The 2012 assimilated observation data sets are real-time along-track
altimeter SLA data from SSALTO/DUACS (Segment Sol multi-missions dALTimetrie, d'orbitographie
et de localisation précise/Data unification and Altimeter combination system; Dibarboure et al., 2011). Mean dynamic topography (MDT), used as a reference for SLA data
assimilation, is based on the “CNES-CLS09” MDT derived from observations
and described in Rio et al. (2011). The assimilated SST observations are the
NCDC/NOAA daily high-resolution SST analysis at 0.25
The 2012 in situ data set is extracted from the real-time Coriolis database, where automated quality controls are applied. Coriolis is the in situ component of the French operational oceanography infrastructure. It provides real-time and qualified ocean in situ measurements to the European MyOcean project (Copernicus Marine Service) and to research and climate communities. Coriolis collects, controls and standardizes temperature and salinity profiles from different types of instruments including Argo floats, CTDs from research vessels, expendable bathythermographs (XBTs), moorings, sea mammals, gliders and drifting buoys. In term of data number, Argo is currently by far the most important source of information for in situ temperature and salinity profiles.
Argo floats provide measurements of temperature and salinity from the surface to 2000 m every 10 days at the global scale. The XBT network provides temperature measurements mostly along the main shipping routes from the surface to 800 m. Moorings are mostly in the tropical oceans with TAO/TRITON moorings for the Pacific, PIRATA for the Atlantic and RAMA for the Indian Ocean. Typically, the buoys sample the ocean from the surface down to 500 or 750 m with 10–15 levels. Other moorings sample specific regions such as the Drake Passage or the Labrador Sea. CTDs carried by sea mammals are located in high-latitude regions such as the Svalbard Islands, the French Southern and Antarctic Lands, the Ross Sea and Kerguelen region. Gliders are used to sample temperature and salinity from the surface to a given parking depth in specific areas of interest. There is not yet a global measurement strategy for such an observing system.
The 2012 coverage of the in situ data set is shown in Fig. 1. Figure 1a and b each correspond to 50 % of the Argo data sets. To ensure that all Argo profiles were selected, we sorted data from the instrument type variable, WMO_INST_TYPE, specified as “VERTICAL PROFILING: observation” for Argo floats in the Coriolis data set. The most realistic way of dividing up the Argo data set and of keeping coherent spatial and temporal resolution was therefore to sort it by platform numbers. Odd-numbered Argo platforms are shown in Fig. 1a, even ones in Fig. 1b. One of the most striking features of these two plots is the global and dense coverage of the oceans. The sparse distribution of the No Argo data set (green dots) is also remarkable. Some regions are rather more densely sampled by No Argo platforms than others. For example, the Kuroshio and North Atlantic areas are highly sampled compared to the southwest Pacific. Figure 2 represents the time coverage of temperature (Fig. 2a) and salinity (Fig. 2b) profiles from the surface down to 2000 m for the last 6 months of 2012. The time distribution is fairly regular and no specific feature should impact our conclusions.
Spatial distribution of 2012 in situ data set divided into three sub-data sets. Red dots are the Argo profiles from odd WMO (World Meteorological Organization) platform numbers, blue dots are Argo profiles from even WMO platform numbers, green dots are the other in situ observations.
Time series of the number of 2012 in situ temperature
The OSEs presented here focus on the impact of the Argo observing system on
temperature and salinity analysis and forecasts. Three experiments were
performed from 18 January 2012 to 26 December 2012. This corresponds to 50
analyses with an assimilation cycle of 7 days. The three experiments
assimilate SLA and SST data and differ only as regards the in situ
assimilated data sets:
The experiment entitled Run-Ref assimilates SLA, SST and all in situ data (Argo The experiment entitled Run-Argo2 assimilates SLA, SST, 50 % of the Argo
data and all the “other No Argo in situ data”. The experiment entitled Run-NoArgo assimilates SLA, SST and all “other No
Argo in situ data”.
