5.1. ASI algorithm tie point adaption
The ASI algorithm requires characteristic 85 GHz polarization differences for open water and complete ice coverage, respectively. One possible method to obtain these tie points for the Baltic Sea is to choose two areas of examination. One of them is supposed to be completely ice covered during the period of examination and the other is supposed to be ice-free. We selected one area located in the northern Bay of Bothnia and used Finnish and Swedish ice charts to determine days in 2008 and 2009, when the ice concentration within this area was 100%. All together 12 days in March 2008 and 30 days in February/March 2009 matched this criterion. The second area is located further south and was ice-free at the examined days.
For the resulting tie point of open water, we obtained an average value of 45.0 K with a standard deviation of 9.2 K. Hence the tie point found for open water is very close to the one used for Arctic conditions. Spreen et al. (2008) give 46 K as an optimal tie point.
For the ice-covered region averaging leads to a mean value of 21.7 K with a standard deviation of 15.0 K. Shokr et al. (2009) investigated the behaviour of microwave radiation emitted by thin artificial sea ice grown in an outdoor tank. They found that the 85 GHz polarization difference increases, for example when the bare-ice surface starts to melt or immediately after snow starts to fall. The ice charts for the examined days indicate that the sea ice distribution during the examined days was very variable with alternating periods of melting and freezing sea ice. Thus, according to the results obtained in Shokr et al. (2009) the mean value of 21.7 K is not necessarily the best value representing 100% ice concentration, but rather an upper boundary for the tie point.
To find a more representative ice tie point, we used SSM/I brightness temperatures from several days to generate ASI ice concentration maps. The ice tie point was varied and the resulting ice concentration fields were compared to high-resolution satellite images and ice charts. We obtained the best results for an ice tie point value near 16 K. This value appears reasonable as it is between the value Spreen et al. (2008) found to be representative for Arctic conditions (7.4 K) and the upper boundary obtained from averaging all measured polarization differences (21.7 K).
5.2. Adjusting the weather filter thresholds
To obtain optimal weather filter thresholds for Baltic Sea applications, we chose one area that was always ice-free during the examined time period and one area containing at least partly ice during that period. The chosen areas are roughly the same as were used for the tie point determination. The period of examination was again selected by studying Swedish and Finnish ice charts for the winter seasons 2008 and 2009. The SSM/I swath data for the period January 30 to May 1, 2008 and January 15 to May 20, 2009 were then included in the analysis.
The swath data was used to calculate the gradient ratios GR(37V, 19V) and GR(22V, 19V). Simultaneously, the mean polarization difference P(85) was determined. As the weather filters are used to reduce ice concentration overestimation, only cases where P(85) was below the open water tie point of 45 K were analysed. In the strict sense this approach is restricted to the ASI algorithm usage. However, we assume that constraining the gradient ratio analysis to low polarization differences does not introduce serious limitations concerning the weather filter usage in connection with the NASA Team algorithm.
The optimal threshold was defined to be the value associated with the lowest number of false assignments. If the minimum error rate could not be determined uniquely, the average value of the gradient ratios associated with the minimum number of false assignments was selected.
When the ice concentration algorithms were applied to SSM/I data using the new weather filter thresholds some difficulties arose in areas where melting conditions were found. The weather filters caused an underestimation of ice occurrence under these circumstances. Consequently we analysed the time period when melting occurred in the ice covered area separately. In the last days of the respective winter season the ice cover was obviously retreating towards ice-free conditions. We thus chose the time periods 20 April to 1 May 2008 and 1 May to 20 May 2009 to represent melting conditions. However, as the ice coverage in the Baltic Sea is quite variable, melting and freezing occur certainly as well in the earlier stage of the winter season, but possibly less pronounced.
Fig. 4 shows the results. The optimal threshold values and the respective occurrences of false assignments are given in Table 2. Under melting conditions the optimal threshold values were found to be higher than for non-melting conditions. This is especially true for the GR(22V, 19V) values.
Figure 4. Gradient ratios GR(22V, 19V) and GR(37V, 19V) as functions of the 85 GHz polarization difference P(85). The subfigures show measured GR(22V, 19V)-values for freezing (a) and melting conditions (b) and the measured GR(37V, 19V)-values for freezing (c) and melting conditions (d). The blue dots refer to data points obtained in the open water area, red dots refer to the at least partly ice-covered examination area. The black line indicates the chosen optimal threshold value. For the weather filter based on GR(22V, 19V) the analysis yields an optimal threshold value of 0.027 for freezing conditions and 0.043 for melting conditions. For the GR(37V, 19V) weather filter an optimal threshold value of 0.054 for freezing conditions and 0.057 for melting conditions was obtained.
