Cross-platform application of a sea ice classification method considering incident angle dependency of backscatter intensity and its use in separating level and deformed ice

Wide-swath C-band synthetic aperture radar (SAR) has been used for sea ice classification and estimates of sea ice drift and deformation since it first became widely available in the 1990s. Here, we examine the potential to distinguish surface features created by sea ice deformation using ice type classification of SAR data. To perform this task with extended spatial and temporal coverage, we investigate the cross-platform transferability between training sets derived from Sentinel-1 Extra 5 Wide (S1 EW) and RADARSAT-2 (RS2) ScanSAR Wide A (SCWA) and Fine Quad-polarimetric (FQ) data, as the same radiometrically calibrated backscatter coefficients are expected from these two C-band SAR platforms. For this, we use a novel sea ice classification method developed based on Arctic-wide S1 EW training, which considers the ice-type-dependent change of SAR backscatter intensity with incident angle (IA). This study focuses on the region near Fram Strait north of Svalbard to utilize expert knowledge of ice conditions from co-authors who participated in the Norwegian young sea ICE (N-ICE2015) 10 expedition in the region. Separate training sets for S1 EW, RS2 SCWA and RS2 FQ data are derived using manually drawn polygons of different ice types, and are used to re-train the classifier. Results show that although the best classification accuracy is achieved for each dataset using its own training, different training sets yield similar results and IA slopes, with the exception of leads with calm open water, nilas or newly formed ice (the ’leads’ class). This is found to be caused by different noise floor configurations of S1 and RS2 data, which lead to different IA slopes of this class. This indicates that dataset-specific re-training 15 is needed for leads in the cross-platform application of the classifier. Based on the classifier thus re-trained for each dataset, the classification scheme is altered to target the separation of level and deformed ice, which enables direct comparison with independently derived sea ice deformation maps. The comparisons show that the classification of C-band SAR can be used to distinguish areas of ice divergence occupied by leads, young ice and level first-year ice (LFYI). However, it has limited capacity in delineating areas of ice deformation due to ambiguities in ice types represented by classes with higher backscatter intensities. 20 This study provides reference to future studies seeking cross-platform application of training sets so they are fully utilized, and we expect further development of the classifier and the inclusion of other SAR datasets to enable image classification-based ice deformation detection using only satellite SAR data. 1 https://doi.org/10.5194/tc-2021-119 Preprint. Discussion started: 26 April 2021 c © Author(s) 2021. CC BY 4.0 License.


Introduction
The general thinning of Arctic sea ice in recent decades has led to reduced internal strength (Landrum and Holland, 2020), 25 which together with increased wind forcing (as indicated by atmospheric reanalyses) has caused accelerated ice drift speed (Spreen et al., 2011) and hence increased ice deformation (Rampal et al., 2009(Rampal et al., , 2011Itkin et al., 2017). As surface features created by ice deformation, e.g. lead edges, rafted ice and pressure ridges, are the primary snow-trapping sea ice surface types (Liston et al., 2018), the shifting regime of Arctic sea ice deformation will directly impact snow accumulation on ice and thus affect heat fluxes through the ice layer, thereby influencing winter sea ice growth (Sturm et al., 2002). Also, sea ice deformation 30 influences sea ice surface and bottom roughness and thus affects the transfer of momentum between atmosphere and ocean (Cole et al., 2017;Martin et al., 2016), preconditions the ice layer for more lateral melt (Arntsen et al., 2015;Hwang et al., 2017;Graham et al., 2019), and increases ice drift speed due to reduced floe sizes following ice break-ups (Toyota et al., 2006;Steer et al., 2008;Asplin et al., 2012). Additionally, ice deformation has a significant impact on ice primary productivity as it provides sheltered growth environment for ice flora and fauna in deformed ice (Gradinger et al., 2010; Fernández-Méndez reduces radar penetration depth and result in dominant backscatter from the air-snow or dry-and-wet-snow interface (e.g . Gill 125 et al., 2015). Data covering the N-ICE2015 campaign is used as the primary dataset. SAR scenes with optical imagery overlap from 2016 to 2019 are also used to expand the dataset for re-training and validation. In-situ observations show that the sea ice area investigated during the N-ICE2015 campaign primarily consisted of a mixture of FYI and second-year ice (SYI, which belongs to the MYI category for the purpose of SAR-based sea ice classification), while other thinner ice types including nilas and young ice in leads also existed. Frost flower coverage of young ice was observed for the entire drift period of the campaign 130 Johansson et al., 2017;Granskog et al., 2017Granskog et al., , 2018Singha et al., 2020). The terminology of sea ice classes used for classification in this study is defined in detail in Section 2.3.
Specifically, the following selection and masking processes are conducted on RS2 scenes: 1. The original GIA classifier has limited separating capacity between open water and sea ice, as IA slopes and backscatter intensities of open water in different sea states vary significantly, creating ambiguities with ice surfaces (Lohse et al., 2020). 135 To mitigate this known issue and also to serve our purpose of separating level and deformed ice within pack ice, daily sea ice concentration data generated from SSMI/S (Special Sensor Microwave Imager/Sounder, DMSP F18 satellite) is used to filter out areas in RS2 SCWA scenes with ice concentration values of lower than 87%. This is an empirically derived value for the removal of large, contiguous open water areas and the marginal ice zone. 2. The RS2 scenes are further selected based on the availability of overlapping S1 data and optical imagery. For 2015, RS2 scenes with at least one overlapping S1 scene are kept for analysis. From 2016 to 2019, RS2 scenes with at least one S1 and one optical (S2 or L8) scene with significant near-coincident overlap (overlapping area ≥ 30% of the masked RS2 SCWA scene) are selected to ensure optical reference, resulting in only 2 RS2 SCWA scenes in 2019 (Table 1). The maximum time difference between overlapping RS2 data and S1, S2 and L8 data are 1 hour, 5.3 hours and 8 hours, respectively. These selection parameters are empirically determined judging from the data availability of each dataset corresponding to the RS2 scenes. The 145 overlap analysis is conducted in Google Earth Engine (Gorelick et al., 2017), from which overlapping S1 scene IDs are derived and used for downloading using the Sentinelsat Python API, while RGB composites of S2 and L8 data (both from bands 4, 3 and 2) are directly generated and downloaded.
Pre-processing of RS2 and S1 data is performed using the SNAP software package (European Space Agency, 2020). All scenes are radiometrically corrected and calibrated to σ 0 values, so that RS2 and S1 data are directly comparable. For RS2

