Validation of Sentinel-3 SAR Level-2 and Level-3 Products in the Baltic Sea and Estonian lakes

Multimission satellite altimetry (e.g. ERS, Envisat, TOPEX/Poseidon, Jason) data have enabled a synoptic view of ocean variations in the past decades, including sea-level rise and mesoscale circulations. Since 2016, the Sentinel-3 mission has provided better spatial and temporal sampling compared with its predecessors. The Sentinel-3 Ku/C Radar Altimeter (SRAL) is one of the synthetic aperture radar altimeters (SAR Altimeter) which is more precise in coastal and lake observations. In this study, we validate Sentinel-3 Level-2 products in Baltic Sea coastal areas and two lakes in Estonia. Moreover, the Copernicus Marine Environment Monitoring Service (CMEMS) Level-3 sea-level anomaly data and the Nucleus for European Modelling of the Ocean (NEMO) reanalysis model outcomes are compared with measurements from a tide gauge network. A dense in situ water level network deployed along the coast for geodetic observation was utilised to provide ground truths for validating altimetry results. Three validation methods were used for Level-2 data: (i) collocated Sentinel-3 and GNSS ship measurements; (ii) a national geoid model (EST-GEOID2017) with sea-level anomaly correction; (iii) collocated Sentinel-3 and buoy measurements. The validations were carried out in seven Sentinel-3A/B overpasses in 2019. Our results show that the uncertainty of the Sentinel-3 Level-2 altimetry product is below decimetre level on the Estonian coast and the targeted lakes. Results from CMEMS Level-3 showed a correlation of 0.8 (RMSE 0.19 m) and 0.91 (RMSE 0.27 m) when compared against tide gauge measurements and NEMO model, respectively.


Introduction
Tide gauges (TG) have been widely deployed to register sea-level variations on the coast, and their time series are considered longer and more reliable than other remote sensing measurements. However, the records might be contaminated by several effects, such as glacial isostatic adjustment (GIA), neotectonic movements and local land subsidence. Tide gauge networks are usually sparsely distributed along the shoreline and their measurements are related to the land with surface deformation they are connected to. To avoid the above-mentioned influences and to transform sea surface heights (SSHs) into a common height system, it is recommended to use GNSS tide gauges or satellite altimetry to calculate the absolute SSH above a reference ellipsoid [1,2]. A radar altimeter is a precise ranging tool designed for measuring SSH over the open ocean at an uncertainty level of around 3.5-5 cm [1,3,4]. Satellite altimetry has been used in sea-level variation studies for more than 30 years; it was evolved from the experimental Seasat mission in the late 1970s and gained popularity in the 1990s when TOPEX/Poseidon and ERS-1 missions were launched. Nowadays, it is a very important tool for understanding the topography of mesoscale eddies and the multidecadal trend of eustatic (global) sea-level rise.

Study sites
The Baltic Sea is a semi-enclosed water body dotted with thousands of islands and islets. Our study area is in the eastern part of the Baltic Sea ( Figure 2), including the Gulf of Finland and the Gulf of Riga. The topography of the southern coast of the Gulf of Finland is varying. There is a 55-metrehigh cliff in the east of Kunda (cf. tide gauge number 2 in Figure 2) and a flat seashore along the entire western coast. Steep terrain near the coast can cause peaky waveforms and hinder the performance of retracking algorithms. The Gulf of Riga has a very shallow and sandy seashore. There are many islands in the northern part of the Gulf which also make the altimetry measurements challenging along the fragmental pass-overs.
Lake Peipus (Lake Peipsi) and Lake Võrtsjärv are the two largest lakes in Estonia (Figure 2), with a surface area of 3555 km 2 and 270 km 2 , respectively. Lake Peipus is the fifth largest lake in Europe, with an average depth of 7.1 m and a max depth of 15.3 m. Võrtsjärv is very shallow with an average depth of 2.7 m and a maximum depth of only 6 m.
