Download as pdf or txt
Download as pdf or txt
You are on page 1of 31

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228733653

Surface salinity in the Atlantic Ocean (30 S–50 N)

Article  in  Progress In Oceanography · April 2007


DOI: 10.1016/j.pocean.2006.11.004

CITATIONS READS

109 1,578

4 authors, including:

Gilles Reverfdin Thierry Delcroix


LOCEAN Institute of Research for Development
330 PUBLICATIONS   9,631 CITATIONS    155 PUBLICATIONS   6,655 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

sea surface salinity View project

MEOP: Marine Animals Exploring the Oceans Pole to Pole View project

All content following this page was uploaded by Thierry Delcroix on 31 January 2018.

The user has requested enhancement of the downloaded file.


Progress in
Oceanography
Progress in Oceanography 73 (2007) 311–340
www.elsevier.com/locate/pocean

Surface salinity in the Atlantic Ocean (30S–50N)


a,*
G. Reverdin , E. Kestenare b, C. Frankignoul a, T. Delcroix b

a
Laboratoire d’Océanographie et de climat, Expérimentation et approches numériques, Institut Pierre Simon Laplace,
Université Pierre et Marie Curie, case 100, 4 pl. Jussieu, 75252 Paris, Cedex 05, France
b
LEGOS/OMP, IRD, Toulouse, France

Accepted 19 November 2006


Available online 4 May 2007

Abstract

Sea surface salinity (SSS) data in the Atlantic Ocean is investigated between 50N and 30S based on data collected
mostly during the period 1977–2002. Monthly mapping of SSS is done to extract the large-scale variability. This mapped
variability indicates fairly long (seasonal) time scales outside the equatorial region. The spatial scales of the seasonal anom-
alies are regional, but not basin-wide (typically 500–1000 km). These seasonal SSS anomalies are found to respond with a
1–2 month lag to freshwater flux anomalies at the air–sea interface or to the horizontal Ekman advection. This relation
presents a seasonal cycle in the northern subtropics and north-east Atlantic indicating that the late-boreal spring/summer
season is less active than the boreal winter/early-spring season in forcing the seasonal SSS variability. In the north-eastern
mid-latitude Atlantic, SSS is positively correlated to SST, with SSS slightly lagging SST. There are noticeable long-lasting
larger-scale signals overlaid on this regional variability. Part of it is related to known climate signals, for example ENSO
and NAO. A linear trend is present during the first half of the period in some parts of the basin (usually towards increasing
salinities, at least between 20N and 45N). Based on a linear regression analysis, these signals combined can locally rep-
resent up to 20% of SSS variance (in particular near 30N/60W or 40N/10–30W), but usually represent less than 10% of
the variance.
 2007 Elsevier Ltd. All rights reserved.

1. Introduction

Sea surface salinity (SSS) is thought to play an important role at high latitudes in contributing to the var-
iability of surface density, and therefore deep water formation in the northern North Atlantic and neighbour-
ing seas, hence affecting the intensity of the meridional circulation (THC) (Rahmstorf, 1995; Häkkinen, 1999).
Because of the THC, part of the SSS variability in the high latitudes of the North Atlantic is expected to orig-
inate from lower latitudes. For instance, the modelling hindcast study of Häkkinen (2002) suggests that
changes in THC and therefore transport of high SSS from the south plays a role in decadal variability of
SSS in the North Atlantic subpolar gyre.

*
Corresponding author. Tel.: +33 1 44272342; Fax: +33 1 44273805.
E-mail address: gilles.reverdin@lodyc.jussieu.fr (G. Reverdin).

0079-6611/$ - see front matter  2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.pocean.2006.11.004
312 G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340

Near surface salinity also plays an important role at lower latitudes where its stratification can control the
depth of the mixed layer therefore potentially regulating the heat and momentum exchanges between the
ocean and the atmosphere (Lukas and Lindstrom, 1991). It will also contribute to changes in the properties
of thermocline and mode waters in regions of subduction, for example in the eastern subtropics or just south
of the Gulf Stream.
Information on recent SSS variability in the subtropical or tropical Atlantic is available from a few analyses
of data and/or from model simulations. Using hydrographic data, Levitus (1989) illustrates higher SSS in the
North Atlantic subtropical gyre in the early 1970s than in the late 1950s. This data set is however not sufficient
to construct a continuous time series and identify the time scales of the variability. Based on the same data
grouped by pentads, Curry et al. (2003) illustrate the increase of upper ocean salinity in the northern tropical
and subtropical regions of the Atlantic Ocean in the 1960s and 1970s. This contrasts with a decrease observed
in the North Atlantic subpolar gyre in the 1970s to mid-1990s near the surface (Reverdin et al., 2002) and
through the water column (Curry et al., 2003). The similarity between surface and upper ocean salinity var-
iability is also found in the analysis of linear trends for 1955–1998 by Boyer et al. (2005). Dessier and Donguy
(1994) (later on, DD) were the first to combine sufficient data to resolve the interannual variations in tropical
Atlantic SSS. They mostly reconstructed the mean seasonal variability, but illustrated a low-frequency deca-
dal-like variability in the northern part of the tropical Atlantic during 1977–1989, reminiscent of the one iden-
tified in sea surface temperature (SST). The surface data in the subpolar gyre (Reverdin et al., 2002) also
resolve shorter term, but often multi-annual variability, which can be to some extent related to air–sea fresh-
water fluxes (Josey and Marsh, 2005), or to influx of freshwater from the Arctic or from the fresher western
part of the subpolar gyre (the Great Salinity anomaly of the late 1960s and 1970s (Dickson et al., 1988; Häkki-
nen, 1993), as well as a later freshwater pulse (Belkin et al., 1998)).
We wish to extend the analysis of SSS variability for the recent period to lower latitudes of the North Atlan-
tic with a nearly seasonal time resolution. This paper presents an attempt at doing that for the period 1977–
2002 using a new product of mapped SSS starting in 1970 in the Atlantic Ocean between 50N and 30S. We
will first try to evaluate the usefulness of this mapped product. We have two major scientific objectives.

1. to seek the scales of the seasonal signals, and find to which extent regional variability results from local
processes (air–sea freshwater fluxes or Ekman advection). This will be done separately for different seasons.
2. investigate larger scales of the variability (often also at lower frequencies) and find to which extent they are
associated to known modes of climate variability.

The data and mapping methods are briefly outlined in Section 3, and some characteristics of these fields are
then presented in Section 4. Then (Section 5) we investigate to which extent the regional variability on seasonal
time scales result from the forcing terms, using primarily a stochastic forcing framework outlined in earlier
studies (Hall and Manabe, 1997; Mignot and Frankignoul, 2003). Finally (Section 6), we identify low-fre-
quency variability in the data and investigate whether some of it is directly related to known modes of vari-
ability (in particular the North Atlantic Oscillation (NAO) and the El Niño/Southern Oscillation (ENSO)).

2. Background

Limited information exists on the time and space scales for sea surface salinity variability in the Atlantic
Ocean, and often is constrained to well-sampled shipping routes (Delcroix et al., 2005). Model studies are cur-
rently our main source of information on the spatial and temporal characteristics of the salinity signals.
Häkkinen (2002) discusses the SSS variability during the last 50 years in the northern Atlantic Ocean. Her
Fig. 4 suggests a large variability at all frequency ranges (from interannual to decadal) exists, with particularly
large anomalies at periods less than 4 years just south of 20N between 40W and 60W (east of the lesser
Antilles), and at lower frequencies between 20N and 28N across the Atlantic, but also further south between
40W and 60W. Further investigation on the mechanisms of SSS variability was carried in the fairly realistic,
but low resolution Bergen coupled climate model (Mignot and Frankignoul, 2003), identifying in large parts of
the tropical/subtropical North Atlantic a variability resulting from NAO and ENSO through both freshwater
forcing (E  P: evaporation minus precipitation) and anomalous advection, in particular by Ekman currents.
G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340 313

The spatial distribution of variability in this model has also been shown to be dependent on the frequency
range, with signals associated with ENSO and NAO dominating at interannual frequencies in the tropical/
subtropical Atlantic (Mignot and Frankignoul, 2004). This model study also suggested that the spatial scale
of variability of SSS is somewhat smaller than for SST.
Studies of the response of SSS to stochastic forcing (Spall, 1993; Hall and Manabe, 1997; Mignot and
Frankignoul, 2003) illustrate how SSS responds to stochastic air–sea fluxes, in particular through E  P forc-
ing, but also through horizontal advection by Ekman currents (in the presence of an horizontal gradient in the
SSS field). These studies illustrate that away from the equator these forcing terms (there is no direct local feed-
back of SSS on air–sea fluxes) usually explain reasonably well the magnitude of SSS variability and its spectral
distribution at interannual or higher frequencies. These studies however do not explicitly take into account the
seasonal cycle in stratification and vertical mixing, nor the effect of geostrophic horizontal advection, and do
not pretend to be appropriate for the larger-scale variability that has often multi-annual time scales.
At the lower frequencies, horizontal advection by the currents is expected to strongly modify the patterns
imprinted by the atmospheric forcing (Mignot and Frankignoul, 2003) and may induce some spatial homo-
geneity over gyre scales. The imprint (time integration) of the forcing might however still be found, as dis-
cussed for the eastern part of the North Atlantic subpolar gyre according to Josey and Marsh (2005). Near
the equator, the interplay of forcing and dynamics may result in a different spectrum of the signal than in
the forcing, and we expect there a strong seasonal dependence of the SSS variability. For example, Ferry
and Reverdin (2004) indicated a large seasonal modulation of the anomalies in the region off the South Amer-
ican shelves north of the equator based on an analysis of an eddy-permitting resolution model. Off-shelf anom-
alies appeared in boreal spring due to anomalous advection of shelf water and disappeared in boreal winter as
a result of vertical mixing and advection with little contribution of anomalous freshwater flux with the atmo-
sphere (also found in longer simulations presented in Ayina, 2002).
Part of this large-scale variability in the air–sea fluxes is directly related to known modes of climate vari-
ability. Some are specific to one season, but have exhibited some interannual persistence (for example, the
North Atlantic Oscillation in winter). Others have a multi-season presence, although with seasonally changing
forcing patterns and no interannual persistence (ENSO, for example). The modes of variability of the air–sea
fluxes in the tropical Atlantic are also typically large scale. Some have little persistence beyond one season (the
equatorial mode, Wang, 2002), whereas others seem not have some persistence from 1 year to another (the
gradient mode). Our intention is not to discuss these different modes of climate variability already presented
in numerous studies, but to point out that because the air–sea fluxes associated to these modes are large scale,
they imprint large scale signals in SSS that may persist and propagate over more seasons and that need to be
identified and illustrated. There have also been trends in the air–sea fluxes in the North Atlantic, for instance
associated with NAO (Hilmer and Yung, 2000), as well as changes in the nature and structure of ENSO and its
connection to the Atlantic sector (Chiang et al., 2000) that may result in changes in the associated signatures in
SSS, and possibly in trends in SSS. In some modelling studies of the effect of anthropogenic CO2 increase on
the THC (Latif et al., 2000; Thorpe et al., 2001), this last effect is present: a salinity increase in the Atlantic is
attributed to changes in the net freshwater budget over the Atlantic sector associated with global warming and
an increase in the frequency of El Niño/Southern oscillation (ENSO) events in the equatorial Pacific. This
modeled salinity increase in the Atlantic counterbalances the reduction in THC intensity due to the increase
in temperature of the surface waters.

