Improved estimates of water cycle change from ocean salinity: the key role of ocean warming

Changes in the global water cycle critically impact environmental, agricultural, and energy systems relied upon by humanity (Jiménez Cisneros et al Climate Change 2014: Impacts, Adaptation, and Vulnerability (Cambridge: Cambridge University Press)). Understanding recent water cycle change is essential in constraining future projections. Warming-induced water cycle change is expected to amplify the pattern of sea surface salinity (Durack et al Science 336 455–8). A puzzle has, however, emerged. The surface salinity pattern has amplified by 5%–8% since the 1950s (Durack et al Science 336 455–8, Skliris et al Clim. Dyn. 43 709–36) while the water cycle is thought to have amplified at close to half that rate (Durack et al Science 336 455–8, Skliris et al Sci. Rep. 6 752). This discrepancy is also replicated in climate projections of the 21st century (Durack et al Science 336 455–8). Using targeted numerical ocean model experiments we find that, while surface water fluxes due to water cycle change and ice mass loss amplify the surface salinity pattern, ocean warming exerts a substantial influence. Warming increases near-surface stratification, inhibiting the decay of existing salinity contrasts and further amplifying surface salinity patterns. Observed ocean warming can explain approximately half of observed surface salinity pattern changes from 1957–2016 with ice mass loss playing a minor role. Water cycle change of 3.6% ± 2.1% per degree Celsius of surface air temperature change is sufficient to explain the remaining observed salinity pattern change.


Introduction
Rates of rainfall and evaporation are tightly linked to the vapour pressure of water, which increases by approximately 7% per degree Celsius of warming, raising concerns global warming will substantially enhance the hydrological cycle (Held and Soden 2006). Global climate models project overall increases in evaporation and precipitation, rainfall extremes, and the implied net transport of moisture from net precipitation to net evaporation zones (Durack et al 2012, Liu et al 2012, Lau et al 2013. At regional scale, increased vapour pressure may be compensated by changes in atmospheric circulation (Chadwick et al 2013) and subject to influences on evaporation other than warming (Laîné et al 2014).
Given limitations in the global network of rainfall measurements (Hegerl et al 2015) and substantial discrepancies between atmospheric reanalysis products, model-based reconstructions of past rainfall and water cycle change avoid proper scrutiny and projections of future change are highly uncertain (Schanze et al 2010, Skliris et al 2014, Levang and Schmitt 2015, Grist et al 2016, Yu et al 2017. It has been proposed that ocean salinity be used to estimate past water cycle change via its imprint on surface salinity patterns. Net evaporation increases salt concentration, while net precipitation lowers salt concentration. Surface salinity changes are indeed dramatic, with saline regions of the globe becoming more saline and fresh regions becoming more fresh (Durack and Wijffels 2010, figures 1(a)-(c)). Because surface signals are rapidly advected by ocean currents, salinity measurements at individual locations are not akin to 'rain gauges' (Yu 2011). Full depth salinity observations combined with a three dimensional picture of ocean circulation could provide strong constraints on water cycle change. However subsurface observations remain sparse. Bulk measures of global salinity change, such as salinity pattern amplification (PA hereafter, see methods), can filter out and/ or integrate over some of the rearrangement effects caused by variability in ocean currents (Helm et al 2010, Durack et al 2012, Zika et al 2015. PA has been used to infer historical trends in water cycle change (Durack et al 2012. However, the relationship between surface salinity pattern changes and water cycle change is puzzling (Vinogradova and Ponte 2017). While the surface salinity pattern has amplified by 5%-8% since 1950 (Durack et al 2012, Skliris et al 2014, changes in full depth salinity suggest a water cycle change of only 2%-3% . Surface salinity PA also exceeds amplification of the water cycle in climate model projections (Durack et al 2012).
To address the above puzzle, we undertake perturbation experiments using a realistic global ocean model. We find that, in addition to water cycle changes and ice mass loss, surface ocean warming is a key and previously overlooked process driving sea surface salinity amplification. We then consider observations of salinity pattern amplification, glacial melt and ocean warming and use these to estimate historical water cycle changes.