Four GNSS transects were selected for Sentinel-3 validation, as shown in Figure 2. The following criteria were taken into consideration in order to choose the transects to conduct GNSS vessel measurements: (i) at least one transect has to run over the selected waterbodies (Gulf of Finland, Gulf of Riga, Lake Peipus and Võrtsjärv); (ii) both Sentinel-3A (S3A) and Sentinel-3B (S3B) transects would be used; (iii) the GNSS measurements on the boat and the buoy measurements could be carried out at the transect. S3B pass 321 crosses over Lake Peipus (leaving the 35 km GNSS transect to the west). One S3B track (pass 528) runs over Lake Võrtsjärv and the pass is close (<3 km) to the reedy coast.
In addition, S3A pass 186 (Gulf of Riga) and S3B pass 625 (Gulf of Finland) in the Baltic Sea were selected for our GNSS vessel campaigns. Three additional transects were added to the selection in order to increase the reliability of the validation results: S3A pass 397 and S3B pass 397 in the Gulf of Riga and S3A pass 186 in the Gulf of Finland ( Figure 2). This selection was based on the following criteria: (i) the same transect (S3A pass 186) has to run over two water bodies (Gulf of Riga, Gulf of Finland); (ii) both S3A and S3B have to pass over the same waterbody (S3A_186 and S3B_397 over the Gulf of Riga); (iii) the crossing point of ascending and descending tracks should emerge. This choice helps verify the validation results over various areas and missions in the same water body.

Geoid model EST-GEOID2017
Since sea-level variations of tidal origin are less than 10 cm in the study area of the Baltic Sea [15], the precise geoid model is one of the best reference surfaces for validating altimetric heights. The Estonian geoid model (EST-GEOID2017) is a quasi-geoid model covering 57°N-60°N and 20°E-30°E, with a spatial resolution of 1' x 2'. Its long-wavelength information refers to the GOCO05s global gravity model [16]. Additionally, nearly 50,000 gravity points were used with an average formal uncertainty of 0.75 mGal, whereas within Estonian mainland and islands, the uncertainty of gravity data is mostly within 0.5 mGal (see [17] for more details). The EST-GEOID2017 model is very smooth in the mainland ( Figure 2) and its GNSS-levelling-based accuracy remains within ±0.005 m [17]. This means that the accuracy of the geoid model in the coastal zones of the lakes should also be around ±0.005 m. Note that the estimated discrepancy of ±0.028 m between the model and the GNSS measurements is estimated in the Gulf of Finland [18]. Having such superb accuracy, the use of the model for satellite altimetry validation in the coastal zones is justified.
Since the validation result depends not only on the accuracy of the altimetry data but also on other factors, the accuracy of the geoid, tide gauge data and GNSS-data should also be taken into consideration. Note that the pressure sensor-based tide gauges are affected by time-dependent drift and need to consider employing control readings from the staff gauge. Therefore, the uncertainty of automatic tide gauges remains in 1-2 cm level [19]. The accuracy of the GNSS height component depends on several factors, such as the antenna and receiver, the method of measurement and processing. However, the GNSS height component accuracy on the vessel, based on the kinematic method, remains within a ±2 cm level as well [18].

Sentinel-3 data products
Sentinel-3 satellites carry a dual-frequency SAR altimetry payload. The main frequency used for range measurements is Ku-band (13.575 GHz), while the C-band (5.41 GHz) is used for ionospheric correction. The SRAL altimeter has high resolution in the along-track (300 m) and across-track (1.64 km) directions, which gives better results in the coastal areas [20] compared with the conventional pulselimited radar altimeter. Sentinel-3 is operating in a sun-synchronous orbit with a repeat cycle of 27 days. Combining S3A and S3B, the repeat cycle between two satellites on a regional scale becomes 13.5 days and the interval between interleaved groundtracks is 27 km near the Equator, which offers a better temporal and spatial resolution and also increases the chance of passing over small waterbodies.