3. Data and mapping methods

3.1. Data sources

The salinity data originate from a variety of sources, mostly from underway collection on research vessels
and voluntary observing ships (VOS), with the addition of profile data (preferentially at the uppermost level
located between 5 m and 15 m depth), either from CTD or bottle casts (from NODC), autonomous profiling
floats (PALACE, 1997–2002), from moorings in the tropical Atlantic (PIRATA in 1998–2002), and from four
CARIOCA drifters (in 2001 during the POMME experiment, Mémery et al., in press). This is a rather inho-
mogeneous set, and issues of data origin and consistency will be discussed in a companion paper (see also,
314 G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340

Reverdin et al., 1994, for the early VOS data). Tests on CTD and XCTD profile data indicate that the mix of
data originating from different depths (between 2 m and more than 15 m) can result in a noise exceeding 0.10
PSS-78 in areas of large precipitation (in particular the ITCZ) or river outflow. This can even be a cause of
long-term trends as the data sources vary in time. Except for those areas, we find that the artificial variability
induced by these changes in data depth is small (on the order of 0.02 PSS-78) compared to the signals that we
investigate. We will now briefly list these different data sets.
Underway data from research vessels are mostly from thermosalinographs (TSG), usually collecting water
at a depth of 3–5 m. Research vessel data were extracted from the archive at the WOCE sea surface salinity
data assembly centre at IRD for the period 1989–1995. Other research vessels have contributed significantly to
the data base, in particular five French research vessels during dedicated cruises, and since late 2000 also dur-
ing transits, and three vessels from Antarctic research institutes during their transits between Europe and the
South Atlantic. Reduced (hourly) data from thermosalinographs (TSG) on board NOAA research vessels and
a few VOS were also obtained in near-real time (GOES transmission) from NOAA/AOML for 1996–2000.
The most extensive data set of VOS data is derived from a program managed by IRD (formally ORSTOM).
It provided data across a large part of the Atlantic Ocean south of 50N since 1977 on regular lines between
northern Europe and various destinations in Africa, the Antilles and South America (DD). In addition, data
have been collected mostly since 1991 along a line calling in particular in Mediterranean ports, northern Eur-
ope, eastern North America, and Panama. Surface data were also collected on French navy and hydrographic
vessels for the French Hydrographic Office (SHOM) and on VOS for the US fisheries (NMFS). Additional
surface data were obtained from various research projects (most recently in 2002 with the E.U. funded CAV-
ASSOO project). Initially, data were from water samples, with a progressive installation of TSGs between
1993 and 1997, associated with a strong reduction in the collection of water samples. The TSGs usually get
sea water from an intake at a depth between 3 m and more than 15 m, depending on the vessel and how it
is loaded. Part of these TSG data has been presented and used for a statistical description of the variability
in Delcroix et al. (2005). Most of the discrete surface data originate from a water sample drawn from a bucket,
although a subset of them was collected from an intake (in particular since 1991 on the CS Rimbaud and CS
London).
The PALACE float data were either from validated sets provided by the originators or from invalidated
sets for the most recent floats deployed within the ARGO project. The PIRATA mooring data have been
screened for errors and artificial trends removed, and near-surface (2 m) data complemented with subsurface
data when the temperature profiles indicated deep mixed layers and there was no rain.

3.2. Data gridding

The data are grouped in monthly, 1 square bins in latitude · longitude for the period 1970–2002, and
monthly deviations calculated with respect to a mean seasonal cycle which is based on all available data
for 1895–1990 (Reverdin et al., 1994; DD). The data probability distribution function within a bin is typ-
ically non Gaussian, both because of the occurrence of erroneous data and of the nature of SSS variability
(in particular freshwater lenses producing a low SSS tail in the distribution (Bingham et al., 2002)). When
there is enough data within a bin, the distribution can be characterized by the median, and the 25% and
75% percentiles. However, this is rarely the case, in particular before the advent of TSGs, with a common
occurrence of just 1 or 2 data in a bin. Therefore, only the average and rms standard deviation, estimated
after screening for outliers, are retained to characterize the SSS distribution in a bin. The screening is usu-
ally done on data which differ from the seasonal cycle by more than 3 · (ria + ris) or 0.3 (0.5 in the Gulf of
Guinea; note that we report salinity without units in the practical salinity scale PSS-78), whatever is the
largest of the two, ria and ris being estimates of the interannual and intra-bin variability (these are them-
selves estimated after a first screening of the data, and might underestimate the real variability). This typ-
ically results in removing 5–10% of the data in areas of large variability (western tropical Atlantic, Gulf of
Guinea, North American shelves), and somewhat less (except for specific data sets) elsewhere. The choice of
removing a large number of data in the areas of large variability was taken in order to reduce the influence
of very low SSS data in near-coastal areas when mapping SSS. This has the effect of introducing a positive
bias in the average SSS in those areas.
G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340 315

10000

5000

0
1970 1975 1980 1985 1990 1995 2000

50N
40N
30N
20N
10N
0
10S
20S
30S
90W 70W 50W 30W 10W 10E
Month

0 48 96 144 192 240 288

Fig. 1. Top panel indicates the number of 1 · 1 grid points with data in a month as a function of year (there are on the order of 65,000
grid points in a year). The lower panel provides the spatial distribution of the number of months with data in a 1 · 1 box for 1977–2002
(312 implies that every month in 1977–2002 had data).

The overall data density (Fig. 1, bottom panel) mapped as the number of months with binned data presents
a wide range. The central southern Atlantic south of 5S is virtually devoid of data which contrasts with large
density in areas corresponding to regular ship lines. The western part of the Caribbean Sea/Gulf of Mexico
and off the south-eastern USA is also irregularly sampled. In time, the number of bins with data has also var-
ied greatly (Fig. 1, top panel). It is lowest before 1977, but much higher thereafter, although there is a low in
the late 1980s and early 1990s. In specific regions, the sampling often shows different characteristics, with fairly
long data voids. This is for instance the case in the Caribbean Sea between 1987 and 1992, or in the Gulf of
Guinea in 1980–1982 and 1986–1988. Even along the regular ship lines, it is quite common to have data gaps
of 2 or more consecutive months, in particular since the installation of TSGs whose technical problems are
only fixed during ship calls at irregular intervals.

3.3. Spatial and temporal coherency in gridded SSS

The binned deviations from the average seasonal cycle (1977–2002) are used to evaluate the space–time cor-
relation scales of the low-frequency variability, which are likely to vary geographically and seasonally. To esti-
mate the scales reliably, sub-domains should be considered that include enough grid points while being, when
possible, homogeneous in the SSS properties. For this purpose, we retained eight regions shown in Fig. 2 (top).
Out of these, only the regions ST and IG (interior subtropical gyre and north-eastern Atlantic) are close to
satisfy the homogeneity requirement. The year was divided into two seasons, which roughly correspond in
the northern tropics and mid-latitudes to a season with strong salinity near-surface stratification (June–
November) and a less stratified season (December–May). The spatial scales are estimated from cross-correla-
tion functions of the binned data as a function of longitude (Lx) and latitude (Ly). An exponential function is
fitted to estimate the scales Lx and Ly (excluding the point at 0-separation and separations larger than 7). The
fit is usually good as illustrated on the middle and bottom panels of Fig. 2. The value of the fitted function at
the origin, in the range 0.2–0.6, provides the portion of the variance in binned data corresponding to a large-
scale signal (the other part of the variance resulting from intra-bin variability and data errors).
316 G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340

50N
40N W IG
30N
W ST
20N
WT
10N
E
0 GG
10S
20S S

30S
90W 70W 50W 30W 10W 10E
0.6
Winter
0.4
Longitude
Latitude
0.2

0.6
Summer
0.4

0.2

0 2 4 6 8 10 12
Distance ˚

Fig. 2. Top panel, the regions selected. Lower panels: average correlation functions of the gridded (1 · 1 · 1month) deviations from the
average seasonal cycle for regions IG + ST. The curves represent exponential fits to the function between lags 1 and 7, and the values in
latitude (dashed line) and longitude (full) are separated. Winter is the season December through May and summer, from June–November.

The variables Lx, Ly (Table 1) are usually well defined, except for region WT (near South America in the
tropics) where they are small and poorly derived from the data. In most regions outside of the equator band,
Lx and Ly are comparable, while near the equator (in particular E), Lx is much larger than Ly. They are largest

Table 1
Correlation scales Lx, Ly and Lt of the gridded data for different regions and seasons
Region Parameter
Lx, winter Lx, summer Ly, winter Ly, summer Lt
E (equ. Atl.) 3 4.5 2 2.5 3
W (W Atl.) 3.5 3 3 2.5 5
ST (N subtropics) 2.6 3.8 2.6 3.1 6.2
WT (W trop. Atl) 2 2 2 2 2.5
IG (Intergyre) 3.8 3.7 4.2 3.6 6.0
The scales are estimated from an exponential fit to the correlation functions at lags 1–7 for Lx and Ly (in degrees), and 1–9 months for Lt
(in months). The letters refer to the regions presented on Fig. 2, and two seasons are distinguished S for June–November, and W for
December–May. Results are most accurate for regions ST and IG where they are reported with one decimal digit, and less so for the
others.
G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340 317

at mid-latitudes (region IG) where winter values exceed summer values (the non-reported signal/noise ratio is
also larger). In the subtropics (ST) and in the tropics (E), however, the scales are larger in summer.
A time scale Lt is estimated similarly from the gridded anomaly data, but without differencing the seasons
(Table 1). It is large in the subtropical region as well as in the north-east and north-west Atlantic (5–6 months),
and smaller in the tropics (3 months or less). Interestingly, these values are in the bulk range of what was
found in the Bergen Climate Model (Mignot and Frankignoul, 2003), as well as when using along-track obser-
vations and a Gaussian instead of exponential fitting function in the tropical region (Delcroix et al., 2005).

3.4. Forcing data

To interpret the surface salinity (SSS) variability, one usually considers its tendency equation:
d g SSS=dt ¼ 1=hfSSSðE  P Þ  k  ðs=f Þ:rSSSg þ R; ð1Þ
where dg/dt corresponds to the tendency following the surface geostrophic current (includes horizontal geo-
strophic advection), the first terms on the right hand side can be thought as external forcing terms because
they depend on air–sea exchanges that are not directly controlled by SSS, h being a scaling depth, sometimes
identified with a mixing depth (see Mignot and Frankignoul (2003) for further discussion or De Boyer-Monté-
gut et al. (2004) for a discussion of the concept of mixed layer). The first term corresponds to the freshwater
flux E  P, where E and P are the net evaporation and precipitation; the second term will be referred to as the
Ekman flux, where k is the unit vector pointing upward, f the Coriolis parameter, s the wind stress. R corre-
sponds to entrainment, as well as vertical and horizontal mixing terms.
Focusing on the better sampled 1977–2002 period, we have considered a variety of data sets, and, on the
basis of the best comparison with the observed SSS, have retained the following combination. E, and s were
taken from the monthly reanalysis product ERA40 of ECMWF (Simmons, A.J., J.K. Gibson, 2000, ERA-
40 Project Report Series No. 1, unpublished document available on www.ecmwf.int/research/era). P is from
a combination of satellite and in situ products constructed mostly from the CMAP (1979 to August 2002)
blended product (Xie and Arkin, 1997) to which we added the 1977–1978 in situ COADS estimates (Da Silva
et al., 1994) after adjusting its seasonal cycle to the CMAP one over the common period (1979–1989), and the
September–December 2002 data from GPCP (Huffman et al., 1997; CMAP and GPCP present a close variabil-
ity over the Atlantic Ocean for the common period). The CMAP set retained is a recent 1 · 1 gridded product
computed over the ocean mostly from satellite-derived precipitation. Using instead precipitation from reanal-
ysis products (either from the NCEP or ECMWF models) resulted in slightly degraded comparisons to SSS.