Observational data
Observed multi-decadal salinity changes are investigated here using the UK Met Office Hadley Centre Enhanced Ocean Data Assimilation and Climate prediction (ENACT) archive version 4 (EN4, subversion En4.2.2 (Good et al 2013)) objectively-analysed monthly dataset covering the 1950-2016 period, at 1 • × 1 • horizontal grid and a vertical grid of 42 levels.
PA of sea surface salinity is defined as the slope of the linear regression by basin zonal averages of SSS change versus the climatological mean SSS anomaly from the global climatological sea surface salinity (Durack et al 2012).
Ocean heat content time series for the upper 2000 m between 1957 and 2016 are obtained from data.nodc.noaa.gov and Boyer et al (2013). Although standard error estimates are provided with these, we add an uncertainty of 15% to be consistent with uncertainties in long term ocean heat content change recently reported (e.g. 288 ± 44 ZJ; ZJ = 10 21 Joules) for the period 1960-2015 (Cheng et al 2017). For the ocean deeper than 2000 m and period 1991-2010 an additional 33 ± 22 TW of ocean heat uptake has been reported (Desbruyères et al 2016). To create a time series for this deep component we assume there is a linear trend in heat uptake from 1956 reaching 33 PW in the year 2000 and increasing linearly there after. The resulting ocean heat uptake between 1957 and 2016 is 345 ± 75 ZJ in close agreement with recent estimates (e.g. 335 ± 70 ZJ for the period 1960-2010 (Cheng et al 2017)).
Church et al (2013) document estimates of fresh water input into the ocean for the period 1993-2010 attributed to mass loss from glaciers and the Greenland ice sheet (1.2 ± 0.4 mm yr −1 ), mass loss from Antarctic ice sheets (0.27 ± 0.11 mm yr −1 ) and land water storage changes (0.38 ± 0.12 mm yr −1 ). This fresh water input totals 1.7 ± 0.4 mm yr −1 or 19 ± 5 mSv (Sv = 10 6 m 3 s −1 ). The contrast between estimates of northern and southern hemisphere ice mass loss motivated the 2:1 ratio of fresh water input into the North Atlantic Ocean and Southern Oceans respectively in our ice mass loss numerical ocean modelling experiments discussed below.
An estimate of the rate of multi-year sea-ice mass loss of 3.1 ± 1.0 km decade −1 (9.8 ± 3.2 mSv) is obtained from http://psc.apl.uw.edu/research/ projects/arctic-sea-ice-volume-anomaly/ and was derived using the method described by Schweiger et al (2011). For simplicity, we assume that fresh water input from both land-ice mass loss and sea-ice mass loss increased linearly from zero in 1957 to reported values at the midpoint of their respective observational period (i.e. 2001 for land-ice and 1998 for sea-ice), increasing linearly thereafter.

Numerical ocean model perturbation experiments
We use NEMO (Nucleus for European Modelling of the Ocean) (Madec 2008) v3.5 in the global ORCA1 configuration. The model grid has a 1 • longitudinal resolution at the equator with refined meridional resolution in the tropics transitioning to a tri-polar isotropic mercator grid elsewhere. The model has 75 z coordinate vertical levels increasing in thickness from 1 m at the surface to 250 m in the abyssal ocean. This model configuration is the National Oceanography Centre contribution to a recent model intercomparison project (Danabasoglu et al 2014) where it is described in further detail.
The control simulation starts from rest and is integrated for a spin-up period of 400 years. It is then continued for a further 100 years to allow comparison with the control simulation conducted with explicit flux forcing. Surface forcing is provided by CORE-2 normal year forcing (Large and Yeager 2009), interfaced to the model through CORE bulk formulae. Ice is represented by the Louvain-la-Neuve ice model version 2 (LIM2) sea ice model (Timmermann et al 2005). Climatological initial conditions for temperature and salinity were taken in January from PHC2.1 (Steele et al 2001) at high latitudes, MEDATLAS (Jourdan et al 1998) in the Mediterranean, and elsewhere using the World Ocean Atlas (Boyer et al 2013).