For this study, Sentinel-3 SRAL Non Time Critical Level-2 data (SR_2_LAN) from the Estonian National Copernicus Hub (ESTHub) mirror site (https://ehdatahub.maaamet.ee/dhus/#/home) in the period from Dec 2018 to Nov 2019 were used. Note that the radar range (R) measured between the satellite and the sea surface observed near nadir is affected by several error sources when the signal passes the atmosphere. Additionally, several tidal effects are involved in measuring the sea level. Therefore, the following atmospheric and geophysical corrections were used: where the meanings of corrections are ionospheric correction Δhiono, dry and wet tropospheric corrections Δhdry and Δhwet, sea state bias correction Δhssb, sum of tide heights ∑ht (solid earth, geocentric ocean and geocentric pole tide height), inverted barometer height correction ha and high frequency fluctuations of the sea surface topography correction hf, respectively. The corrected range between the satellite and the sea level was calculated: where R is the directly measured distance or pseudorange between the satellite and the sea level ( Figure  3). Sea surface height above the ellipsoid (SSHalt) based on the off-centre-of-gravity (OCOG) retracker, which is empirically fit various types of waveforms [21; 22], was calculated: where hsat is satellite height above ellipsoid (WGS84), which can be calculated using satellite ephemeris. SSHalt values were calculated for all transects ( Figure 2) in the period from 1 January 2019 to 1 November 2019, equivalent to 10-11 cycles for each pass. The SAR mode of Sentinel-3 effectively reduces land-induced contamination in the radar signal by increasing the spatial resolution along track. However, the problem still exists when the satellite azimuth direction and the coastline are not perpendicular. Therefore, Sentinel-3 measurements within 2 km of the shoreline were removed in order to reduce potential noisy data, as demonstrated by the blue points appearing between the grey and red lines in Figure 4. Normally, the differences (ΔSSH) between the sea-level anomaly corrected geoid model (SSHgeoid) and satellite altimetry (SSHalt) remained within 0.5-1 m. However, some ΔSSH outliers could be up to 2 m or more on open water due to reflections from small islets or vessels, which is known as the 'hooking effect' in off-nadir measurements. The grey and red lines indicate the coast and the distance (2 km) from the coast, respectively.

In Situ Data
The Estonian coastline has a dense tide gauge network, from which 21 stations were selected for the study (Figure 2). Among them, 13 stations are managed by the Estonian Environmental Agency (EEA) and 8 stations are operated by the Department of Marine Systems (MSI) of Tallinn University of Technology. All of these tide gauges are equipped with pressure sensors and have operated for more than ten years already [19]. In 2017-2018, the automatic tide gauges as well as tide gauge rods were reconnected to the national levelling network. Today, the automatic tide gauges record the hourly sealevel height in the Estonian height system EH2000, which is a national realisation of the European Vertical Reference System EVRS [23].
The tide gauge data were used to estimate sea-level anomaly along each altimetry track. Sea-level anomaly from tide gauges (SLATG) was calculated from the four tide gauge records nearest to the satellite track at the time of the satellite overpass. The average of SLATG was used to correct the geoid height (N) from the EST-GEOID2017 model above the ellipsoid GRS80. The geoid-based sea surface height (SSHgeoid) at each high-frequency Sentinel-3 footprint can thus be presented as: For Sentinel-3 validation, the difference between altimetry-based sea surface height (SSHalt) and geoid-based sea surface height (SSHgeoid) was calculated using the following equation: Residual difference ΔSSHgeoid was further analysed. The ΔSSHgeoid residuals larger than 2.5 standard deviation were considered outliers. Residuals were plotted against distance from the coast ( Figure 4). Correlation plots between SSHalt and SSHgeoid were compiled as well.
