Evaluation of Microwave Radiometer Sea Ice Concentration Products over the Baltic Sea

Abstract

Sea ice concentration (SIC) monitoring in the Arctic using microwave radiometer data is a well-established method with numerous published accuracy studies. For the Baltic Sea, accuracy studies have not yet been conducted. In this study, we evaluated five different SIC products over the Baltic Sea using MODIS (250 m) and Sentinel-2 (10 m) open water–sea ice classification charts. The selected SIC products represented different SIC algorithm types, e.g., climate data records and near-real-time products. The one-to-one linear agreement between the radiometer SIC dataset and the MODIS/Sentinel-2 SIC was always quite poor; the slope of the linear regression was from 0.40 to 0.77 and the coefficient of determination was from 0.26 to 0.80. The standard deviation of the difference was large and varied from 15.5% to 26.8%. A common feature was the typical underestimation of the MODIS/Sentinel-2 SIC at large SIC values (SIC > 60%) and overestimation at small SIC values (SIC < 40%). None of the SIC products performed well over the Baltic Sea ice, and they should be used with care in Baltic Sea ice monitoring and studies.

1. Introduction

In this study, we evaluated selected sea ice concentration (SIC) products derived from satellite microwave radiometer data over the Baltic Sea. This evaluation was conducted using MODIS 250 m pixel size and Sentinel-2 (S-2) 10 m pixel size open water–sea ice (OWSI) charts.

Over the Arctic, numerous SIC product validation studies have been conducted, e.g., [1,2,3,4,5], but none so far have been over the Baltic Sea. The usage of SIC products over the Baltic Sea has been limited; no SIC interannual trend studies have been published. This is due to the following reasons. First, over the Baltic Sea, radiometer SIC estimation is challenging due to its typical rugged coastline, with many islands which manifest as a land spillover effect on the measured brightness temperatures and further as erroneous SIC values if not properly corrected or filtered [6]. Global SIC estimation algorithms are tuned with tie points for the Arctic sea ice, and it is poorly quantified how they fit for low-salinity (0.2–2 psu) and thin (<0.8 m) Baltic Sea ice [7]. Another error source is variable weather in the Baltic Sea, with mid-winter air temperatures sometimes well above 0 °C making snow cover on the sea ice wet, which accelerates snow metamorphosis and creates surface and internal layering in the snow, and, therefore, a SIC algorithm tuned to the Arctic’s cold conditions may give false SIC values. Finally, the Baltic Sea is quite small compared to the resolution of some SIC products, e.g., those with a 25 km resolution. It is envisaged that validation data on global SIC products will show their usefulness for Baltic Sea ice monitoring.

The validation of SIC products follows methods used in [5] where Landsat 30 m imagery was used as reference data. Landsat OWSI charts were aggregated to SIC charts with the pixel sizes of radiometer SIC products by calculating the fraction of sea ice pixels to the total number of pixels in each radiometer grid cell. Next, various statistical parameters between the Landsat and radiometer SIC charts were calculated. Validations were also conducted separately for first-year ice (FYI)-, mixed FYI and multiyear ice (MYI), or MYI-dominated cases as well as for different ice regimes, e.g., ice-edge, freeze-up and melt conditions.

The validation study with the Landsat imagery in [5] had ten different SIC products including OSI SAF [8], ESA SICCI (Sea Ice Climate Change Initiative) [8,9], and NOAA [10,11] climate data records (CDR). A superior linear agreement between the radiometer SIC and Landsat SIC was found for the 25 and 50 km SICCI-2 products. We do not aim to conduct as comprehensive a validation study as that in [5]. We mainly include SIC products from the EUMETSAT OSI SAF (Ocean and Sea Ice Satellite Application Facility) and ESA SICCI (Sea Ice Climate Change Initiative) which represent both climate data records (CDR) and near-real-time products and also high–low resolution hybrid products by pansharpening.

In the following, Section 2 first briefly describes our study area, the Baltic Sea. Next, the chosen SIC products are introduced, followed by description of the MODIS and S-2 OWSI charts and methods used in their calculation. Section 4 presents our validation studies and results, and Section 5 presents our discussion and conclusions.

2. Study Area

The Baltic Sea is a semi-enclosed brackish sea water basin in northern Europe. The Baltic Sea ice cover usually begins to form in November and has its largest extent between January and March [12,13,14]. The normal ice break-up starts in April and the ice melts completely by the beginning of June. The maximum annual ice cover ranges from 9% to 100% of the whole Baltic Sea area, and the average is 50% [14,15]. The Baltic Sea ice extent and duration of the ice season depend on the indices of the North Atlantic Oscillation (NOA) and Arctic Oscillation (AO); for example, the annual maximum ice extent generally decreases with the increasing indices of the AO and NAO [13]. The ice in the Baltic Sea is categorized as drift ice and landfast ice (LFI). LFI occurs in the coastal and archipelago areas and usually extends to the 5–15 m isobath. Drift ice has a dynamic nature due to being forced to move by winds and currents. The motion of drift ice results in an uneven and broken ice field with distinct floes up to several kilometers in diameter, leads and cracks, brash ice barriers, rafted ice, and ice ridges. In the Bay of Bothnia, the annual maximum ice thickness is typically 0.65–0.80 m, and it reaches 0.3–0.5 m even in mild winters [12,13]. The measured all-time maximum is around 1.2 m. In the southern Baltic Sea, in the coastal areas of Germany and Poland and the Danish straits, the annual maximum ice thickness seldom exceeds 0.5 m [16]. The thickness of ice ridges (sail height plus keel depth) is typically 5 to 15 m [17]. The salinity of the Baltic Sea ice is typically from 0.2 to 2‰ depending on the water salinity, growth rate, and ice age [7].

