3.1. Validation with In Situ Salinity
We aggregated in situ salinity data collected by Argo floats, ships, gliders, and in OMG field campaigns to create a gridded in situ data set on 0.25° × 0.25° grids daily from 1 April 2015 to 31 March 2018. Measurements of different instruments, such as CTD (Conductivity, Temperature and Depth), XCTD (Expandable CTD), and TSG (Thermosalinograph) taken within 5 m of the sea surface are used. Figure 1
shows the distribution of in situ measurements North of 50°N on 0.25° grids, as long as there is as least one daily record in three years on the grid point. Areas South of 60°N are mostly covered by Argo floats, complemented by a few ship tracks. Areas between 60°N and 80°N in Greenland area are well covered combining ships, gliders, and OMG data. However, above the Arctic Circle, particularly North of the Bering Strait, there are almost no in situ salinity data.
The three-year mean of SSS averaged over collocated grid points from April 2015 to March 2018 are illustrated in Figure 2
. In situ data confirms the large-scale salinity feature that was observed in SMAP SSS, i.e., the much saltier seawater on the Atlantic side than on the Pacific side of the Arctic Ocean. The biases with respect to in situ salinity are less than 1 psu in a majority of the areas, which are masked out in white so areas with large biases stand out more clearly (Figure 2
, right column). It can be seen that SMAP SSS shows positive biases in a few places near the coast, and negative biases along the east coast of Greenland and near Davis Strait (Figure 2
d). On the other hand, HYCOM SSS shows mostly positive biases east of Greenland (Figure 2
e). The contrast between SMAP and HYCOM east of Greenland is particularly interesting. It suggests that SMAP SSS retrieval may contain un-detected sea ice effect, while HYCOM is not able to catch the freshening signal due to ice melting where no Argo floats are available for HYCOM’s operational data assimilation.
The statistical results of comparing in situ salinity with collocated SMAP SSS and HYCOM SSS are presented in Table 1
and Figure 3
. All data are gridded on 0.25° grids daily, covering the period from April 2015 to March 2018. North of 50°N, the bias, standard deviation, and Root Mean Square Difference (RMSD) between 19,738 pairs of SMAP SSS and in situ are 0.442, 2.391 and 2.431 psu with correlation of 0.81. The statistics between HYCOM SSS and in situ is slightly better. The open circles in the scatter plot indicate outliers defined as those pairs where the absolute difference exceeds three times the standard deviation. The number of such defined outliers is less than 1% of the total collocated data set. The entire set of collocated data can be visually divided into two groups: one group of high SSS (>25 psu) and the other of low SSS (<10 psu). It is interesting to note that in the low SSS group, in situ salinity are distinct at three sub-groups that are centered around 2, 5, and 7 psu, with collocated SMAP SSS spread from 0 to 10 psu, while HYCOM SSS almost perfectly aligned with the in situ data. Again, this is expected because most of the in-situ data being used for the evaluation here are from Argo profiling floats, and HYCOM assimilates the Argo float data operationally.
Significant outliers of SMAP SSS are seen as being associated with the high SSS group in two branches, with positive and negative biases, respectively, where positive bias may indicate an over estimation of the surface roughness correction, and the negative bias may suggest un-detected ice contamination. This will be considered for the next version of SMAP geophysical model function to improve the SSS retrieval algorithm. Note that there is no negative bias in HYCOM for the high SSS group. However, there is similar positive bias in the HYCOM versus in situ. After excluding the outliers, the bias, standard deviation, and RMSD between 19,543 pairs of SMAP SSS and in situ are 0.385, 0.987, and 1.060 psu with correlation of 0.82.
North of the Arctic Circle, the number of collocated points drops more than 70%. Table 1
also provides of statistical results North of 65°N. The bias and RMSD between 5712 pairs of SMAP and in situ after removing outliers are 0.339 and 1.227 psu, respectively.