In addition, GNSS kinematic measurements on the vessels were used to validate SSHalt over the study sites. Multiple-frequency Trimble R8s and Trimble R4-3 receivers were mounted on top of the vessel (Figure 5b). Antenna reference points were connected to the sea level using a total station ( Figure  5a). Two antennas were used to eliminate the risks when one receiver stops working and also to filter out vessel fluctuations [24,25]. Measurements were carried out on the same day of the satellite overpass and the weather was calm during all measurements (Table 1). GNSS measurements were recorded at 2-second intervals along the altimetry tracks (see Figure 2). The first data quality control was performed on board the vessel with the program TEQC [26]. The GNSS base stations data from the Estonian Permanent GNSS Network ESTPOS [27] and the Trimble Business Centre software were used for kinematic data processing in double difference mode. The International Geodetic Service (IGS) precise ephemeris and absolute antenna calibration data were used. First, both receiver data were processed separately and sea-level heights above the ellipsoid from two receivers were compared. When the difference between the sea-level heights from two receivers on the same time tag was larger than ±10 cm, the measurement was excluded from further processing. Cleaned heights from two receivers on the same time tag were averaged. Next, the temporally moving-averaged sea-level height was calculated in order to smooth out the fluctuations of the vessels. The window size for the moving average filter was selected based on the velocity of the vessel and the GNSS data rate (2 s) in order to obtain window size, which is close to the altimeter along-track resolution (300 m). For example, if the vessel's velocity was 10 m/s and the GNSS sample rate was 2 s, moving average window 15 was chosen, since 15 x 2 x 10 = 300 m. Spatially moving-averaged GNSS sea surface heights (SSHGNSS) were compared with SSHalt for altimetry validation:

Copernicus Marine Environment Monitoring Service (CMEMS) data products
One of the purposes of this study was to compare the accuracy of CMEMS Level-3 (L3) along-track altimetry products (SLAL3) [28] and NEMO reanalysis model (SLANEMO) [29] outcomes against the sealevel anomaly data from coastal tide gauge stations (SLATG). The CMEMS Baltic Sea Physical Reanalysis product provides a physical reanalysis for the entire Baltic Sea area. It is produced using the ice-ocean model NEMO-Nordic [29]. The reported mean correlation between SLATG and SLANEMO for the entire Baltic Sea is 0.95 with an RMSE of 7 cm, with lower values in highly dynamic marine areas [30].
The analysed SLA data covered the period from 2014 to 2017 when the sea-level measurements of all methods were available. The SLAL3 altimetry data was collected only from the altimeters that were active in 2019 (Table 2). An overview of tide gauges is given in Altimetry SLAL3 measurements were collected at a maximum distance of 10 km offshore from the listed gauges, and the maximum time difference between SLATG and SLAL3 measurements was one hour. For the NEMO, SLANEMO at the closest grid point to the station within an hour's time difference was extracted. SLA data from all seasons (including winters with ice cover) was used in the comparison. Finally, a selection was made from all three sources to match the measurements by time. The number of common, collocated measurements was 625.
A separate comparison was carried out with the data from the buoy in the Gulf of Riga (cf. green circle on Figure 2) to measure the accuracy of SLAL3 along-track data as well as the reanalysis model outcomes. The data was collected using the same scheme as that of the other stations described above, and the final analysis consisted of 33 unique data points in all three sources.
Data analysis results were presented as standard descriptive statistics: Pearson correlation coefficient (r), Root Mean Square Error (RMSE) and Mean Error (ME). The statistics were calculated between SLA values since the accurate translation of modelled water levels into a common reference surface is an as yet unresolved technical issue [31]. Table 3. Altimetry SLAL3 measurements were collected at a maximum distance of 10 km offshore from the listed gauges, and the maximum time difference between SLATG and SLAL3 measurements was one hour. For the NEMO, SLANEMO at the closest grid point to the station within an hour's time difference was extracted. SLA data from all seasons (including winters with ice cover) was used in the comparison. Finally, a selection was made from all three sources to match the measurements by time. The number of common, collocated measurements was 625.
A separate comparison was carried out with the data from the buoy in the Gulf of Riga (cf. green circle on Figure 2) to measure the accuracy of SLAL3 along-track data as well as the reanalysis model outcomes. The data was collected using the same scheme as that of the other stations described above, and the final analysis consisted of 33 unique data points in all three sources.