Download figure to PowerPoint
Table 2. Weather filter thresholds and rate of false assignments associated with the chosen threshold values
|Threshold||Error rate||Threshold||Error rate|
|GR(22V/19V)||0.043||5 of 44 (11.4%)||0.027||30 of 358 (8.4%)|
|GR(37V/19V)||0.059||8–9 of 44 (18.2–20.5%)||0.053||24 of 358 (6.7%)|
5.3. Ice concentration fields
To test the developed separation method it was applied to SSM/I data as described in Section 4. The area under examination lies within a rectangular box in the Sea of Bothnia and is depicted in Fig. 5.
Figure 5. Examination area in the Baltic Sea. The larger rectangle indicates the area in the Sea of Bothnia for which the separation method was tested. The smaller rectangle depicts the region shown in Fig. 11.
Download figure to PowerPoint
As a first step we identify the swath tracks containing data located in the area of investigation and convert the data point positions into polar-stereographic coordinates. The swath track with the maximum number of valid brightness temperature measurements for any examined day is then selected.
For each channel, the area in which land surfaces have an impact on the measured brightness temperature is determined. This includes the coastline and any islands in the area of investigation and the surrounding area within a distance of twice the main-axis of the considered footprint, corresponding to four times the main-axis of the −3 dB ellipse given in Table 1. This land affected area is computed automatically using a Sobel filter to detect edges in the land mask and broadening the obtained line to satisfy the requirements. In the Bothnian Bay, this proceeding is not necessary for the lower frequency channels with their large footprint sizes but in case of the 85 GHz channels and their considerably smaller footprints this pre-selection of affected regions can be useful to avoid unnecessary calculations. This can also be true for the remaining channels when applied to retrieve ice concentrations in the Arctic.
The separation method described in the previous section is then performed for all pixels with 0.05 ≤α≤ 0.95.
To compare our results to high-resolution satellite images, we selected some example days in the winter season 2009 for which ASAR images were available and the weather conditions were favourable concerning the MODIS records. The resulting ice concentration fields for two example days are shown in the Figs 6 and 8. The depicted ice concentration fields are interpolated to a 2.5 km × 2.5 km grid. However, it should be kept in mind that decreasing the pixel size for depiction purposes does not provide any additional information.
On 27 March 2009 (Fig. 7), the ASAR and MODIS images show that the Bay of Bothnia was almost completely covered by ice except for a tongue of open water extending from the south-western coast northwards. At the eastern coast the ice reaches further south in the Gulf of Bothnia. The Finnish archipelago off Vaasa located approximately at the same geographical latitude as the Swedish island Holmön is surrounded by fast ice. In addition, the Holmön island that is depicted in Fig. 2 is connected to the mainland by fast ice.
For the ice concentration retrieval on 27 March, the weather filters for freezing conditions are used. Applying the NASA Team algorithm to the uncorrected SSM/I data roughly reveals the open water tongue, but produces a rather broad band of ice occurrence along the entire coastline. The ice concentration near the south-western coast of the Bothnian Bay is slightly underestimated as is the fast ice extent in the vicinity of Holmön and the archipelago off Vaasa.
When we insert the corrected SSM/I brightness temperatures in the NASA Team algorithm, the coastal ice band resulting from the land impact disappears. The ice occurrence near Holmön and the archipelago off Vaasa are still indicated, but the ice concentration is underestimated. In the northernmost part of the Bothnian Bay, the separation method causes a slight decrease of ice concentration.
Due to the higher resolution of the 85 GHz channels, the effect of applying the separation method to the SSM/I data is less obvious for the ASI algorithm ice concentration fields. Using the ASI algorithm with the uncorrected SSM/I brightness temperatures again leads to a continuous band of ice occurrence along the coast, whereas the land impact is narrower compared to the NASA Team algorithm. Using only the ocean contribution for coastal data points removes this erroneous detection of ice, while ice actually present in coastal areas is maintained, as for example the fast ice surrounding Holmön and the archipelago off Vaasa.
In the MODIS image of this day it is difficult to distinguish between dark sea ice areas and open water areas. The MODIS image seems to indicate a larger open water area than the ASAR image. The ice edge retrieved with the separated SSM/I data resembles the ice edge seen in the ASAR image. This example day indicates how difficult it is to retrieve ice concentrations from MODIS and ASAR images.