150
FQ scenes which are single look, 2×2 multi-looking is performed to reduce speckle and reach similar number of effective looks compared to S1 and RS2 SCWA scenes, while considering the preservation of linear features of interest, especially leads, within the scenes. Then, speckle filtering (Boxcar, 3×3) is applied on all SAR scenes, and backscatter intensities are converted to dB.
Typically several S1 scenes overlap with each RS SCWA scene, but only one is selected to avoid redundancy. The first sub-155 swath of each S1 scene is removed for more reliable and inter-comparable classification results, as radiometric variations are especially pronounced between the first sub-swath and others (e.g. Park et al., 2019). The same masking process to erase pixels with low ice concentration is also conducted on S1 scenes. No further processing for S2 and L8 RGB composites is conducted, as these product are used qualitatively as reference data. The final list of satellite data used in this study is summarized in Table 1 (Gatti and Bertolini, 2015;Northrop, 2015;MacDonald Dettwiler Assoc. Ltd. (MDA), 2016, 2018, and the image 160 boundaries of RS2 SCWA and S1 scenes are shown in Figure 2. Subsequent data analyses are performed using MATLAB (The Mathworks Inc., 2020) and Python (Van Rossum and Drake, 2009).

Cross-platform application of the GIA classifier
The original GIA classifier aims for Arctic-wide applicability, and thus has been trained on S1 scenes spread across the entire Arctic region from 2015 to 2019, in 'winter and early spring' months when ice is under freezing conditions (Lohse et al., 165 2020). Therefore, the class-specific IA dependencies in the classifier is produced from a generalized training set, and in theory encompass all IA dependencies of these classes found within this spatial and temporal domain. This study focuses on the transferability of training between S1 and RS2 data, for which we do not target Arctic-wide generalization. Instead, we focus on the N-ICE2015 region in winter 2015 to provide confidence to the validity of training and validation with input from expert knowledge, as mentioned above. Thus, we investigate the applicability of local training sets separately derived from S1 and 170 RS2 scenes to both platforms, through which the between-sensor difference of sea ice backscatter can be assessed, and an optimal way of applying the GIA classifier on RS2 data can be derived. For classes usually with small spatial coverage (leads and young ice) where criteria 2 and 3 cannot be both satisfied (i.e. small spatial coverage leading to inevitable selection of polygons in small areas), preference is given to criterion 3 to yield 190 more polygons. Examples of reference polygons for S1 and RS2 SCWA scenes are shown in Figure 3. Reference polygons in especially essential for RS2 FQ scenes where their small extents (spatial coverage: 25 km by 25 km; IA range: 1.24 • to 1.94 • for our scenes) necessitate the combined investigation of IA dependencies across multiple scenes.
To evaluate cross-dataset transferability of training, the original GIA classifier is firstly applied to all datasets, providing a baseline for further analyses. Secondly, the training sets for S1, RS2 SCWA and RS2 FQ scenes are used to re-train the classifier to the study area for their corresponding datasets. The classifier with these different training sets are then applied to 200 each dataset, and the results are evaluated using the validation polygons. Also, IA dependencies of different ice types as seen from validation polygons and classification results are examined. Based on these, the transferability of training between the datasets is assessed. Finally, based on this assessment, an optimal approach of applying the GIA classifier on RS2 data will be determined, supporting subsequent classification of ice types and also comparison to ice deformation.