3. Materials and Methods

3.1. Radiometer SIC Products

Five radiometer SIC products selected for this study are summarized in

Table 1. Microwave radiometer SIC products evaluated over the Baltic Sea.

Product	Input data	Grid	References
OSI-408-a	AMSR2: 18.7 GHz V-pol, 36.5 GHz	PolarStereo 10 km	[19,22]
OSI-450-a	SSMIS: 19.35 GHz V-pol, 37 GHz	EASE2.0 25 km	[8,20]
OSI-458	AMSR2: 18.7 GHz V-pol, 36.5 GHz	EASE2.0 25 km	[8,20]
SICCI-HR-SIC	SSMIS: 19.35 GHz V-pol, 37 GHz, 91.7 GHz	EASE2.0 12.5 km	[21,23]
ASI-AMSR2	AMSR2: 89 GHz	PolarStereo 3.125 km	[18,24]

. Instead of including all available SIC products in our study, we selected products representing different SIC algorithm types. One of the products, the ARTIST (Arctic Radiation and Turbulence Interaction Study) Sea Ice (ASI) SIC algorithm using AMSR2 (ASI-AMSR2) 89 GHz data [18], represents high-resolution SIC products retrieved with 90 GHz brightness temperature ($T_B$) channels in combination with low-resolution weather filters. The OSI-408-a is near-real-time SIC product based on the AMSR2 18.7 and 36.5 GHz $T_B$ data used in many SIC algorithms [19]. The OSI-450-a and OSI-458 are climate data records based on SSMIS and AMSR2 data, respectively [8,20]. The SICCI-HR-SIC (HR means high resolution) is a resolution-enhancing SSMIS product where SIC with 19.35 and 37.0 GHz $T_B$ data (the same SIC as in the OSI-450-a product) is pansharpened using a SIC with finer-resolution 91.7 GHz data [21]. These products have pixel sizes from 3.125 to 25 km. Some of the products are oversampled, like the ASI-AMSR2 3.125 km product. Products with larger pixel sizes are taken to be too coarse for the Baltic Sea, especially for the Bay of Bothnia and Gulf of Finland where sea ice forms every winter.

Descriptions of the SIC retrieval algorithms, e.g., tie point selection and applied weather filters, can be found in the references shown in Table 1. All products, except the ASI-AMSR2, use the daily dynamic tie points method from Maass and Kaleschke [6] to correct the land spillover already occurring at the $T_B$ level, a dynamic open water filter, and the atmospheric correction of $T_B$s [8]. The ASI-AMSR2 is based on fixed tie points and fixed open water filters [18], and $T_B$s are not corrected for atmospheric effects. All SIC products have SIC values between 0 and 100%, i.e., they are truncated to the range between 0% and 100%.

3.2. MODIS OWSI Chart

For the validation of the radiometer SIC products over the Baltic Sea SIC, we derived an open water—sea ice (OWSI) chart using MODIS band 1 reflectance data which have a 250 m spatial resolution. The total number of the MODIS OWSI charts is 62, and they cover February to April 2019. This OWSI chart derivation follows our earlier study for the Barents and Kara Seas [25]. In the following, the used MODIS datasets and their processing are first described, including the automatic and manual cloud masking of the MODIS reflectance data. Next, the OWSI chart derivation is described.

The MODIS spectrometer measures reflected solar and emitted thermal radiation in 36 spectral bands ranging from the visible to the thermal infrared (TIR) (0.415 to 14.235 μm) [26]. The MODIS instrument is aboard NASA Terra and Aqua satellites. In this study, we derived an OWSI chart for the Baltic Sea using band 1 reflectance ($R_1$) data at a 250 m spatial resolution and the MODIS cloud mask product [27]. MODIS sea ice surface temperature (IST) data and an RGB image with a 2-1-1 band combination (RGB211) were used in finding cloud-free MODIS swath datasets. The following Terra and Aqua MODIS products were used: the MOD/MYD/02QKM-MODIS/Terra Calibrated Radiances 5-Min L1B Swath 250 m, MOD/MYD/03-MODIS Geolocation Fields 5-Min L1A Swath 1 km, MOD/MYD/35_L2-MODIS Cloud Mask and Spectral Test Results 5-Min L2 Swath 250 m and 1 km, and MOD/MYD/29-MODIS Sea Ice Extent 5-Min L2 Swath 1 km. The MOD/MYD29 product contained IST data, and the cloud mask was derived from the MOD/MYD35_L2 unobstructed field-of-view flag [28]. In the following, these products are only referred to with the ‘MOD’ prefix.

The MODIS products were rectified to the Mercator projection used for the Finnish Ice Service (FIS) chart using the MODIS Swath Reprojection Tool developed by the NASA LP DAAC (Land Processes Distributed Active Archive Center). In the rectification of the MODIS products, original pixel sizes (250 or 1000 m) were kept, and fixed Mercator grids were used. The size of the rectified data was 1330 km in the northing and 1120 km in the easting. The MODIS sensor scan angle (${\mathsf{\theta }}_s$) was calculated from the sensor zenith angle, Earth radius, and satellite orbit altitude. Raw data from the bands 1 and 2 were converted to top-of-the-atmosphere (TOA) sun angle corrected reflectances ($R_1$ and $R_2$). The land mask for the Baltic Sea study area was retrieved from the MOD44W (Land Water Mask Derived L3 Global 250 m) product.

