In recent years, the Arctic Ocean has been under significant changes as shown by numerous in situ and remotely sensed measurements. The temperature of the upper layer of the Arctic Ocean has been increasing and more solar heat has been absorbed by the increasing ice-free areas [1
Latest observational and modeling studies have documented changes in the upper Arctic Ocean hydrography [4
]. In particular, an increase of liquid freshwater content over both the Canadian Basin and the central Arctic Ocean has been observed. This increase of freshwater has been linked to an intensification in the large-scale anticyclonic winds as well as sea level pressure changes [5
]. An increased Bering Strait freshwater import to the Arctic Ocean, a decreased Davis Strait export, and the enhanced net sea ice melt could play an important role in the observed freshwater trend [6
Rivers are important sources of freshwater and heat to the Arctic Ocean and changes in the river runoff or temperature could have a strong impact on the Arctic system. An increment of the global mean annual temperature will produce an increase in the discharge of Arctic rivers [7
The 2015 update of the Arctic Report Card alerts that, in 2014, the combined discharge of the eight largest Arctic rivers was 10% greater than their average discharge during the 1980–1989 period [9
]. However, the impact of this increase of freshwater runoff on the Arctic ocean dynamics remains unknown due to the lack of available salinity measurements in the Arctic.
Unfortunately, the number of surface salinity measurements is very scarce at high latitudes, especially in the Arctic Ocean. In such context, the three L-band missions—the Soil Moisture and Ocean Salinity (SMOS) mission [10
]; the Aquarius mission [13
]; and Soil Moisture Active Passive (SMAP) observatory [15
]—can provide an unprecedented source of salinity information over the Arctic Ocean, which can help to improve the models.
The retrieval of sea surface salinity (SSS) from microwave radiometric measurements is based on the emissivity of the ocean surface, which depends on the dielectric constant of sea water that is a function of temperature and salinity, and on the sea surface roughness. The SMOS radiometer operating frequency (1.43 GHz, in the L-band) provides good sensitivity of the ocean-surface brightness temperature (
) to SSS in the tropics and subtropics [16
]. In cold waters, however, the sensitivity of the
to salinity decreases rapidly [17
]. As shown in [18
], such sensitivity drops from 0.5 K/psu to 0.3 K/psu, when SST decreases from 15
C to 5
C. Moreover, some undesired effects in SMOS
and to lesser extent in Aquarius and SMAP
measurements, such as the land–sea and ice–sea contaminations, and Radio Frequency Interference (RFI) [19
] make the Arctic region one of the most challenging regions for SMOS SSS retrieval.
Some previous works assessed the quality of SMOS SSS at high latitudes. For example, in Köhler et al. [20
] he authors performed a comparison of previous versions of SMOS (salinity maps computed from the L2OS v550) and Aquarius products with in situ measurements and models for the north Atlantic region, but they did not perform any comparison inside the Arctic Basin. Despite the large biases (mainly produced by land–sea and ice–sea contaminations) that affected the SMOS L2OS v550, in Matsuoka et al. [21
] this product was used to develop an algorithm for identifying surface water sources in the southern Beaufort Sea by using Aqua/MODIS ocean color along with SMOS SSS L2. Recently, the potential and challenges of monitoring the Arctic Ocean SSS by using SMAP data have been demonstrated in [22
A recently developed SSS retrieval algorithm [23
] has noticeably improved the coverage of the global SMOS SSS leading to retrievals in some critical areas where no-valid or few salinity retrievals were available before (for example in the Mediterranean Sea [24
]). The Barcelona Expert Center (BEC) team used this methodology for the generation of three-year time series of SMOS SSS at high latitudes. In [26
], a comparison of these SMOS SSS maps and three other SSS products provided by Aquarius with in situ data is performed. The authors concluded that SMOS SSS maps are consistent with ship and CORA5.0 data, although they also pointed out that the sea ice mask should be improved.
In this work, we generate seven years of SMOS SSS maps at high northern latitudes (beyond
N) by using the methodology described in [23
]. Additionally, we improve the methodology in terms of the seasonal bias. The objectives of this work are the following: (i) to present seven-year time series of this new SMOS SSS product at high northern latitudes; (ii) to assess the quality of these new SMOS SSS maps at high latitudes by comparing them to different sources of in situ data; (iii) to compare the SMOS SSS with other available products in this region (model and other remotely sensed SSS products); and (iv) to show the potential of SMOS SSS to capture the SSS variability in the Arctic region.