For the three experiments above, the strategy is to start from the same
initial conditions of the PSY3 operational system that assimilates all the
data and then withdraw part of the Argo data set for the OSEs. Lastly, a free
run (i.e., where no data at all are assimilated), hereafter called Free Run,
was also carried out to assess the overall improvement of the PSY3 system.
The free run starts from the same initial conditions as for the three
above-mentioned experiments. Table 1 summarizes the experiment strategy.
List of OSEs carried out as part of this study.
This section is organized as follows. The first part is an independent comparison of the Run-NoArgo analyzed fields with Argo observations. The second part compares analyzed temperature and salinity fields from the different OSEs. This quantifies the amplitude and the spatial distribution of the changes bring by the assimilation of Argo profiles. We then verify that those changes in the analyzed fields correspond to a decrease of the misfit to in situ observations when Argo profiles are assimilated. The last part compares Argo observations with co-located profiles from forecasted fields to assess the impact of the Argo data assimilation. In each subsection, temperature and salinity results are discussed separately. The impact on the sea surface height and sea surface temperature innovations is also briefly discussed.
Spatial distribution of the RMS and the mean temperature differences
between Run-NoArgo and Argo observations in the 0–300 and 700–2000 m
layers over the last 6 months of the experiments.
Spatial distribution of the mean and RMS salinity differences
between Run-NoArgo and Argo observations in the 0–300 and 700–2000 m
layers over the last 6 months of the experiments.
19 December 2012: analyzed temperature fields – differences between
Run-Ref and Run-NoArgo at 100 m
The RMS of temperature differences between Run-Ref and Run-NoArgo in the
0–300 m layer
Spatial distribution of the mean and RMS temperature differences
between Run-Ref and in situ observations in the 0–300 and 700–2000 m
layers over the last 6 months of the experiments.
Statistics are done over the last 6 months of each year of OSE experiments in
order to avoid the spin-down period due to the initialization of the OSEs
with an analyzed field where all Argo observations were previously
assimilated. This period appears to be sufficient for the temperature field to
reach a stable state compared to observations, but the deep ocean still shows
a small drift in salinity. This spin-down time is still difficult to
evaluate in 1-year simulations where the evolution of the
Figure 3 shows the spatial distribution of the mean and RMS (root mean square) of the
temperature difference between Argo observations and the Run-NoArgo analysis.
RMS and mean statistics are calculated in 2
Heat content anomaly time series for
the 0–2000, 0–300 and 700–2000 m layers of the Run-Ref (blue),
Run-Argo2 (light blue) and Run-NoArgo (black):
Figure 3a shows that, in the 0–300 m layer, our model without Argo data
assimilation fails to correctly represent temperature fields over large
regions. As expected, errors are also larger in western boundary currents and
in the thermocline in the tropics where a small misplacement leads to large
temperature errors due to the sharpness of the thermocline. The RMS of the
differences between analyzed fields and Argo observations in these regions
reaches 1.5
19 December 2012, analyzed salinity fields – differences between
Run-Ref and Run-NoArgo at 100 m
RMS of salinity differences between Run-Ref and Run-NoArgo in the 0–300 and 700–2000 m layers for the last 6 months of the experiment.
Spatial distribution of the RMS and the mean salinity differences
between Run-Ref and in situ observations in the 0–300 and 700–2000 m
layers over the last 6 months of the experiments.