For the ice concentration retrieval on 23 April 2009, the weather filter for melting conditions was used, as the ice was retreating gradually, especially in the southern part of the examined area. Figure 8 shows the strong land influence on ice concentration retrieval in actually ice-free parts of bays being embedded by land from both sides. This is especially true for algorithms using the lower SSM/I channels. When the uncorrected SSM/I data is used, the NASA Team algorithm for instance reveals no open water where the Sea of Bothnia is narrowest. Although this area is actually completely ice-free. Applying the separation technique leads to ice concentration fields rather similar to the ones obtained with MODIS and ASAR (see Fig. 9) for both algorithms. Due to the higher resolution of the ASI algorithm, the open water in the south-western Bothnian Bay is depicted even more detailed. The erroneous ice concentration values along the coast vanish when the separation method is applied. Only one small area near the south-western coast of the Sea of Bothnia is assigned an ice concentration of approximately 40% that cannot be found in the MODIS image. In this area, the ASAR image shows a large pattern of high backscatter values that is hard to interpret. But as the ice edge is far north and the surroundings show merely open water areas, this seems to be an erroneous indication of sea ice. It is remarkable that the small band of ice along the western coast of the Bothnian Bay is maintained. As is the fast ice around the archipelago off Vaasa. Although the fast ice concentration is slightly underestimated by both algorithms.
For some cases where the ice concentrations appeared to be underestimated it should be considered whether these low concentrations may also be allocated to the coarse resolution of the involved channels. If the large footprint size of the lower resolution channels is imagined to be overlied to such a pixel the given ice concentration values might be reasonable.
Figure 10 shows only the corrected coastal ice concentration values for 23 April 2009. Within this context the coastal pixels are defined as the pixels with 0.05 ≤α≤ 0.95, that is all pixels included in the process of correction for land spillover. In these figures the ice concentration values are shown without applying any interpolation. Moreover, no land mask is overlaid to the obtained ice concentration values. The size of the square boxes corresponds to the size of the involved SSM/I channels' sample spacing. It should be noted that this is only the grid spacing of the measurements and not the resolution which is given by the footprint size. The square boxes are declined according to the orientation of the corresponding footprints. It exposes the grid cell distribution as measured with SSM/I instead of interpolating the data points to the rectangular coordinate system specified by the polar stereographic projection.
Figure 10. Corrected coastal ice concentrations for 23 April for the NASA Team (left-hand side) and the ASI (right-hand side) algorithm. The size of the square boxes corresponds to the sample spacing of 25 km × 25 km and 12.5 km × 12.5 km for the lower frequency channels and the 85 GHz channels, respectively. They are declined according to the declination of the footprint major axis.
Download figure to PowerPoint
In this depiction, the corrected brightness temperatures are not overlayed by a land mask and the values retrieved for pixels whose centre points are located over land are visible. Because only the pixels corresponding to α-values greater than 0.95 are included in the correction processing, all pixels shown in the figures have footprints with an ocean contribution of at least 5%. However, it should be noted that the separation method may introduce quite large errors for high α-values. Thus pixels not located in the direct vicinity of the coast should be treated carefully. The area where the obtained brightness temperatures can be included in the interpretation of the images depends on the footprint size. It is of course smaller for the 85 GHz pixels than for the lower resolution pixels. Considering the sample spacing and the footprint sizes, as a first approximation all pixels whose margins do not cross the shoreline can be ignored.
The remaining pixels located over land can be interpreted as follows. Each pixel represents the brightness temperature distribution within the corresponding footprint. In the land spillover correction process the land contribution is supposed to be subtracted. The remainder is then assumed to represent the ocean contribution for this footprint. Due to the much larger footprint size as compared to the pixel size, this signature may be related to an area lying outside the pixel in the adjacent sea.
For some areas where coastal ice concentration retrieved with the corrected SSM/I data was underestimated compared to the MODIS and ASAR images some interesting features are visible in the figures not overlayed with the landmask. The red arrow in Fig. 10b indicates an ice concentration of 100% retrieved with the ASI algorithm. For the same region the NASA Team algorithm as well indicates ice (Fig. 10a). In Fig. 8 these pixels were hidden by the landmask. Figure 11 shows an enlarged ASAR image of the Finnish archipelago off Vaasa as depicted in Fig. 5. It clearly indicates ice in the considered region. Possible explanations are discussed in the next section.