After excluding the outliers, we performed a linear regression of the SMAP (or HYCOM) vs. in situ data. The regression slope for SMAP is close to 0.9 if only the data for in situ greater than 25 psu are considered (Figure 3
). If all data are included, then the regression slope is 1.06. Interestingly, HYCOM SSS seems to have a systematic positive bias for in situ SSS in the range of 25 to 30 psu, resulting in a reduced regression slope of 0.467 for the high SSS group of data. The SMAP SSS conditioned on the in situ within this range has a smaller bias, although being more noisy.
3.2. The Arctic Ocean SSS and Sea Ice
illustrates the monthly mean SSS from 2015 to 2017 in August, when the Arctic Ocean ice has the minimum extent with maximum ice-free surface areas for satellite salinity retrieval. The large-scale salinity feature, i.e., saltier Atlantic and fresher Pacific is consistently observed in three consecutive years and it agrees with the known contrast between the two ocean basins. North of the Arctic Circle (~66°N), however, SMAP SSS shows large interannual variations, not only in magnitudes of SSS retrieved, but also in areas with a valid retrieval. For example, North of the Bering Strait, in August 2015, SSS is retrieved in the large areas of the Chukchi Sea and the East Siberian seas; in August 2016 in the Beaufort Sea and the East Siberian Sea; and, in August 2017 areas with valid retrieval extended further north in all three marginal seas. This is due to the differing sea ice extent for the same months among three years, because SMAP SSS retrieval is only possible in open water. For the regions with SSS retrieval in all three years, e.g., in the Hudson Bay or Kara Sea, SMAP SSS shows large interannual differences. In contrast, HYCOM SSS shows a much small variation in the marginal seas within the Arctic Circle. It is known that HYCOM uses climatology river discharge and its SSS is relaxed to a seasonal climatology; both may suppress the magnitude of interannual variations of SSS in HYCOM [71
SMAP SSS in marginal seas near the ice edge often show very low values. These fresh signatures should be examined and validated very carefully because it could be a mixture of real signal and ice contamination. It is known that in seawater near the ice edge, sea ice formation, and melt have significant contributions to the mixed layer salt budget of the ambient waters, with growing importance toward the ice edge [59
]. Sea ice concentration (SIC) averaged in the areas North of 65°N changes dramatically through the seasonal cycle, from 20% in October or November to more than 70% in winter (January–May) (blue curve in Figure 5
). It appears that the seasonal cycle of SSS averaged North of 65°N closely follows that of SIC, with the lowest SSS being observed within one or two months of the minimum SIC. This covariance should be taken with a grain of salt though, since the low SSS in summer also includes the effect of river runoff, and the high SSS in winter largely comes from the Atlantic side when most marginal seas are frozen. However, it is evident that SMAP SSS (black curve in Figure 5
) reveals large interannual variations that are consistent with the sea ice conditions. Note that while averaging of SIC was over the entire region including zero or non-zero SIC values, the averaging of SSS is obtained over ice-free areas, which itself changes with time. We found the ice-free area during the ice melting seasonal peak of 2017 is 5% and 10% more than the two previous seasons of 2015 and 2016 (red curves in Figure 5
). Correspondingly, the Arctic Ocean averaged SMAP SSS is the lowest in August 2017 despite HYCOM SSS showing essentially no changes. To alienate the possibility that summer SSS minimum was dominated by a few extremely low SSS retrievals or river discharges, Figure 5
also shows the time series of SSS after excluding outliers, which is identified as where the absolute differences from the daily mean exceeding three times of the standard deviation in the averaging domain. After removing the outliers (about 1 to 5% of valid retrievals of the day), the domain averaged SSS increases by 1–2 psu in its summer low for both SMAP and HYCOM, indicating that most outliers are in the low end of SSS. It is important to note that with outliers removed, SMAP SSS still shows clear interannual variability with minimum of 2017 season more than 2 psu lower than that of 2016, in contrast to HYCOM SSS. The consistent seasonal variation of SSS and SIC suggests that SMAP SSS retrieval correctly characterized the effect of sea ice changes.