In all MODIS data analyses herein, the retrieval of the MODIS OWSI chart ${\mathsf{\theta }}_s$ was limited to a maximum of 40° (the overall maximum is 55°). At a ${\mathsf{\theta }}_s$ of 40° the MODIS spatial resolution in the across-track direction decreased by a factor of around two, e.g., from 250 m at the nadir to ~500 m at 40° [29]. In the MOD29 and MOD35 products, daylight data were defined as data collected with a ${\mathsf{\theta }}_{zs}$ less than 85° [27,28]. We used here a somewhat smaller limit of 80° as at high sun zenith angles the $1/\mathrm{c}\mathrm{o}\mathrm{s}\left({\mathsf{\theta }}_{zs}\right)$ term in the TOA reflectance calculation is highly variable, and, thus, the accuracy of the estimated ${\mathsf{\theta }}_{zs}$ is critical. This ${\mathsf{\theta }}_{zs}$ limitation excluded all the MOD02 data from December 2018 to January 2019 in the Bay of Bothnia. Therefore, we used here only MODIS data acquired over the Baltic Sea from 1 February 2019 onwards to end of April 2019.

The following procedure was used to select mostly cloud-free MOD02 data over the sea ice-covered part of the Baltic Sea. First, all available MODIS IST data were processed over the Baltic Sea from 1 February to 30 April 2019. Visual analysis of the IST charts with the MOD35 cloud mask was conducted to select times for the MOD02 data, which possibly had large cloud-free areas over the sea ice-covered part of the Baltic Sea. Next, RGB211 images with and without s cloud mask were used in the further screening of the MOD02 datasets for the OWSI chart derivation. We ended up with 20 MOD02 swath datasets for February, 21 for March, and 21 for April. Here, a MOD02 dataset was either one MOD02 swath or two swaths stitched together. For some days, there were two datasets (Terra and Aqua) and their temporal difference was mostly 10 or 15 min and at maximum 100 min.

Cloud masking was started by designating pixels that had ‘confident clear’, ‘probably clear’, or ‘probably cloudy’ values in the unobstructed field-of-view quality flag of the MOD35 product as cloud-free. This logic followed the MOD29 IST product generation with the target to increase the number of retrievals balanced against the cloud conservative nature of the cloud mask [28]. A visual analysis of the resulting cloud mask over the RGB211 image showed that sometimes cloud-free pixels of open water or the thin ice within led to larger open water/thin ice areas being classified as cloudy, and cloud-free sea ice, in general, was classified as cloudy. This is because NSIDC’s Near-real-time Ice and Snow Extent (NISE) product with a 25 km resolution was used to initialize the sea ice background flag for directing the cloud masking algorithm processing flow [28]. Open water or thin ice when the NISE sea ice flag is set may be classified as cloudy, and sea ice when the flag is not set may also be classified as cloudy. Methods have been developed for the correction of these cloud mask-induced errors in the OWSI classification, e.g., in [30], but we decided to correct these errors by the manual editing of the cloud mask. For the sea ice-covered Baltic Sea, which has a rather small area, the manual cloud mask editing for one MODIS dataset was not an overly laborious task, and it also served as a quality control step in the OWSI chart processing.

Before the manual cloud masking, all single or groups of two cloudy pixels within the NISE sea ice flag set area were first removed. These were assumed to be erroneous cloud detections over the open water or thin ice. Next, the cloud mask was aggregated into 5 km blocks (5 by 5 1 km pixels), and if there were more than or equal to three cloudy pixels (12% of all 25 pixels), then the whole block was flagged as cloudy. Using 5 by 5 km blocks for the cloud mask, we could better identify large cloud-free areas and discard areas where the cloud mask was tattered. Small cloud-free areas (‘holes’ in the mask) less than ten blocks in size were flagged as cloudy. Finally, the following manual editing procedures could be conducted: filling holes, removing erroneous cloud mask elements, and masking arbitrary polygonal areas as cloudy or clear. The manual editing was conducted using the RGB211 image. An example of the original and manually edited cloud masks is shown in Figure 1.

We aimed to develop a simple OWSI classification method based only on the threshold $R_1$ (bandwidth 620–670 nm, red). We only used the $R_1$ data as it had the best possible MODIS resolution, 250 m. The $R_2$ data (841–876 nm, near-infrared (NIR)) also had a 250 m resolution, but we evaluated that it was not needed in the OWSI classification. In addition, we have previously determined an $R_1$-based OWSI classification for the Barents and Kara Seas [25]. In the Baltic Sea, MODIS $R_1$ data have been used to map the sea ice extent in the Gulf of Riga [31]. In this study, the threshold for the OWSI classification was determined automatically for each swath dataset based on a sampled $R_1$ histogram which was bimodal, showing OW and SI peaks. More accurate OWSI classification is also possible using other MODIS reflectance bands and shortwave and thermal infrared bands, e.g., as in [28,30,32], but this is at the cost of decreasing the resolution to 500 m or 1 km. In the MOD29 product, for daytime, there is sea ice extent data based on various reflectance tests, including the threshold $R_1$, at 1 km resolution [28]. An algorithm called IceMap500 has been developed to generate 500 m sea ice extent maps with a systematic accuracy over 90% [30]. Machine learning methods have also been developed to classify OW, SI, and clouds; for example, Jiang et al. [33] constructed a training sample library and used a Multi-Feature Level Fusion Random Forest classification algorithm that integrated multiple features from the reflectance bands 1, 2, and 7. Their SI chart had a 250 m pixel size, but this required the interpolation of the band 7 data at a 500 m resolution to 250 m.

