Remote sensing of sea surface salinity: challenges and research directions

ABSTRACT

Salinity is a key parameter that affects the surface, deep circulations, and heat transport of oceans. Sea surface salinity (SSS) represents the salinity at the ocean surface and impacts atmosphere – ocean interactions and vertical ocean circulation. To monitor SSS, three passive microwave radiometers with an L-band (1.4 GHz) have been launched since 2009. The scientific need for SSS retrieval and estimation has grown in recent years; however, the operational retrieval of SSS via satellite remote sensing still faces significant challenges. This study provides a review of satellite-based SSS retrieval methods and guidelines to encourage future research. This paper introduces satellite-derived SSS research trends and summarizes the representative SSS satellite sensors and their retrieval methods. The limitations and challenges of satellite-derived SSS are then discussed. The errors from the retrieval algorithms, discrepancies in the spatio-temporal scales of in situ and remote sensing, and limitations of the satellite-derived SSS are then detailed. Finally, our paper provides suggestions for the future directions of SSS remote sensing in five ways: mitigation of measurement errors, improvement of currently available SSS products, enhancement of the usage of in situ data, reconstruction of three-dimensional salinity information, and synergetic uses of multi-satellite missions.

KEYWORDS:

• Sea surface salinity
• salinity
• passive microwave
• Ocean color
• Remote sensing
• L-band

1. Introduction

Salinity, along with temperature, is a key parameter that affects the surface and deep circulation and heat transport of oceans (Barale, Gower, and Alberotanza 2010; Dinnat et al. 2019; Martin 2014; Park 2015). Ocean salinity is also an important indicator that characterizes the global water cycle because it is impacted by various factors, such as precipitation, evaporation, the freezing and melting of sea ice, and river runoff; approximately 85% of global evaporation and 80% of global precipitation occur above oceans (Martin 2014). Sea surface salinity (SSS) represents an ocean’s salinity at its surface; SSS affects atmosphere-ocean interactions and vertical ocean circulation (Dinnat et al. 2019; Durack et al. 2016). SSS monitoring is essential because it can provide key information on the global water cycle and ocean dynamics (Barale, Gower, and Alberotanza 2010; Martin 2014).

SSS monitoring methods include in situ measurements, model-derived reanalysis data, and satellite observations. The Array for Real-time Geostrophic Oceanography (ARGO) is a global project which involves in situ measurements of oceanographic parameters (i.e. sea temperature and sea salinity) that began in the 2000s (Dinnat et al. 2019). ARGO collects these parameters using a sparse array (average of 3° × 3°) every 10 days (Vinogradova et al. 2019). In addition to ARGO, many countries also monitor seas around their territories. However, in situ measurements are relatively sparse, which makes it difficult to monitor spatiotemporally continuous SSS over vast areas. Using reanalysis data and satellite-derived data mitigates this limitation. The Hybrid Coordinate Ocean Model (HYCOM), which assimilates in situ measurements using vertical and horizontal interpolation algorithms (Cummings and Smedstad 2014), provides SSS at a resolution of 1/12° every 3 hours. However, the accuracy of the HYCOM is relatively low in regions with rapidly changing low salinity water (Jang et al. 2022) and the application of the HYCOM in real-time SSS monitoring is limited because the HYCOM does not include satellite data. Therefore, satellite sensors play a crucial role in SSS monitoring because they enable real-time SSS observations.

Since 2009, three passive microwave radiometers with an L-band (1.4 GHz) have been launched to monitor SSS via the National Aeronautics and Space Administration (NASA) Aquarius mission (August 2011 to June 2015), European Space Agency (ESA) Soil Moisture and Ocean Salinity (SMOS) mission (since May 2010), and NASA Soil Moisture Active Passive (SMAP) mission (since April 2015). The three satellite sensors provide SSS estimates based on a dielectric constant model (Reul et al. 2020); a detailed description of their retrieval algorithms is shown in Section 2 and Reul et al. (2020). In addition, various satellite sensors, such as the Advanced Microwave Scanning Radiometer Earth observing system (AMSR-E), Moderate Resolution Imaging Spectroradiometer (MODIS), Landsat series, and Geostationary Ocean Color Imager (GOCI), have been used to estimate SSS (Chen and Hu 2017; Kim et al. 2020; Reul et al. 2009; Wang and Deng 2018; Zhao, Temimi, and Ghedira 2017). Despite the relatively short history of satellite observations of SSS (approximately 10 years), there has been an increase in the need of these observations in recent years (Wu et al., 2018; Vinogradova et al. 2019; Reul et al. 2020). Satellite-derived SSS has broad research applications, including in the fields of ocean dynamics, hydrological cycles, climate change prediction, and the relationship between SSS and environmental phenomena, such as hurricanes and flooding (Vinogradova et al. 2019). However, the operational retrieval of SSS data via satellite remote sensing still faces significant challenges.

Such challenges have been discussed in previous studies over the last decade. Some studies summarized the currently available satellite-derived SSS products for Aquarius, SMOS, and SMAP, and enumerated the limitations of the SSS products, such as uncertainties in retrieval algorithms, coarse spatio-temporal resolution, and the need for further validation in specific regions from which limited in situ data are available (Pope et al. 2017; Vinogradova et al. 2019; Dinnat et al. 2019; Reul et al. 2020; Le Vine and Dinnat 2020). However, Vinogradove et al. (2019) suggested further applications of SSS remote sensing in hurricane monitoring, El Niño-Southern Oscillation (ENSO) forecasting, and terrestrial flood and drought prediction. Recent reviews have highlighted the major scientific achievements of satellite-derived SSS (Reul et al. 2020) and described the trade-offs of operating SSS satellites in terms of their frequency and spatial resolution to provide a design guideline for future SSS satellites (Le Vine and Dinnat 2020). However, studies on guidelines to minimize the limitations and improve the availability of satellite-derived SSS products on both spatial and temporal domains are lacking. Our study provides a review of satellite-based SSS retrievals and guidelines to encourage future research in various fields by improving the scientific understanding of satellite-derived SSS.

Our study aims to (1) summarize satellite-derived SSS research trends, (2) examine the limitations and challenges of satellite-derived SSS, and (3) discuss the future directions of remote sensing of SSS. The remainder of this paper is structured as follows: Section 2 briefly summarizes commonly used satellite data and methods for SSS estimation, Section 3 discusses the limitations and challenges of remote sensing of SSS, Section 4 discusses future directions and potential approaches for remote sensing of SSS, and Section 5 presents the conclusions.

