Antarctic sea ice types from active and passive microwave remote sensing

Polar sea ice is one of the Earth’s climate components that has been significantly affected by the recent trend of global warming. While the sea ice area in the Arctic has been decreasing at a rate of about 4% per decade, the multi-year ice (MYI), also called perennial ice, is decreasing at a faster rate of 10%–15% per decade. On the other hand, the sea ice area in the Antarctic region was slowly increasing at a rate of about 1.5% per decade until 2014 and since then it has fluctuated without 5 a clear trend. However, no data about ice type areas are available from that region, particularly of MYI. Due to differences in physical and crystalline structural properties of sea ice and snow between the two polar regions, it has become difficult to identify ice types in the Antarctic. Until recently, no method has existed to monitor the distribution and temporal development of Antarctic ice types, particularly MYI throughout the freezing season and on decadal time scales. In this study, we have adapted a method for retrieving Arctic sea ice types and partial concentrations using microwave satellite observations to fit the 10 Antarctic sea ice conditions. The first circumpolar, long-term time series of Antarctic sea ice types; MYI, first-year ice and young ice is being established, so far covering years 2013-2019. Qualitative comparison with synthetic aperture radar data, with charts of the development stage of the sea ice, and with Antarctic polynya distribution data show that the retrieved ice types, in particular the MYI, are reasonable. Although there are still some shortcomings, the new retrieval for the first time allows insight into the evolution and dynamics of Antarctic sea ice types. The current time series can in principle be extended 15 backwards to start in the year 2002 and can be continued with current and future sensors.


ECICE -Environment Canada's Ice Concentration Extractor
ECICE can take passive and active microwave satellite measurements as input. Possible passive microwave input data are from 90 the satellite microwave radiometers Special Sensor Microwave/Imager (SSM/I, 1987(SSM/I, -2009, Advanced Microwave Scanning Radiometer for EOS (AMSR-E, 2002(AMSR-E, -2011 and Advanced Microwave Scanning Radiometer 2 (AMSR2, 2012-present). The measured quantities are the brightness temperatures at different frequencies and at vertical (V) as well as horizontal (H) polarisation. Possible active microwave input data include scatterometer measurements from QuikSCAT (1999-2009) and Advanced Scatterometer (ASCAT, 2007-present). The measured quantity is the backscattering coefficient (normalised radar 95 backscattering cross section) at one or two polarisations (HH and VV). The number of input parameters to ECICE must be equal or greater than the number of the surface types to be distinguished. Here we use four surface types; namely open water, young ice (YI), first-year ice (FYI), and multiyear ice (MYI). The input parameters are listed in Table 1 and explained in Section 2.3.
Most methods that retrieve the concentration of sea ice or sea ice types from radiometer or radar data use representative 100 values of each input parameter for each surface type (ice or ice types, open water) -these representative values are known as tie points. In contrast, ECICE uses the statistical distribution of all possible values of each input parameter for each surface type.
Such distributions are obtained by sampling data of the given input parameter from the given surface obtained under different meteorological, dynamic and freezing conditions. The distributions have been established for the application of Arctic sea ice (Shokr et al., 2008;Ye et al., 2016a) and re-established here for the Antarctic application (see details in Section 2.3). With the 105 distributions, which can be interpreted as probability densities, a number n of possible realisations of tie points for all surfaces is selected using a random number generator. Here, we use 1000 realisations. For any given observation in remote sensing data, the observation is considered to represent a linear mixture of typical values (tie points) for of each surface type, weighted by the area fraction of that surface type in the observation cell. Therefore, a set of linear equations (equal to the number of observations) in which each equation represents decompositions of each input observation into its components from the given 110 surface types is then constructed. The equations are solved simultaneously for the area fraction of each surface type under the constraint that all fractions must add up to one (equality constraint) and that each area fraction must be between 0 and 1 (inequality constraint). This formulates the problem into an inequality constrained optimisation problem, which is solved to find the ice concentrations that minimise a given coast function. Then, the median of the n solutions for the area fractions (i.e., concentrations) is produced from which the final answer is generated (a set of concentrations of the given ice types). In 115 addition, the spread of the n solutions around the median is used as a measure of confidence of the result for each surface type (Shokr et al., 2008).