However, this agreement is encouraging but insufficient to completely exclude the effect of un-detected ice-contamination impacting SMAP SSS. The appearance of frazil-pancake ice maybe increasing in the margins of the Arctic during autumn, as identified in a recent field campaign [16
]. This type of new sea ice forms in the presence of incoming surface waves and it commonly occurs in the wave-dominated marginal ice zone of the Southern Ocean. With the retreat of the Arctic summer ice edge, the increased fetch has led to the increase of surface waves in the summer and fall, particularly in the Beaufort-Chukchi seas. Frazil-pancake ice may present a challenge to identify, say with passive microwave, as the sea ice concentrations may be quite variable over short time scales, which may lead to contamination in the SSS fields. Frazil-pancake ice has long been identifiable in SAR imagery. One recent study [72
] describes the validation of a method to estimate the thickness of this young ice type based on wave dispersion. Being able to identify this form of new ice in the Arctic along with other sea ice conditions (e.g., melt, thin ice thickness, eddy formation, presence of waves) and the extent within the marginal ice zone will improve the accurate detection of ice and reduce the uncertainty in the satellite SSS retrieval.
3.3. SSS and River Discharge
As indicated in Section 3.1
, there are very few in situ salinity data in the marginal seas within the Arctic Circle; for example, in the Hudson Bay and Kara Sea, no in situ data during the three years of SMAP were identified. In this section, we explore an alternative validation by examining the change of SMAP SSS in response to independently measured river discharge data. Massive northern rivers transport huge quantities of water from the continents to the Arctic Ocean. The freshwater inputs that are associated with river runoff should be reflected in the SSS field, particularly near the river mouths and over the shallow shelf areas. Due to its proximity to the two major Arctic rivers (Ob’ and Yenisey), the Kara Sea provides an ideal case to examine SSS response to river discharge.
illustrates the evolution of SMAP SSS over the Kara Sea for the warm season (May to October) of 2015 and 2016. No SSS were retrieved in the Kara Sea before June and after October because it was completely covered by sea ice. The first valid SSS retrievals appeared in June, but at different locations in 2015 and 2016 due to the different spatial coverage of sea ice. The differences between these two years became more dramatic throughout the season. In the summer of 2015, the freshwater patch first appeared in June, in the area east of the northern tip of the Novaya Zemlya archipelago. It grew and spread to cover almost half of the Kara Sea in August. In contrast, in the season of 2016, the freshwater signature was limited to areas near Ob’ and Yenisey Bay. Because the whole region has become almost ice-free after July, the impact of new freshwater inputs from sea ice melt is likely to be minimal. Therefore, the dramatic freshening signature spreading through middle of the Kara Sea from July to September in 2015 and along the Siberia coast in 2016, must have originated from the other freshwater source—river discharge. Below, we show that the contrast in SMAP SSS clearly reflects the differences in river runoff in those two years.
As a reference, Figure 7
illustrates the SSS evolving patterns in the Kara Sea from HYCOM SSS. The anomaly in HYCOM SSS is an order of magnitude smaller. As mentioned before, HYCOM is forced by climatological river discharges and its SSS is relaxed to a seasonal climatology [71
]. These two climatological forcing would suppress the magnitude of interannual variations of SSS in HYCOM for regions without in-situ data or where in-situ data are insufficient to constraint the model SSS, such as the Kara Sea.
In the surface layer of Kara Sea ocean currents carry waters from the River Ob’ and River Yenisey to the north and northeast. Based on the SMAP SSS observations in the Kara Sea (Figure 6
), we find that river discharge adjacent to the Kara Sea must be much larger in 2015 than in 2016. The additional freshwater in 2015 was transported northward and further spread along the path, while the spreading of freshwater discharge in 2016 was limited along the coast.
Indeed, the daily discharge data from the Ob’ River and the Yenisey River (Figure 8
) show differences that are consistent with SMAP observation. In the first two months of the warm season (May and June), discharge from Ob’ shows a similar magnitude for the two years, while Yenisey discharge peaked early and injected 93.4 km3
more water in 2015 than 2016. In the following four months from July to October, Ob’ became the main player, putting 90.9 km3
more freshwater into the Kara Sea, while Yenisey added another extra 31.5 km3
). Combining the discharges from the Ob’ and Yesiney during May to October together, the Kara Sea received more than 210 km3
extra freshwater in 2015 relative to 2016.