2. Trends in satellite-derived SSS research

This section introduces trends in recent research on remote sensing of SSS and summarizes the satellite sensors and methods used to retrieve and estimate SSS. First, we searched the SCOPUS database using the keywords “remote sensing” and “sea surface salinity” and restricted our review to articles published during the last 21 years (2001–2021); we gathered 531 studies, which were further examined by sensor and mission types used. The number of publications on the remote sensing of SSS per year has increased over time (Figure 1). However, although three L-band passive microwave radiometers have been monitoring global SSS since 2009, research on the remote sensing of SSS was actively conducted before 2009. Most articles published before 2009 address issues regarding mission configuration, scientific requirements, and technical constraints of the L-band remote sensing of SSS.

Figure 1. Trends in the number of publications on remote sensing of sea surface salinity by year.

Satellite-based SSS retrievals and monitoring can be divided into 3 groups in terms of sensor types: L-band (1.4 GHz) passive microwave radiometers, C- (6.6 GHz) and X-band (10.7 GHz) passive microwaves, and Ocean Color (OC) satellite sensors (Reul et al. 2020). SSS retrievals using L-band microwave data are based on the temperature-dependent sensitivity of the L-band to water salinity (Barale, Gower, and Alberotanza 2010; Li, Liu, and Zhang 2019; Tang et al. 2017; Reul et al. 2020).

With respect to sensor type (Figure 2a), other than the general keyword “microwave radiometer,” “L-band” was the keyword that most frequently appeared in the publication database, followed by “Ocean color” and “X-band” (91 (80%), 21 (19%), and 1 (1%), respectively). As shown in Figure 2b, most papers were related to L-band passive microwave radiometers (Aquarius (154 articles; 38%), SMOS (157 articles; 39%), and SMAP (38 articles; 9%)), followed by those involving OC missions, such as MODIS (32 articles; 8%) and Landsat (7 articles; 2%).

Figure 2. Number of studies by (a) sensor type and (b) satellite missions of the 531 articles published between 2001 and 2021.

The dielectric constant of seawater is strongly influenced by salinity (Lang et al. 2016). The dielectric constant affects the sea surface emissivity, which is measured using radiometers, in terms of their brightness temperature (T_b), as a function of SSS and sea surface temperature (SST) (Dinnat et al. 2019). However, T_b is mainly affected by sea surface characteristics (Yueh et al. 2010); a rough sea surface under windy conditions and the existence of sea foams may cause an increase in the surface emissivity (Sasaki et al. 1985; Yin et al. 2014, 2016). To retrieve SSS from T_b data, excessive emissivity should be accurately assessed via roughness correction using a geophysical model function (Freedman, McWatters, and Spencer 2006; Tang et al. 2017). Unlike SMOS and SMAP, Aquarius can correct for sea surface roughness using its own L-band active radar. SMOS and SMAP correct for roughness effects using ancillary data, such as the European Centre for Medium-Range Weather Forecasts (ECMWF) data (Freedman, McWatters, and Spencer 2006).

Several studies have attempted to correct biases of satellite-retrieved SSS products resulting from T_b uncertainties in various ways. Olmedo et al. (2019) and Qin et al. (2020) conducted the bias correction of SMOS and SMAP SSS products through empirical modeling. Others adopted machine learning approaches to improve the quality of satellite-retrieved SSS products (Rajabi-Kiasari and Hasanlou 2020; Jang et al. 2021, 2022). Jang et al. (2022) improved global SSS collection based on seven machine learning approaches (i.e. K-nearest neighbor, support vector regression, artificial neural network, random forest, extreme gradient boosting, light gradient boosting model, and gradient boosted regression trees) through the synergistic use of SMAP and HYCOM SSS. Several studies have been conducted to correct the SSS retrieved by the three commonly used microwave sensors, namely Aquarius, SMOS, and SMAP, by modifying their dielectric models and input data (González-Gambau et al. 2017; Kolodziejczyk et al. 2016; Liu, Wan, and Hong 2020; Olmedo et al. 2018; Sharma 2019).

Microwave sensors with C-band (6.6 GHz) and X-band (10.7 GHz) (e.g. Water Cycle Observation Mission Interferometric Microwave Imager (WCOM/MI), AMSR-E, and HaiYang (HY)-2A) have also been used to estimate SSS (Li, Liu, and Zhang 2019; Reul et al. 2009; Song and Wang 2017). The C-band and X-band in vertical polarization (V) can be applied to SSS estimation using the dielectric constant model because they reduce the influence of sea surface roughness and thermal impacts more effectively than the L-band microwave sensors (Reul et al. 2009; Song and Wang 2017). However, both bands (i.e. C- and X-bands) lower the sensitivity of T_b to changes in salinity by a factor of 10 to 20 (approximately 0.06 K/psu and 0.03 K/psu for C- and X-bands, respectively) (Reul et al. 2009; Picart et al. 2008).

The use of microwave sensors for SSS remote sensing has several limitations. First, the observed T_b is susceptible to contamination by Radio Frequency Interference (RFI) from land, particularly in coastal regions. In addition, passive microwave-retrieved SSS products typically have coarse spatial (25–100 km) and temporal (3 days or more) resolutions. These limitations make it difficult to use microwave radiometers in coastal areas, which have complex coastlines and high SSS variability. OC or optical sensors [i.e. MODIS, Sea-viewing Wide Field-of-view Sensor (SeaWiFS), Visible Infrared Imaging Radiometer Suite (VIIRS), GOCI, and Landsat-8 Operational Land Imager (OLI)] have been used in open oceans and coastal regions as substitutes for microwave sensors (Chen and Hu 2017; Kim et al. 2020; Sun et al. 2019; Vandermeulen et al. 2014; Wang and Deng 2018; Zhao, Temimi, and Ghedira 2017). SSS itself does not directly emit color signals, but they can be indirectly estimated from OC signals (Gholizadeh, Melesse, and Reddi 2016; Kim et al. 2020).

Moreover, SSS typically has a negative relationship with Colored Dissolved Organic Matter (CDOM) (Bowers et al. 2000; Yu et al. 2017). As an important optical property of the ocean, CDOM has been used, as a proxy, in the retrieval of SSS (Bricaud, Morel, and Prieur 1981; Yu et al. 2017). However, CDOM signals are mainly caused by terrestrial sources that experience the inflow of freshwater, particularly that of river plumes in coastal regions (Kim et al. 2020). Thus, satellite-derived CDOM can be effectively used to estimate SSS in river-dominated regions adjacent to coasts (Bai et al. 2013; Keith, Lunetta, and Schaeffer 2016; Nakada et al. 2018; Sasaki et al. 2008; Wang and Deng 2018; Yu et al. 2017). Because the relationship between CDOM and SSS varies by region and season, it is difficult to generalize the SSS estimation algorithms developed locally based on CDOM concentrations (Chen and Hu 2017). In addition, because satellite-derived CDOM is a function of remote sensing reflectance (Rrs) of water, CDOM retrieval using satellite-derived Rrs is additionally uncertain due to the indirect estimations involved (Wang and Deng 2018). OC sensors provide a finer spatial resolution (0.25–1 km) than passive microwave radiometers (25–100 km). For example, the Geostationary Ocean Color Imager II (GOCI-II) can be used to periodically monitor the diurnal cycle of SSS in northeast Asia (2,500 × 2,500 km) every 1 hour from 8 AM to 5 PM with a spatial resolution of 250 m. SSS estimation using OC sensor data can minimize the limitations of microwave sensor-derived SSS, which does not cover coastal oceans and marginal seas, thereby improving the ability to study land – sea linkages (Vinogradova et al. 2019). However, cloud contamination and the varied sensitivity of surrogate variables (e.g. CDOM) are major limitations of estimating SSS using OC sensors.