Salt content anomaly time series for the 0–2000, 0–300 and
700–2000 m layers from Run-Ref (blue), Run-NoArgo (black), and Run-Argo2
(light blue) and for the global ocean
In the 700–2000 m layer, errors are more spatially concentrated. The RMS
differences between Run-NoArgo-analyzed fields and Argo observations in the
western boundary currents is over 0.25
Figure 4 is similar to the previous figure but concerns salinity. Figure 4a shows the RMS differences between analyzed salinity and Argo observations in the 0–300 m layer. At that depth, mid-latitude oceans, northern Indian Ocean, North Pacific, Atlantic, western boundary current regions and part of the Southern Ocean show differences larger than 0.1 psu and could be considered as regions very sensitive to salinity observations, unlike the South Pacific, southern Indian Ocean and part of the Southern Ocean. The distribution of the mean salinity differences (Fig. 4c) shows particular patterns in the Southern Ocean between South Africa and Australia, in the eastern Indian Ocean and also in the South Atlantic. The analyzed ocean in these regions displays a strong positive salt bias.
From 700 to 2000 m (Fig. 4b), the North Atlantic basin, outflow regions, western boundary current regions and part of the Southern Ocean between South Africa and Australia have large RMS misfits. In this layer, the mean salinity misfit is greatest in the Southern Ocean and Mediterranean outflow region (Fig. 4d). In this depth range, the analysis is likely to be sensitive to Argo assimilation as there are very few other in situ data to constrain it.
In the next two subsections, we discuss the differences in analyzed temperature and
salinity fields from the surface down to a depth of 2000 m due to
the Argo data set assimilation. In each subsection, we study a snapshot of
the daily difference between OSEs with and without Argo data assimilation to
illustrate the effect of Argo data assimilation. We then use the spatial RMS
of the daily differences in the 0–300 and 700–2000 m layers to provide
quantitative information on the realism of the analyzed
Figure 5 shows the temperature differences between Run-Ref and Run-NoArgo for 19 December 2012 at 100 and 1000 m. We chose that date to illustrate the state of the analyzed ocean at the end of a year-long experiment.
The impacts of Argo observations at 100 m (Fig. 5a) are widely but unequally
distributed in the global ocean. Many regions show differences higher than
0.3
At 1000 m the influence of Argo observations is much more localized:
analyses for the North Atlantic, Agulhas Current, south of Australia and
southern Indian Ocean are strongly affected by Argo assimilation. In these
highly dynamic regions, differences between the two experiments reach
0.5
Figure 6 shows the RMS of the temperature differences between Run-Ref and Run-NoArgo. It is calculated from the daily differences of the last 6 months of the experiment. It quantifies the spatial distribution of the impact of Argo assimilation on the PSY3 data assimilation system. We chose to focus on the 0–300 m layer, where SST and SLA assimilation also plays a major role, in order to evaluate the importance of Argo temperature measurements in the upper ocean, and on the 700–2000 m layer to assess the impact of Argo profiles in this specific layer, where Argo is almost the only in situ observing system available.
Figure 6a shows that global ocean analyses are impacted by assimilating the
Argo data in the 0–300 m layer. The RMS of the temperature differences
between experiments reaches 0.5
In the 700–2000 m layer (Fig. 6b), the impact of Argo on analyzed
temperature fields is more localized in regions with high variability. In the
Gulf Stream, the Agulhas Current and around South Africa the RMS of the
temperature differences reaches 0.5
The map of RMS differences in the layer 300–700 m (not shown here) exhibits similar patterns to the 0–300 m layer but with smaller amplitudes.
Figure 7 shows the spatial distribution of the mean and RMS of the
temperature differences between Argo observations and the Run-Ref analysis
which assimilates Argo observations with satellite and other in situ
observations. The RMS and mean statistics are calculated in
2
Figure 8 shows the time series of heat content anomaly estimates in different
oceans and for different depth ranges. Heat content changes are calculated as
described in Von Schuckmann et al. (2009). Anomalies are obtained by
subtracting the 3-year mean (2011–2013) of PSY3-analyzed fields from the
OSE's analyzed temperature. Time series for the global ocean, the North Atlantic
(20–60
The global ocean heat content (hereafter referred to as GOHC) anomaly is an important diagnostic measure of changes in Earth's climate system (Levitus et al., 2005; Hansen et al., 2005). This diagnostic measure is often derived from Argo observations (e.g., Von Schuckmann et al., 2011, 2009; Willis et al., 2009; Trenberth and Fasullo, 2010) or other observing systems such as altimetry or through the closure of Earth's energy budget (e.g., Domingues et al., 2008; Cazenave and Llovel, 2010; Trenberth and Fasullo, 2010).