For the determination of the $R_1$ threshold for the OWSI classification in the Baltic Sea, we manually selected samples in the cloud-masked RGB211 images from February 2019 for OW, thin ice (THI), and snow-covered sea ice (SSI). These samples represented winter conditions in the Baltic Sea. The size of the sample windows for the OW and SSI was 5.25 by 5.25 km (21 by 21 pixels), and it was 2.75 by 2.75 km (11 by 11 pixels) for the THI. The number of sample windows was 210 for the OW, 278 for the THI, and 128 for the SSI. Another set of samples were selected in April 2019, which represented springtime melting sea ice conditions. Here, we selected windows for the OW (219 windows), SSI (35 windows), and melting sea ice (likely mostly snow-free) (MSI, 385 windows, 11 by 11 pixels window size).

The resulting $R_1$ probability density function (pdf) for the OW is very narrow (see Figure 2); for example, the 95th percentile of the $R_1$ is only 0.070 in the winter data. The SSI is well separated from the OW, and the 5th percentile of the $R_1$ is around 0.55. In the winter data, the THI has a large variation, from 0.16 to 0.51 by the 5th and 95th percentiles. The MSI likewise has a large variation, from 0.073 to 0.42. Setting the $R_1$ threshold to 0.10, only 0.12% of the OW samples in the winter data are classified as THI, and only 1.0% of the THI is classified as OW. In the spring data, the OW and MSI misclassifications are 0.46% and 4.5%, respectively. Unfortunately, it is not possible to separate fully the OW and very thin ice which both have small $R_1$ values. In addition, the $R_1$ data (TOA) here do not have atmospheric corrections, and in the springtime, water turbity close to the coast and river estuaries may raise $R_1$. Therefore, it is better to use somewhat too high than too small of a threshold. We decided to use the threshold of 0.10 for all $R_1$ data as it gives very small OW misclassifications in our manually selected sample data:

$$\mathrm{sea} \mathrm{ice} \mathrm{if} R_1>0.10.$$

(1)

This threshold was manually determined. In the resulting OWSI charts, very thin ice (thickness: a few cm) and very wet sea ice may be classified as OW. Fog and thin unmasked clouds over OW can also be misclassified as sea ice. In addition, the mixed-pixel effect affects the OWSI classification as even a small fraction of sea ice, especially that of snow-covered ice, within the 250 m pixel leads to assigning the SI class to the pixel. This would further lead to the overestimation of SIC when the MODIS data are used to calculate SIC at radiometer data resolutions. An example of the resulting OWSI charts in shown in Figure 3 for the same date and time as the RGB211 images in Figure 1.

3.3. Sentinel-2 OWSI Chart

Sentinel-2A and Sentinel-2B (S-2A/B) carry the MultiSpectral Instrument (MSI) with 13 different spectral bands from VNIR (visible–near-infrared) to SWIR (shortwave infrared) with band central wavelengths varying from 443 to 2202 (S-2A) / 2186 (S-2B) nm [34,35,36]. The spatial resolution for different bands is either 10, 20, or 60 m. The swath width of MSI is 290 km. The Sentinel-2 mission supports Copernicus land, marine, security, emergency, and climate change services and studies, including the monitoring of vegetation, soil, and water cover, as well as the observation of inland waterways and coastal areas. The MSI data can also be used for the remote sensing of sea ice, e.g., for open water–sea ice (OWSI) classification, but MSI lacks longwave infrared bands for good cloud masking over sea ice, e.g., 3.9 and 11 µm TIR bands, as in MODIS.

There are two end-user MSI products: Level-1C top-of-atmosphere (TOA) reflectance and Level-2A bottom-of-atmosphere (BOA) reflectance. We used here the L1C product with the processing baseline 02.07 version. The L1C product had 100 by 100 km fixed tiles (or granules) in the UTM/WGS84 projection. The L1C product had opaque (dense clouds) and cirrus cloud masks [37]. The cloud masks were processed with data sampled at a 60 m spatial resolution for all spectral bands. The land masking of the L1C tiles was conducted here with the ESA SNAP software (version 9.0) (fractional land/water mask). The fixed land mask was based on the ESA Global Land Cover Map data. It did not identify all the small islands in the Bay of Bothnia.

We aimed to use the 10 m resolution bands (best possible resolution)—B2 (490 nm; blue), B3 (560 nm; green), B4 (665 nm; red) and B8 (842 nm; NIR)—from the L1C product for the OWSI classification over the Bay of Bothnia. B4 was equal to MODIS band 1 (bandwidth of 620–670 nm) which was used here for the MODIS OWSI classification.

As S-2 cloud detection over sea ice may be inaccurate, we searched S-2 data over the Bay of Bothnia only for the dates when the MODIS data were mostly cloud-free and the sun zenith angle was mostly below 80°. The Bay of Bothnia was covered with five 100 km tiles in the UTM34 projection. Our data search for February-April 2019 resulted in S-2 imagery for 13 dates; see an example in Figure 4.