3.5. Mapping of the seasonal cycle

The binned anomaly data are noisy and distributed irregularly. Furthermore, they are initially estimated
with respect to a climatological seasonal cycle based on earlier data (a combination of Reverdin et al.
(1994) with DD climatology for 25N–25S). To update the seasonal cycle, we proceed as follows. First,
the anomalies are averaged for each calendar month and spatial grid point. Assuming that individual years
are independent, the error variance is 1=N ia r2ia , where r2ia is the estimated interannual variance and Nia the
number of binned years. Then, objective mapping (Bretherton et al., 1976) is applied to the averaged devia-
tions of each calendar month, using the spatial correlation scales of Section 3.3 and, as guess field, the mapped
corrections for either the previous or the following month (i.e., forward or backward in time). Indeed, we carry
the analysis both forward and backward in time over two seasonal cycles. The last seasonal cycles of the two
analyses are averaged to correct the seasonal cycle and recompute monthly deviations. Mapping of the
monthly deviations is then performed for each year by a similar objective mapping as discussed in Section
3.6. An average seasonal cycle of the analysed deviations for 1977–2002 is then constructed, which is added
to the earlier estimate of the seasonal cycle, thereby providing a seasonal cycle for the period 1977–2002.
An illustration for March and September is presented on Fig. 3(top). The SSS seasonal cycle has been
described in various papers (Levitus, 1986; DD) and so will not be detailed here.
The deviations of this seasonal cycle from the initial climatology are on the order of 0.1 over large parts of
the Atlantic (Fig. 3 bottom for March and September). In the North Atlantic subtropical gyre and near 50N,
318 G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340

50N
40N
30N
20N
10N
0
10S March September
20S
30S
90W 70W 50W 30W 10W 10E 90W 70W 50W 30W 10W 10E

34 34.4 34.8 35.2 35.6 36 36.4 36.8 37.2 37.6 38

50N
40N
30N
20N
10N
0
10S March September
20S
30S
90W 70W 50W 30W 10W 10E 90W 70W 50W 30W 10W 10E

—0.45 —0.35 —0.25 —0.15 —0.05 0.05 0.15 0.25 0.35 0.45

Fig. 3. Upper panels: average SSS for March and September; lower panels: deviation of the new climatology with respect to the guess
fields we used based on earlier climatologies illustrated for the months of March and September.

they are negative (less salty), whereas in the South Atlantic subtropical gyre or near 40N, they are positive
(more salty). These large-scale differences are probably the signature of very low-frequency variability, the ini-
tial climatology being based to some extent on earlier data. In the near-coastal areas with low SSS, the new
analysis is more salty than the earlier one, mostly a result of removing low outliers in the data. The seasonal
cycle on the 1 grid also presents small-scale inhomogeneities that originate mostly from the earlier
climatology.

3.6. Objective mapping of monthly SSS anomalies

The sequential approach adopted is very similar to the objective mapping used for the average seasonal
cycle with the addition that we take into account the time correlation of the signal. The guess is the mapped
anomaly for the previous step (a month ahead or after) weighed by exp(1/s), where s is taken as 2 months
which is slightly less than the time scale Lt of the interannual signal (Table 1). Details of the analysis approach
are given in Appendix A. Two analyses are performed over 1970–2002, one forward and one backward, which
are then averaged. The average of the forward and backward (in time) analyses provides a more reliable esti-
mate of the SSS anomalies (SSS 0 ) which minimizes possible lags with the real variability. The differences
between the two are usually smaller than the anomalies portrayed (Fig. 4 top) and usually consistent with
the estimated errors (relative to ria, Fig. 4 bottom). Large errors appear in the southern Atlantic subtropical
gyres, a region where Fig. 4 (top) shows low values because there are so few data that the analyzed values are
nearly null most of the time. The comparison of Fig. 4 with the bottom panel of Fig. 1 indicates, as expected,
that anomalies are most reliable in areas with the highest data density. In addition, the error map presents the
discontinuity between regions (Fig. 2 top) associated with the different scales specified in each region (Table 1).
G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340 319

50N
40N
30N
20N
10N
0
10S
20S
30S
50N
40N
30N
20N
10N
0
10S
20S
30S
90W 70W 50W 30W 10W 10E

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

Fig. 4. Upper panel: ratio of the rms difference between the forward and backward analyses and the rms variability in the period 1977–
2002; lower panel, estimated average rms error normalized by the signal amplitude (boundaries of the domain in Fig. 2 are noticeable, as
the statistics (Table 1) used in the mapping are defined separately for each domain).

Examples of anomaly fields are presented for two typical months corresponding to two different data dis-
tributions but rather typical of what is witnessed for the beginning and end periods of the sampling (Fig. 5).
These fields suggest large scale anomalies at least in the subtropical gyres or in the northern inter-gyre region.
Elsewhere, they present a complicated pattern that to some extent results from the insufficient sampling (esti-
mated errors not shown, but corresponding roughly to the data density distributions).
We illustrate the analysis by a few time series in places where the data coverage is usually fairly good
(Fig. 6). In those regions the objective mapping monthly analysis fits fairly closely to the original data (assem-
bled in a 2 · 5 latitude · longitude box centred on the analysis point), even though large spikes in the ori-
ginal data time series are not reproduced and the month-to-month variability is much reduced as expected.
An extreme (and unusual) example is at 0W/10S where the very low salinity signal encountered during
the first part of 2000 is not reproduced in the analysis. This happens because the raw signal was of small merid-
ional extent (on the order of 1), a spatial scale that cannot be captured in the analysis.
In most other locations on Fig. 6, the objective mapping confirms the significance of the low-frequency
events. For example, there is a large jump (over 1.5) at the end of 1989 at 18N/55W between earlier positive
deviations and the later negative ones that last throughout 1990. These anomalies and other data in the area
suggest a particularly strong northward extension throughout 1990 of the freshwater tongue found east of the
lesser Antilles that does usually not extend that far north in the earlier part of the year. Smaller low frequency
fluctuations in the North Atlantic subtropical gyre or mid-latitudes are also well captured in the analysis (for
example, negative values in 1977–1978 at 44N/15W, as expected in this region from hydrographic data, Pol-
lard and Pu, 1985).
When the data coverage is insufficient, for example at 6S/35W or 10N/25W before 1977, the analysis
produces values too small compared with the later period. The variability in SSS 0 is less than the real one
320 G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340

50N
40N
30N
20N
10N
0
10S FMA 1978 JAS 1997
20S
30S
50N
40N
30N
20N
10N
0
10S March 1978 August 1997
20S
30S
90W 70W 50W 30W 10W 10E 90W 70W 50W 30W 10W 10E

—0.45 —0.35 —0.25 —0.15 —0.05 0.05 0.15 0.25 0.35 0.45

Fig. 5. Two analyses with data distribution during the 3-month period centered on the analysis date. Left panels, March 1978; right
panels, August 1997.

(a property of objective mapping, also verified by comparison with the a priori ria) by a factor that depends on
the data coverage. As we commented earlier, even for the better sampled period since 1977, the time series
reconstructed in areas too poorly sampled are not realistic (or present large gaps) because of the uneven data
distribution. The influence of data coverage is also seen in the spatial distribution of the rms variability of
SSS 0 . However, this effect is somewhat removed when considering only points for which the estimated analysis
relative error is relatively small (Fig. 7). Empirically, we find that mapped SSS 0 with a relative error of 0.8
seems to be retaining part of he signal, albeit with reduced amplitudes (possibly by up to a factor 2). Relative
errors in the South Atlantic subtropical gyre are often close to that level, which might explain why mapped
SSS 0 rms variability is smaller there than in the better sampled North Atlantic subtropical gyre.

4. SSS variability

4.1. Standard deviations

With the afore-mentioned caveat in mind, we will comment the most salient features in the standard devi-
ation of interannual variability and in its seasonal dependency (Fig. 7). During DJF, very large interannual
variability is found off the north-east coasts of South America (from the equator to the lesser Antilles), which
bulges to the northwest during MAM. During JJA and SON it spreads further offshore over the deep ocean: to
the east along 5–10N (the area of the retroflection of the North Brazil Current and North Equatorial Counter
Current (NECC)) and to the north-east near 60W east of the lesser Antilles, extending all the way to 20–
25N. In SON, this northern maximum is separated from the one on the shelf, and corresponds to a region
with very large SSS gradients (Fig. 3). In the tropical Atlantic interior, there is a maximum variability that
migrates from a southern position near the equator in MAM to a northern position in SON close to 10N
and is most pronounced in JJA and SON, seemingly following the average seasonal displacements of the
G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340 321

0.5
44N-15W

—0.5
0.5
20N-40W

—0.5
2
18N-55W
1

—1

—2
1
10N-25W
0.5

—0.5

—1
1
6S- 35W
0.5

—0.5

—1
1
10S-0E
0.5

—0.5

—1
1970 1972 1974 1976 1978 1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002
Year

Fig. 6. Time series of the gridded data (red) and the mapped data (blue) after removing the average seasonal cycle (the gridded anomalies
have been averaged over a 2 · 5 latitude · longitude box surrounding the chosen position (For interpretation of the references to colour
in the figure legend, the reader is referred to the web version of this article).

Inter-tropical Convergence Zone (ITCZ). In the eastern Gulf of Guinea, the region of large variability is the
smallest in JJA with minimum variability near the equator. These tropical regions of large interannual vari-
ability tend to mirror somewhat the regions of low SSS in the average seasonal cycle (Fig. 3, top).
322 G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340

50N
40N
30N
20N
10N
0
10S DJF MAM
20S
30S
50N
40N
30N
20N
10N
0
10S JJA SON
20S
30S
90W 70W 50W 30W 10W 10E 90W 70W 50W 30W 10W 10E

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

Fig. 7. Interannual rms variability in the analyses for four seasons (period 1977–2002, but only individual months with relative error less
than 0.8 are considered).

In MAM near South America there is an equatorial maximum in variability spreading to 5–10S. In this
south-western equatorial region, the lowest variability is found in SON-DJF, mimicking somewhat the rainfall
seasonal cycle in this region. In JJA and SON there is a relative minimum of variability between the equator
and 8S with larger values near 8–10S, which experiences no rainfall during this season but is located in the
SSS gradient to the north of the region of maximum SSS. This mirrors somewhat what is found near 18–20N
east of 50W except in boreal winter where there is a relative (weak) maximum, located southwest of the max-
imum SSS of the North Atlantic subtropical gyre (Fig. 3 top).
The subtropics north of 20N, as well as the regions further north away from the shelves and from the vicin-
ity of the Gulf Stream and North Atlantic Current, present usually low interannual variability with less indi-
cation of a seasonality in the interannual standard deviation. Notice however the band of a relative maximum
in variability cutting partially the region of low values of the subtropics and particularly noticeable in JJA. It is
located in the western and north-western flank of the region of maximum salinity. In the North Atlantic sub-
tropics west of 30W, there is a suggestion for a slightly smaller variability in DJF-MAM, whereas, further
east, at least for the well-sampled regions between 25N and 50N, the minimum occurs in MAM-JJA.