Figure 8a shows an overestimation of the GOHC anomaly calculated without Argo
observations that is significant compared to the variability of the system.
In the 0–2000 m layer, the GOHC anomaly difference between Run-Ref and
Run-NoArgo is around 0.8
RMS time series of the temperature innovations for Run-Ref
Vertical structure of the RMS of temperature innovations
RMS time series of the salinity forecast field and in situ salinity
differences for Run-Ref
Focusing on different ocean regions (Fig. 8b, c, d) shows that the impact of
Argo differs depending on regions and hemispheres. In the North Atlantic
Ocean, in the 700–2000 m layer, ocean analysis without Argo observation
gives warmer results than our best ocean estimate. The heat content anomaly
reaches 1.4
Vertical structure of the RMS of salinity innovations
These experiments show the high sensitivity of the ocean heat content estimation from the PSY3 analysis to the assimilation of the Argo observation array. This is especially true for depths below 700 m, where the variability is smaller than in the surface layer. Estimates differ if only half of the Argo floats are assimilated.
Figure 9 shows the salinity differences between Run-Ref and Run-NoArgo for 19 December 2012 at 100 and 1000 m. At 100 m, significant differences are spread all over the global ocean. Many regions are heavily affected by Argo assimilation, such as the tropical oceans and the Gulf Stream area. Around the Equator, the maximum amplitude of the daily differences is larger than 0.3 psu. The spread and amplitude of differences show the sensitivity of the PSY3 analysis to the assimilation of Argo salinity data.
At 1000 m, North Atlantic regions are very sensitive to the assimilation of Argo profiles, the differences reach 0.1 psu. The impact is greatest in the Mediterranean and Red Sea outflows and in the Gulf Stream areas. Regions with high variability are also significantly impacted: western boundary currents, southern Indian Ocean, Antarctic Circumpolar Current (ACC) and Arabian Sea. There is no significant difference in the Pacific Ocean at that depth for this date.
Figure 10 shows the RMS of the salinity differences between Run-Ref and Run-NoArgo. Similar results are obtained for salinity and temperature. In the 0–300 m layer, the impact of the assimilation of Argo salinity profiles is spread over all of the world's oceans. The RMS of the salinity difference between OSEs reaches 0.1 psu in most of the oceans. Tropical oceans and western boundary current regions are the most affected: RMS exceeds 0.2 psu in these areas. The tropical Atlantic is the most sensitive region. The Amazon outflow is extremely active and the sensitivity of the analyses to the assimilation of Argo profiles is clear to see.
It is also noticeable that the Labrador Sea, mainly along its shelf break, is very sensitive to Argo profile assimilation. The analyses for the “boundary” between the Indian Ocean and the Antarctic, where the subtropical front approaches the Agulhas Front, are also highly affected. The analysis of these different water masses is obviously very sensitive to Argo salinity and temperature profile assimilation.
In the 700–2000 m layer, the greatest impact is found in the North Atlantic. The RMS of the salinity differences reaches 0.1 psu along the European coast, due to the ill-positioning of the Mediterranean outflow in the model forecasts, and is around 0.05 psu elsewhere in the basin. As for temperature, high-variability regions are strongly affected by Argo profile assimilation: Arabian Sea, Agulhas Current region, South America west boundary region and southern Indian Ocean. Again, the Pacific Ocean is far less affected.
Figure 11 shows the spatial distribution of the mean and RMS of the salinity
difference between Argo observations and the Run-Ref analysis which
assimilates Argo observations together with satellite and other in situ
observations. The RMS and mean statistics are calculated in
2
Temperature forecast skill improvement derived from the reduction of the RMS temperature innovations.