3. Challenges of satellite-derived SSS

Various methods for estimating SSS based on satellite data have been investigated over the past decade, starting with the dielectric constant model, which is a basic retrieval algorithm. Although SSS estimation based on satellite data has substantially improved in recent years, there are still several major challenges, including the uncertainty of retrieval algorithms and ancillary products, land-sea contamination (LSC), and the difficulty in validating satellite-derived SSS using in situ measurements due to discrepancies in their spatial scales as summarized in Table 1. This section discusses the major challenges associated with deriving SSS from satellite data and the recent advances in addressing them.

Table 1. Summary of the five challenges in remote sensing of sea surface salinity (SSS).

(1) Errors from retrieval algorithms	Extraterrestrial sources (i.e. cosmic background, galaxies, sun, and moon) Attenuation in the atmosphere and ionosphere (Faraday rotation) Environmental factors affect Tb retrieval (detailed in the 2^nd and 3^rd challenges)
(2) Errors from ancillary data	Uncertainties of ancillary SST and wind speed data that affect surface emissivity
(3) Difficulties in coastal regions	Contamination of Tb influenced by RFI and LSC Coarse spatial resolution of passive microwave radiometers in coastal regions Uncertainties from the indirect estimation of SSS using CDOM as a proxy and of cloud contamination of OC sensors in coastal regions
(4) Discrepancies of in situ and remote sensing	Retrieval errors from the discrepancies in the spatiotemporal scales of point-wise in situ SSS and pixel-wise satellite observations
(5) Limitation of spatio-temporal resolution	Limits in spatiotemporally continuous SSS products (i.e. daily global SSS coverage) using current SSS-observing L-band passive microwave radiometers

3.1. Errors from retrieval algorithms

Observational errors caused by sensors, antennas, and various environmental factors affect surface emissivity and received radiance during the retrieval of SSS from satellite sensor data. Environmental factors include extraterrestrial sources (i.e. cosmic background, galaxies, sun, and moon), attenuation in the atmosphere and ionosphere (Faraday rotation), and other environmental conditions that affect surface emissivity, including heavy precipitation, low sea surface temperature, and RFI (Barale, Gower, and Alberotanza 2010; Li, Liu, and Zhang 2019; Martin 2014; Park 2015). Although SSS retrieval algorithms contain corrections for environmental disturbances (Dinnat et al. 2019; Reul et al. 2020), it is challenging to eliminate them.

Moreover, apart from observational errors, conventional SSS retrieval algorithms based on the dielectric model have internal limitations. Empirically, the dielectric model is based on numerous instances of varying temperatures and salinity levels. However, the model is severely limited because it cannot account for all possible environmental conditions. A few studies have estimated SSS by empirically recalculating dielectric models based on the different physical conditions of SST, wind speed, and other atmospheric parameters for a given area (Dinnat et al. 2019; Reul et al. 2009; Song and Wang 2017). For example, Dinnat et al. (2019) reprocessed the Aquarius SSS by modifying three components: the dielectric constant, ancillary SST data, and atmospheric attenuation model.

Another significant source of errors in satellite-retrieved SSS is changes in sea surface emissivity caused by wind speed (Barale, Gower, and Alberotanza 2010; Martin 2014). To correct for wind-affected ocean surface roughness, Sharma (2019) developed a regression equation between wind speed and ΔSSS (the difference in SSS between flat and rough ocean surfaces). A higher wind speed results in a more turbulent sea surface, which decreases the accuracy of the correction (Park 2015; Reul et al. 2009). Liu et al. (2020) modeled SSS using Aquarius data for strong winds (i.e. 7–22 m/s) with coefficients different from those used in the model for low speeds (<7 m/s), which increased the accuracy when compared to the model with the same coefficients for all wind speeds.

The sensitivity of microwave sensors to SSS is significantly affected by SST (Barale, Gower, and Alberotanza 2010; Dinnat et al. 2019; Park 2015). In cold waters, the SSS retrieval error increases as the sensitivity of sensors to the heat emitted by the sea surface decreases and sea ice contamination increases. Supply et al. (2020) modified the dielectric constant model using a novel empirical correction and calibrated the SMOS SSS to remove a global bias by adding a constant. The corrected model performed well in retrieving SSS in the Arctic Ocean and the proposed retrieval algorithm demonstrated the potential to increase the accuracy of retrieved SSS at low SST. The current operational SMOS and SMAP SSS retrieval algorithms use external ancillary wind speed and SST data, which are considered two major external error sources. These two error sources are further discussed in Section 3.2 and other error sources in coastal areas are discussed in detail in Section 3.3.

3.2. Error from ancillary SST and wind products

SST and wind speed affect the T_b detected by microwave sensors, which is closely related to the accuracy of the SSS products (Dinnat et al. 2019). Many studies have evaluated multiple sources of ancillary data (i.e. SST and wind speed) to determine whether they can improve the accuracy of SSS retrievals. Meissner et al. (2016) examined the sensitivity of retrieved SSS to different ancillary SST datasets. They compared the Aquarius SSS retrieval performance with four SST data: the National Oceanic and Atmospheric Administration (NOAA) Optimally Interpolated SST (OISST), as an original ancillary dataset of Aquarius; weekly SST averages from WindSat; Canadian Meteorological Center (CMC) daily SST; and Multi-Scale Ultra-High Resolution (MUR) SST from the Jet Propulsion Laboratory (JPL). Although SST retrieved from thermal infrared sensors is heavily influenced by clouds, high water vapor, and aerosols, microwave-derived SST is rarely affected by these factors (Meissner et al. 2016). Dinnat et al. (2019) and Olmedo et al. (2019) used the Operational Sea Surface Temperature and Sea Ice Analysis (OSTIA) SST produced by the United Kingdom Met office as ancillary data to improve the satellite-retrieved SSS products.