The paper is structured as follows: In Section 2
, we describe the different datasets that are used. In Section 3
, the methodology used for the generation of the SMOS salinity maps is briefly presented. The assessment of the SMOS salinity maps is presented in Section 4
. Variations of SSS shown by SMOS, Aquarius, SMAP and the model outputs from TOPAZ close to the mouth of the major Arctic rivers are shown in Section 5
. A final discussion is provided in Section 6
3. Methodology Used for the SMOS SSS Product Generation
Seven years (2011–2017) of the SMOS L1B data product (v620), provided by the ESA, are processed to generate salinity maps at high latitudes (from N to N).
The galactic [35
], sun glint [36
] and surface roughness [37
] contributions are corrected using auxiliary information provided by ECMWF [38
], similar to what is done in the official ESA SMOS L2 SSS products. The dielectric constant model proposed by Meissner and Wentz (M&W) [39
] is used instead of the model defined by Klein and Swift (K&S) [40
], which is used in the official SMOS Ocean Salinity Level 2 product. The work presented in [41
] shows that, when analyzing SSS from Aquarius, differences between M&W and K&S are small at low and mid latitudes, but they increase at high latitudes, i.e., in cold waters. The authors concluded that, for very cold waters (colder than 3
C), retrieved salinities using M&W model are significantly closer to in situ floats measurements than those retrieved using K&S.
measurements are geo-referenced using a 25-km resolution Equal-Area Scalable Earth (EASE) North Pole grid [42
]. To account for the SMOS residual spatial and temporal systematic errors, the SSS retrieval methodology presented in [23
] is used. This methodology introduces important changes with respect to the standard processing [16
] used in the ESA SMOS L2OS processor:
Individual retrievals: The retrieval follows a non-Bayesian scheme, that is, for each SMOS a single value of SSS is retrieved.
Characterization of the systematic errors: All the SSS retrieved under the same acquisition conditions, i.e., the same geographical location, incidence and azimuth angles and satellite overpass direction (ascending/descending). throughout this seven-year period are accumulated in a SSS distribution. The systematic error associated to each acquisition condition is estimated by computing the central estimator of the corresponding SSS distributions. We use the mode of the distribution as the central estimator, i.e., as the SMOS climatological value for each specific acquisition condition. In this aspect, a relevant difference with respect to the official SMOS L2OS processor is that the
used for the ESA SMOS L2OS SSS retrieval are previously corrected by Ocean Target Transformation (OTT) [44
]. The OTT is computed as the mean of the difference between the measured and modeled
s (applying the Geophysical Forward Model) at a particularly stable region of the ocean. We do not apply an OTT since systematic errors are already accounted for, point by point, with the new methodology.
Filtering criteria: The statistical properties of those SSS distributions are also used for filtering the non-accurate measurements. Two types of filters are applied to remove questionable values and outliers in the SSS retrievals: (i) all the SSS belonging to distributions having a large standard deviation (std larger than 10), defined by too few measurements (less than 100), or with a large skewness (larger than 1 in absolute value) or kurtosis (lower than 2) are all excluded (i.e., the distribution is marked as “bad” distribution, and all its salinities are discarded); and (ii) an additional outlier criterion is applied to the remaining retrieval values by further excluding any value that is farther than 10 (in absolute value) from the SMOS climatological value (see more details in [23
Computation of SMOS anomalies: The SMOS-debiased SSS anomalies are computed by subtracting to each individually retrieved SSS value (corresponding to a specific acquisition condition) the corresponding SMOS climatological value (computed as explained in (b)), thus effectively removing local biases, especially those produced by the land–sea (or ice–sea) contamination and permanent RFIs.
Computation of SMOS SSS: In [23
], the SMOS SSS are generated by adding an annual SSS reference (annual WOA SSS, [45
]) to the SMOS anomalies. This is an issue for the SMOS SSS values of the Arctic Ocean, since there are many zones in the Arctic with very few measurements of SSS (as shown in the bottom-right plot of Figure 1
), and any reference could provide non-accurate SSS values there. For this reason, before generating the seven years of SMOS SSS, we analyze two test datasets by using two annual references: WOA; and the Polar science center Hydrographic annual Climatology (PHC) (version 3) [46
] which is the usual reference for Arctic regions. We assess the quality of these two datasets by comparing the resulting SMOS SSS products—the SMOS SSS computed from WOA annual reference (SMOS woa) and SMOS SSS computed from PHC (SMOS phc)—with TARA SSS. In Section 4.1
, a full discussion of this assessment is given. The conclusions of these comparisons are summarized in Figure 2
. SMOS SSS has lower RMS with respect to TARA SSS than the corresponding annual references used for its generation (i.e., the blue and red lines are below the green and black lines, respectively, except in Buffin Bay where the SMOS-PHC has slightly larger RMS than the PHC product). The actual RMS values, together with the bias and standard deviation values, can be found in Table 1
. However, many regions in the Arctic Ocean present large RMS values. These regions correspond to the areas where few or no in situ data were taken into account in the generation of the annual reference. In Section 4.1
, we describe this analysis in more detail. Since both references provide similar results (WOA is slightly better), we use WOA for the generation of SMOS SSS product to be coherent with the global SMOS SSS product distributed by BEC.