It might be useful to roughly estimate the effect of this extra amount of freshwater on the SSS anomaly. About 3 cm of freshwater are needed to dilute 1 m of seawater with a change of salinity by 1 psu. Assuming the extra ~210 km3
freshwater spread over half of the Kara Sea (total surface area 926,000 km2
), it may produce 15 psu salinity changes within top 1 m surface water layer, or 7.5 psu within top 2 m. As seen in Figure 6
(right column), the areas with positive SSS differences that are exceeding 10 psu covering about half of the Kara Sea areas, which is in the same order of magnitude as the freshening effect that was possibly produced by the river discharge difference in the two years. It is understood that large differences in SSS anomaly may depend on how the discharge freshwater transported horizontally and vertically, in terms of depth and spread of the diluted water body. Nevertheless, the agreement in order of magnitude is encouraging.
3.4. SSS Variability at Arctic Ocean Gateways
We have demonstrated that SMAP SSS retrieves reasonably good quality data in the Arctic Ocean in terms of comparisons with in situ salinity data (Section 3.1
) and in response to river discharge (Section 3.3
). Here, we further explore the feasibility of using satellite SSS to monitor the surface salinity variability at Arctic Ocean Gateways. We examine if the SSS variability at the gateways are significantly greater than the retrieval accuracy of 1 psu.
The Arctic Ocean exchanges freshwater with the sub-oceans through four major gateways, as shown schematically in the currents map (Figure 9
a from Figure 2
a in [14
]): the Bering Strait inflows of relative fresh Pacific waters; the Barents Sea Opening (BSO) and part of the Fram Strait inflow of the salty Atlantic water; and, through the Davis Strait and the part of the Fram Strait between Greenland and Svalbard, which comprise the major outflow locations of water modified by the Arctic Ocean in addition to freshwater that is associated with sea ice flux and subsequent melt [12
]. Because the general directions of the ocean currents at gateways are known, the surface salinity variability observed at the gateways carries the information of the freshwater passing through the regions in the upper ocean.
We first describe the procedure to extract SSS information at each of the Arctic gateways. The exact locations of the gateways defined for this study are given in Table 3
and are indicated in Figure 9
b. To ensure currents flow through the passage in a roughly unique direction, we divided the Fram Strait into two parts: one between east Greenland coast and 0° longitude (named EG), the other from 0° longitude to the west coast of Svalbard (named Fram Strait). We also added a section to extend the narrow Bering Strait from 62°N to 68°N so sufficient grid points can be included in the calculation. According to this definition, a maximum of around 250 to 500 grid points can be extracted from SMAP L3 daily maps with 0.25° spacing, as listed in Table 3
. The actual number of grid points with valid SSS retrieval changes with time depending on the ice situation, which varies with the gateway. This is illustrated in terms of the percentage of the ice-free area at each gateway (Figure 10
, right column). We note that BSO and the Fram Strait are mostly ice-free year round, while EG and the Davis Strait are only open for about three months in summer, with more than 50% ice cover the rest of the year. The Bering Strait, which is the only gateway connecting the Arctic Ocean and Pacific Ocean, opens about half of the year from June to December. The time series of SSS are obtained by daily averaging over all grid points with valid retrievals at each gateway. In the extremely challenging environment of the Arctic Ocean, we designed two criteria to avoid possible false retrievals. First, we include in the time series calculation only if >30% of the grid points are ice-free. The daily mean and standard deviation were obtained from 30% or more grid points at the gateway with valid SSS data, named SSSGateway
(t) and δGateway
(t). Examining the time series, we found extremely large values of δGateway
(t) often occurred at seasonal transitions of ice melting or freezing, suggesting the possibility of undetected ice contamination. Therefore, we defined the second criteria to exclude those outliers. We calculate the standard deviation of SSSGateway
(t) over the whole period (σGateway
), and defined outliers as those with δGateway
(t) exceeding twice of σGateway.