With respect to wind ancillary data, SMOS uses European Center of Medium-Range Weather Forecasts (ECMWF) wind data (Supply et al. 2020) and National Center for Environmental Prediction (NCEP) Global Data Assimilation System (GDAS) wind data (Fore et al. 2020) in their retrieval algorithms. However, these ancillary data have some limitations. For example, the NCEP GDAS wind data is used after preprocessing of bilinear interpolation in longitude and latitude to the SMAP. Furthermore, SMAP uses the NCEP Global Forecast System (GFS) to overcome the limited availability of NCEP GDAS wind data (available every 6 h). However, ancillary data have not been completely collocated with satellite observations in both spatial and temporal domains. As an alternative, Cyclone Global Navigation Satellite System wind speed data are useful under low wind speed conditions. The data improved the accuracy of SMAP and SMOS SSS retrieval algorithms based on their unique sensitivity to the smooth ocean surface at low wind speeds (Liu et al. 2020).

3.3. Difficulties of remote sensing in coastal regions: RFI, LSC, and resolutions

The T_b measured at the land surface is higher than that of the ocean and is influenced by anthropogenic RFI from the land, which increases the error of T_b measurements along the coast (Kolodziejczyk et al. 2016; Park 2015). Contamination of T_b reduces the retrieval accuracy (González-Gambau et al. 2017). Together with the issue of RFI, mitigation of LSC in microwave signals is a major issue in the field of microwave satellite-based SSS retrievals (Dinnat et al. 2019). Many efforts have been made to reduce systematic errors in satellite-retrieved SSS measurements, such as those posed by LSC near coastal regions (González-Gambau et al. 2017). Kolodziejczyk et al. (2016) and Boutin et al. (2018) proposed LSC corrections considering the biases of SSS in both the spatial and temporal domains. Gonzalez-Gambau et al. (González-Gambau et al. 2017) modified the T_b correction techniques for coastal regions in two ways. They reduced T_b contamination using residual multiplicative error correction at the calibration level and the nodal sampling method at the imaging level. Olmedo et al. (2018) improved SMOS SSS in the Western Mediterranean Sea through the synergistic use of debiased non-Bayesian retrieval and Data Interpolating Empirical Orthogonal Functions (DINEOF). Nevertheless, owing to the coarse spatial resolution of microwave sensor data, there are still high uncertainties in T_b measurements in coastal regions that often show rapid SSS changes in a short period. As discussed in Section 2, to compensate for these limitations, OC sensors have been used to monitor SSS in coastal regions.

In contrast to previous studies, recent studies have estimated SSS using not only CDOM but also Rrs signals based on statistical models, which include simple linear regression, polynomial regression, multivariate nonlinear regression, and principal component analysis (Bai et al. 2013; Choi et al. 2021; Kaffah and Damayanti 2020; Liu et al. 2017; Marghany, Hashim, and Cracknell 2010; Nakada et al. 2018; Qing et al. 2013; Sun et al. 2019; Vandermeulen et al. 2014; Yu et al. 2017; Zhao, Temimi, and Ghedira 2017). In addition, machine learning approaches (i.e. decision trees, random forest, support vector regression, and artificial neural networks) have been widely used to model the complex relationships between input parameters and SSS (Bao et al. 2019; Chen and Hu 2017; Geiger et al. 2013; Kim et al. 2020; Liu et al. 2015; Wang and Deng 2018; Urquhart et al. 2012). Machine learning models based on OC sensor data generally show better estimation accuracy than statistical models but exhibit relatively poor accuracy in open oceans compared to L-band remote sensing. It is difficult to apply these locally developed empirical algorithms to other regions with distinct characteristics and must be customized for different regions (Chen and Hu 2017).

3.4. Discrepancies between spatio-temporal scales of in situ and remote sensing

The expected errors of in situ SSS measurements vary by instrument sensitivity but are generally on the order of 0.01 practical salinity units (psu) (Delcroix et al. 2005). Because of the low signal-to-noise ratio of satellite observations, the measurement errors of satellite-derived SSS are relatively large (of the order of 0.1 psu scale) (Boutin et al. 2018; Kao et al. 2018). As previously stated, measurement errors are influenced by uncertainties resulting from systematic errors, empirical retrieval algorithms, and inaccuracies of ancillary data (Lagerloef and Font 2010; Reul et al. 2020).

Another major error source comes from the discrepancies in the spatiotemporal scales of point-wise in situ SSS and pixel-wise satellite observation data (Wang, Sun, and Zhang 2019). In situ point-wise measurements and satellite pixel-wise observations (i.e. spatially averaged footprints) cannot be resolved at the same scale (Wang, Sun, and Zhang 2019). For example, satellite SSS retrievals represent the Gaussian-weighted average within the satellite footprint (40 km for SMOS and SMAP and 150 km for Aquarius). In contrast, in situ data represent point-by-point measurements and can observe small-scale variations in SSS (Wang, Sun, and Zhang 2019). The variability within a satellite pixel is smoothed over in the satellite footprint (about 40–100 km for L-band and 0.5–4.0 km for OC), resulting in different signals (T_b or Rrs) compared to those from point-based signals, particularly in regions with large spatio-temporal SSS variations due to frequent heavy rainfalls or massive river inflows (Stoffelen 1998; Vinogradova and Ponte 2013; Boutin et al. 2016). Previous studies have examined the representativeness error of SSS products. The global average error of SMOS and SMAP Level (L3) SSS products compared to that of the ARGO floats data was approximately 0.093 psu (Wang, Sun, and Zhang 2019). Some studies have validated satellite-based SSS products by focusing on regional SSS variations. D’Addezio and Subrahmanyam (2016) compared the Aquarius and SMOS L3 SSS and ARGO float data for the Agulhas Return Current region (ARC, 20°E–100°E, 45°S–20°S) in the South Indian Ocean, which revealed the strongest western boundary current experienced between 2010 and 2014 in the Southern Hemisphere. Both SSS products were accurate, resulting in a correlation coefficient of 0.9 and Root Mean Square Deviation (RMSD) less than 0.2 psu. Tang et al. (2017) validated the SMAP SSS product using various in situ-collected data at global and regional scales. The study found an RMSD of 0.2 psu between SMAP L3 SSS and ARGO global gridded data in the tropics and mid-latitudes between 40°S and 40°N. The accuracy of SMAP L3 SSS regionally varied, resulting in RMSDs of 0.56, 0.26, and 0.5 psu for the Mediterranean Sea, Tropical oceans, and Bay of Bengal, respectively. Although representative errors in satellite-derived SSS have been quantified (Wang, Sun, and Zhang 2019, 2021), measurement errors in satellite-derived SSS due to sampling errors and region-specific uncertainties remain the most significant limitations.