Correction schemes
The determination of the ice type concentrations (i.e., the area fractions of the three ice types) is only based on their radiometric and backscattering properties. During atmospheric warm spells, commonly occurring in the transition seasons, snow 120 wetness develops. The return of cold temperatures may cause snow metamorphism. Both effects cause anomalous microwave observations that make MYI appear as FYI and vice versa. Another process that causes anomalous observations is ice surface deformation at floe edges (e.g., pancake ice), which makes brightness temperature and backscattering from FYI look similar to those from MYI. Such errors are reduced by the two corrections schemes described in the following sections.

125
Warm air advection can occur during fall and spring seasons. When air temperature rises to near melting or beyond melting conditions, snow wetness develops and MYI will have lower backscatter and higher brightness temperature, typical of FYI.
Therefore, it is misclassified by ECICE as FYI (more precisely: the retrieved MYI concentration is too low and the FYI concentration too high). After the end of the warm event, which typically takes between one to a few days, the correct classification is resumed. In order to account for this error, the so-called temperature correction scheme (Ye et al., 2016a) examines the 2-m 130 air temperature (T2m) data to identify warm episodes of up to N days. Each episode starts with temperatures rising above a threshold T 1 (near freezing temperature) and ends with temperatures falling below a threshold T 2 . If the MYI concentration drops at any location during a warm spell by more than a specified threshold ∆C t and later rises again, such MYI concentrations are replaced by values linearly interpolated from before and after the warm episode. The values of the parameters N , T 1 , T 2 and ∆C t used here are specified in Section 2.3 (Table 2).

Drift Correction
Since the warm temperatures in the spring may progress for the rest of the season, the above temperature correction may not hold because it depends on the returning to the normal cold winter temperatures. For this reason, another correction was developed to identify locations of estimated MYI which are not realistic. This correction is based on the fact that MYI starts by definition as all remaining ice at the end of the melting season, i.e., at the onset of freeze-up. After that, MYI can only drift 140 and its concentration can only be changed by divergence, convergence, and melting, but it cannot be generated during the cold season. Therefore, MYI is unrealistic if it appears at locations to which it cannot have drifted. To identify such locations, daily ice drift data are used to implement what is called "drift correction" (Ye et al., 2016b). This correction starts with defining the boundary of MYI cover from the map of a given day. The boundary is then adjusted according to the ice drift map of the same day to predict its contour in the next day. This is done by applying ice motion vectors (obtained from the sources 145 mentioned below) to all pixels inside the boundary. This domain is then further extended by one grid cell to the outside in order to account for uncertainty of the ice drift product. Any MYI that has been retrieved outside of this domain cannot be multiyear ice. Therefore, all non-zero MYI concentrations in grid cells outside the mentioned domain are set to zero. However, since this spurious/erroneous MYI is in fact FYI or YI that has anomalous radiometric and scattering properties (so that ECICE has classified it as MYI), it is kept as a new pseudo-ice-type, which we call Ex-MYI. In addition, this correction scheme includes 150 correction for some effects of snow metamorphosis for pixels inside the MYI contour of the given day which would make FYI radiometrically similar to MYI. It looks for sudden (within one day) rises ∆C d of MYI concentration concurrent with sudden reductions of ∆T 37 of T B,37V or reductions ∆T 19−37 of T B,19H − T B,37H . The latter difference is also called horizontal range, HR, and is used by Drobot and Anderson (2001) to identify the onset of snow melt. The use of this parameter in the drift Table 1. Input parameters ("channels") used in Antarctic sea ice type retrieval with ECICE. Note: TB is brightness temperature, GR is gradient ratio (see Equation (1)), σ • is normalised radar backscattering cross section. correction is explained in detail in Ye et al. (2016b). In cases of such anomalies, the MYI concentration at the given pixel is 155 replaced by the value of the previous day. The values of the parameters ∆C d , ∆T 37 , and ∆T 19−37 used here are specified in Section 2.3 (Table 3).

Adapting ECICE algorithm to the Antarctic sea ice
The Antarctic implementation of ECICE at the Institute of Environmental Physics (IUP) at the University of Bremen uses as input microwave radiometer data of the sensors AMSR-E (Advanced Microwave Scanning Radiometer for EOS) on the NASA  Table 1. The gradient ratio at 19 and 37 GHz, V polarisation is defined as