Data analysis results were presented as standard descriptive statistics: Pearson correlation coefficient (r), Root Mean Square Error (RMSE) and Mean Error (ME). The statistics were calculated between SLA values since the accurate translation of modelled water levels into a common reference surface is an as yet unresolved technical issue [31].

Sentinel-3 Level-2 validation with GNSS campaigns
The statistics of the residuals from the Eq. (6) along altimetry tracks are presented in Table 4. According to the results in multiple locations, the two datasets in the Baltic Sea correspond better than those of inland cases; the mean (MEAN) and standard deviation (STD) were 0.11 ±0.08 m and 0.14 ±0.05 m in the Gulf of Finland and the Gulf of Riga, respectively. These results also match well with the validation by the geoid model for S3B pass 625 in the Gulf of Finland on 25 April 2019 (0.11 ±0.06 m) (Figure 7) and for S3A pass 186 in the Gulf of Riga on 20 June 2019 (0.09 ±0.05 m) ( Figure  8). In Lake Peipus and Lake Võrtsjärv, the MEAN and STD were 0.16 ±0.13 m and 0.31 ±0.39 m, respectively. Again, these GNSS validation results agree with the validation results using the geoid model, showing a discrepancy of 0.15 ±0.14 m and 0.28 ±0.30 m for the Lake Peipus on 19 June 2019 and for Lake Võrtsjärv on 13 July 2019, respectively (Figure 9). Compared with marine conditions, the worse performance on the lakes could be attributed to several factors, such as vegetation coverage near shore and the hooking effect from the surrounding land [32]. It is also arguably due to the fact that the geophysical corrections (Eq. 1) for continental waters are not as accurate as for the open sea [33].
The other error sources for the determination of ΔSSHGNSS are: (i) errors in the determination of the GNSS antenna height from the water level; the uncertainty of the sea surface height from the GNSS antenna is at least ±0.020 m and it is possible that the GNSS antenna connection to the water level could also cause a systematic shift; and (ii) the accuracy of the GNSS solution itself. Multiplefrequency Trimble R8s and Trimble R4-3 receivers were used for the GNSS kinematic measurements on the boat but the vertical uncertainty could be ±0.015 m + 1 ppm RMS according to the manufacturer's specification. It is also noted that the mean ΔSSHGNSS is biased in all waterbodies, especially in Lake Võrtsjärv.

Sentinel-3 validation with the SLA-corrected EST-GEOID2017
We studied the waterbodies separately for validation with the geoid model. Two transects were chosen for the Gulf of Finland ( Figure 2) and the residual of differences for S3A and S3B (ΔSSHgeoid, Eq. 5) are presented in Figure 7. The mean ΔSSHgeoid was calculated for each 25-km-long transect when the satellite passed over the gulf. The time difference between SA3 and S3B overpasses is two days. It is observed that the validation results and trends for both passes (S3A_186 and S3B_625) are very similar in the Gulf of Finland, which implies a similar accuracy level for both missions. The average ΔSSHgeoid for S3A and S3B was 0.02 ±0.08 m. Note that there is no ΔSSH value for S3B for May 2019 (Figure 7, cycle 25) because the ΔSSHgeoid at 0.31 ±0.55 m in this cycle is clearly an outlier. Such a big residual difference could have been caused by a local wind-driven water pile up against the coast. After removing ΔSSHgeoid from cycle 25, an average residual difference of 0.01 ±0.06 m was obtained for the Gulf of Finland (Table 5).    (Figure 2). The results of S3A and S3B fit quite well in the Gulf of Riga (Figure 8). S3A 186 showed a better fit due to the shorter transect (54 km). Larger deviations were observed in winter from December to April when stormy weather conditions prevailed. Wind set-up near the coast is proportional to tangential wind speed component squared and can reach up to ca 1 m in certain Estonian coastal locations during strong storms [34]. The average residual differences between the sea surface height of Sentinel-3 (SSHalt) and geoid-based sea surface height (SSHgeoid) was 0.01 ±0.07  Table 5). The validation results for the lakes show much larger discrepancies compared with the results in the gulfs (Figure 9). The main reasons for this are the size of the water body and the distance of the track from the coastline. In winter, the results are also affected by the sea/lake ice. In February and March, both lakes were covered by ice and snow up to 0.4 m thick and hummock ice occurred in the coastal areas as well. Thus, the average residual differences between the sea surface height of Sentinel-3 (SSHalt) and geoid-based sea surface height (SSHgeoid) for Sentinel-3B in Lake Peipus and Lake Võrtsjärv were 0.14 ±0.16 m and 0.13 ±0.27 m, respectively (Table 5).