Salinity forecast skill improvement derived from the reduction of the RMS salinity innovations.
Figure 12 shows the evolution of the daily salt content anomaly in different
regions of the global ocean and for different layers. The salt content
anomaly is calculated by subtracting the 3-year mean (2011–2013) salinity of
the Run-Ref from the OSEs' analyzed salinity. Time series for the global
ocean, North Atlantic (20–60
The global estimate of the salt anomaly (Fig. 12a) shows differences depending on whether it is calculated with Run-Ref, Run-Argo2 or Run-NoArgo. This masks some larger differences in regional estimates. Even the salt anomaly estimate in the surface layers from 0 to 300 m shows a strong sensitivity to the assimilation of Argo profile data, especially in the Southern Ocean. At depths between 700 and 2000 m, where Argo is nearly the only in situ observing system available, the impact on the estimated variability of the salt anomaly is significant compared to the natural variability, even on 1-year experiments. The results for the North Atlantic are the most heavily affected. The contribution of the 700–2000 m layer to the 0–2000 m layer salt anomaly is not negligible.
In most of the regions, salt content estimation differs depending on whether only half or the full Argo array is assimilated. The estimates obtained with half of the Argo array are, in most cases, closer to the estimate obtained with the full Argo array than the simulation without Argo, but the differences are still significant compared to the anomaly itself.
In this section, the impact of Argo data assimilation on short-term forecasts (< 7 days) is evaluated using the innovation (observation values minus model forecast values) statistics. OSEs forecast fields are compared with the Argo and other in situ observations. The statistics are computed over the last 6 months of the experiment.
Figure 13 shows the global average RMS of the temperature innovation from 0 to 2000 m for the last 6 months of the Run-Ref, Run-Argo2, Run-NoArgo and Free Run experiments.
For each experiment, the RMS is greatest at approximately 100 m and decreases with depth. The amplitude of the RMS temperature innovations below 1000 m is very low compared to the mixed layer depth values, but global variability at that depth is obviously also very low. The RMS of temperature innovation decreases with increasing quantity of Argo data assimilated. Atmospheric forcing and assimilation of SST and SLA may explain the good surface results for Run-Ref, Run-Argo2 and Run-NoArgo.
Figure 14a shows the temporal mean RMS temperature innovation profile of the
previous 6-month time series from the surface down to 2000 m. This is a
standard procedure for characterizing the performance of a forecasting system
and evaluating the impact of an observing system (Oke and Schiller, 2007;
Vidard et al., 2007; Fujii et al., 2014; Lea et al., 2014; Guinehut et al.,
2012). The maximum RMS is found at 100 m for the four experiments, but the RMS of
the innovations ranges from 1.4
On Figure 14b, each RMS temperature innovation profile shown in Fig. 14a is normalized with the Run-Ref RMS innovation profile, which represents our best forecast, shown in grey in Fig. 14a and b. We can then quantify the degradation of system performance in terms of temperature RMS error forecast due to the decrease of the number of Argo profiles assimilated. From those normalized profiles, we deduce an estimation of the percentage of degradation of the system performance for different depth ranges, summarized in Table 2. Coarsely, improvements range from 10 % in the 0–300 m layer to 50 % in the 700–2000 m layer. Assimilation of the first half of the Argo array improves the performance of the system by 15 % from the surface to 300 m depth and from 15 to 30 % in the 700–2000 m layer. The assimilation of the second part of the array improves the performance of the system by around 5 % in the 0–300 m depth and by 10–20 % in the 700–2000 m layer.
The same calculations are performed for salinity as for temperature. Figure 15 is a time series of the RMS of salinity profile innovations for the last 6 months of experiments. Run-Ref, Run-Argo2, Run-NoArgo and Free Run time series are represented here.