165
For all input parameters, we use daily gridded data, projected on a polar stereographic grid with a nominal resolution of 12.5 km. This is the common grid used by the National Snow and Ice Data Center, NSIDC (see more details in Melsheimer and Spreen, 2019). The grid spacing is close to the resolution and the sampling interval of the original swath data of the used AMSR2 channels.
As Antarctic sea ice is different from Arctic sea ice (see Section 1), the probability distributions of the input parameters 170 needed by ECICE for the Antarctic application had to be derived.  Markus and Burns, 1995) in the implementation by Kern et al. (2007), using combinations of the 36 GHz and 85 GHz, H and V polarisation channels of the Special Sensor Microwave/Imager (SSM/I). The data contain a surface class    Validation of sea ice type concentration is a challenging task not only because there are few in-situ data sets of ice types 205 available for the Antarctic region but also because point observations do not represent the large footprints of the satellite data.
Validation studies using data from research cruises in the Antarctic have just started recently.
In this study we have compared the sea ice type concentration output from ECICE against three related satellite-based data sets: The new Antarctic uncorrected FYI and YI data and the corrected MYI data have been compared to data from (1) synthetic aperture radar (SAR) images, (2) charts showing the stage of development (SoD) of the ice, (3) maps of thin ice from 210 the polynya data set mentioned above. Examples are presented in the following.

Synthetic Aperture Radar (SAR) Images
The MYI concentration from ECICE has been checked against high resolution SAR images acquired by Sentinel-1A/B at 40 m grid resolution. In SAR images, MYI usually looks considerably brighter than FYI and YI mainly because of the high backscatter triggered by volume scattering from the bubbly sub-surface layer. We have examined a few cases in the Weddell

215
Sea and near the Antarctic Peninsula. An example is presented in Figure 2, showing an area in the Weddell Sea on 11 Aug, 2016. The Sentinel-1 SAR image (grey shade) is overlaid on the MYI concentration (colour scale). In addition, the dominant ice type is indicated by the coloured contours; white: YI=50%; purple: FYI=70%; black: MYI=70%.
The black and purple contours, which mark the transition from MYI to FYI, coincide well with a clear boundary between bright and dark radar backscatter which marks the boundary between dominating MYI and dominating FYI. Note also that  ("old ice") brown (see also colour legend in appendix, Table A1). The triangle symbols denote large icebergs.
July, 2017; namely (top to bottom), SoD, corrected MYI, YI and FYI. The corrected MYI concentration (second row) has a 250 sharp inner (Southern) boundary that agrees well with the boundary between MYI (brown) and YI (purple) or FYI (green and yellow) classes in the SoD charts (top row). The MYI and its inner boundary move towards North and North-East from March to July (left to right). Note the similarity to the FYI-MYI boundary in the previous Section 3.1 (Figure 2), and the two strips of MYI separated by FYI in the right part of the rightmost two SoD charts: The MYI strips correspond to similarly shaped areas with MYI concentration above about 30% to 40% in the MYIc maps (second row). The FYI areas in the SoD maps, in 255 turn, correspond rather to FYI concentrations (forth row) above about 60% to 70%. This hints at the difficulty of comparing a sea ice classification where each point is assigned exactly one sea ice type class with a sea ice type fraction retrieval where in each grid cell, more than one surface type can coexist with the summation of their fractions equal to 100%. The YI and FYI concentration in Figure 5 (third and fourth row) also match reasonably well with the SoD charts, noting that here as well, areas with YI concentration above about 30% correspond to the class YI in the SoD charts, and only FYI concentrations above 260 about 70% correspond to the FYI classes in the SoD charts. This is best visible on the first two maps from the left. Another noteworthy detail are the small patches of YI apparently in the lee of grounded iceberg A23A (third and forth column). Large icebergs are often erroneously retrieved as FYI, but this does in principle not cause problems as the position and extent of such icebergs is usually well known and monitored, so they can be masked out using ancillary iceberg data (e.g., https://usicecenter.gov/Products/AntarcIcebergs).