Figure 9.
Residual differences (ΔSSHgeoid) between the sea surface height of Sentinel-3 (SSHalt) and geoid-based water surface height (SSHgeoid) in Lake Peipus (blue dots) and Lake Võrtsjärv (red dots). The zero line denotes the reference SSHgeoid. Coloured numbers indicate the cycle's number.

Sentinel-3 validation with buoy data in the Gulf of Riga
The buoy placed on the bottom of the Gulf of Riga was collecting sea surface height data (SSHbuoy) from June 2019 until November 2019. During this period, satellite tracks S3A_186 and S3A_397 passed the buoy six and five times, respectively. For the validation, the mean SSHalt was calculated using the altimetry data within a radius of 5 km from the buoy. It was compared with the buoy SSHbuoy for the same time, cf. Figure 10. According to the results, the agreement between the two datasets is very good (Pearson correlation coefficients r for S3A_186 and S3A_397 were 0.88 and 0.95, respectively). The mean residual differences (ΔSSH) and STD were 0.05 ±0.06 and -0.03 ±0.10 m for S3A_186 and S3A_397, respectively. The CMEMS SLAL3 along-track altimetry data comparison with SLATG measurements yielded accurate results with the Pearson correlation coefficient r = 0.80 for all stations in sum (Altimetry SLAL3 measurements were collected at a maximum distance of 10 km offshore from the listed gauges, and the maximum time difference between SLATG and SLAL3 measurements was one hour. For the NEMO, SLANEMO at the closest grid point to the station within an hour's time difference was extracted. SLA data from all seasons (including winters with ice cover) was used in the comparison. Finally, a selection was made from all three sources to match the measurements by time. The number of common, collocated measurements was 625.
A separate comparison was carried out with the data from the buoy in the Gulf of Riga (cf. green circle on Figure 2) to measure the accuracy of SLAL3 along-track data as well as the reanalysis model outcomes. The data was collected using the same scheme as that of the other stations described above, and the final analysis consisted of 33 unique data points in all three sources.
Data analysis results were presented as standard descriptive statistics: Pearson correlation coefficient (r), Root Mean Square Error (RMSE) and Mean Error (ME). The statistics were calculated between SLA values since the accurate translation of modelled water levels into a common reference surface is an as yet unresolved technical issue [31]. Table 3, Figure 11a). The RMSE for all stations together was 0.19 m, the mean error was 0.12 m and the STD between SLATG and SLAL3 was 0.14 m. These results are similar to previous findings for the area even though the RMSE is larger [35]. The correlation between SLATG and SLANEMO is higher (r = 0.91, Figure 11b, Altimetry SLAL3 measurements were collected at a maximum distance of 10 km offshore from the listed gauges, and the maximum time difference between SLATG and SLAL3 measurements was one hour. For the NEMO, SLANEMO at the closest grid point to the station within an hour's time difference was extracted. SLA data from all seasons (including winters with ice cover) was used in the comparison. Finally, a selection was made from all three sources to match the measurements by time. The number of common, collocated measurements was 625.
A separate comparison was carried out with the data from the buoy in the Gulf of Riga (cf. green circle on Figure 2) to measure the accuracy of SLAL3 along-track data as well as the reanalysis model outcomes. The data was collected using the same scheme as that of the other stations described above, and the final analysis consisted of 33 unique data points in all three sources.