After filtering with these two criteria, we apply a 30 days moving average on SSSGateway
The time series of SSS at Arctic gateways that were observed by SMAP during the past three summers reveal very rich information (Figure 10
, left column). First, it correctly reflects the typical characteristics of the freshwater sources from different ocean basins: relative salty water from the Atlantic and fresh water from the Pacific, with more than 3 psu differences between mean SSSBSO
is similar to SSSBSO
in terms of mean and variability; both bring warm and salty Atlantic water into the Arctic Ocean. Also note, SSSFram
is slightly fresher than SSSBSO
, possibility due to its vicinity to Greenland. On the other hand, SSSEG
, although being geographically adjacent to each other, bear dramatic contrast in the properties of water transport out from (SSSEG
, mean 31.2 psu) and into (SSSFram
, mean 34.8 psu) the Arctic Ocean. The much fresher signature and larger dynamical range of SSSEG
suggest the impact of Arctic ice. We note the exceptionally large drop in SSSFram
(~3 psu) and SSSEG
(~6 psu) from late 2016 to early 2017. It is unlikely that this large interannual change is caused by the SMAP algorithm or long-term calibration changes because the changes at BSO are much smaller. At this moment, we do not have enough data to pin point the exact mechanism causing such freshening. Also, the magnitude of SSSEG
may need further calibration due to less coverage (~40% ice-free area). We speculate that this freshening and recovering in the later part of 2017 is influenced by the Greenland sea ice melting and freezing. This hypothesis may be supported by the observation that the Davis Strait, another passage like EG where outflows of Arctic Ocean modified water mixed with sea ice drift to the northern Atlantic. Different from SSSEG
shows similar seasonal cycles for the three summers that were captured by SMAP.
Consistent seasonal cycles are also observed over the Bering Strait, with a peak-to-peak range of about 4 psu. It is also interesting to note the inter-annual difference. SSSBering
in June–December 2016 (red curve in Figure 10
i) is exceptionally low, about 2 psu less than the same period in 2015 (black curve in Figure 10
i); while it recovered partially one year later, as seen in Figure 10
i, the green curve (June–December 2017) falls between the previous two years. Year 2016 was the strongest El Niño
in recent two decades. One well-known characteristic of an El Niño
event is the zonal displacement of the western equatorial Pacific warm/fresh pool. The edge of the pool extends eastward during El Niño
, retreats westward during La Niña
Moorings are installed at all major Arctic Ocean gateways, measuring salinity at depth ~50 m year round. At the close of this study, we only obtained mooring data at the Bering Strait up to the summer of 2016 [24
], with about one year overlapping with SMAP data. Figure 11
illustrates the time series of SMAP SSSBering
(from Figure 10
i) along with salinity form June 2015 to June 2016 at three mooring sites: A2 (66°19′N, 168°57′W) and A3 (65°46′N, 168°34′W) in the US channel of the strait, and A4 (65°44′N, 168°15′W) close to the Alaskan coast. All three mooring sites are within the white polygon, defined as the Bering Strait (Figure 9
b). They are within 90 km from each other. During August–November, the salinities from the three moorings were mostly within 1 psu from each other, indicating a relatively small spatial variability. However, the A4 mooring data reduced by about 5 psu from November to December, whereas the other two moorings had smaller changes. The SMAP SSSBering
at 1–2 cm depth depicted a change of about 5 psu, well resolved by the 1 psu accuracy (Section 3.1
), and appeared to have a consistent trend with the A4 mooring on the amplitude and timing of the freshening peak in December 2015. Nevertheless, the differences with the other two moorings were clearly larger. A2 and A3 moorings revealed less variability relative to SMAP, which is likely due to the vertical and horizontal dilution from the surface to 50 m depth. A longer time series of data is required to bring a more definite conclusion on the consistency or discrepancy, possibly resulting from differences in sensing depth or spatial variability.