S-2 data have been used previously in sea ice remote sensing studies either as a validation dataset or to classify sea ice for further analyses, like for lead statistics, e.g., in [38,39,40]. Ludwig et al. [38] evaluated the accuracy of their MODIS and AMSR-E data-based SIC product with a 1 km pixel using 79 cloud-free (visually determined) S-2 L1C products acquired mainly over the Arctic FYI. They investigated the $R$s of the visible spectrum bands 2–4 for open water (OW), thin ice, and thick ice classification. These bands did not show significant differences in the classification, and there was a clear distinction between the thin and thick ice and a clear distinction between the OW and sea ice (SI) if only thick ice was present. They decided to use only the band 4 reflectance ($R_4$) for the surface type classification and selected specific classification thresholds (OW vs. thin ice and thin vs. thick ice) for each $R_4$ image instead of using global thresholds for all images. Muchow et al. [39] investigated lead-width distribution for the Antarctic sea ice using 20 cloud-free L1C products. Lead detection was conducted by thresholding $R_4$. For the threshold determination, they selected manually sample data for OW and the following four ice types: nilas (NI; thickness < 10 cm), gray sea ice (GI, 10–15 cm), gray–white ice (GWI; 15–30 cm) and sea ice covered with snow. The sample data were used to calculate $R_4$ histograms for each surface type, and then a summation of Gaussian functions was fitted to them. The threshold for each surface type was then determined from the intersection of two curves adjacent to each other. For the lead identification, two different thresholds were used: 0.10 for leads covered with OW and 0.17 for leads covered with OW and NI. The threshold of 0.10 classified 29% of NI as OW and the threshold of 0.17 classified 11% of GWI as NI. Wang et al. [40] developed a decision tree classification of OW and various ice types in the Liaodong Bay of the Bohai Sea using S-2A/2B data. Using one S-2B image, they selected manually sampled areas for different surface types: OW, NI, GWI, and landfast ice (LFI). Utilizing mean reflectance spectra ($R_2$ to $R_{12}$), a decision tree classification was determined for these surface types. This classification tree included 10 m and 20 m bands. The total accuracy of the classification was 88% according to visual validation. These previous sea ice classification studies demonstrate that OWSI classification in the Baltic Sea is possible with the S-2 data.

We started the development of the OWSI classification method by first investigating the pdfs of the 10 m reflectances of $R_2$ (490 nm; blue), $R_3$ (560 nm; green), $R_4$ (665 nm; red), and $R_8$ (842 nm; NIR)), as well as the normalized difference water index (NDWI) in the Bay of Bothnia. The NDWI is calculated as follows [41]:

$$\mathrm{N}\mathrm{D}\mathrm{W}\mathrm{I}={\displaystyle \frac{R_3-R_8}{R_3+R_8}}.$$

(2)

The NDWI detects open water features against soil and terrestrial vegetation features (they have a smaller NDWI than OW). It is noted that the normalized difference snow index (NDSI) could not be calculated from the S-2 10 m reflectance data. The pdfs were calculated using imagery of one tile in the northern Bay of Bothnia (tile 34WFT) acquired on 5 March and on 6 April 2019. The first date represented winter conditions (the air temperature $T_a$ at the Kemi 1 lighthouse was −10 °C), and there were large areas of thin ice (<15 cm) according to the Finnish Ice Service (FIS) ice chart; the second date had spring, early melting conditions (the $T_a$ was around 0 °C) when there was no thin ice in the Bay of Bothnia. For 5 March, all four 10 m $R$ pdfs show a distinct peak at small $R$ values, representing OW, and another peak at high $R$ values, representing snow-covered SI, mostly LFI; see Figure 5. There are a few smaller peaks, although they are not very distinct, representing different types of level, deformed, and snow-free thin ice. There are no clear $R$ thresholds between the OW and thin ice. For $R_4$ and $R_8$, the OW peak is somewhat narrower than for $R_2$ and $R_3$. The $R$ pdfs on 6 April also have a OW peak, but they have much a smaller snow-covered SI peak as large areas of LFI snow had melted; see Figure 5. The snow melting resulted in large fractions of $R$ values in the 0.2–0.3 to 0.6 range. There is now a clear $R$ local minimum after the OW peak, as there was no new thin ice on 6 April. The NDWI pdf shows a clear LFI peak and smaller OW peak but no better discrimination of surface types than the single $R$ pdfs.

Likely, there was not much difference that affected which $R$ band we would use for the OWSI classification, but $R_4$ and $R_8$ could have been better than $R_2$ and $R_3$ due to their narrower OW peaks. Therefore, we selected $R_4$ for further analyses as it has also been used previously for lead detection and SIC estimation [38,39].

First, we investigated the $R_4$ statistics for the OW on 5 March and 6 April by visually selecting 195 OW rectangles from four S-2 tiles acquired on 5 March and 194 rectangles from five tiles on 6 April. The size of the rectangles was from 0.15 to 870 km². On 5 March, the tile-based mean $R_4$ varied only from 0.047 to 0.056, the mode was either 0.05 or 0.06 when the pdf bin width was 0.01, and the 99th percentile was from 0.057 to 0.065. On 6 April, the tile based $R_4$ statistics were slightly larger: the mean $R_4$ varied from 0.048 to 0.059, and the 99th percentile varied from 0.056 to 0.078. The mode was again either 0.05 or 0.06.