The spatial coverage of in situ observations is another challenge. There are spatial variations in SSS by region type (e.g. marginal ocean, open ocean, and coastal region). Furthermore, there are regional differences in the number of in situ SSS observations from the ARGO float. The Western Pacific near East Asia, the south Atlantic, the southern Pacific near north Oceania, and the polar seas are examples of regions where few in situ observations are available (Figure 3a). Owing to the lack of in situ observations, the accuracy assessment of satellite-retrieved SSS in these regions is difficult to conduct. The RMSD of satellite-retrieved SSS in coastal and marginal seas is relatively higher than that in open oceans. Because some coastal areas have large river outlets, it is crucial to accurately estimate the SSS for low-salinity waters (e.g. the Bay of Bengal and the East China Sea). In particular, there are no in situ observations in the East China Sea (i.e. Yellow Sea), which has relatively low-salinity water due to the Yangtze River outflow. A low-salinity event affects the environment by altering the biological or physical characteristics of seawater. For instance, low salinity causes sea surface warming by impeding the vertical heat exchange between the warm surface and cold thermocline below (Moon et al. 2019). However, these regions are usually near land, which is susceptible to issues such as LSC and RFI. To validate satellite-derived SSS products and further improve their retrieval algorithms, the spatial coverage of in situ SSS observations should be expanded in the future, particularly for coastal regions, marginal oceans, and high latitudes.

Figure 3. (a) Density of ARGO floats in situ SSS observations per 1° x 1° square collected from April 2015 to December 2020 and (b) spatial distributions of the difference between SMAP and in situ data (SMAP SSS–ARGO SSS).

Because empirical SSS retrieval algorithms based on the dielectric constant concentrate on normal seawater that has salinity from 34 to 36 psu with average conditions of temperature and roughness, the retrieval accuracy for high and low saline water regions is relatively low (Zhou et al. 2017; Vinogradova et al. 2019; Reul et al. 2020). High-salinity regions include the subtropical North and South Atlantic and tropical southeast Pacific, while low-salinity regions are found in the Southern Pacific Convergence Zone, Eastern Pacific Fresh Pool, and North Pacific (Figure 3b).

3.5. A limitation of satellite-derived SSS: spatio-temporal resolution

As a well-known limitation of sun-synchronous satellite observations, the three SSS-observing L-band satellites cannot provide daily global SSS coverage (Figure 4). Level-3 (L3) products provide global SSS coverage, but their revisit time is 7–8 days. Aquarius has a spatial resolution of approximately 150 km with a 7-days’ revisit time, whereas SMOS and SMAP provide products with a spatial resolution of 40 km and a revisit time of 3–5 days, resulting in a global coverage of 8 days.

Figure 4. Daily coverages of three L-band radiometer missions on 31 May 2015: (a) Aquarius, (b) SMOS, and (c) SMAP.

Spatiotemporally continuous SSS products are often required for various regional and global applications. Since ocean circulation has changed in recent years due to ongoing climate change, more frequent global SSS monitoring has become an important issue in ocean and climate science (Zika et al. 2018). However, the relationship between SSS variation and the global water cycle remains unclear (Vinogradova and Ponte, 2017). Compared to the 1950s, the global SSS pattern amplification (PA) has increased by 5–8% (Durack, Wijffels, and Matear 2012; Skliris et al. 2014), implying that the subpar spatiotemporal resolution of current SSS products should be improved.

In the field of ocean remote sensing, motivated by the need for spatiotemporally continuous SSS products, the spatial and/or temporal interpolation of SST has been extensively studied. For example, the NOAA OISST has provided daily global SST data at 0.25° grids since September 1981 (Reynolds et al. 2007). The OISST is an analysis developed by combining observations from various platforms (satellites, ships, buoys, and ARGO floats). Since September 1981, the OSTIA system has produced daily SST data for oceans around the globe with a spatial resolution of 5 km (Donlon et al., 2012). The system combines in situ observations with various satellite data provided by the Group for High-Resolution Sea Surface Temperature (GHRSST). The analysis is based on a variant of optimal interpolation (Martin et al., 2014; Donlon et al., 2012). Both products (i.e. OISST and OSTIA) have been widely used in many applications, from fundamental climate applications to climate studies (i.e. El Niño), owing to their reasonable accuracy (Yang et al. 2021).

However, previous attempts to generate SSS maps based on optimal interpolation have yielded low levels of accuracy (over 0.2 psu on average). Melnichenko et al. (2014) generated weekly optimal interpolation-based SSS maps from the Aquarius L2 product in the North Atlantic region (0°–40°N) with 150 km grids, resulting in an RMSD of approximately 0.28 psu. Buongiorno Nardelli, Droghei, and Santoleri (2016) introduced the multidimensional interpolation of SMOS SSS using SST and in situ SSS over the Southern Hemisphere (10°S–65°S). They generated daily SMOS SSS maps with 0.25° grids, yielding Root Mean Square Error values ranging from 0.13 (in the North) to 0.40 (approximately 60°S) psu. This optimal interpolation technique is cost-effective with light computational demand but insufficient for operational SSS mapping and analysis because of its relatively low accuracy. Consequently, more advanced techniques that can produce accurate spatiotemporal continuous SSS data should be explored.

4. Future directions

There are still large gaps in the estimation of SSS using satellite remote sensing. Considering the limitations and challenges mentioned in the previous section, we suggest future directions to 1) mitigate measurement errors, 2) improve currently available SSS products, 3) enhance the usage of in situ data, 4) reconstruct three-dimensional salinity information, and 5) synergistically use multi-satellite missions for SSS estimation.

4.1. Mitigation of retrieval errors

The retrieval errors of SSS vary by algorithm, depending on the calibration method and ancillary data used. One of the biggest sources of uncertainty for SSS retrievals is temperature (i.e. SST), implying that large errors in SST lead to high uncertainty in SSS (Dinnat et al. 2019). Several studies have attempted to correct the SST-dependent bias that arises when using empirical dielectric models. Dinnat et al. (2019) addressed the SST-dependent bias and revised atmospheric attenuation model, resulting in a significant reduction in the retrieval bias of SSS (less than 0.1 psu). The improvement of internal models used for SSS retrieval can also increase the accuracy of SSS retrieval. However, these issues remain in SSS retrieval from cold waters such as polar seas, where in situ samples and reliable input variables are very limited (Dinnat et al. 2019). Multiscale Ultrahigh Resolution 0.25° (MUR25) SST analysis data might be useful because MUR25 integrates in situ SST measurements from the NOAA iQuam dataset (Xu and Ignatov 2014) to improve SST parameterization in polar regions (Yang et al. 2021).

In addition, SSS in coastal regions retrieved from L-band microwave sensors still has high uncertainty owing to LSC and RFI; thus, many L-band-derived SSS products do not provide data in coastal areas (~40 km; Gabarró, Martinez, and Font 2012). Satellite-derived SSS errors are high in coastal regions and low in open oceans (Reul et al. 2020). It is necessary to consider other satellite sensors to estimate SSS in coastal regions, as data are not available up to 100 km off the coast because of the high uncertainty of L-band microwave radiometers. One solution is to use microwave sensors with frequencies higher than that of the L-band. In particular, the effects of sea surface roughness and temperature are effectively minimized by combining C- and X-bands in vertical polarization under moderate wind speeds (Reul et al. 2009). C- and X-bands are more sensitive to higher SST (Klein and Swift 1977) and the difference between the two channels in T_b showed little dependence on sea surface roughness under normal wind speed conditions (4 to 10 m/s) (Reul et al. 2009).