Objectively analyzed maps: Objectively analyzed nine-day SSS maps at 25-km resolution are generated daily. In [23
], the same correlation radii used for the computation of WOA products were proposed for the generation of the SMOS SSS products: 321 km, 267 km, and 175 km (see [45
]). These correlation radii do not seem to be the most appropriate for describing the dynamics in the Arctic region [47
]. For this reason, we assess the impact on SSS quality of using different correlation radii, by means of the following experiment:
We consider a finite set of candidates for the first correlation radius, : 175 km, 200 km, 225 km, 250 km, 275 km, 300 km, and 325 km.
For each one of the previous values, we consider the second
radii of convergence such that:
taking each one of the following values:
For each set of three convergence radii computed as before, we generate seven months of SMOS nine-day SSS maps: from the 1 April to the 9 November of 2013, one map every five days.
We apply the time-bias correction proposed in [23
], by removing the (spatial) mean anomaly between each SMOS SSS map and the annual WOA SSS.
We compare the resulting SMOS SSS with Argo SSS. We use as a metric the global RMS of the difference between SMOS and Argo SSS, averaging the seven months for each one of the previous choices of and .
The configuration which provides the lowest error (
) is the one with
, which is the closest one to the configuration proposed in [23
]: 321 km, 267 km and 175 km. Notice that the global L3 SMOS SSS maps distributed by BEC (http://bec.icm.csic.es/ocean-experimental-dataset-global/
) use a set of smaller correlation radii (175 km, 125 km and 75 km) for better describing the mesoscale. The results raised from this experiment suggest that, at high latitudes, the larger is the correlation radii, the larger is the smoothing effect, and therefore the lower is the noise. This is probably because individual SMOS SSS retrievals at high latitudes are noisier than in other regions of the globe. In other words, the generation of SMOS SSS maps with smaller correlation radii and the same level of noise as in the case of the global SSS maps requires SMOS SSS retrievals less noisy. In this sense, improvements at
level as the ones introduced in [48
] and assessed at salinity level in [49
] are providing promising results in terms of noise reduction in the SSS retrievals. The application of this technique will probably help to retrieve more accurate SSS in those regions and therefore to generate SMOS SSS maps with smaller correlation radii (more appropriate to capture the dynamics of this region).
Mitigation of the seasonal bias: An additional time-dependent bias correction is needed to mitigate the effect of seasonal and other time-dependent biases which affect the SMOS
]). In [23
], the authors proposed subtracting the global mean of the SMOS SSS anomaly for each nine-day map. This assumption is appropriate for global SSS maps, as it implies that the total content of salt remains constant in time. However, the application of this hypothesis regionally, in particular at high latitudes, produces seasonal biases. In other words, there are net exchanges of salinity across region boundaries. In a recent study [25
], a multivariate analysis is used to characterize and mitigate the time-dependent bias in the SMOS SSS maps in the Mediterranean Sea. In this work, we include a simpler time-dependent bias correction:
We consider the Argo SSS available for the same nine-day period used in the generation of the nine-day SMOS SSS maps.
We compute the median of the differences between the collocated SMOS SSS fields and the Argo SSS.
We subtract this median from each nine-day SMOS SSS map.
shows the time-dependent correction resulting from this procedure, which has been applied to each map.
5. Sea Surface Salinity Variability Observed by SMOS at the Mouth of the Main Arctic Rivers
The largest intra-annual variability observed by SMOS is located near the mouth of the main Arctic rivers. In Section 4.3
, we show that the inter annual variations of SSS described by SMOS agree with the ones described by TSG SSS in sub-Arctic regions. However, inside the Arctic basin, very few in situ measurements are available. In this section, we show that SMOS SSS variability is consistent with the SSS dynamics of the region, in the Arctic basin. In particular, we analyze the SSS variability close to the mouth of the Mackenzie and Ob Rivers. We compare the SMOS SSS maps with the output of the TOPAZ model and with the remotely sensed SSS provided by Aquarius and SMAP. We also look at the discharge data provided by Arctic Great Rivers project to correlate the freshening observed by SMOS to the river discharge events.