Drift in temperature and salinity is negligible below 1000 m (supplementary figures S1(a) and (b) available at stacks.iop.org/ERL/13/074036/mmedia). The upper 1000 m shows a gradual warming, which stabilises at 0.7 • C after around 200 years. There is a small surface fresh bias in the upper 200 m, with a saline bias subsurface between 500-800 m. Again these appear stable after 200 years. The mean surface temperature and salinity bias from the EN4 1980-1999 average (figures S1(c) and (d)) is minimal over most of the global ocean. The largest temperature biases are in the North Atlantic and Nordic Seas. Salinity biases are largest in the Arctic. Rigorous tuning of NEMO ORCA2, forced by the CORE2 normal year, identified a persistent warm bias in the upper ocean which was largest around 200 m depth, and spatial distributions of surface T and S anomalies (similar to those in figure S1), suggesting that the biases may be inherent to the forcing, arise from the use of the EN4 climatology, or are structural to NEMO (Williamson et al 2017).
Surface fluxes for the decade spanning years 391 and 400 of the spin-up integration, including those at the base of the ice (surface of the liquid ocean) and those due to the weak salinity restoring, were stored at 6 hourly intervals. The model was then adjusted to remove the interactive ice model and to read surface fluxes directly instead of computing them through the CORE bulk formulae. A 100 year explicit flux forced control experiment, starting at year 391, was then conducted using 10 repeat cycles of this decade of explicit surface fluxes. Although the effect of surface salinity restoring is captured in the saved fluxes from the spinup, no salinity restoring is used throughout the 100 year control and perturbation experiments.
The difference in drift between the bulk formulae forced and explicit flux forced control simulations is small, with differences in global mean temperature and salinity both in the region of two hundredths of a unit (degree/pss) (figure S1(e) and (f)). This shows the model to be in a stable state close to equilibrium, and forms the baseline on which to assess heat and water flux anomaly experiments.
Flux anomaly patterns are introduced temporally in two ways: as a step-increase with respect to time (e.g. following Marshall et al 2015, Gregory et al 2016 and as a linear trend. Three types of perturbation patterns were added to the saved fluxes to force perturbation experiments: the first was a global amplification of the mean water flux pattern (10% step change and 1%/decade linear trend); the second was a globally constant heat flux anomaly (2 W m −2 step change and 0.2 W m −2 decade −1 linear trend); and the third was an idealised glacial and seaice mass loss fresh water flux two thirds of which is distributed over the sub-polar North Atlantic between 50 • N and 70 • N (i.e. following Stouffer et al 2006) and one third of which was evenly distributed south of 50 • S (the total flux 0.03 Sv for step change experiment and 0.003 Sv decade −1 in the linear trend case).
Our forcing anomaly patterns for water fluxes, surface heat fluxes and glacial and sea-ice mass loss are highly idealised and simplify their complex dynamics. Climate model inter-comparison project (CMIP5) experiments with one per cent per year atmospheric CO 2 radiative forcing increases, show water flux pattern changes qualitatively similar to an amplification of the mean water flux pattern (Gregory et al 2016). Warming patterns tend to be associated with larger downward heat fluxes at high latitudes and weaker downward (or even upward heat fluxes) at low latitudes (Gregory et al 2016). The latter is likely closely coupled to an individual model's circulation change (Marshall et al 2015). We consider our uniform heat flux approach to be useful since it does not assume or force particular ocean circulation changes and is straightforward to reproduce. To our knowledge, there is no consensus on how high latitude ice mass loss will be distributed as the climate warms. Most likely, some will remain near coasts and be influenced by boundary currents while a substantial fraction will likely be distributed further offshore via iceberg drift and melting, as in the present day Adcroft 2010, Marsh et al 2015). We have chosen our ice mass loss flux strategy so as to maintain simplicity and consistency with other model studies.
In addition we have run an experiment with all three forcings applied (in the linear-trend forcing cases) to the same experiments to assess the linearity of the responses. All data from perturbation experiments shown in the manuscript are deviations from the control experiment.