Another option is to use OC sensors for SSS monitoring, which have higher spatial resolutions than passive microwave sensors, as discussed in Section 2. Especially, the use of geostationary satellite sensors, such as GOCI, can alleviate the cloud contamination problem, which is often severe when optical sensors are employed (Kim et al. 2020; Liu et al. 2017; Nakada et al. 2018; Sun et al. 2019), because of their higher temporal resolution. Further, GOCI-II with higher spatiotemporal resolutions is expected to monitor SSS more precisely in coastal regions in the future. In addition, OC sensors can mitigate the retrieval errors of passive microwave sensors caused by LSC and RFI. The blending of optical and passive microwave radiometer data may enhance the capabilities of the current SSS monitoring system. Both types of sensor data can be blended by linking OC signals to passive microwave signals. Jin et al. (2021) showed that the blended approach successfully enhanced the spatial details of the retrieved SSS in estuarine areas, the East China Sea, and the Mississippi River Estuary. The study combined SMAP (40 km) emissivity and VIIRS-derived CDOM (i.e. a_cdom (443)) and SST (4 km), and then derived a linear equation model to generate high-resolution SSS with a 4 km resolution.

Combining emissivity, a_cdom, and SST data can reduce the impact of SST and surface roughness on SSS derivation (Jin et al., 2021). However, Jin et al. (2021) reported some limitations, including cloud contamination and instability of the model to derive the a_cdom (443) variable in the presence of phytoplankton. However, these limitations can be mitigated in several ways. First, more accurate spatial and/or temporal OC sensors, such as Sentinel-3 and GOCI-II (Jin et al., 2021) can enhance the spatial resolution of the final SSS by reducing cloud contamination and transferring SSS models to other river-dominated regions (Jang et al. 2021). Second, additional OC-derived variables that might have non-linear relationships with SSS can be incorporated with machine-learning approaches rather than by conducting a simple linear regression. In addition to Rrs data, other OC-derived inputs, such as the difference or ratio of Rrs signals, chlorophyll-a concentrations, and turbidity, can be considered. Machine learning approaches can model non-linear relationships between OC signals and SSS without any prior knowledge (Haykin and Network 2004; Wang et al. 2009; Rajabi-Kiasari and Hasanlou 2020; Jang et al. 2021). However, the uncertainties in OC signals caused by atmospheric correction and retrieval errors of OC-derived products should be carefully examined.

In short, it is essential to employ accurate dielectric models with high-quality input data and improve empirical models by incorporating additional data that account for the limitations of region-specific samples exhibiting high or low water salinity. Both microwave sensor-based dielectric constant models and OC Rrs-based models are empirical (Chen and Hu 2017). The models need to be continuously improved by amassing as much data, densely distributed in time and space, as possible, through the cooperation of a number of countries and institutions. In addition, the blended approach discussed above can provide a new capability to monitor SSS with a higher spatial resolution in both open oceans and coastal seas.

4.2. Improvement of remote sensing SSS using spatio-temporal interpolation

As introduced in Section 4.1, L-band radiometers have coarse spatiotemporal resolution. Interpolations in both spatial and temporal domains have been carried out to fill the gaps in satellite-derived SSS (Wang, Sun, and Zhang 2019). However, several limitations have been pointed out with respect to using spatiotemporal interpolation methods. For example, the SMAP L3 SSS generated by the Gaussian weighted averaging method with an 8-day running temporal window is not provided near-real-time, resulting in a 4-day data delay. Furthermore, previous gap-filling interpolation methods are based on the region-specific temporal cycle of the SSS, which might not be applicable for different areas (Delcroix et al. 2005; Bingham, Foltz, and McPhaden 2010; Melnichenko et al. 2016). Thus, an advanced interpolation technique with spatiotemporal transferability is required to fill the gaps in SSS retrieval without future observations (Buongiorno Nardelli, Droghei, and Santoleri 2016).

Because SSS has a significant seasonal pattern, regional time-series analysis, which is a type of temporal context analysis, is useful for generating daily global SSS maps. For example, time-series analysis methods, such as DINEOF, recurrent neural networks (RNN), and long short-term memory (LSTM), will be helpful for gridded SSS data interpolation. The time-series analysis methods mentioned above have shown successful interpolation results for SST retrieval with a clear seasonal pattern (Huynh et al., 2016; Xiao et al. 2019). In addition, the most significant benefit of time-series analysis is the ability to extend the development of a forecast model based on a large number of general time-series patterns (Yang, Wang, and Wang 2017). A relatively short collection of SSS data (spanning approximately 10 years) compared to SST data (over 50 years) is a major weakness of time-series analysis.

The spatial context of SSS is helpful for filling data over unobserved areas because of the limited satellite coverage. Various deep learning approaches for interpolation have been successfully applied in oceanographic fields. For example, several types of deep-learning approaches based on spatial patterns of SST exhibited higher accuracy in the reconstruction of missing data, including the convolutional autoencoder (CAE) (Barth et al. 2020; Liu et al. 2020; Barth et al. 2022), convolutional neural networks (CNN) (Barth et al. 2020; Xu et al. 2020; Izumi et al. 2022), and generative adversarial networks (GAN) (Zhang, Stanev, and Grayek 2020; Izumi et al. 2022). In addition, taking into account the spatial context and temporal trends simultaneously is helpful in SSS retrieval in specific regions that have dynamic changes, such as frequent heavy rainfall, large amounts of river discharge, and mixing events in oceans (Du, Zhang, and Shi 2019; Tzortzi et al. 2016; Vinogradova and Ponte 2013).

To fill the gaps in SSS data, their spatiotemporal patterns should be concurrently considered, especially for regions exhibiting higher variability to effectively model low or high SSS. Such regions need to be monitored on a much finer scale than the mesoscale (spatial scale of 50–500 km and temporal scale of 10–100 days) (Boutin et al. 2021). Thus, region-wise models may be better than a single global model in addressing these gaps. For example, Convolutional-LSTM (ConvLSTM) might be appropriate for filling the gaps in coastal areas within the basin scale (~10 km). ConvLSTM can effectively handle the time series of spatial (image) information. ConvLSTM and its variants have demonstrated successful forecasting skills for oceanic variables including SSS, SST, and surface waves (Zhang, Geng, and Yan 2020; Zhou et al. 2021; Song et al. 2022; Zhao, Zhou, and Du 2022). Alternatively, end-to-end models, such as the Four-Dimensional Variational Network (4DVarNet) and Data Interpolating Convolutional Auto-Encoder (DINCAE), may be appropriate for SSS reconstruction. These recently introduced approaches have exhibited excellent performance for gap-filling of several oceanographic parameters, such as SST (Barth et al. 2020, 2022), suspended sediments (Vient et al. 2022), sea surface currents (Fablet et al., 2022), and sea surface height (Febvre et al. 2022). However, these deep learning approaches are yet to be fully explored for gap filling; thus, further research is needed.