Data analysis results were presented as standard descriptive statistics: Pearson correlation coefficient (r), Root Mean Square Error (RMSE) and Mean Error (ME). The statistics were calculated between SLA values since the accurate translation of modelled water levels into a common reference surface is an as yet unresolved technical issue [31]. Table 3). However, the RMSE and mean error are much higher at 0.27 m and 0.26 m, respectively, which indicates a bias in the model fields. Standard deviation on the other hand was 0.10 m. The comparison with Narva-Jõesuu station (tide gauge number 3 in Figure 2, Altimetry SLAL3 measurements were collected at a maximum distance of 10 km offshore from the listed gauges, and the maximum time difference between SLATG and SLAL3 measurements was one hour. For the NEMO, SLANEMO at the closest grid point to the station within an hour's time difference was extracted. SLA data from all seasons (including winters with ice cover) was used in the comparison. Finally, a selection was made from all three sources to match the measurements by time. The number of common, collocated measurements was 625.
A separate comparison was carried out with the data from the buoy in the Gulf of Riga (cf. green circle on Figure 2) to measure the accuracy of SLAL3 along-track data as well as the reanalysis model outcomes. The data was collected using the same scheme as that of the other stations described above, and the final analysis consisted of 33 unique data points in all three sources.
Data analysis results were presented as standard descriptive statistics: Pearson correlation coefficient (r), Root Mean Square Error (RMSE) and Mean Error (ME). The statistics were calculated between SLA values since the accurate translation of modelled water levels into a common reference surface is an as yet unresolved technical issue [31]. Table 3) shows the worst performance regarding its low correlation and high RMSE (0.35 m) and mean error (0.30 m). A closer look at the data shows that ca 30% of the data were collected in winter, which may be one of the reasons the satellite results are underestimated. In addition, the exact measurement point in Narva-Jõesuu is at the mouth of Narva River. The location may also hinder the quality of in situ measurements, as the dynamics of the river itself are not recorded in the satellite results within a 10 km radius.   Figure 2). Red line is the linear regression between the two datasets.
The good correlation between the SLA measurements and the SLANEMO (Figure 11b, Figure 12b) proves that the reanalysis model NEMO calculates water level dynamics very well. This is in accordance with the accuracy of the CMEMS reanalysis product in the relatively dynamic coastal sea area [30]. However, the RMSE in the current study is even larger than in the model product quality analysis (7 cm) [30]. The higher RMSE value is clearly influenced by the bias (ME = 0.26 m). The large Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 26 August 2020 doi:10.20944/preprints202008.0586.v1 RMSE can be seen on the frequency distribution of water level anomalies in Figure 13. The maximum frequency distribution of model results is around zero water level, whereas it is around +0.2 m for SLATG measurements. However, the small STD (±0.10) shows the good fit of the model with the observations when bias is removed and is comparable with the RMSE found in [30]. The STD between SLATG measurements and SLAL3 (Figure 11a) is higher than that of model outcome comparison (Figure 11b). This could be explained by (i) land contamination which exists in altimetry results, (ii) altimetry processing in general, (iii) tide gauge measurement errors, or (iv) the fact that SLAL3 data were collected at a maximum distance of 10 km from the tide gauge station and even by the time difference between compared measurements. The reasoning above is confirmed by a comparison of SLAL3 with SLAbuoy measurements from the buoy in the Gulf of Riga (Figure 12). Even though there should be very little land influence if any on the altimeters that measured over the station, a distinguishable offset is still present between the remotely sensed and modelled sea level (Figure 12a and b). This indicates possible errors in altimetry processing. However, the smaller ME, RMSE and STD on Figure 13b compared with 12b indicate that the performance of the NEMO model is better on the open sea compared with the coastal Baltic Sea.
In general, validation of altimetry products with tide gauge data in the coastal area, although constantly improving, still remains a challenge [31,36]. This is because of various coastal effects as well as possible geoid connection issues. Moreover, a mismatch of characteristic (temporal and spatial) scales inevitably exists when sampling the SSH variability with different methods, such as coastal tide gauges, relatively coarse-gridded models and remote sensing products.