These OW $R_4$ statistics suggest that fixed thresholds for the OWSI classification are not feasible. A tile-based threshold is a better approach. This tile-based OW threshold takes into account variation in the sun zenith and sensor zenith angles (TOA reflectance does not correct for atmospheric effects, which depend on the sensor zenith angle, i.e., the path length). In the above rectangle selection, we observed that it was sometimes difficult to visually discriminate between OW and very thin ice, i.e., dark nilas and new ice (including grease ice and slush), which had same kind of ‘dark’ appearance (in level and texture) in the $R_4$ image. There was also sometimes thin fog over the OW and thin ice, which made their identification more difficult and raised the $R_4$. Most notably, there were areas where both the OW and very thin ice had a similar $R_4$ level (and no texture). The OW and new ice also had very similar S-2 spectra in [40]. These observations show that full discrimination between OW and very thin ice is not possible, and a $R_4$ threshold for the OWSI classification will classify some unknown part of thin ice to the OW class. This is likely also the case for OW and melting wet ice in late spring.

Next, we needed to find a way to determine the $R_4$ threshold for each L1C tile. One way to conduct this is to manually sample OW data from a tile and to calculate the 99th (or a lower 95th) percentile of $R_4$. Unfortunately, this manual sampling is time consuming. We tested fitting a mixture of Gaussian pdfs to the $R_4$ OW and SI data in a tile and used the intersection between the pdf with the smallest mean (representing OW) and the pdf with the second smallest mean as the threshold. The fitting seemed to work fine when there was a local minimum after the OW peak but not when this was not the case. Therefore, we decided to set the threshold manually with the help of the $R_4$ pdf. The manually set threshold varied from 0.06 to 0.11, and it was larger in winter, from 0.09 to 0.11, than in the melting season (12 April onwards), from 0.06 to 0.08. The manual setting was easier in the melting season as new thin ice was not present.

Before the OWSI classification, the opaque (dense clouds) and cirrus cloud masks were combined, and the combined mask was manually edited. This editing was conducted with the help of an RGB image with $R_4$-$R_3$-$R_2$ bands (approximating a true color image); see an example in Figure 4. In the manual cloud masking editing we rather masked too much than too little to cover for certain all cloudy and cloud shadow areas. A visual comparison of the cloud masks of the $R_4$ and RGB images showed that the opaque cloud mask was sometimes showing erroneously cloudy areas over LFI. In some $R_4$ images from later in April, there were increased $R_4$ values over open water at river mouths due to water turbidity. These areas were included in the manual cloud mask as they would otherwise be classified as SI.

An example of the resulting OWSI charts is shown in Figure 6. Here, the S-2 tile data are the same as in Figure 4. All tile-based OWSI charts for the same S-2 overpass (max. five tiles) were combined to create an OWSI mosaic over the Bay of Bothnia.

3.4. Coastal Weather Station Data

Air temperature data from three coastal weather stations in the Bay of Bothnia (see Figure 7) were used to identify cold winter conditions (air temperature below 0 °C) and melting conditions for the SIC product evaluations. There were cold conditions roughly on 1–15 February, 17–23 February, 28 February–15 March, and 25–28 March; otherwise, there were warm melting conditions.

3.5. Co-Location and Comparison

For comparison with the MODIS OWSI chart, the SIC products were rectified to the Mercator projection of the MODIS chart with nearest neighbor sampling. Inside each SIC product pixel, the MODIS SIC was calculated as the fraction of sea ice pixels to the total number of pixels. To reduce the land spillover effect, only those radiometer SIC pixels which had, at maximum, 10% land in the MODIS land mask were included in the comparison.

In similar way, the SIC products were rectified to the UTM34 projection of the S-2 OWSI chart, and then the S-2 SIC was calculated within the SIC product grid cells which had, at maximum, 10% land in the S-2 land mask.

In selecting radiometer data pixels for comparison, the status flag of the SIC retrieval in the OSI-450, OSI-458, and SICCI-HR-SIC products was utilized. This flag showed when an open water filter and a land spill over filter had been applied. We only selected pixels with no filters applied in order to only have ‘true’ retrieved open water pixels in the comparison. There was also a status flag in the OSI-408-a product, but it did not have the needed information. The ASI-AMSR2 product did not have a status flag at all. For these two products, we excluded pixels with a radiometer SIC of exactly zero. These pixels were assumed to be mostly the result of applied filters. Keeping the rejected pixels would have had a large effect on the comparison statistics.

The comparison between the gridded MODIS or S-2 SIC and co-located radiometer SIC was conducted by calculating the mean difference of the radiometer SIC minus the MODIS/S-2 SIC, standard deviation of the difference, and linear regression line and coefficient of determination of the regression ($R^2$).

4. Results

4.1. Comparison with the MODIS OWSI Chart

Statistics on the comparison between the MODIS SIC and different radiometer SIC products are shown in

Table 2. Statistical parameters between the radiometer SIC and MODIS SIC using all data. The mean difference is the radiometer SIC minus the MODIS SIC. The slope and intercept are the coefficients of linear regression, and R² is the coefficient of determination. N is the number of data pairs.