4.3. Enhancement of satellite-derived SSS estimation using in situ data

In situ measurements have been used to improve SSS products based on periodic validation processes in the operation of satellite missions. The Aquarius, SMOS, and SMAP satellites have been updating their retrieval algorithms using the ARGO float SSS measurements. Unfortunately, the ARGO floats do not cover the extent of the Earth’s oceans. Thus, we need to collect more in situ datasets to further improve SSS retrieval products. A simple solution for expanding observation coverage is to integrate in situ SSS measurements collected from a variety of instruments while minimizing measurement errors caused by instrument differences (Delcroix et al. 2005). Numerous in situ SSS observation platforms, including ARGO floats, exist including ship-based observations, such as thermo-salinograph (TSG) sensors, drifters, and mooring (Alory et al. 2015). We can also utilize observation data from marine mammals carrying Conductivity-Temperature-Depth (CTD) sensors (Treasure et al. 2017). Data from marine mammals can be used particularly in regions located at high latitudes, where in situ measurements are scarce. Nearly 400,000 SSS profiles from 1,234 marine mammal tags were available in 2017 (Treasure et al. 2017).

Lastly, the other sources of SSS observations, including national, institutional, and even personal measurement data, are important for region-specific algorithm development. However, in the integration of multi-source data, a robust method to account for the uncertainties and inter-errors of each dataset (Vinogradova et al. 2019) is required. After the integration of various in situ datasets, future studies must calibrate and validate SSS estimation models with integrated in situ SSS datasets.

The most difficult step is the integration of various in situ SSS sources. Commonly used ARGO, TSG, and mooring data have been frequently integrated because of the reliable quality control processes of each monitoring system (Wang, Sun, and Zhang 2019, 2021). However, animal movement data are prone to errors (Jonsen, Flemming, and Myers 2005; Albertsen et al. 2015; Auger-Méthé et al. 2017; Jonsen et al. 2020). To integrate in situ salinity data dynamically collected via marine mammals, additional data processing models may be necessary such as the continuous-time correlated random walk model (Jonsen et al., 2020), fast fitting of non-Gaussian state-space models (Albertsen et al. 2015), and continuous-time state-space model (Jonsen et al. 2020).

To enhance the current empirical salinity retrieval algorithms, future research should focus on waters with very high or low salinity (i.e. extreme cases). An ensemble model might benefit from incorporating regional SSS characteristics with a focus on extreme salinity. High-salinity areas are located in the subtropical north and south Atlantic and tropical southeast Pacific, while low-salinity waters are found in the Southern Pacific Convergence Zone, Eastern Pacific Fresh Pool, and North Pacific.

4.4. 3-D salinity reconstruction

The vertical profile of sea salinity provides crucial information for the understanding of global ocean circulation and climate change (Bourgain and Gascard 2011; Laurie et al. 2012) and changes in the Arctic (Polyakov, Pnyushkov, and Carmack 2018). The halocline structure can vary by region and season (Katsura et al., 2020). The salinity profile has a close relationship to the biosphere not only in the remote ocean but also in coastal area ecosystems, where humankind is directly associated (Smyth and Elliott 2016; Whittle et al., 2022). The coastal environment is likely to be impacted by the tidal effect, coastal warming, freshwater income, and typhoon events, which increase its spatial and temporal variability (Weller et al. 2019; Park et al. 2021; Stammer et al. 2021). Moreover, Madsen, Høyer, and Tscherning (2007) reported the low accuracy of calculating mean dynamic topography over coastal areas from satellite altimetry from international achieves sustainable business development support, which can result in high uncertainty when calculating vertical sea salinity structure. Despite these limitations, it is important to reconstruct 3-D salinity information, especially for coastal regions, to investigate the effect of anthropogenic and natural events on marine ecosystems.

To construct the vertical temperature and salinity (T-S) profiles, gridded Argo datasets with a coarse resolution of 1° × 1° were generated (Hosoda, Ohira, and Nakamura 2008; Roemmich and Gilson 2009; Zhang et al. 2022). However, the mesoscale grid could not provide sufficient spatial information to understand the interaction between anthropogenic/natural events and salinity profiles near the coast. Using surface information, such as SSS and SST, several methods have been proposed to extrapolate the vertical profile of salinity with higher resolution. Tang et al. (2022) reconstructed the 3-D temperature structure using the historical T-S profiles, SST, and sea surface height (SSH) data. Based on the correlation between SST, SSH, and vertical temperature, they reconstructed the vertical salinity profile with a 0.25° spatial resolution. An extrapolation approach using a multivariate empirical orthogonal function was suggested by Buongiorno Nardelli and Santoleri (2004) using the CTD dataset. The authors extended their study for vertical salinity reconstruction with 0.1° spatial resolution using surface satellite observations and in situ vertical profiles based on LSTM (Buongiorno Nardelli 2020). They emphasized the impact of climatological patterns of salinity in vertical reconstruction. These studies demonstrated the feasibility of reconstructing the vertical salinity profile from satellite-derived surface observations with a higher spatial resolution.

Owing to the shallower depths, the vertical salinity structure has a closer relationship with SSS in coastal regions than in the open ocean. Due to active mixing by surface wind (~50 m of Ekman depth with 10 m/s wind speed in latitude ~30°N), there is a greater likelihood for SSS to represent the majority of the vertical salinity profile in coastal regions than in the open ocean (Meira et al. 2017). Reconstructed T-S profiles can suffer from high errors in the upper mixed layer and coastal regions due to wind-driven mixing, precipitation events, and high uncertainty from coarse input data (Jeong et al. 2019; Buongiorno Nardelli 2020). For example, Kubota, Tada, and Kimoto (2015) showed that the massive discharge of freshwater from the Yangtze River into the East China Sea affected the SSS in the adjacent marine environments of Korea and Japan, which contain a mixture of Kuroshio Waters and Taiwan Strait Water. Abrupt changes in salinity profiles caused by an external source can affect coastal ecosystems (Velasco et al., 2019; Pereira et al., 2019; Whittle et al., 2022). This implies that SSS measurements and their application to reconstructing the vertical salinity structure are crucial in coastal regions, but little research has been conducted. Although satellite-retrieved SSS can significantly contribute to obtaining the vertical salinity structure over coastal areas, the high uncertainty due to land-corrupted signals in passive micrometers with a coarse spatial resolution (e.g. SMOS) should be carefully examined to avoid the accumulation of errors during the reconstruction process. In addition, recent advances in deep learning techniques and their applications for reconstruction (He et al. 2022; Zhou et al. 2022) can further enhance 3-D salinity reconstruction.