In Figure 11
, monthly SMOS SSS maps (July, August and September) close to the mouth of the Mackenzie and Ob Rivers are shown (left and right plots, respectively). In Figure 12
, TOPAZ SSS maps are displayed for the same months and regions. In Figure 13
, the same regions and months are also used for representing Aquarius SSS maps (for years 2012 and 2014) and SMAP SSS maps (for years 2016 and 2017 in the case of the maps close to the Mackenzie River and 2015 and 2017 in the case of the maps close to the Ob River). Ice mask thresholds are different: the SMAP products use the limit of 3% of Sea Ice Concentration (SIC) computed with the Bootstrap algorithm [52
], above this threshold that pixel is not considered water and is filtered out; Aquarius uses a threshold of 15% of the same algorithm; SMOS considers water pixels those with a SIC lower than 15% by using the EUMETSAT Ocean and Sea Ice Satellite application Facility (OSI-SAF) product.
shows the daily river discharge of the Mackenzie and Ob Rivers. The Mackenzie River presents the maximum discharge by the end of May, except in 2012 when two maximums were observed (Figure 14
left). At that time, there is still a high percentage of sea ice in the region. Sea ice can be considered melted (less than 30% of ice concentration) by mid-July, except for 2013, when it melted slightly later (not shown). Due to the strong density stratification of the Arctic Ocean, the newly supplied fresh water tends to stay at the surface in the absence of enough strong wind-driven stirring. The persistence of the river plume on surface is well observed in the temporal evolution (July–August) of SMOS SSS maps in the Mackenzie River (left plots of Figure 11
). The output SSS from the TOPAZ model in the mouth of the Mackenzie River (left plots of Figure 12
) shows a smaller plume than the one displayed by SMOS. It does not change noticeably during the different years, while SMOS shows important inter annual differences. Moreover, variability for different months is not observed with TOPAZ outputs, while it is observed very clearly with SMOS. Aquarius SSS (2012 and 2014 of the left plots of Figure 13
) suffers from strong positive biases with respect to the SSS captured by SMOS and TOPAZ at the ice edge. Although the ice mask in SMAP is more restrictive than for SMOS, both satellites observe coherent plumes structures. However, SMAP SSS gradients are larger than the ones captured by SMOS.
Daily river discharge data illustrates that the maximum discharge of the Ob River occurs by the end of May (in Figure 14
right), and that the greatest discharge happened in 2015. Since the region is almost melted around the beginning of July, the maximum Ob River discharge occurs before the sea water is free of ice.
Despite of the differences in the grid resolution: 12.5 km for TOPAZ and 25 km for SMOS, the plume of the Ob River described by TOPAZ (right plots of Figure 12
) seems more consistent with the one described by SMOS (right plots of Figure 11
) than in the case of the Mackenzie plume (in terms of SSS variability and the spatial coverage of the plume). Aquarius maps (right plots of Figure 13
) display saltier SSS than SMOS, SMAP and TOPAZ close to the Ob River mouth, which does not seem geophysically reasonable. The inter annual variations shown by SMAP and SMOS (right plots of Figure 11
and Figure 13
) are also coherent. For example, 2015 is the year that both satellites display the major extension of fresh water in July and August, even though SMAP shows a larger plume than SMOS. This is also coherent with the in situ river discharge data shown in Figure 14
(right). TOPAZ maps also show a large extension of fresh water in 2015.
In Figure 15
, we compare monthly SMOS SSS anomalies with respect to monthly river discharge anomalies. These anomalies are computed as follows. First, the daily river discharge data are monthly averaged. Since these regions are frozen in winter months, and we are interested in comparing these data with SSS, we consider as reference the average of the monthly discharge of June, July, August and September for the seven years of study (2011–2017). The monthly discharge anomalies (green points in Figure 15
) are generated by subtracting this reference to the monthly discharge data. Consistently, we consider as SMOS SSS reference the average of the monthly SMOS SSS maps of June, July, August and September for 2011–2017. The SMOS SSS anomalies shown in Figure 15
are the differences between the monthly SMOS SSS maps and the mentioned reference. The purple points represent the spatial average of these SMOS SSS anomalies in two regions close to the mouth of the Mackenzie River (latitudes between 69
N and 73
N and longitudes between 150
W and 130
W) and the Ob River (latitudes between 73
N and 77
N and longitudes between 60
E and 90
E). The relationship between the SSS anomaly and the discharge anomaly is not straightforward. The geophysical phenomena that play a role on the modification of the SSS in those regions are diverse and complex, for example the melting of the ice, the mixing due to wind, water advection, etc. A complete understanding of the interactions of these phenomena and the salinity are out of the scope of this work. At this stage, a more qualitative analysis is carried out, i.e., to check whether decreasing trends of SMOS SSS anomalies are linked to positive trends of river discharge anomalies and vice versa. As already mentioned, in both cases (Mackenzie and Ob Rivers), the maximum discharge happens in May when the region is usually frozen. That is why the green lines in Figure 15
usually show a decreasing trend. Consequently, increasing trends of SMOS SSS anomalies (purple lines) are expected. Figure 15
shows that, typically, SMOS SSS anomalies increase from July. However, from June to July, both the SMOS SSS and river discharge anomalies mainly show a decreasing trend. This can be explained by several mechanisms: on the one hand, the discharge measurements are not performed in the mouth of the rivers, therefore a delay between the discharge variations and the SSS response is expected, while, on the other hand, as already discussed, these regions are typically strongly stratified and, when the wind is not strong enough, the newly supplied fresh water tends to stay at the surface. Additionally, in July, the ice melting still occurs and refreshes the water in those regions. For the rest of months, an anti-symmetric (anti-correlated) behavior of the SSS with respect to the river discharge is observed (as expected).