	OSI-408-a	OSI-450-a	OSI-458	SICCI-HR-SIC	ASI-AMSR2
Mean difference [%]	1.1	10.6	0.8	5.3	−12.4
Std of difference [%]	19.7	21.0	15.5	17.8	26.8
Slope	0.71	0.58	0.54	0.77	0.66
Intercept	17.0	29.5	37.0	11.6	14.2
R2	0.80	0.79	0.64	0.80	0.50
N	8101	2249	912	18,291	114,733

. Examples of the MODIS SIC and OSI-408-a SIC images, both with a 10 km pixel size, are shown in Figure 7. The images also show the coastline to illustrate how far the land mask extends from the coast. Here, the land mask extension is typically 2–3 pixels, at which distance there can be some leftover land spill effect after correcting or filtering the measured $T_B$s. A scatterplot between the MODIS and OSI-408-a SIC datasets is shown in Figure 8.

For the OSI-408-a SIC, the overall mean difference is small, only +1.1%; however, at high SICs, the OSI-408-a SIC underestimates the MODIS SIC. For example, in the 85 to 95% MODIS SIC range, the average OSI-408-a SIC is only 76.0%. At small MODIS SICs, there is on the contrary overestimation; for example, in the 5 to 15% MODIS SIC range, the average OSI-408-a SIC is 37.0%. The standard deviation of the difference is large, at 19.7%. Dividing the data into winter and melting cases does not change statistics much. For example, the standard deviation is 20.5% for the winter case and 18.1% for the melting one. The slope of the linear regression is 0.71 and the $R^2$ is 0.80 shows a rather poor one-to-one linear agreement between the datasets.

The mean difference for the OSI-450-a SIC is 10.6%. At a high SIC, there is an underestimation of the MODIS SIC; in the 85 to 95% MODIS SIC range, the average OSI-450-a SIC is only 78.8%. Also, at a small SIC, we find overestimation; in the 5 to 15% MODIS SIC range, the average is 41.9%. The standard deviation of the difference is large here, at 21.0%, and it is almost the same for the winter and melting conditions. The slope is only 0.58 and the $R^2$ is 0.79.

For the OSI-458 SIC product, the mean difference is very small, 0.8%. There is a good match with the average when SIC ≥ 60%, but there is overestimation at SIC ≤ 50%. The standard deviation of the difference is modest, at 15.5%, ranging between 13.4% for the melting conditions and 16.2% for the winter. The slope is small, only 0.54, and the $R^2$ is 0.64.

For the SICCI-HR-SIC product, the scatterplot against the MODIS SIC also shows SIC underestimation at a high SIC; in the 85 to 95% MODIS SIC range, the average SICCI-HR-SIC is 77.5%. Also, there is overestimation at a small SIC; in the 5–15% MODIS SIC range, the mean SIC is 26.0%. The mean difference is 5.3% and the standard deviation of the difference is 17.8% overall, taking a larger value for the winter case, at 19.4%, and a smaller value for the melting case, at 16.4%. The slope is the largest of the products investigated, at 0.77, and the $R^2$ is 0.80.

The ASI-AMSR2 SIC has a very poor match with the MODIS SIC. The overall mean difference is −12.4%, the standard deviation of the difference is 26.8%, and there is typically underestimation of the MODIS SIC when the radiometer SIC is larger than 40; for example, in the 85 to 95% MODIS SIC range, the average ASI-AMSR2 SIC is only 54.6%. For the winter and melting conditions, the standard deviation of the difference is also large, at 28.2% and 24.0%, respectively. A poor match between the datasets is reflected in the particularly low value for the $R^2$ of only 0.50, while the slope of the linear regression is only 0.66.

4.2. Comparison with the Sentinel-2 OWSI Chart

Comparison statistics between the S-2 SIC and different SIC products are shown in

Table 3. Statistical parameters between the radiometer SIC and Sentinel-2 SIC using all data. The mean difference is the radiometer SIC minus the Sentinel-2 SIC. The slope and intercept are the coefficients of linear regression, and R² is the coefficient of determination. N is the number of data pairs.

	OSI-408-a	OSI-450-a	OSI-458	SICCI-HR-SIC	ASI-AMSR2
Mean difference [%]	−0.6	1.3	0.4	2.5	−12.0
Std of difference [%]	15.5	22.0	20.2	22.1	24.0
Slope	0.70	0.50	0.40	0.67	0.76
Intercept	19.9	32.7	46.5	20.6	7.8
R2	80.0	0.59	0.26	0.68	0.48
N	1293	242	139	1386	18,755

. Figure 9 depicts a scatterplot between the S-2 SIC and OSI-408-a data. In the comparison, we again rejected samples where the radiometer SIC was exactly zero or status flags showed the application of open water and land spillover filters.

The mean difference, the OSI-408-a SIC minus the S-2 SIC, is small, only −0.6%, but there is a slight underestimation of the S-2 SIC when the S-2 SIC is larger than 60; for example, in the 85 to 95% S-2 SIC range, the average OSI-408-a SIC is 82.8%. At SICs below 50%, the OSI-408-a SIC overestimates compared to the S-2; in the 5 to 15% S-2 SIC range, the average OSI-408-a SIC is 31.0%. The standard deviation of the difference is modest, at 15.5%, ranging between 13.5% for winter conditions and 16.1% for melting conditions. The linear one-to-one agreement between the SIC datasets is not very high; the slope is 0.70 and the $R^2$ is 0.80, almost identical to the values obtained for the comparison to the MODIS SIC for this product.