4.2. Lagged-correlation of the anomalies

The patterns of persistence or lagged correlations are estimated for regions with small estimated errors and
the results are plotted only when it includes at least 13 years or half the series length with estimated relative
errors less than 0.8.
Late boreal winter anomalies (FMA) have a large persistence (Fig. 8 top). Correlations usually remain posi-
tive over 12 months, except in the NW tropical Atlantic (5N–20N) where they tend to be lost after 6 months,
as well as near 30–35N east of 30W (the Azores Current) or in the North Atlantic Current (north of 45N).
G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340 323

50N
40N
30N
20N
10N
0
10S L= 6 (FMA/ASO) L= 9 (FMA/NDJ)
20S
30S
90W 70W 50W 30W 10W 10E 90W 70W 50W 30W 10W 10E

1
ST
SSS FMA IG
E
0.5 S

1
SSS FMA ST
IG

0.5 SSS JAS ST


IG

0 2 4 6 8 10 12 14 16 18
Lag months

Fig. 8. Seasonal auto-correlation of the SSS anomalies (the 0-lag refers to the season FMA). The top panels present maps at 6 and 9
month lags (the thick lines indicate non-zero correlations at the 90% confidence level, with the assumption that different years are
independent; only months with an estimated relative error less than 0.8 are included and a value is reported when there are at least 13 years
included). The lower panels present average auto-correlation functions of the mapped deviations for four domains of Fig. 2 and the
comparison of the auto-correlation function in two reference seasons (FMA and JAS) for domain ST (north Atlantic subtropical gyre) and
IG (inter-gyre region). Positive lags refer to later months in the seasonal cycle.

Areas where the correlation remains high at large lags include the north-east Atlantic near 10W/45N, the
band 20–25N east of 60W and areas inshore of the Gulf Stream (GS at 6–9 months) or just offshore of it
(at 12 months). The south-east tropical Atlantic presents large correlations up to 6 months disappearing then
during austral summer (9–12 months). The mapped correlations are usually only marginally within the 90%
confidence interval, but they remain when averaging over the different regions with the largest correlations
being for regions ST and IG (Fig. 8, middle panel). The regionally averaged lagged correlations obtained
for other seasons (for instance JAS, Fig. 8, lower panel) are quite similar with slightly lower correlations at
5–9 months lag (winter relative to the previous summer) between 16N and 35N (region ST), and to a lesser
extent for region IG further north. This drop of persistence is probably associated with the autumnal deepen-
ing of mixed layer and entrainment of subsurface water.
The SSS 0 persistence is usually higher than the one in SST 0 (here from the NCEP-NCAR reanalysis, not
shown) with much higher correlations from winter to summer (for SST 0 , these correlations are usually very
low south of 50N, not shown). At larger lags, the difference is less with the March–December correlations
being roughly comparable for SSS 0 and SST 0 in the north-west Atlantic or near 20–30N at least east of
324 G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340

50W (Fig. 8 for SSS 0 ; not shown for SST 0 ). In these areas, the correlation for SST 0 at those lags was related to
re-emergence of subsurface anomalies during the autumn–early winter deepening of the thermocline (Alexan-
der and Deser, 1995; de Coetlogon and Frankignoul, 2003). This increase of correlation for SST 0 is usually not
observed for SSS 0 , with the exception of the region of the NW Atlantic south-east of the Gulf Stream, the only
one for which there is an increase in correlation with previous winter anomalies from the late summer to the
early winter. This suggests that the re-emergence mechanism does not influence as strongly SSS 0 than SST 0 .

4.3. Spatial correlations

Is the persistence suggested on Fig. 8 associated with the ‘‘local scale’’ variability that dominates the cor-
relation functions (Fig. 2) or with larger scales? We will approach that by looking at lag-correlations patterns
in SSS 0 fields with respect to time series of SSS 0 at specified locations with low estimated errors. Large scale
patterns are often suggested by these analyses that tend to be relatively stationary in time for lags up to 1 year.
This is for example the case for a reference at 20N/30W, to the south-east of the subtropical salinity max-
imum (Fig. 3). The correlation at a 12 months lag remains significant in roughly the same areas as at lag 0
(Fig. 9, left panels) with a slight north-westward shift of maximum correlation (still larger than 0.5) to
23N/34W and the appearance of negative correlations further west in the subtropical Atlantic that were
not clearly observed at 0-lag. Note that the same characteristics are found (not shown) when the reference
point data are selected for a particular season or when the data are averaged over a year, reflecting the long
(multi-annual) time scales of that lagged pattern.
We found a few instances where the lagged correlation patterns change considerably in time, for example
when selecting 32N/65W, an area well sampled close to Bermuda south-east of the recirculation region of the
Gulf Stream (Fig. 9, right panels). In that case, the maximum correlation at 6 months lag seems to have shifted
to the northeast (30–40N/45W). There is also a pattern of positive correlation in the north-eastern Atlantic
which tends to extend further to the southwest at 12 months and 24 months. At those lags, it constitutes the
only positive correlation pattern. This evolution would be consistent with propagation around the subtropical
gyre with the average currents. The very rapid evolution in the western Atlantic is also consistent with its
stronger currents and active dynamics. However, the time series are too short and too marginally significant
for drawing possible mechanisms for these observed evolutions.
Commonly, the 0-lag correlation patterns presents significant correlations over a large part of the basin,
probably a result of common forcing mechanisms (either dynamical or external), and its deformation in time
is therefore hard to interpret. With the 20N/30W reference point, as well as others in the North Atlantic
subtropical gyre or in the north-east Atlantic, the positive correlations extend over a wide range of latitudes
from 15N to 45N (less so around 30–35N or north of the GS). The region of significant positive correlation
rarely extends zonally across the northern subtropical gyre. Strikingly, these point correlations using the
mapped product suggest spatial scales larger than the ones in the correlation function of Fig. 2 (although
the low correlation coefficients in these maps are not inconsistent with Fig. 2).

4.4. Correlation with SST

The scales in these point-correlation maps seem also somewhat shorter than the ones typical of SST (not
shown). This could be interpreted as resulting from large noise in the SSS maps, as the estimated errors suggest
that this contributes to it. On the other hand, the difference between the scales in SST and SSS is also found in
analyses of the variability in a coupled ocean–atmosphere model, where it is interpreted in terms of differences
in the forcing and in advection (Mignot and Frankignoul, 2003). Interestingly, the lagged correlations between
SST and SSS show significant correlation outside in the north-east Atlantic north of 36N (significantly non-
zero positive correlation at lags 1 to 4 months at the 95% level for IG, Fig. 10). This points out to joint pro-
cesses in SST and SSS variability in the north-east Atlantic with a correlation maximum when SSS lags SST by
1–2 months. The positive correlation at 0-lag cannot be caused by evaporation anomalies that would result in
negative SST 0 combined with positive SSS 0 , and therefore in a negative correlation at 0-lag. It is more likely
that the positive correlation at 0-lag results there from horizontal advection because of the horizontal SSS–
SST relationship relating high SST with high SSS. Frankignoul et al. (1998) pointed out that the negative feed-
G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340 325

SSS 20N30W SSS 32N65W

40N
30N
20N
10N
0
10S L= 0 L= 0
20S

40N
30N
20N
10N
0
10S L= 6 L= 6
20S

40N
30N
20N
10N
0
10S L= 9 L= 9
20S

40N
30N
20N
10N
0
10S L= 12 L= 12
20S

40N
30N
20N
10N
0
10S L= 24 L= 24
20S

90W 70W 50W 30W 10W 10E 90W 70W 50W 30W 10W 10E

–0.45 –0.35 –0.25 –0.15 –0.05 0.05 0.15 0.25 0.35 0.45

Fig. 9. Lagged-correlation maps of SSS at different lags (in months) with respect to SSS in a reference location. Left column for the
reference point at 20N/30W; right column for the reference point at 32N/65W. Only data in the time series with estimated relative
errors less than 0.8 are retained, and the correlation is presented when it is based on at least half the time steps. The thick contour
corresponds to the 90% confidence level of non-zero correlation (assuming that the number of degrees of freedom is half the number of
months in the time series).
326 G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340

0.5
0.4 SSS/SST ST IG E S
0.3
0.2
0.1
0

0 2 4 6 8 10 12
Lag months

Fig. 10. Cross-correlation functions between SST and SSS deviations from the average seasonal cycle (the lag is positive when SSS lags
SST). In IG, the lag-correlations are significantly non-zero at the 95% level at lags 1 to 4 months, considering correlation functions
obtained for different seasons as independent samples of the same function.

back of SST 0 on the latent heat fluxes which damps SST 0 anomalies would act in reinforcing the SSS 0 anomaly,
which would contribute to a positive lagged correlation. This would contribute to a higher persistence in SSS 0
than in SST 0 . The correlation in the equatorial Atlantic is negative and maximum at 0 lag, in particular when
SSS is chosen in DJF (not shown). There, either vertical or horizontal circulation or even evaporation could
result in this negative correlation maximum at 0-lag.

5. Local relation of salinity with the atmospheric forcing

This chapter is devoted in trying to identify a part of the SSS variability that can be related to air–sea exchanges
of freshwater and wind momentum. Wind stress contributes to Ekman currents, and therefore Ekman advection.
It also contributes to vertical mixing and vertical velocities, two mechanisms that we will not investigate.
Mignot and Frankignoul (2003) suggested in their model study an important contribution to SSS variabil-
ity by the freshwater and Ekman advection terms. We expect a signature of this forcing mechanism at fairly
high frequencies (seasonal or less; see also Delcroix et al., 2005), and that it contributes in particular to the
part of the SSS variance that was fairly local (Fig. 2). To the extent that the forcing terms have less persistence
than the SSS signal, they can be represented to a first approximation as white noise on the monthly time scale
(Hall and Manabe, 1997; Mignot and Frankignoul, 2003). Assuming that geostrophic advection and other
oceanic processes (vertical mixing, entrainment) can be considered as a source of dissipation for the local sig-
nal forced by freshwater fluxes and Ekman advection, the correlation between the forcing terms and SSS
would be positive at a positive lag (forcing leads). The 0-lag correlation is then expected to be smaller and
to result from the persistence of the forcing, which causes monthly means of the forcing to be slightly
auto-correlated. The slope of the regression function between SSS and forcing at 0-lag should be equal to
the forcing variance (assuming advection and dissipation to be small). Because the forcing term includes a
depth of penetration of the fluxes, this is in principle a way to estimate that depth (see [for an estimate of
the depth of penetration for SST], Cayan, 1992). This depth is expected to vary seasonally.
We adopted this stochastic approach and decided to smooth further the forcing terms with a 3-month run-
ning mean to make it equivalent to the SSS 0 smoothing caused by the analysis technique (i.e. data from three
successive months taken into consideration). This increases the correlations as the number of degrees of free-
dom is reduced, but should not change the basic results. The lagged correlation between SSS and the forcing
terms (lag-1 on Fig. 11) is usually positive between lags 0 and 2 months (SSS follows), which tends to support
this ‘stochastic forcing model’. We also investigated directly the correlation between dSSS/dt and the forcing
terms in Eq. (1), and found usually peak positive correlation at negative lags (between 2 and 0 month) with
negative (weak) correlations at positive (4–6 months) lags (not shown). This again supports the stochastic
model approach. The spatial lag-1 correlation pattern (lower panel of Fig. 11) shows mostly positive values
(not always significant) with a few blobs of negative values, most notably near 30N/70W and 40N/70W
on either sides of the Gulf Stream, as well as near 10–15N in the central Atlantic. The low correlations in
areas of strong currents (Gulf Stream or near north-eastern South America) could result from a large contri-
bution of advection to SSS 0 variability which could swamp ‘local’ forcing.
G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340 327

40N
30N
20N
10N
0 L= 1
10S SSS / (E-P)
20S

40N
30N
20N
10N
0
10S SSS / - (UE′. ∇S)/S

20S

40N
30N
20N
10N
0
10S SSS / (E - P) - (UE′. ∇S)/S

20S

90W 70W 50W 30W 10W 10E

–0.45 –0.35 –0.25 –0.15 –0.05 0.05 0.15 0.25 0.35 0.45

Fig. 11. Correlation between SSS and different forcing terms at lag 1 (SSS follows the forcing by 1 month). For the Ekman term, the 5N–
5S band has been removed. The thick contour is for the 90% confidence interval (same selection rule as for Fig. 9).