The above experimental design was inspired by Marshall et al (2015). Our approach differs from theirs in the following ways however: (i) in addition to the surface heat flux we also perturb fresh water fluxes; (ii) we do not include any feedbacks between the ocean state and the surface flux (both heat and fresh water flues are explicitly controlled); and (iii) 6 hourly data were used for the explicit flux forced experiments rather than daily data.

Results
In water cycle experiments, 1% per decade amplification over 100 years leads to simulated surface salinity changes ( figure 1(d)) that are qualitatively consistent with the observed pattern of change ( figure 1(c)). Some differences can be identified, such as freshening in the sub-polar North Atlantic in our model-which is not seen in observations. This region does not appear to conform to the wet gets wetter dry gets dryer paradigm in observations. Some re-analysis products suggest a large increase in net evaporation over the whole North Atlantic and even a reduction in net precipitation over the North Atlantic sub-polar region (Skliris et al 2014). Such changes have been associated with Atlantic Multidecadal Variability (Zhang 2017) which we do not account for in our forcing anomalies.
Idealised ice mass loss experiments apply a freshening in the sub-polar North Atlantic (0.002 Sv per decade) and Southern Ocean (0.001 Sv per decade). These led to surface freshening at high latitudes, in particular the North Atlantic, with limited changes at lower latitudes (figure 1(e)), the latter likely due to teleconnected changes in ocean circulation.
In warming experiments with surface heat flux anomalies increasing by 0.2 W m −2 per decade, surface salinity changes after 100 years (figure 1(f)) are also qualitatively similar to that in the water cycle change experiments. Differences are most prominent in the equatorial Pacific where water cycle amplification drives freshening, while warming leads to only a muted response. Furthermore some local changes, such the a patch of freshening at the centre of the South Atlantic, are likely related to a circulation response unique to this model configuration and experimental set-up. We now compare PA per experiment (figure 2). A water cycle trend of 1% per decade yields PA of 5.8% after 100 years. A 10% step change in the water cycle leads to a rapid initial PA adjustment approaching quasi-equilibrium after 40 years at 5.9%. A linear trend in ice mass loss of 0.003 Sv decade −1 leads to a linear change in PA reaching 0.6% after 100 years, while a step change of 0.03 Sv per (sub-polar North Atlantic: 0.02 Sv; Southern Ocean: 0.01 Sv) leads to PA of 0.8% which plateaus after 40 years.
Following heat flux linear trends of 0.2 W m −2 per decade (figure 2), PA reaches 7.3% after 100 years, indicating considerable PA response to warming. Under a step change warming of 2 W m −2 , PA increases monotonically throughout the century, reaching 13.4% at year 100.
The response of surface salinity to these different forcing anomalies is approximately linear. A simulation combining water cycle change, ice mass loss and surface warming forcing anomalies yields approximately the same surface salinity response (figures 1(g) and 2) as the sum of the individual simulated responses to each forcing anomaly (figures 1(h) and 2). There are regions where the response to warming and fresh water fluxes are not linear. West of Australia is a prime example where a wave like pattern of salinity change is seen. These patterns differ from experiment to experiment and are likely associated with nonlinear circulation processes (such as variability of the Leeuwin Current and meanders of Antarctic Circumpolar Current fronts) which are sensitive to differences in initial and boundary conditions.
The perturbation experiments suggest that, while the surface salinity pattern amplifies with the water cycle, the ratio of PA to water cycle amplification on multi-decadal timescales is less than unity (in terms of % change), in the absence of other forcing anomalies. Ocean warming however drives substantial salinity pattern amplification.

Discussion
We now address two outstanding questions: (i) by what mechanism does ocean warming lead to PA?; and (ii) can historical ocean warming plausibly explain a substantial portion of observed PA?

How does ocean warming lead to salinity pattern amplification?
To address this question, we employ a method based on the water mass framework. Specifically we will consider changes in the distribution of ocean volume as a function of salinity (Zika et al 2015). The volume of water contained between surfaces of constant salinity only varies if there are changes in surface water fluxes or changes in mixing. The latter acts in a fairly predictable way, always mixing salt from high to low salinity regions. In contrast surface salinity is influenced by ocean circulation changes and variability and the effects of ocean mixing on surface salinity are not as straightforward (Yu 2011).