4.5. Synergistic use of multi-frequency remote sensing

L-band passive microwave radiometer-retrieved SSS is now routinely integrated into the global ocean and coupled ocean-atmosphere models (Hackert, Busalacchi, and Ballabrera‐poy 2014) and has been demonstrated to significantly improve forecasting of El Niño Southern Oscillation (ENSO) (Zhu et al. 2014), terrestrial precipitation and drought (Li et al. 2016), and intra-seasonal variations, such as Madden-Julian Oscillations (Subrahmanyam, Trott, and Murty 2018), which have direct societal relevance. While L-band radiometry is relatively sensitive to SSS variations in tropical and subtropical oceans, it has poor sensitivity in high-latitude oceans (for SST <5 C°; Jang et al. 2022). A multi-frequency approach incorporating low-frequency radiometers (1.4–6.9 GHz) is expected to significantly improve the sensitivity to SSS in high-latitude oceans and thus the retrieval accuracy of SSS (Boukabara, Garrett, and Kumar 2016). Future satellite missions with L-band, C-band (6.6 GHz), X-band (10.7 GHz), and OC sensors (Reul et al., 2020) should be reviewed to promote integrated methods on satellite-derived SSS.

Furthermore, L-band passive radiometers have shown great potential for measuring extreme wind events owing to their relatively longer wavelength compared to traditional scatterometers with C- and Ku-band (18.7 GHz) radars (Yueh et al. 2016). The Ku-band is significantly affected by rain and both the C- and Ku-bands exhibit significant saturation in their backscatters (Stiles and Yueh 2002). The synergetic use of multi-frequency (C-band, Ku-band, or higher) passive microwave data will reduce the uncertainty of ocean surface wind vectors and wind-speed data. By combining X- with C- and L-band channels, it is possible to propose a method to jointly retrieve sea surface wind speed and sea surface temperature, particularly in high-wind regions such as mid-high-latitude oceans. In such an approach, the existing retrieval algorithm can be modified by replacing the T_b from a single channel (L-band) with that from multiple channels (L-band from SMOS or SMAP together with X- and C-bands from AMSR-2). However, it is well known that observations made by sensors contain temporal gaps. Therefore, this type of synergy is optimal for regions with low variability, such as open oceans.

Future satellite missions with higher spatial and spectral resolutions are expected to enable detailed observations of tropical instability waves, intertropical convergence zone dynamics, and climate variability (Vinogradova et al. 2019). For example, the Copernicus Imaging Microwave Radiometer (CIMR) mission was designed to observe the ocean, sea ice, and salinity particularly in the Arctic environment. This satellite mission is currently in the preparatory phase and is expected to be launched by 2025 (Kilic et al. 2018). The CIMR will supplement the SSS records provided by SMOS, Aquarius, and SMAP. Moreover, the CIMR can retrieve SSS with a higher accuracy than SMAP, even at high latitudes (Kilic et al. 2018). As previously mentioned, the history of remote sensing of SSS is relatively short (less than a decade). To encourage and conduct synergetic studies for satellite-derived SSS, there is a need to explore potential new satellite missions, including L-band, C/X-band, and OC sensors.

5. Conclusions

Against the short history of satellite salinity studies, which spans less than a decade, scientific demand for SSS retrieval has increased in recent years. However, the retrieval of SSS via satellite remote sensing still faces significant challenges. This study reviewed satellite-based SSS retrievals and suggested guidelines for future research. Several methods exist to mitigate satellite SSS retrieval errors. First, improving internal empirical models in regions with cold waters can increase SSS retrieval accuracy, and the multi-frequency approach can mitigate the limitations of data availability in coastal regions. Second, the spatiotemporal availability of operational satellite SSS products can be enhanced using state-of-the-art interpolation methods that incorporate spatiotemporal patterns. For example, the current SMAP L3 SSS has apparent limitations in terms of near-real-time analysis. Recent deep-learning approaches that consider spatial and temporal patterns will be helpful for interpolating gridded SSS data. Third, we need to mitigate the bias and variance of SSS products by synergistically using global in situ SSS datasets. Current L-band radiometers update their retrieval algorithms by incorporating only ARGO float SSS data. The combination of multi-source, in situ data (e.g. ARGO floats, TSG, drifters, mooring SSS, and CTD measurements from marine mammals) can expand the observation coverage for improving satellite-derived SSS products. Fourth, the vertical profiles of ocean salinity can be reconstructed at a fine resolution using satellite-retrieved SSS and deep learning techniques in coastal regions. Although these profiles may contain high uncertainties caused by the coarse spatial resolution of passive microwave radiometers and dynamic environments, they can contribute to understanding the effect of anthropogenic and natural events on marine ecosystems. Finally, the synergistic use of multifrequency remote sensing for SSS retrievals should be carefully examined. For example, combining X- with C- and L-band channels may jointly retrieve sea surface wind speed and sea surface temperature, especially in high-wind regions such as mid-high-latitude oceans. The future CIMR mission will expand the currently available SSS records provided by SMOS, Aquarius, and SMAP, and have significant applications in fields such as physical oceanography, marine biogeochemistry, environmental monitoring, ocean and climate forecasts, and water cycle studies. Future satellite missions with L-band, C/X-band, and OC sensors should be thoroughly reviewed to promote and implement synergistic satellite-derived SSS.

Acknowledgements

This research was supported by “Technology development for Practical Applications of Multi-Satellite data to maritime issues (20180456)”, “Development of Advanced Science and Technology for Marine Environmental Impact Assessment (20210427)”, and “Development of technology using analysis of ocean satellite images (20210046)” of Korea Institute of Marine Science & Technology promotion (KIMST) funded by the Ministry of Oceans and Fisheries, South Korea.

Disclosure statement

No potential conflict of interest was reported by the authors.

Data availability statement

The raw data used in this study are publicly available from the sources provided in following links. The processed data and codes generated from this study are available upon request to the corresponding author.

- ARGO in situ data: https://argo.ucsd.edu/data/

- Aquarius L2 data: https://podaac.jpl.nasa.gov/dataset/AQUARIUS_L2_SSS_CAP_V5

- SMOS L2 data: https://smos-diss.eo.esa.int/oads/access/collection

- SMAP L2 data: https://podaac.jpl.nasa.gov/dataset/SMAP_JPL_L2B_SSS_CAP_V5

- SMAP L3 data: https://podaac.jpl.nasa.gov/dataset/SMAP_JPL_L3_SSS_CAP_8DAY-RUNNINGMEAN_V5