For the OSI-450-a and OSI-458 products, the accuracy of the results of the comparison is limited by the small number of samples (max. 242). The mean difference is 1.3% for the OSI-450-a and 0.4% for the OSI-458. The standard deviation of the difference is large, at 22.0% and 20.2%, respectively. The slope is only 0.50 for the OSI-450-a and 0.40 for the OSI-458, which is considerably worse than that for the MODIS SIC.

For the SICCI-HR-SIC, the mean difference is small, at 2.5%, but the S-2 SIC is underestimated when SIC > 70% and overestimated when SIC ≤ 40%. The standard deviation of the difference is again large, at 22.1% for all data, with a substantial variation between the winter conditions (18.6%) and melting conditions (23.6%). The one-to-one linear agreement is quite poor, with a slope of 0.67 and $R^2$ of 0.68.

The ASI-AMSR2 underestimates the S-2 SIC when SIC ≥ 50%. The mean difference is large, at −12.0%. The standard deviation of the difference is again large, at 24.0% for all data, 26.6% for the winter, and 20.8% for the melting conditions. The slope of the linear regression is the highest of the products compared to the S-2 SIC, at 0.76, but the $R^2$ is only 0.48.

5. Discussion and Conclusions

Compared to the Arctic Sea ice statistics in [5], the standard deviations of the differences are here larger. For example, in [5], the largest observed standard deviation was 11.7% in the Arctic FYI case. Here, the smallest standard deviation was 15.5% and the largest was 26.8%. A part of the large differences in the standard deviation could be due to the usage of Arctic tie points for cold conditions under Baltic Sea melting conditions. Further, the Arctic tie points in the SIC algorithms may not fit with the low-salinity Baltic Sea ice, but this is likely a minor error source.

A common feature of the OSI SAF and SICCI-HR-SIC datasets was the typical underestimation of the MODIS SIC at large SIC values (SIC > 60%) and its overestimation at small SIC values (SIC < 40%). A similar underestimation of the S-2 SIC was also present when it was larger than 60 or 70%, and its overestimation typically happened when SIC ≤ 40 or 50%. In the case of the ASI-AMSR2 SIC, the underestimation of the MODIS and S-2 SIC occurred over a wider SIC range, SIC > 50%. This underestimation was partly due to a possible positive bias in the MODIS/S-2 SIC, because a pixel with even a tiny fraction of highly reflective sea ice, i.e., snow-covered sea ice, was classified as sea ice with the reflectance thresholding method. This underestimation effect was larger for finer-resolution SIC products, as the same number of MODIS/S-2 pixels classified as sea ice resulted in a larger aggregated MODIS/S-2 SIC. For example, 100 such MODIS pixels resulted in a SIC of 1% in the 25 km grid, but the 10 km grid gave a SIC of 6.25%. Another source for the underestimation is the occurrence of thin ice [2,42,43]. The SIC overestimation at a small SIC could be partly due to remnant land contamination, i.e., the land contamination correction was not perfect on the measured $T_B$ data.

The one-to-one linear agreement between the radiometer SIC dataset and the MODIS/S-2 SIC was always quite poor; the slope of the linear regression was from 0.40 to 0.77 and the $R^2$ ranged from 0.26 to 0.88.

Dividing the data into winter and melting conditions did not have a clear consistent effect on the statistics. In the case of the comparison against the MODIS SIC data, the standard deviation value of the difference typically slightly increased for the winter conditions compared to the value for all data, and it slightly decreased for the melting conditions. However, we did not always observe such a behavior in the comparison against the S-2 SIC data. This could be due to different data amounts in the MODIS and S-2 SIC cases and complex changes in snow and sea ice properties from winter to melting conditions and vice versa.

It is difficult to rank the SIC products based on the comparison results. For example, the OSI-458 had a smaller standard deviation of the difference (15.5%) than the MODIS SIC data, but against the S-2 SIC data, the standard deviation of the difference was considerably larger than the smaller OSI-408-a one (20.2% vs. 15.5%).

The two CDRs, the OSI-450 and OSI-458, were retrieved with the same algorithm. Only the input data were different, with SSMIS for the OSI-450 and AMSR2 for the OSI-458. The standard deviation of the difference was smaller for the OSI-458 product when against both the MODIS and S-2 data, as it should have been based on results in [5], because the OSI-458 SIC was computed from finer-resolution AMSR2 input data.

The worst overall performance was surprisingly observed for the ASI-AMSR2 product, which had the finest resolution of all the SIC products. Finer resolution meant larger spatial variation in the SIC and a larger effect from the spatial mismatch of an ice area between the radiometer SIC and MODIS/S-2 products (the radiometer SIC is a daily product whereas the MODIS/S-2 SIC is an instantaneous swath product). The ASI-AMSR2 product, using 89 GHz channels, was sensitive to atmospheric effects, which were not corrected for. However, cloud liquid water does not have role here, as clear-sky SIC data were used and, thus, the atmospheric effects were likely not that large. The other radiometer SIC products included the atmospheric correction of the $T_B$ data, and less sensitive lower-frequency channels were used in the SIC retrieval.

In summary, none of the SIC products performed well over the Baltic Sea ice, and they should be used with care in Baltic Sea ice monitoring and studies. The accuracy of the products could be improved by diminishing any further land spillover effects on the measured $T_B$ data and using Baltic Sea-specific tie points.