The water mass framework permits a balance to be defined between the rate of change of the mean deviation of salinity over the whole ocean (W; the average deviation of salinity, S, from its global mean, ) variations in the water cycle F , which transports fresh water from saline regions (where >̄) to fresh regions (where <̄) and a diffusive flux D which transports salt back from saline to fresh regions (Zika et al 2015). This balance is given by where V 0 is the total volume of the ocean. The above equation is exact in the Boussinesq approximation and when there is no net fresh water flux into or out of the ocean (i.e. the total volume of the ocean remains constant) which is the case in our experiments where no water cycle or water flux change is applied. In our control and linearly increasing heat flux anomaly experiments̄= 34.72 pss (this value does not change with the perturbation since the total fresh water flux is identical in both runs). In the final decade of the control experiment W = 0.18 pss and in the perturbation W increases by = 0.0021 pss. Although the geographical pattern of water fluxes do not change between these two runs, the pattern of salinity does change, implying different salinity layers are exposed to different fresh water fluxes and hence different rates of fresh water transport from saline to fresh regions. This leads to an increase in F of 11 mSv between the control and perturbation. Using (1) and changes in W and F we infer a reduction in D of 8.45 × 10 2 kg s −1 between the control and linearly increasing heat flux experiment. This is equivalent to a fresh water flux anomaly of 24 mSv from saline to fresh regions (i.e. a reduction in the mean fresh water flux due to mixing which is from fresh to saline waters). This change in diffusive flux due to the warming anomaly has the same effect on the full depth salinity contrast (i.e. not only the surface salinity pattern) as an increase in the water cycle of approximately 1% on average over the experiment.
We hypothesise that the reduction in the diffusive flux of fresh water from fresh to saline regions in our warming experiments is at least in part explained by changes in near-surface stratification. Surface warming ( figure 3(a)) changes surface stratification and thus increases the stability of surface layers. More stable surface stratification inhibits the formation of deep mixed layers and isolates surface salinity anomalies from dilution with the ocean interior (Capotondi et al 2012). This hypothesis is corroborated by generally shallower mixed layer depths in warming experiments ( figure 3(c)).
Salinity can also strongly influence ocean stratification ( de Boyer Montégut et al 2007). However, in our simulations we find that warming is a far stronger driver of stratification change than water fluxes. For example with and without water flux changes (holding other forcing patterns fixed) we see near negligible differences in surface temperature changes between simulations. The North Atlantic is an exception. There high latitude freshening causes weakening of the AMOC and consequent SST changes (not shown).

How does ocean warming affect estimates of water cycle change?
Here to attribute observed salinity PA to the three discussed processes (water cycle change, ice mass loss and ocean warming), we use a linear transient response approach (Hasselmann et al 1993). Using this approach we can diagnose the time-dependent response of a particular variable (such as PA) based on the timedependent change in a particular forcing (water cycle change, ice mass loss and warming), using the timedependent response of the system to step changes in those forcings.
If R PA is the linear transient response (LTR) of PA to a step increase in a specified forcing (F; e.g. the water cycle), then the linear response of PA to time-variable forcing (F(t)) follows as: Our step change perturbation experiments allow us to diagnose R PA for each forcing: water cycle change; ice mass loss; and surface warming.
A fundamental assumption in the linear transient response approach is that the relationship between forcing effects and their response is linear. We assess this assumption by comparing the linear trend forcing experiment results to predictions made by LTR which exploit output from the step change experiments. That is, we are asking how well can we predict the linear trend response given output from the step change experiments. Figure S2 shows PA between each linear trend experiment and its LTR reconstruction. The period of interest for our observational reconstruction is between 1958 and 2017 so we will focus on validation of 60 year reconstructions. After 60 years PA for the water cycle amplification experiment matches the LTR reconstruction to within a factor of 0.03 (i.e. PA is 3.22% for the simulation and 3.31% based on LTR). For the ice melt case the difference is a factor 0.1 and in the warming case a factor of 0.12. For all the forcings combined the discrepancy is less than a factor of 0.11.