Adaptation to separate level and deformed ice 205
Based on results from the above section, ice type classification is conducted on S1 and RS2 data in the study area. To separate level ice from areas of ice deformation and allow for direct comparison with ice deformation maps (see Section 2.4), the 5class classification scheme of the GIA classifier is altered to a 3-class one: deformed ice, level ice and others. This is hereafter referred to as the classification of 'deformation states,' and the correspondence between classes in the two schemes is listed as follows: 210 1. DFYI and MYI: deformed ice. The 5-class scheme involves both ice age and deformation state. An ideal classification of level and deformed ice requires separation between deformation states in every ice age category (mainly: young ice, FYI, and MYI). However, the GIA classifier does this only for FYI, such is common practice of SAR-based sea ice classification (e.g. see review by Zakhvatkina et al. (2019)). Specifically, the DFYI class refers to rough FYI with stronger C-band backscatter due to either ridging or the presence of other rough surface features, e.g. pancake ice.

215
MYI surface is usually rougher than younger ice types due to more accumulated deformation. The separation of MYI deformation states using SAR-based classification is challenging due to stronger presence of volume scattering from the desalinated and porous upper ice layer (increasing backscatter from level ice) as well as weathering of deformation features (decreasing backscatter from deformed ice) (Dierking and Dall, 2007;Dierking, 2008;Casey et al., 2014). As this study uses only C-band HH and HV intensities, this volume scattering component in MYI is significant, and the capacity of fully polarimetric data to 220 distinguish between surface and volume scattering is not available (Moen et al., 2013;Casey et al., 2014). For the same reason, the contribution from volume scattering and ice deformation to strong SAR backscatter (characteristic of both DFYI and MYI) cannot be perfectly separated, creating ambiguities also between DFYI and MYI. Thus, although MYI is not necessarily physically associated with ice deformation, this study labels MYI together with DFYI as 'deformed ice.' 2. Young ice and LFYI: others. Among the many forms and stages of growth, the young ice class defined in the GIA 225 classifier correspond to rough (in small-scale roughness) young ice covered by frost-flowers (mostly in re-frozen leads), thus having similar backscatter intensities to MYI. It has been demonstrated that typical deformation in young ice, i.e. ice rafting, is difficult to distinguish using C-band SAR (Dierking, 2010), and therefore cannot be included in the analysis. The experience from co-authors who participated in the N-ICE2015 campaign also shows that young ice in the study area experienced little rafting and were predominantly level with frost flower coverage, while ridging primarily occurred where young ice was in 230 contact with thicker ice. Example photographs of young ice from the campaign are shown in Figure 3. Level young ice is not part of the 5-class scheme, as the LFYI class does not exclude level ice less than 30cm thick due to similar backscatter intensities (Dierking, 2008(Dierking, , 2010. As we are interested in ice deformation occurring on thicker ice, i.e. FYI and MYI, young ice is not labeled as deformed ice in the 3-class scheme, and is instead labeled 'others' along with LFYI.
3. Leads: leads. The leads class in the GIA classifier corresponds to ice openings occupied by calm open water, nilas or 235 newly formed ice, thus having the lowest backscatter intensities. This class is of direct interest in the second goal of this study, and is labeled 'leads' in the 3-class scheme.