This paper demonstrates the capability of SMOS-derived SSS to follow the vast salinity spatial and temporal variability in Arctic and sub-Arctic regions. This new product is produced using the retrieval algorithm proposed by Olmedo et al. [23
] and includes some improvements with respect to the initial data set distributed by the Barcelona Expert Center. In particular, an enhanced time-dependent bias correction is applied. We also analyze the impact of considering different annual SSS references for the generation of the product. The conclusion of this analysis is that there are regions in the Arctic Ocean where the analyzed references are of poor quality, and, therefore, the corresponding SMOS SSS products suffer from spatial biases (which are constant in time). However, the dynamical range of the SSS described by these annual references (and also by the SMOS product) is geophysically consistent. Additionally, we analyze which is the most appropriate correlation radii for the generation of the objectively analyzed SSS fields. The conclusion of this analysis is that correlation radii greater than what is expected by the dynamics of the region provide the lowest error with respect to Argo floats, probably because the SMOS SSS retrievals have a larger error than in other regions (of warmer waters).
Validation of SMOS salinity maps at high latitudes against Argo floats shows that the product has a std in the range of 0.24–0.35. We have also used a dataset of 86 TSG provided by Copernicus. Besides the statistics of the differences between SMOS and TSG, we analyze the inter annual variations described by each one of the two SSS sources. SMOS and TSG SSS are in agreement in the major features of the SSS dynamics, although, as expected, the spatial scales that are resolved by SMOS are blurred with respect to the ones resolved by the TSG. It is important to notice that most of the measurements used in these comparisons are located outside the Arctic Ocean. Therefore, this comparison cannot be used to project a realistic quality assessment of SMOS SSS maps inside the Arctic Circle.
To complete the overview of the quality of this SMOS SSS product in the Arctic regions, we compare the SMOS SSS with TOPAZ, Aquarius and SMAP SSS close to two of the main rivers of the Arctic Ocean: the Mackenzie and Ob Rivers. We observe that the output from TOPAZ underestimates the plume of the Mackenzie River. Aquarius SSS maps are probably affected by ice–sea (and maybe also by some residual land–sea) contamination, showing saltier waters than the other satellite datasets do. Despite the differences in the ice mask, the inter- and intra-annual variations described by SMAP and SMOS are quite consistent. Although a more extensive work on the comparison between the remote sensed salinity products is still required, these results suggest that both SMOS and SMAP have great potential to routinely monitor the extension of the surface freshwater fluxes in the Arctic Ocean.
Even though the original resolution of the SMOS SSS is ≈40 km, the selected correlation radii are likely too large to monitor some mesoscale features of these Arctic and sub-Arctic regions. The reduction of the correlation radii will be driven by the reduction of the noise of the salinity retrievals. Future improvements of this SMOS SSS product are aimed at reducing the error of the salinity retrievals at high latitudes. Enhanced image reconstruction techniques as the one introduced in [48
] have been assessed in [49
] reaching promising results. The decrease of the error in the salinity retrieval would allow generating SMOS SSS maps with smaller correlation radii. Then, a richer mesoscale dynamics is expected to be captured by the future SMOS SSS products.
Finally, we would like to underline the need of implementing more in situ measurements in Arctic regions. On the one hand, they are absolutely required for a proper assessment of the satellite products in these regions. On the other hand, they are also required for the computation of a better annual reference in the Arctic Ocean. There are still many regions in the Arctic Ocean where very few measurements have been used for the computation of the salinity climatology. This in particular means that the knowledge of the salinity of those regions is really limited.