Conclusions
This study discussed the quality of the Sentinel-3A/B Level-2 product and the Copernicus Marine Environment Monitoring Service (CMEMS) Level-3 product. To validate the Level-2 product, the SSH from altimetry (SSHalt) in Estonian coastal waters (Gulf of Finland and Gulf of Riga) and larger lakes (Lake Peipus and Lake Võrtsjärv) was compared with the SSH obtained from GNSS (SSHGNSS) and buoy (SSHbuoy) measurements, in situ coastal tide gauges and national geoid model EST-GEOID2017 (SSHgeoid). In addition, to validate the along-track Level-3 product (SLAL3) and NEMO reanalysis outcomes (SLANEMO) from the CMEMS database, the water levels were compared with in situ measurements at 21 Estonian coastal tide gauge stations. The comparison was made with data from six satellites from 2014 to 2017.
The validation results of the Sentinel-3 Level-2 product showed that the altimetry accuracy is much higher (~3x) on the open sea than on lakes. The average difference between SSHalt and SSHgeoid (residual difference ΔSSH, Eq. 5) was 0.07-0.10 m in the Gulf of Finland and the Gulf of Riga. The corresponding numbers were 0.23 and 0.27 m in Lakes Peipsi and Võrtsjärv, respectively. Based on the average residual differences, it was noted that the altimetry SSH values in inland waters were systematically shifted: SSHalt was 0.14 m and 0.13 m higher than SSHgeoid values in Lake Peipus and Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 26 August 2020 doi:10.20944/preprints202008.0586.v1 Lake Võrtsjärv, respectively. In the coastal waters, positive and negative residual differences were more equally distributed, i.e. average residual differences were close to zero. Validation with GNSS showed good consistency with validation using the geoid model. The best fit with the altimetry height was found using buoy measurements in the Gulf of Riga (average difference 0.01 ±0.09 m). In summary, considering all validation methods, our results show that satellite altimetry can determine the height of the water level in Estonian coastal areas with an average accuracy of 0.08 ±0.07 m and in inland waters 0.20 ±0.26 m. The Level-3 along-track altimetry products from six active altimeter missions and the modelled results from the NEMO reanalysis product (available at CMEMS database) were compared with in situ tide gauge measurements between 2014 and 2017. The sea-level anomaly comparison between the three data sources was carried out with concurrent measurements at 21 stations around the Estonian coast (from 2014-2017) and a buoy station in the open part of the Gulf of Riga (in 2019).
The validation of the CMEMS Level-3 data shows that the altimeter water level data and Sentinel-3A/B Level-2 product have similar accuracy. The residual difference between the NEMO reanalysis model and the satellite data was approximately twofold near the coast compared with the open sea. The Pearson correlation coefficient on the other hand was better for the SLANEMO than the SLAL3 data near the coast. The mean correlation between the NEMO fields and the in-situ measurements was 0.91 and the mean difference was 0.26 m. The respective statistics between the satellite and the in-situ measurements were 0.80 and 0.12 m. This shows that satellite measurements near the coast have, on average, a mean error twice as low as that of the NEMO model. However, the offshore water level data from the NEMO model is very accurate (r = 0.99, mean error -9 cm), while the satellite sea level remains on a similar accuracy level as in the coastal areas. Funding: This study was partially supported by the Estonian Research Council grants PUT 1553 ('Joint Estimation of Geocentric Sea-Level Rise and Vertical Crustal Motion of the Baltic Sea Using Multi-Mission Satellite Altimetry Over the Last Seven Decades') and PUT1439 ('Future marine climate and ecological risks in the Baltic Sea') and by the European Regional Development Fund within the National Programme for Addressing Socio-Economic Challenges through R&D (RITA1/02-52-08, RITA1/02-52-04): 'Use of remote sensing data for elaboration and development of public services'. It was also supported in part by the Ministry of Science and Technology (MOST), Taiwan, under projects 108-2621-M-008-008 and 108-2911-I-008-507.