The RMS error is greatest at the surface and decreases with depth. From 0 to 2000 m the more salinity data are assimilated the closer to the observation the forecasts become. There is no significant increase of the innovation RMS during the 6-month experiment for Run-Ref and Run-Argo2 as there is in Run-NoArgo and Free Run. This increase becomes visible at around 300 m. This result demonstrates the importance of Argo observations for constraining salinity in the PSY3 system. Figure 16a shows the global mean absolute and normalized profiles of the RMS of salinity innovations for the different experiments. In the 0–300 m layer, the RMS innovation improvement depends on the quantity of Argo data assimilated by the system. Figure 16b shows that the RMS innovation is reduced by 20 % when the first half of Argo profiles is assimilated, compared to the RMS innovation without Argo data assimilated. The assimilation of the second half of the Argo data set reduces it by a further 5–10 % relative to the best scores (Run-Ref in blue). In the 700–2000 m layer, the increase of the quantity of Argo data assimilated, together with SLA and SST, induces a decrease of the RMS misfit in salinity.
This again shows the need for a good coverage of in situ profiles to estimate
a coherent
To validate the assimilation process, we briefly look at the impact of the assimilation of temperature and salinity profiles on other model-forecasted variables. It allows checking the physical consistency of the increment with the model physics. An “unbalanced” increment will destroy model equilibriums.
The global RMS of the SST innovations does not differ in the simulations with
and without Argo data assimilated. The mean RMS is close to 0.6
The global RMS of the SSH innovations also does not differ significantly
between the OSEs with and without assimilated Argo
Observing system experiments were carried out with the Mercator Ocean
0.25
The quality of the 3-D temperature and salinity analyses without Argo observations was first assessed. This highlighted the system's weaknesses when only SST, SLA and non-Argo in situ data are assimilated. Without Argo data assimilated, large errors are found in the western boundary currents, Antarctic Circumpolar Current, the Mediterranean and Red Sea outflows and in the tropics.
The effect of Argo data assimilation was then assessed through the comparison
of the analyzed temperature and salinity fields over the last 6 months from
the different experiments. The comparison of the Run-Ref and Run-NoArgo
experiments highlights the high sensitivity of the analyses to Argo data
assimilation. The 6-month RMS differences of daily fields between these
experiments easily reach 1
We show that the changes seen in the analyzed PSY3 temperature and salinity fields when Argo is assimilated correspond to an improvement of the analysis and forecast fields in terms of innovation and residuals to in situ observations. This shows the ability of the data assimilation system to take advantage of the Argo observations. The progressive improvement of the system's forecasting skills from assimilation of half of the Argo array to the full Argo array also indicates that all observations are needed to constrain our system. These results highlight the major importance of Argo data assimilation for operational oceanography. A decrease in the existing coverage of the Argo array would lead to a degradation of the PSY3 global ocean analysis and forecasts.
Finally, it is important to bear in mind that results from OSEs depend on the modeling and data assimilation system used. Our study takes place in the context of short-term real-time ocean analyses and forecasts. The conclusions could not be generalized to other ocean reanalyses without further investigation. The impact of Argo data assimilation on the other model variables also has to be further investigated. Here we focus on the impact of Argo observations on the reconstruction of 3-D temperature and salinity fields. Furthermore, we did not keep any in situ independent data sets as we stayed close to the real-time PSY3 system. General statements about observing systems should only be made with caution unless consistent results based on several systems are obtained. This is the approach promoted by the GODAE OceanView OSE/OSSE (observing system experiment/observing system simulation experiment) task team.
The research leading to these results received funding from the European FP7 program under the E-AIMS project. The authors would like to thank Jean-Michel Lellouche and Olivier Le Galloudec for providing support on the use of the operational system, Charly Régnier for providing the diagnostic tools and Gilles Garric and Yann Drillet for comments on an earlier version of the manuscript.
The altimeter products were produced by Ssalto/Duacs and distributed by Aviso
with support from CNES (
MDT_CNES-CLS09 was produced by the CLS Space Oceanography Division and
distributed by Aviso, with support from CNES
(