Polynya Data Set PSSM
The total area of MYI in the entire Antarctic should not increase during one cold season because MYI originates as the remaining ice at the end of the melting season and hence cannot be generated after freeze-up. However, the total MYI area 285 derived form our data shows large fluctuations, and often an increase around July. This can be seen in Figure 10 which shows the total MYI area for all Antarctic seas for the cold season 2018, smoothed with a 21 day running mean.
The main reason for the increase in July seems to be a large offshore area of MYI in the outer Ross sea and off Wilkes Land, East of the Ross Sea (roughly between 160 • E and 140 • W, and between 65 • and 70 • S) that often seems to grow during the cold season and is not eliminated by the drift correction. Figure 11 shows such an area of MYI that appears in late July, 2018, 290 and disappears in September. While, according to the NIC/AARI ice type (SoD) charts, a considerable area of MYI persists far offshore in the outer Ross Sea in many years (though not in the year 2018 shown here), the retrieved offshore MYI areas are larger, in different locations, and sometimes grow quickly within days, which cannot be correct and is not shown on the SoD charts. As MYI cannot be generated during the freezing season, this is clearly spurious identification of MYI. This area of spurious MYI is in the marginal ice zone. A similar phenomenon can be observed at around the same time in that region in 295 all years of the current data record 2013 to 2020 though it is less pronounced in 2019. The most likely reason for this is that in the course of the cold season, the snow layer on FYI in particular in the marginal ice zone changes. Usually, snow backscatter increases and emissivity decreases, making it resemble MYI in that respect. In addition, pancake ice has higher backscatter than FYI and might also be mistaken for MYI (see also Willmes et al., 2011;Arndt et al., 2016;Arndt and Haas, 2019).
Elsewhere in the Antarctic seas, such areas of spurious MYI ice can also occur but are generally much smaller. In principle, (within one day's drift) will then not be removed; such seeding points can be caused by "land contamination" of the satellite measurement in footprints directly at the coast or near small islands that are not properly masked. Since removing coastal MYI pixels and extending the land mask before applying the correction had no effect on the spurious MYI pixels, the latter reason can be ruled out. Thus, the reason is most likely inaccuracies in the drift data which accumulate over the season (strange drift vectors have been observed in particular in the Ross sea near the date line; Ted Maksym, priv. comm., 2020).

310
Instead of trying to correct erroneous MYI, preventing the misclassification that leads to its retrieval in the first place is worth while. To do this, the distributions of MYI and FYI in the different channels, specifically in the outer Ross Sea and off Wilkes Land, could be investigated. If these distributions differ significantly from those currently used (see Section 2.3), the latter can be adapted.
Towards the end of the freezing season, in September and October, the retrieved, corrected MYI concentration declines 315 strongly in most years (see Figure 10), which is not seen in the weekly SoD charts. However, SoD charts seem to become unstable in the sense that the ice type seems to oscillate between FYI (WMO type 2.5 "first-year ice" (JCOMM Expert Team on Sea Ice, 2015)) and MYI (WMO type 2.6 "old ice") from one week to the next, or from AARI chart to the almost simultaneous NIC chart. The most likely reason for the too strong decline of MYI in our retrieval scheme are temperatures rising to near melting conditions which causes MYI to be misclassified as FYI as described above in Section 2.2.1 about the temperature cor-320 rection. Since these near-melting or melting conditions are not episodic any more, they cannot be corrected by the temperature correction.
Note that a first analysis of the new OSI-SAF ice type classification data likewise does not show the expected decrease of Antarctic MYI area in the course of the season, but instead a steady slow increase followed by a rapid decline toward the end of the freezing season in September/October (Aaboe et al., 2021b, Fig. 14), which is similar to the time series derived from our 325 data ( Figure 10).

Summary and Conclusions
The sea ice type retrieval method ECICE (Shokr et al., 2008;Shokr and Agnew, 2013) and the subsequent correction schemes for MYI (Ye et al., 2016a, b) developed for the Arctic can be adapted for the Antarctic, given samples of Antarctic ice types.
Input satellite data are microwave radiances at several channels as well as scatterometer backscattering measurements. Daily 330 maps of uncorrected YI, FYI and MYI, and of MYI corrected for effects of melt-refreeze and snow metamorphosis, can be retrieved, outside the melt season, at spatial resolution of 12.5 km. The results look reasonable in the sense that they show agreement with SAR images, with remote-sensing-based polynya data, and with weekly charted sea ice stage of development (so far the only source of ice type information in the Antarctic). In particular, the general distribution of the Antarctic MYI at the beginning of the freezing season is well captured. The subsequent time evolution of the MYI concentration in the Weddell 335 Sea, as far as the AARI/NIC stage of the development charts can tell, is captured as well in the months after freeze-up, showing the effects of advection and melt. The retrieved distributions of YI and FYI concentration are reasonable as well. However, comparing our ice type concentrations, i.e., area fractions of ice types, with weekly charts that assign only one ice type to each location is inherently problematic. A more detailed validation study that also includes in-situ observations from research cruises is planned. The most problematic area is the outer Ross Sea and the sea off Wilkes Land, where large and growing areas