Then, we consider separately the contributions of the two forcing terms (freshwater and Ekman, upper and
middle panels of Fig. 11) to the pattern of lag-1 correlation. The anomalies of the advection by Ekman cur-
rents (estimated poleward of 5 of the equator) have often a larger magnitude than the freshwater fluxes (as in
model results, Mignot and Frankignoul (2003), in particular equatorward of 20N–20S). However, we find a
main contribution from the freshwater fluxes to the correlation pattern. It almost everywhere contributes to
positive correlations at lag 1–2 months (as found in the model study of Mignot and Frankignoul, 2003) with
low values south of the Gulf Stream, very close to Africa north of the equator, or off north-west South Amer-
ica between 0N and 15N. It also presents a large positive correlation near the equator, in particular in the
western Atlantic, but peaking at lag 0, and therefore should not be thought after in the framework of a sto-
chastic model. The contribution of advection by Ekman currents at lags 0–2 months (or to SSS tendency, not
shown) is generally less than for the freshwater fluxes, except off North Africa or near the Iberian peninsula.
Between 5N and 16N, the correlation for the Ekman term is mostly negative, as is the case of areas off North
America, in particular west of 60W north of the Gulf Stream. In this last region, it might coincide with the
impact of NAO to the Gulf Stream position. Positive NAO is associated with a northward displacement of the
328 G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340

Gulf Stream (Joyce et al., 2000; Frankignoul et al., 2001), and therefore an increase of SSS north of the Gulf
Stream, but is also associated with an anomalously negative contribution of Ekman advection (higher winds).
In the western subtropical Atlantic (west of 50W), including the Ekman contribution resulted in a decrease of
the correlation with the forcing terms, whereas a slight increase in correlation is usually found further east.
To get a clearer sense of the time series of lagged correlations, the local correlation coefficients are then
averaged in the different boxes in Fig. 2. The curves in the lower panel of Fig. 12 have a similar shape, either
peaking at lag 0 or 1 month (near the equator: regions E, S) or at lag 1 or 2 months (SSS follows forcing term)
with a slow decrease at longer lags; the correlations are higher for these higher latitude regions. The results are
robust, as indicated by similar curves when averaging the lagged correlation functions over latitudinal bands
or only over the points that present a significant correlation at lag 1, or even one half of each domain. How-
ever, there probably is a strong seasonal dependence in the correlation function masked by this averaging. For
example, in the northern subtropics (region ST) and more so in the north-east Atlantic (region IG), lagged
correlation with SSS in DJF peaks for the forcing 1 month earlier, whereas in summer (SSS in JJA), the peak
is 3 or 4 months earlier (in March/April). This indicates that in these regions summer SSS is most responsive to
the late winter–early spring forcing, when the ocean re-stratifies. There is a suggestion of a similar seasonality
in the southern hemisphere (with a 6 months shift). These results are coherent with the forcing terms being
weaker at the extra-equatorial latitudes during late spring and summer than during the preceding season,
which would imprint more SSS. Near the equator (E), the correlation seems to be mostly present in boreal
summer, for which we have no explanation.
Although we find a correlation suggestive of an SSS signal forced by freshwater fluxes and Ekman advec-
tion, we also need to find whether the observed magnitudes are compatible with this mode of forcing, i.e., are
the depth h* of penetration that can be estimated in the stochastic model framework, compatible with
observed mixed layer depths? A qualitative estimation of h* is done for different seasons assuming a time scale
of 1 month, and based on the value of the co-variance function at lag 1 (the slope at the origin is too variable
and poorly determined to estimate h* from its forcing variance; we also could not use directly Eq. (1), which

0.5
0.4 SSS DJF ST IG E S
0.3
0.2
0.1
0

0.5
0.4 SSS JJA
0.3
0.2
0.1
0

0.5
0.4 SSS
0.3
0.2
0.1
0

0 2 4 6 8 10 12
Lag months

Fig. 12. Averaged correlation functions for four regions of Fig. 2 between SSS and the forcing term E  P-Ekman (positive lags when SSS
follows the forcing terms). The upper panel is for SSS in DJF; the middle panel is for SSS in JJA; the lower panel includes SSS for all
months.
G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340 329

was too noisy). The values found for the depth h* are not unreasonable outside of the equatorial band, when
retaining only points with large enough correlations. Estimates north of 15N tend to be on the order of 50 m
for JJA and 100 m or more in JFM. The points retained are however too sparse to draw a map that could be
compared with climatological mixed layer depth, a quantity expected to be related to h*. Closer to the equator,
h* is too large and too variable, further indication that the stochastic model is not adequate in these regions.
This chapter indicates that in a large part of the extra-equatorial Atlantic, part of the SSS variance is related
to local forcing through freshwater fluxes. Not surprisingly, the correlations are low, as this includes only a
portion of the forcing mechanisms of SSS variability. Although these correlations are of a similar magnitude
to those found in the model study of Mignot and Frankignoul (2003), we expect also a contribution of errors
in the analyzed fields.

6. Low frequencies

The previous discussion focused on local relationship to forcing. A significant part of the local variance has
little persistence, but when considering the persistence in the field or the response to forcing, there is also an
indication of lower frequencies SSS signals on Fig. 8. Although the atmospheric forcing has local effects on
SSS, we also find large scale signals in the point correlation maps of Fig. 9. This is not surprising as there
is a number of forcing mechanisms that favour those scales. First the atmospheric forcing is also organized
at large scales (NAO, ENSO, tropical Atlantic dipole. . .). Second, at low frequencies the mean circulation will
redistribute the surface water over gyre scales. Thirdly, the low-frequency variability of the circulation is also
expected to have a large-scale component (for example inter-decadal changes in gyre strength described in
Curry and McCartney (2001)); or the trend in subtropical gyre strength during the last decade discussed in
Volkov (2004).
In this section, we will briefly discuss the oceanic signals associated with major known modes of climate
variability (mostly NAO and ENSO). Then, we will present the SSS signals averaged over some parts of
the North Atlantic. We separate the region in oceanographic domains avoiding regions with too large and
poorly resolved variability: (1) an inter-gyre region between 36N and 47N east of 35W (coverage becomes
poorer further north or west); (2) eastern (20N–36N) and (3) western (24N–36N) portions of the North
Atlantic subtropical gyre (the separation between the two at 50W corresponds to a change in the seasonal
cycle of the forcing terms and also to where the relation between forcing and local variability drops
(Fig. 11); we do not include the Gulf of Mexico into the western domain).
A way to assess the SSS associated with a particular climate mode is to regress SSS 0 fields onto index time
series associated with this mode, possibly with a lag. This assumes a linear response of the upper ocean to the
mode. Different indices were adopted for NAO and ENSO, resulting in some differences in the SSS regressed
fields, with however similar major characteristics, suggesting that the analysis is fairly sturdy despite a small
time series length. The signature of these modes in the Atlantic sector is probably modulated by inter-decadal
variability. Fortunately, the period investigated is fairly homogenous in that respect. It is characterized by
NAO centers of action located particularly far west from 1977 to 1998 (Hilmer and Yung, 2000) and a ten-
dency to have a high proportion of positive NAO events (based on the classical station-based indices, Hurrell,
1995). It also corresponds to a period with an anomalous warming of the central equatorial Pacific, and a
clearer connection of the equatorial Atlantic to the Pacific than was observed in the previous 20 years (Chiang
et al., 2000; Giannini et al., 2001).

6.1. Index time series

To investigate the influence of ENSO, we use the Southern Oscillation index (with the opposite sign)
(http://www.cdc.noaa.gov/Correlation/soi.data). The season DJF is retained, which usually corresponds to
the mature phase of ENSO events in 1977–2002. It also corresponds to the season with the strongest lower
atmosphere signal in the northern tropics of the Atlantic sector (Enfield and Mayer, 1997). The index is fairly
correlated in time, so that the choice of alternative seasons did not result in large changes.
A winter ’atmospheric’ index (DJFM) is retained to characterize NAO. Because NAO presents more ener-
getic variability in winter and little season-to-season correlation during the rest of the year, this index is fairly
330 G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340

representative of the annual record. We also expect that the major impact in the northern tropics takes place
during winter (Czaja et al., 2002). The index we use is a principal component associated to a rotated analysis of
sea level pressure (NCEP SLP) and averaged over the DJFM season (http://www.cdc.noaa.gov/Correlation/
nao.data; it will be referred to as the rotated index). Alternatively, we checked the result using the SLP station-
based winter index of Hurrell (1995) which is more weighted towards the eastern Atlantic. Both winter indices
present a large noise due to large month-to-month variability in NAO. The two winter indices are correlated
(0.61), although with large differences in individual years, for example 1998 and 1999 portrayed as weak and
high NAO years respectively for the rotated index, whereas they are close to zero in the station index (index
years correspond to the January month).
In addition, a mode having a large atmospheric signal in the tropical Atlantic is the so-called dipole mode
(Moura and Shukla, 1981; Servain, 1991) that is best defined in boreal spring (Chiang et al., 2000). It is how-
ever partially correlated with ENSO and NAO, and will not be discussed. Instead we select the equatorial
Atlantic mode which overlaps its southern pole. This mode presents an in phase SST pattern maximum in
the eastern equatorial Atlantic (Zebiak, 1993). The pattern is associated with an index (averaged SST
15W–5E 5S–5N from NCEP reanalysis) denoted ATL3 which we average over JFM, which is not the sea-
son with maximum variance of this index, but which is then nearly independent of ENSO.
The three indices are nearly uncorrelated over 1977–2002. Taken together, SSS regressed on the three indi-
ces has a variance often less than 10% of the total SSS variance. The pattern of the variance explained is rel-
atively similar in different seasons (illustrated for JFM on Fig. 13) with patches of large explained variance in
the western subtropics (up to 20% in an area largest for the autumn following the index), as well as in the
north-east Atlantic (near 40N in all seasons, but extending to 50N in winter–spring). The first patch is more
related to ENSO and the second one more to NAO. The variance explained in the equatorial and southern
Atlantic regions experience more seasonal variability, with the two maxima near the equator and at 10-
20S near Brazil being more present in the early part of the year than at later lags. The percentage of SSS var-
iance linearly explained by these climate modes is typically much less than the one found in SST. According to
various authors (see Wang, 2002, for a review), ENSO and NAO can directly explain close to 50% of the SST
variance in a large part of the tropical Atlantic (in particular north of the averaged position of the ITCZ). We
will now consider individually the regressions to each index.