Deformation parcel tracking
Six pairs of S1 scenes from 21 to 26 January, 2015 (one image pair per day) surrounding the N-ICE2015 campaign is used to construct ice deformation histories for drifting ice parcels based on the sea ice drift algorithm developed by Korosov and 240 Rampal (2017). Sea ice deformation is calculated from line integrals as described in e.g. Buollion and Rampal (2015) and Itkin et al (2017) and further filtered for noise. Sea ice drift is calculated on a 400 m-regularly-spaced grid and deformation on the corresponding triangular grid. The deformation is classified to: no deformation, divergence and convergence. The rectangular parcels are initiated on the first day of the sequence on a regular grid with centroids spaced by 300 m and with a size 120 by 120 km. Initially, all parcels are undeformed. For each next day the parcels move with the average velocity of the drift 245 calculated inside of the 300 m radius around each parcel centroid. Each ice parcel accumulates at every step counts of each deformation class, based on the average value inside the 150 m radius from the centroid. Finally, based on the total number of counts, every parcel is classified into undeformed, predominantly convergence, predominantly divergence or a mix of both.
Such Lagrangian parcel product is then gridded onto a 100 m grid, based on the nearest neighbor value. In the six days that the parcels are tracked, the ice pack undergoes vigorous deformation limited to several active linear kinematic features (LKFs).

250
The effects of this recent deformation can then be visually compared with the classification results, thus examining the ability of the classification in recognizing the most recently formed surface features created by ice deformation.  Figure 4). This corresponds well with our expectation of similar sea ice backscatter (σ 0 ) for the two C-band sensors, while suggesting dataset-specific training can be used to achieve optimal overall accuracies. Examples of qualitative comparison between results using the GIA classifier with different training are shown in Figure 5.
It can be seen that the original GIA classifier yields classification maps dominated by DFYI (green) and MYI (blue), while 265 visual inspection of the SAR RGB composites ( Figure 5 (a1-e1)) indicates significantly more prominent existence of the LFYI, rough young ice and leads. This is caused by frequent mis-classification of young ice as MYI and leads as LFYI. The three under-represented ice types (LFYI, rough young ice and leads) are better recognized by the locally re-trained GIA classifier e5)) yields results with similar spatial distribution of different ice types, and varies on the relative amount of pixels between classes.

IA dependencies
Theoretically, the performance of cross-dataset application of training is driven by the different IA dependencies of the ice types they record. To investigate this, scatter plots of HH intensities and IAs for different ice types are derived using the 275 validations polygons (Figure 6). NESZ values across the IAs are generated from associated product files with the SAR data and plotted for comparison. Corresponding least-squares linear regression lines for different ice types are also plotted. HH-IA slopes derived from validation polygons as well as the original and re-trained (using corresponding training sets) GIA classifier are summarized in Table 2, with slope values from previous studies using winter C-band satellite SAR data listed for reference.
For all datasets, the HV signals show less dependency on IA and is much more affect by noise, but its inclusion in the classifier