Using the time series of ocean heat content change discussed in section 2.1, we estimate that ocean warming resulted in a contribution to PA of 2.21% ± 0.46% between 1958 and 2016 ( figure 4(b)). Estimates of fresh water input into the ocean, comprising both total glacial melt (mass input into the ocean from melting land glaciers and ice sheets) and melting of multi-year sea ice, suggest a contribution to PA of 0.75% ± 0.14% up to 2016 ( figure 4(b)).
We have estimated PA based on the most recent objectively analysed hydrographic observations for periods 1957to: 19851995;. Between 1957, the surface salinity pattern amplified by 5% ± 1.1%. Having inferred the contributions to PA from surface ocean warming and ice mass loss, we subtract these from the observed PA up to 2016. The resulting residual PA is 2.04% ± 1.2% over 1957-2016 ( figure 4(b)), which we attribute to water cycle change.
Observational estimates of water cycle change are unreliable (Hegerl et al 2015). Model studies, however, have consistently suggested a tight relationship between measures of water cycle change (e.g. global mean precipitation, meridional moisture transport, etc) and global mean surface air temperature change Figure 4. (a) Global mean surface air temperature change (ΔSAT) relative to the period 1880-1920 (cyan); five-year running mean global ocean heat content anomaly relative to 1957 (red); ice mass loss from glaciers (blue) and sea-ice (light gray). (b) Proposed budget of contributions to salinity pattern amplification including surface warming (red), idealised ice mass loss (blue) and water cycle amplification related linearly to surface air temperature change (cyan). Shown in grey is the sum of these contributions and in black observationally based linear trend estimates (1957-1987, 1957-1995, 1957-2005, 1957-2016). (ΔSAT; see Kirtman et al (2013) and references therein). We thus reduce the problem of quantifying water cycle change to a single multiplying factor for ΔSAT. We use NASA air temperature data (Hansen et al 2010) and note that uncertainties in global air temperature are negligible compared to the range of possible water cycle to SAT amplification factors. Combining the transient linear response approach and the water cycle forced PA inferred above and the ΔSAT time-series we infer an amplification factor of 3.6% ± 2.1% per degree Celsius ΔSAT.
Our 60 year PA estimate based on EN4 is weaker than 50 year estimates from recent authors (5. 4% Boyer et al (2005), 5.5% Hosoda et al (2009), and 8% Durack et al (2012)). Linearly extrapolating those three estimates to 60 years and weighting them equally with our EN4 based estimate yields PA of 6.9% ± 1.9%. Subtracting the effect of ocean warming and ice mass loss, as above, the implied amplification factor is 6.9 ± 3.6%/ • C. Analysis by Durack et al (2012) was distinct in that trends estimated in regions where data coverage was substantial were extrapolated to regions where data coverage was not sufficient to discern a trend. Conversely in other analyses, which used optimal interpolation or objective analysis, trends were only estimated, and therefore contributed to PA, close to regions where observations supported a trend. Excluding Durack et al's estimate yields an amplification factor of 5.4 ± 0.18%/ • C and considering only their estimate and reported uncertainty (0.7%) yields an amplification of 11.6 ± 1.7%/ • C.

Conclusions
Surface salinity observations have previously been put forward as a means to estimate past water cycle amplification and monitor ongoing variability and change (Schmitt 2008, Kerr et al 2010. Our analysis suggests a strong influence of ocean warming on the evolution of surface salinity. In numerical simulations with anomalous surface warming we have shown that the surface salinity pattern amplifies substantially. This change is coincident with a reduction in the diffusive flux of fresh water from fresh to saline waters and reduced mixed layer depths. There is potential for water cycle changes to be monitored using surface salinity observations if these surface warming effects are appropriately accounted for.
Combining both numerical model experiments and observational analysis up to 2016 we report that approximately one third of observed salinity pattern amplification can be explained by ocean warming and one sixth by ice mass loss. The remaining signal can be attributed to a water cycle amplification of 3.6% ± 2.1% per degree Celsius of global surface air temperature rise.