6.2. ENSO

The regression patterns to the DJF ENSO index will be first presented. The external forcing terms (E  P
and horizontal Ekman advection terms) exhibits a large seasonal variability (not presented), as expected for

JFM % variance (NAO+ENSO+ATL3)


50N
40N
30N
20N
10N
0
10S
20S
30S
90W 70W 50W 30W 10W 10E
%

0 2 4 6 8 10 12 14 16 18 20

Fig. 13. Percentage of variance in SSS (average of months JFM) associated with ENSO, NAO and ATL3 modes of variability. Only areas
with at least half the points in the time series associated with a relative error less than 0.8 are plotted.
G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340 331

the seasonally changing atmospheric manifestations of ENSO in the Atlantic Ocean (Enfield and Mayer,
1997). Both E  P and Ekman provide large contributions (lower panels of Fig. 14 for DJF). Near the equa-
tor, E  P presents a large and positive tongue, extending to 10S near South America from December to Sep-
tember (not shown). This is associated with less southward displacement of the ITCZ in these regions in the
early part of the year, and stronger trade winds (Enfield and Mayer, 1997), a feature particularly noticeable
during the 1983 El Niño (Delecluse et al., 1994). Both would contribute to higher SSS near the equator, in
addition to which the intensified trade winds are associated to stronger upwelling that would bring more salty
(thermocline) water in the surface layer. Ekman contribution is positive north of the equator and south of
20N at least until April, and in those regions seems to dominate the more often negative boreal winter
E  P contribution associated in the eastern tropical Atlantic with a reduction of the trade winds (Enfield
and Mayer, 1997) (Fig. 2). After April, the negative E  P band near 5N–10N dominates the forcing field
in this region (not shown). The negative forcing in the 20–35N band remains through much of the year, but
with much reduced intensity in boreal summer and autumn and confined mostly near 30N.
We therefore expect a positive near equatorial SSS 0 signal to be associated with ENSO. This is found in the
regression pattern (Fig. 14), becoming significantly non-zero by March and lasting along the equator until the
end of the year with peak amplitude between April and July (Fig. 14). At 1 month lag, the most significant

ENSO DJF
50N
40N
30N
20N
10N
0
10S SSS SSS
LAG1 (JFM) LAG6 (JJA)
20S
30S
90W 70W 50W 30W 10W 10E 90W 70W 50W 30W 10W 10E

50N
40N
30N
20N
10N
0
′. ∇S)/S
10S E
LAG0 (DJF) LAG0 (DJF)
20S
30S
90W 70W 50W 30W 10W 10E 90W 70W 50W 30W 10W 10E
mm/month

Fig. 14. Regressed fields on the ENSO index. Two upper panels for SSS 0 at lags 1 month and 6 months; two lower panels for the external
forcing terms at lag 0 (E  P on the left, and the sum of E  P and Ekman on the right). Values significant at the 90% level are surrounded
by a contour; blank areas correspond to regions with less than 13 years retained for the regression.
332 G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340

SSS 0 signal is the negative patch near 15–25N/50–60W (Fig. 14). It is somewhat shifted southwest of the
largest negative regressed forcing in this region. This anomaly pattern seems to extend in the following months
towards the northwest. The southern portion of it becomes weaker and at lag 6 months, this is the north-wes-
tern part of the North Atlantic subtropical gyre that presents the most coherent anomalies, with the largest
ones centred along 30N. Close to the coast of South America, there are positive anomalies, initially not sig-
nificant, but with the positive anomalies extending north-westward in the following months, in particular in
JJA, when they become marginally significant along the way and near the lesser Antilles (from 5S to
20N). These features present some coherence with the patterns of the forcing term (shown for DJF on
Fig. 14). Near 20–35N in the western and central Atlantic, the joint negative contribution of Ekman and
E  P which is significant through May seems to have contributed to the development of the negative anom-
alies in the boreal spring. After May, forcing becomes weak south of 30N (although not significantly non-
zero), which would contribute to the reduction of the SSS anomalies, whereas the signal remains strong near
30N. On the other hand, the strong negative DJF forcing in the eastern tropical Atlantic along 25N does not
seem to have imprinted SSS, which steadily remains weakly positive.

6.3. NAO

NAO index is strongest in winter time, with no correlation with its index in the following seasons. The asso-
ciated E  P forcing has a complicated pattern with positive E  P in the eastern North Atlantic between
20N and 45N and a tendency for negative E  P in the western Atlantic at these latitudes (Bojariu and Rev-
erdin, 2002). We therefore expect a SSS signal maximum in late winter and decreasing thereafter. Seasonal
SSS 0 regressed on the winter NAO index confirms this late winter maximum (February–May) (not shown).
This pattern is rather stationary through the following seasons with a tendency for positive regressed anom-
alies, except near north-west South America and the north-west corner of the domain near Newfoundland. By
the following SON, the maximum amplitudes are smaller and the statistically significant areas have diminished
with respect to February–May. The actual pattern is not closely related to the forcing pattern, except in the
eastern part of the basin (not shown).
To summarize this signal associated with NAO, we will in the following only consider annual averages.
During this period, the winter NAO index presents a noticeable correlation from 1 year to the next with a
slightly red spectrum indicative of energetic low frequencies (Jones et al, 2003). We isolate this low-frequency
contribution by applying a 5-year running average on the index (referred to LF), the high-frequency residual
being referred to as (HF). The annual SSS fields are regressed on the two time series, with a larger LF
regressed SSS and smaller signals in the HF regression that are rarely significantly non-zero at the 90% level
(Fig. 15). The areas of significance in HF are patchy and disappear at 1-year lag.
The forcing patterns related to HF and LF present some similarities, although the LF one has larger ampli-
tudes than the HF one, and presents negative values near 15N as well as in a southwest to northeast band
from the Gulf of Mexico to the central North Atlantic. There are much larger differences between the
regressed SSS patterns for HF and LF (Fig. 15). Except north of the GS and in the Gulf of Mexico, the
LF SSS 0 tends to be positive at 0-lag (as we found in the regression to NAO of the monthly time series),
whereas the high-frequency part has more negative cores in the 20N–40N band. The largest values in LF
are in the east, in particular near 40N, near 25N, and near 10N. The values north of the GS near the shelves
are also large (and negative). The large area of positive LF SSS 0 from 25N to 45N, and an area with smaller
positive values near 10N correspond to regions of positive forcing. However, others do not and the patterns
of the forcing bear little resemblance with the ones in SSS, both at low frequencies and high frequencies. This
is suggestive of a strong effect of the circulation on the anomaly pattern.

6.4. Averaged domains

The time series for the inter-gyre and the two subtropical gyre domains have comparable amplitudes
(Fig. 16). In all three regions, successive seasonal averages are highly correlated, although a little less between
OND and JFM, so that the annual-average time series are very similar to time series from individual seasons.
On the other hand the ‘external’ forcing terms present little correlation between successive seasons (whether
G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340 333

4
NAO DJFM
2

1977 1982 1987 1992 1997 2002

50N
40N
30N
20N
10N
0
10S BF HF
20S
30S
90W 70W 50W 30W 10W 10E 90W 70W 50W 30W 10W 10E

E
′. ∇
50N
40N
30N
20N
10N
0
10S BF HF
20S
30S
90W 70W 50W 30W 10W 10E 90W 70W 50W 30W 10W 10E
mm/month

Fig. 15. NAO DJFM index (thick full curve) with a low-pass (BF 5-year running average) in thin full line, and the HF residual in dashed
line (each time series has been separately normalized). The regressed SSS 0 and external forcing terms are presented at lag 0 in the four
lower panels. Values significant at the 90% level are surrounded by a contour; blank areas correspond to regions with less than 13 years
retained for the regression.

the Ekman term is included or not), and the annual average is not dominated by a particular season (although,
OND and JFM contribute more than the two other seasons, in particular when the Ekman term is included).
All three annual salinity time series (Fig. 16) present a positive trend at least until 1989, whereas the external
forcing terms present no trend. Actually, in the forcing term, this comes from a compensation between a neg-
ative trend during summer and a positive trend in winter.
The correlation between external forcing and SSS is weak (but positive) in these regions. It is largest in
domain 3 (western subtropical Atlantic), where the signature of ENSO is clearly visible both in the forcing
334 G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340

1980 1985 1990 1995 2000


60
SSS forcing
0.1 40

20

0 0

1980 1985 1990 1995 2000

1980 1985 1990 1995 2000


60
SSS forcing
0.1 40

20

0 0

1980 1985 1990 1995 2000

1980 1985 1990 1995 2000


60
SSS forcing
0.1 40

20

0 0

1980 1985 1990 1995 2000

Fig. 16. Series of annual salinity and external forcing anomalies for three domains north of 20N. First, monthly time series of SSS 0 are
created retaining only points with relative errors less than 0.8. Then, these time series are averaged from January through December to
create annual time series (on lower panel, arrows for ENSO during DJF season).

terms and this averaged SSS (low SSS during ENSO, the only missing signal being in 1977–1978). The imprint
of NAO forcing is not as clear as expected for 1 and 2: in particular for 2, we expected NAO+ to be associated
with higher SSS, based on the regressed forcing on NAO (see also Bojariu and Reverdin, 2002). This is coher-
ent with the non-winter season contributing also significantly to the annual external forcing. There is however
for the first part of the time series a tendency for a match-up between NAO LF and SSS.

7. Discussion

In the previous section, we commented on large differences between the ‘external’ forcing terms and the SSS
signals either associated with large-scale climate signals (ENSO, NAO) or averaged over large domains. This
contrasts with the local correlation found earlier between the external forcing and SSS. This points out that the
advection and mixing terms play a major role for the larger scale, lower frequency variability, as is also found in
model analysis. For example, at interannual frequencies, Mignot and Frankignoul (2003) point out that the
relative importance of Ekman and E  P forcing becomes smaller than other sources of the variability of
SSS with a large contribution of advection, more so than what is found for SST. This large contribution of
advection is naturally expected to increase towards the lower frequencies (Mignot and Frankignoul, 2004).
G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340 335

When considering the SSS signals attributed to NAO and ENSO, the areas with most obvious discrepancies
with the ‘external’ forcing terms are expected to be areas with a strong influence of advection. For example,
close to Africa near 25N, we expect less upwelling due to weakened trades during ENSO years (or low NAO),
which would contribute to a reduced freshening by the vertical advection. This counteracts the change caused
by E  P and horizontal Ekman advection. In the western tropical Atlantic, we expect a large influence by
advection of the anomalously salty water found earlier upstream along the South American coast. For exam-
ple, north-westward propagation of such anomalies has been diagnosed in a numerical model (Ferry and Rev-
erdin, 2004) throughout spring and summer. The origin of these anomalies along the shelf is less clear, but
with a likely contribution of changes in freshwater input by the rivers. Changes near the Gulf Stream are also
coherent with observed displacements of the Gulf Stream position, which result in large local anomalies,
because of the large spatial gradients of SSS in this region.
As an alternative to the large-domain averages presented in Section 6.4 that contain large spatial SSS gra-
dients, and therefore are more prone to advection terms, we also considered annually averaged SSS in the
region 18–29N/20–49W. This is a reasonably well-sampled region for 1970–2002 that roughly corresponds
to the maximum surface salinity of the northern subtropical gyre, and therefore has small average SSS gradi-
ents (Fig. 3). The correlation between the annually averaged forcing terms and SSS (both averaged over the
region) is higher for this region than for the domain-averages of Section 6.4 or for the monthly time series in
individual locations presented above (Fig. 17). However, the number of degrees of freedom is relatively small
with correlations slightly above the 90% confidence level for null hypothesis (assuming that the number of
degrees of freedom is the number of years, despite a noticeable auto-correlation of SSS at 1-year lag). Results
differ somewhat whether a trend in SSS is removed or not, and on the period considered. It is intriguing that
when cutting the time series into two pieces of equal length (before and after 1986), the results are highly dif-
ferent with a strong correlation between forcing and SSS before 1986 and none after 1986. It is unlikely that
this results from a major change in the sampling or errors in the forcing. There is the possibility however that
the core of the high salinity region wobbles in time (see Dessier and Donguy, 1994) and that the region is not
always positioned in the same way with respect to the high salinity region, so that the budget of the region

50 0.1
E
′. ∇S)/S SSS
mm/mois

0 0

1970 1974 1978 1982 1986 1990 1994 1998 2002


Year

1
′. ∇S)/S ′. ∇S)/S
E E

0.5

0 2 4 6 8 10
Lag (year)

Fig. 17. Comparison of annual SSS and forcing terms in a region of maximum salinity of the northern subtropical gyre. The upper panel
presents the time series (correlation 0.62), the middle panel presents the auto-correlation functions for SSS and the different forcing terms;
the lower panel shows the cross-correlation between SSS and the forcing terms.
336 G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340

might involve rather different balances during the two sub-periods. Altogether, however, the correlation at 0-
lag is significant between SSS 0 and each forcing term. It is largest with E  P, and slightly higher when the
Ekman term is added.

Slope
50N
40N
30N
20N
10N
0
10S
20S
30S
90W 70W 50W 30W 10W 10E

Slope % variance
50N
40N
30N
20N
10N
0
10S
20S
30S
90W 70W 50W 30W 10W 10E
%

0 2 4 6 8 10 12 14 16 18 20

LF % variance NAO
50N
40N
30N
20N
10N
0
10S
20S
30S
90W 70W 50W 30W 10W 10E

Fig. 18. SSS trend in 1977–2002 (upper panel); percentage of total variance associated with the trend (middle panel); percentage of total
variance in the annual mean SSS associated with LF NAO (lower panel). Only areas with at least half the points in the time series
associated with a relative error less than 0.8 are plotted on the last plot.
G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340 337

Because of fairly localized large signals of short duration, persistence is usually not very large in the ana-
lysed fields. Persistence becomes larger when averaging over large domains (for example, in the series of
Fig. 16). How does this compare with model studies? In the coupled model investigated by Mignot and Frank-
ignoul (2003), high persistence is found near 20–30N (auto-correlation coefficient at 1-year lag larger than
0.3). The region 20–30N also presents a dominance of low-frequency variability (period longer than 4 years)
in the model study of Häkkinen (2002) done with realistic forcing, with a pattern fairly reminiscent of the one
observed with two identified maxima (one within 30–60W located a little further north than the other centred
near 20W). A difference between these two models and the observations is that the band 20N–30N corre-
sponds to low observed rms variability (usually less than 0.1, Fig. 7) in comparison to the models. Interest-
ingly, there is a large salinity change in this region between 1970–1974 and 1955–1959 with surface (and
subsurface) differences larger than 0.2 centred on 40W between 15N and 30N, according to Levitus
(1989) and also suggested in the more recent study of Curry et al. (2003). Such a low-frequency variability
is not captured in our analysis which is for the later period 1977–2002. This is probably a real feature and
not aliased high frequencies. This change that happened before 1980 could explain part of the difference with
the results of Häkkinen (2002) for the years 1951–1993.
The salinity trend presented in Fig. 18 (upper panel) indicates usually fairly small positive values, in par-
ticular between 20N and 45N. In isolated regions, the trend in SSS during the 1977–2002 period contains
a significant portion of the variance (Fig. 18 middle panel). It can reach up to 16% of total variance in annual
time series, but it is more commonly in the range 1–2%. Not too surprisingly, because LF NAO also presents a
trend, there are similarities between the patterns associated with the trend and with LF NAO (Fig. 18, lower
panel). However, the patch of variance in the trend near 25N/50–70W is not related to NAO, and is not that
clearly related to ENSO either. It is coherent with a westward extension of the maximum salinity region. Fur-
thermore, we find a positive SSS trend for the three large area averages north of 20N, albeit overlaying ener-
getic higher frequencies (Fig. 16). The trend in these spatial averages actually ends in the late 1980s, which is
rather coherent with the analysis of Curry et al. (2003) from hydrography.

8. Conclusions

The large data sets of sea surface salinity that were collected since 1977 were sufficient to derive a seasonal
cycle of SSS in most of the Atlantic Ocean that is more detailed than what could be done previously (Levitus,
1986). Objective mapping was also used to create fields on a monthly to seasonal time scale from 1970 to 2002.
Clearly, the method is adequate when the coverage is sufficient, but it underestimates the variance in data-poor
regions. Large areas in the north-east Atlantic or subtropical gyre, and along major shipping tracks were suf-
ficiently sampled. In other regions, despite a fairly intensive sampling effort since 1977, the data coverage is
often not sufficient for reconstructing the SSS variability, in particular because the spatial scales of the sea-
sonal variability are found to be rather short for SSS, certainly shorter than for SST.
We should also mention that the uncertainties on the data have often a magnitude comparable to the
expected signal, and that the errors are likely to be correlated in time or spatially (particular vessel or instru-
ment type, . . .). There is certainly the possibility that this could have resulted in spurious mapped signals. It is
only through a more complete observing system that reduction of errors of this type will be achieved. On the
other hand, there is certainly room to improve the analysis of SSS and expand the local information provided
by the data in a better way than by objective mapping (including estimates of the circulation and some knowl-
edge of what are the processes controlling SSS).
The analysed anomaly fields indicate that SSS variability presents a wide range of magnitude with largest
deviations from the seasonal cycle in the tropical Atlantic in the region straddled by the ITCZ and the North
Equatorial Counter Current, near northern South America and in the eastern Gulf of Guinea, as well as just
north of the Gulf Stream. The time scales of the variability are often shorter in those regions, presumably
because of advection, whereas elsewhere there is persistence over more than one season, and in more localised
areas over more than a year (in particular in the central subtropical gyre or north-eastern Atlantic). We do not
find clear indication of a re-emergence of the SSS signal from a previous winter to the next. The lagged auto-
correlation usually indicates more persistence from winter to summer than from summer to winter, consis-
tently with the seasonality of mixed layer depth, and autumnal deepening of the mixed layer causing a change
338 G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340

in SSS anomalies. The lagged-correlation patterns with respect to a given location present often a fairly sta-
tionary SSS pattern.
SSS anomalies present a weak positive correlation with SST anomalies, most pronounced at the higher lat-
itudes of the northern hemisphere where SSS lags SST. This suggests that SSS anomalies are reinforced as SST
decays, as expected from the negative (for SST) heat flux feedback (e.g., Frankignoul and Kestenare, 2002).
SSS presents a correlation both with local freshwater flux and with the Ekman advection term, maximum cor-
relations being usually found at positive lags off the equatorial band (SSS 0 follows these terms), as found in
stochastic models with white noise forcing. The correlations are highest in the subtropical gyre and at higher
latitudes, but the maximum lag correlations are just only on the order of 0.3 which explains a small (10%) part
of the variance. Such low correlations are also found in coupled models with additional identified contribu-
tions to seasonal variability from geostrophic variability and mean advection (for example, Mignot and
Frankignoul, 2003). There is also a large uncertainty in the freshwater term, as assessed by using different
products, most likely due to uncertainties in precipitation. The precipitation product based mostly on data
(in situ or satellite-based) results usually in the highest correlation. In some regions of the equatorial Atlantic,
fluctuations of the river input of freshwater that have not been taken into account should also contribute
strongly to SSS 0 , albeit through advection from the shelves.
With the caveat that the time series are very short (26 years), we attribute SSS signals to ENSO and NAO.
The part of the variance associated with these major low-frequency signals is usually less than 10% of the total
variance. For NAO, the best-defined SSS 0 is at the lower frequencies (the 5-year running mean), when advec-
tion is expected to have a very large influence on the SSS 0 budgets and patterns. We expect that the length of
the time series is insufficient to define well the SSS signatures associated with NAO, ENSO and other climate
modes. On the other hand, these signatures could well be modulated by decadal and lower frequency variabil-
ity. In that respect, the period 1977–1998 (or even later) seems to correspond to a somewhat stationary period
(Chiang et al., 2000; Hilmer and Yung, 2000) which differs from earlier or later periods both for NAO and
ENSO.

Acknowledgements

A large part of the data originates from ships-of-opportunity, and its collection relied on the cooperation of
the crews, captains and companies, without which kind help and often keen interest this would not be possible.
The collection effort in France was originated by scientists from ORSTOM, and then IRD (in particular, Jean-
René Donguy, Alain Dessier), and constitutes a large share of the data used. However, there are many other
sources that we will not list individually, often communicated by individual oceanographers dedicated to these
surface data. This work is a contribution to the Observatoire de recherche en Environnement (ORE) SSS, and
was supported by the French programme PNEDC and TOSCA. Advice to one of us (Elodie Kestenare) by
Jérome Sirven and comments by two reviewers are gratefully acknowledged. Some of the figures were plotted
by Françoise Besset.

Appendix A. Objective mapping

The basic technique we use is the objective mapping, a statistical procedure described in Bretherton et al.
(1976). For performing the mapping, the gridded anomalies are attributed an error variance of r2im =N l , where
Nl is an estimated number of degrees of freedom of the binned data, defined as the number of data with 5 as
upper bound (in order not to underestimate ‘error’ when data originate from finely spaced TSG data), and r2im
is the intra-monthly variance estimated when gridding the data. The signal is attributed to the expected signal
variance r2ia also estimated from the gridded data. The correlation function is chosen as exponential in time
and space (see Table 1 for the scales Lx, Ly, Lt estimated in individual regions), and the data incorporated
in the analysis for a given month are the ones over a 3-month period centred on the date of the analysis
and within an ellipsoid of axes 2.5 · (Lx, Ly). When there is no data within this analysis, the analysed deviation
is the one of the previous (or next) month reduced by the exponential decay (1/e in 2 months), and the nor-
malized error is equal to 1. In those areas, this reduces the persistence of the anomalies below the one reported
in Table 1, but this was preferred to artificially increasing coherence in the anomalies. We also decided not to
G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340 339

incorporate data in areas close to the mouths of the three main tropical rivers, the Amazon, the Orinoco and
the Congo, as variance is high and poorly sampled and retaining those areas resulted in strong contamination
of nearby grid points.

References

Alexander, M.A., Deser, C., 1995. A mechanism for the recurrence of wintertime midlatitude SST anomalies. J. Phys. Oceanogr. 25, 122–
137.
Ayina, H.-L., 2002. Etude des modes de variabilité de l’océan Atlantique tropical et de leur sensibilité à l’impact des décharges fluviatiles et
des précipitations. Thèse de docteur de l’université Pierre et Marie Curie, 255pp.
Belkin, I.M., Levitus, S., Antonov, J.L., Malmberg, S.A., 1998. Great salinity anomalies in the North Atlantic. Prog. Oceanogr. 41, 1–35.
Bingham, F.M., Howden, S.D., Koblinsky, C.J., 2002. Sea surface salinity measurements in the historical database. J. Geophys. Res. 107
(C12), 8019. doi:10.1029/2000JC000767.
Bojariu, R., Reverdin, G., 2002. Large-scale variability modes of freshwater flux and precipitation over the Atlantic. Climate Dyn. 18,
369–381.
Boyer, T.P., Levitus, S., Antonov, J.I., Locarnini, R.A., Garcia, H.E., 2005. Linear trends in salinity for the World Ocean 1955_1008.
Geophys. Res. Lett. 32, L01604.
Bretherton, F.P., Davis, R.E., Fandry, C.B., 1976. A technique for objective mapping and design of oceanographic experiments. Deep Sea
Res. 23, 559–582.
Cayan, D.R., 1992. Latent and sensible heat flux anomalies over the northern oceans: driving the sea surface temperature. J. Phys.
Oceanogr. 22, 859–881.
Chiang, J.C.H., Kushnir, Y., Zebiak, S.E., 2000. Interdecadal changes in the eastern Pacific ITCZ variability and its influence on the
Atlantic ITCZ. Geophys. Res. Lett. 27 (22), 3687–3690.
Curry, R., McCartney, M.S., 2001. Ocean gyre circulation changes associated with the North Atlantic circulation. J. Phys. Oceanogr. 31,
3374–3400.
Curry, R., Dickson, R., Yashayaev, I., 2003. A change in the freshwater balance over the Atlantic Ocean over the past four decades.
Nature 426, 826–829.
Czaja, A., van der Vaart, P., Marshall, J., 2002. A diagnostic study of remote forcing in tropical Atlantic variability. J. Clim. 15, 3280–
3290.
Da Silva, A., Young, A.C., Levitus, S., 1994. Algorithms and procedures. In: Atlas of Surface Marine Data 1994, vol. 1. National
Oceanographic and Atmospheric Administration, p. 83.
De Boyer-Montégut, C., Madec, G., Fisher, A.S., Lazar, A., Ludicone, D., 2004. A global mixed layer depth climatology based on
individual profiles. J. Geophys. Res. 109, C12003, doi:10.1029/2004JC002378.
de Coetlogon, G., Frankignoul, C., 2003. The persistence of winter sea surface temperature in the North Atlantic. J. Clim. 16, 1364–1377.
Delcroix, T., Dessier, A., Gouriou, Y., McPhaden, M., 2005. Time and space scales for sea surface salinity in the tropical oceans. Deep Sea
Res. 52 (5), 787–813.
Delecluse, P., Servain, J., Levy, C., Arpe, K., Bengtsson, L., 1994. On the connexion between the 1984 Atlantic warm events and the 1982–
1983 ENSO. Tellus 46A, 448–464.
Dessier, A., Donguy, J.R., 1994. The sea surface salinity in the tropical Atlantic between 10S and 30N – seasonal and interannual
variations (1977–1989). Deep-Sea Res. I 41, 81–100.
Dickson, R., Meincke, J., Malmberg, S.-A., Lee, A.J., 1988. The great salinity anomaly in the northern North Atlantic. Prog. Oceanog. 20,
103–151.
Enfield, D.B., Mayer, D.A., 1997. Tropical Atlantic sea surface temperature variability and its relation to El Niño-southern oscillation. J.
Geophys. Res. 102, 929–945.
Ferry, N., Reverdin, G., 2004. Sea surface salinity interannual variability in the western tropical Atlantic: an ocean general circulation
model study. J. Geophys. Res. 109, C05026. doi:10.1029/2003JC002122.
Frankignoul, C., Kestenare, E., 2002. The surface heat flux feedback. Part I: estimates from observations in the Atlantic and the North
Pacific. Climate Dyn. 19, 633–647.
Frankignoul, C., Czaja, A., L’Hévéder, B., 1998. Air–sea feedback in the North Atlantic and surface boundary conditions for ocean
models. J. Clim. 11, 2310–2324.
Frankignoul, C., de Coëtlogon, G., Joyce, T.M., Dong, S., 2001. Gulf stream variability and ocean–atmosphere interactions. J. Phys.
Oceanogr. 31, 3516–3529.
Giannini, A., Cane, M.A., Kushnir, Y., 2001. Interdecadal changes in the ENSO teleconnection to the Caribbean region and the North
Atlantic Oscillation. J. Clim. 14, 2867–2879.
Häkkinen, S., 1993. An Arctic source for the Great salinity anomaly: a simulation of the Arctic ice–ocean system for 1955–1975. J.
Geophys. Res. 98, 16,397–16,410.
Häkkinen, S., 1999. Variability of the simulated meridional heat transport in the North Atlantic for the period 1951–1993. J. Geophys.
Res. 104, 10,991–11,007.
Häkkinen, S., 2002. Surface salinity in the northern North Atlantic. J. Geophys. Res. 107 (C12), 8003. doi:10.1029/2001JC00081.
Hall, A., Manabe, S., 1997. Can local linear stochastic theory explain sea surface temperature and salinity variability. Clim. Dyn. 13, 167–
180.
340 G. Reverdin et al. / Progress in Oceanography 73 (2007) 311–340

Hilmer, M., Yung, T., 2000. Evidence for a recent change in the link between the North Atlantic Oscillation and Arctic sea ice export.
Geophys. Res. Lett. 27, 989–992.
Huffman, G.J., Adler, R.F., Arkin, P., Chang, A., Ferraro, R., Gruber, A., Janowiak, J., McNab, A., Rudolf, B., Schneider, U., 1997. The
global precipitation climatology project (GPCP) combined precipitation dataset. Bull. Am. Meteorolog. Soc. 78, 5–20.
Hurrell, J.W., 1995. Decadal trends in North Atlantic circulation: regional temperatures and precipitation. Science 17, 2294–2316.
Jones, P.D., Osborn, T.J., Briffa, K.R. 2003. Pressure-based measures of the North Atlantic Oscillation (NAO): a comparison and an
assessment of changes in the strength of the NAO and its influence on surface climate parameters. In The North Atlantic Oscillation,
Geophysical monograph 134, AGU, Washington, pp. 51–62.
Josey, R.A., Marsh, M., 2005. Surface freshwater flux variability and recent freshening of the North Atlantic in the eastern subpolar gyre.
J. Geophys. Res. 110, C05008, doi:10.1029/2004JC002521.
Joyce, T., Deser, C., Spall, M., 2000. Decadal variability of subtropical Mode Water in the North Atlantic and large-scale patterns of air–
sea forcing. J. Clim. 13, 2550–2569.
Latif, M., Roeckner, Mikolajewicz, E.U., Ros, V., 2000. Tropical stabilization of the thermohaline circulation in a greenhouse warming
simulation. J. Clim. 13, 1809–1813.
Levitus, S., 1986. Annual cycle of salinity and salt storage in the world ocean. J. Phys. Oceanogr. 16, 322–343.
Levitus, S., 1989. Interpentadal variability of salinity in the upper 150 m of the North Atlantic Ocean, 1970–1974 versus 1955–1959. J.
Geophys. Res. 94, 9679–9685.
Lukas, R., Lindstrom, E., 1991. The mixed layer of the western equatorial Pacific ocean. J. Geophys. Res. 96, 3343–3357.
Mémery, L., Reverdin, G., Paillet, J. Oschlies, A., in press. The POMME Program (Programme Océan Multidisciplinaire Méso Echelle).
Thermocline ventilation and biogeochemical tracer distribution in the northeast Atlantic Ocean: impact of meso-scale dynamics. J.
Geophys. Res. 110, C07S01, doi:10.1029/2005JC002976.
Mignot, J., Frankignoul, C., 2003. On the interannual variability of surface salinity in the Atlantic. Clim. Dyn. 20, 555–565.
Mignot, J., Frankignoul, C., 2004. Interannual to interdecadal variability of sea surface salinity in the Atlantic and its link to the
atmosphere in a coupled model. J. Geophys. Res. 109, C04005. doi:10.1029/2003JC00200.
Moura, A.D., Shukla, J., 1981. On the dynamics of droughts in northeast Brazil: observations, theory, and numerical experiments with a
general circulation model. J. Atmos. Sci. 38, 2653–2675.
Pollard, R.T., Pu, S., 1985. Structure and circulation of the upper Atlantic Ocean northeast of the Azores. Prog. Oceanogr. 14, 443–462.
Rahmstorf, S., 1995. Bifurcations of the Atlantic thermohaline circulation in response to changes in the hydrological cycle. Nature 368,
145–149.
Reverdin, G., Cayan, D., Dooley, H.D., Ellett, D.J., Levitus, S., du Penhoat, Y., Dessier, A., 1994. Surface salinity in the North Atlantic:
can we reconstruct its fluctuations over the last 1 hundred years? Prog. Oceanogr. 33, 303–346.
Reverdin, G., Durand, F., Mortensen, J., Schott, F., Valdimarsson, H., Zenk, W., 2002. Recent changes in the surface salinity of the
North Atlantic subpolar gyre. J. Geophys. Res. 107 (C12), 8010. doi:10.1029/2001JC001010.
Servain, J., 1991. Simple climate indices for the tropical Atlantic Ocean and some applications. J. Geophys. Res. 96, 15.137–15.146.
Spall, M.A., 1993. Variability of sea surface salinity in stochastically forced systems. Clim. Dyn. 8, 151–160.
Thorpe, R.B., Gregory, J.M., Johns, T.C., Wood, R.A., Mitchell, J.F.B., 2001. Mechanisms determining the Atlantic thermohaline
circulation response to greenhouse gas forcing in a non-adjusted coupled climate model. J. Clim. 14, 3102–3116.
Volkov, D., 2004. Interannual variability of the altimetry-derived eddy field and surface circulation in the extratropical North Atlantic
Ocean in 1993–2001. J. Phys. Oceanogr., 34.
Wang, C., 2002. Atlantic climate variability and its associated atmospheric circulation cells. J. Clim. 15, 1516–1536.
Xie, P., Arkin, P.A., 1997. Global precipitation: a 17-year monthly analysis based on gauge observations, satellite estimates, and numerical
model outputs. Bull. Am. Meteorolog. Soc. 78, 2539–2558.
Zebiak, S., 1993. Air–sea interaction in the equatorial Atlantic region. J. Clim. 6, 1567–1586.

View publication stats

You might also like