Leads and noise floors
Between the two wide-swath datasets (S1 EW and RS2 SCWA), the IA slopes for young ice, FYI and MYI follow similar trends ( Figure 6, columns (a) and (b)). However, the IA slope for the leads class in the RS2 SCWA training is visibly flatter than that in the S1 training, which is confirmed by their respective slope values shown in Table 2 (-0.12 dB/ • for S1 and -0.065 dB/ • for RS2 SCWA). It can be seen from the plotted NESZs that the leads class for S1 scenes is above the noise floor throughout the 310 IA range (Figure 6(b1)), while for RS2 SCWA scenes (with a flatter noise floor), it is very close to and reaches the noise floor at an approximate IA of 30 • (Figure 6(b2)). This explains the flatter IA slope for leads in RS2 SCWA scenes.
For the RS2 FQ scenes, IA slopes for the young ice, FYI and MYI (Figure 6(c1)) are similar to the wide-swath datasets.
The RS2 FQ scenes are all well within pack ice, and the leads that these scenes additionally recognize due to higher spatial resolution (compared to wide-swath scenes) are very narrow and scarce. Therefore, the multi-looking and speckle filtering 315 processes has mixed the pixel values in narrower leads to surrounding pixels with higher backscatter intensities, resulting in small numbers of reference polygons. This has prevented full evaluation of its IA dependency due to uneven representation across IAs, and compromises the validity of IA dependencies recorded by this class. However, the existing reference polygons show that HH intensities of RS2 FQ leads pixels do not reach the noise floor across the IA range, and therefore does not present the above issue in the RS2 SCWA training. To further investigate the interaction between the leads class and respective noise floors of the sensors, IA slopes for leads from the near range to increasing IA maximums are plotted in Figure 7. F1 scores of the leads class, which combines producer's and user's accuracies, are also shown to evaluate the overall accuracies of this class. It can be seen that S1 validation polygons show relatively constant HH-IA slopes across the range of IA maximums (in black). For RS2 SCWA scenes (red), the slopes are steeper (higher absolute values) for IA maximums in the near range, and stabilizes after reaching a similar slope level to 325 S1 scenes (at approximately -0.16 to -0.12 dB/ • , at IA maximums from 38 • to 44 • , as shown between the dashed lines). The slopes then quickly flattens (lower absolute values) as the IA maximum approaches the far range, eventually reaching a much flatter overall slope of -0.065 dB/ • compared to -0.120 dB/ • for S1 scenes. This confirms the findings from Figure 6(a2) where the leads pixels reach the noise floor at approximately 30 • , and remains at at a similar dB level due to the presence of the noise floor. Thus, using the S1 training on RS2 SCWA scenes will inevitably introduce IA dependency for leads that does not fit the 330 RS2 SCWA dataset, and vice versa.
For the classification results (Figure 7(b-d)), the training used by the original GIA classifier yields different overall IA slopes than the local validation for leads in both datasets, resulting in relatively low classification accuracies, as shown by the F1 scores (Figure 7(b), F 1 leads )). When applied to their corresponding datasets, the local S1 and RS2 SCWA training sets  To inspect this effect in classification maps, an example is given in Figure 8 which shows the difference between the GIA classifier with different training in recognizing leads in different IA ranges. For the near range, the GIA classifier with local re-training (Figure 8(c1) and (d1)) yields very similar spatial coverage of leads. In IAs between 34 • and 39 • , classification using 340 the RS2 SCWA training (Figure 8(d2)) produces a more complete representation of the leads than the S1 training (Figure 8(c2)) where parts of the leads are identified as LFYI. In IAs between 40 • and 45 • , the RS2 SCWA training preserves all visible leads ( Figure 8(d3)) while the S1 training keeps only part of the main ice opening (Figure 8(c3), circled in black) but almost entirely misses the other leads. This gradual increase in mis-classification of leads as LFYI with IA corresponds well with Figure 6 (column (b)): the stronger HH-IA dependency for leads in S1 training (steeper IA slopes than RS2 SCWA training) yields the 345 same leads-LFYI separation in the near range, but shows gradually more mis-classification of leads to LFYI (Figure 6(b3) compared to (b4)) in IAs greater than approximately 37 • . The same is true for RS2 SCWA training when used to classify S1 scenes ( Figure 6(a4) compared to (a3)), where its flatter HH-IA slope leads to mis-classification of LFYI to leads in the far range.
From the above analyses of CAs, HH-IA dependencies and qualitative comparisons, it can be concluded that in our study 350 area and based on the GIA classifier, S1 and RS2 SCWA training sets are not directly transferable, mainly due to the known potential of mis-classification between LFYI and leads. This is caused by the different interactions between backscatter from leads and the noise floor in the two datasets, i.e. the flattened IA slope of leads in RS2 SCWA data due to contact with the higher noise floor. The RS2 FQ scenes yield similar IA slopes for classes other than leads compared to the wide-swath datasets, while full assessment of the leads class is impeded by the lack of reference polygons. Re-training to the study area also increases 355 performance of the GIA classifier when applied to S1 scenes. Based on these assessments, the S1, RS2 SCWA and FQ scenes are classified using the GIA classifier locally re-trained using their corresponding training sets, and are used for the following comparison to ice deformation.

Comparison between classification results and deformation parcels
The five ice types in the classification results are summarized into three deformation states, as mentioned above, and are 360 compared with the deformation parcels derived from S1 data ( Figure 9)  Graham et al., 2019). The storm first pushed ice northward and thus compacting the ice pack, causing ice deformation along re-frozen leads and cracking thicker ice floes. It then transported ice southward towards the ice 370 edge in the second phase, generating strong divergence and opening along the same leads and cracks. Following the storm passage, newly opened leads rapidly re-froze following the returning of dry and cold conditions, creating new ice Graham et al., 2019). Accordingly, the parcels (Figure 9, row (b)) indicate major presence of new and deformed ice concentrated along several LKFs. Divergence zones with new lead ice prevail, but are mixed with convergence zones where It is difficult to interpret direct correspondence between the deformation parcels and the classification maps, as areas of ice divergence and convergence do not directly correspond to specific ice types, and the deformation parcel maps only represent ice motion accumulated in the 6-day period. Still, observations can be made for: 1. whether the classified ice types correctly correspond to ice divergence or convergence indicated by the deformation parcels, and 2. if the classification maps and deformation 380 parcels each identifies visible deformation features not shown by the other.
1. Classification results. An overall examination of classification maps of deformation states (Figure 9(d1-d3)) shows that 'deformed ice' is pervasive within the scenes. The same can be concluded from visual examination of the scenes (Figure 9(a)) and is reported by N-ICE2015 records and observations Granskog et al., 2018;Graham et al., 2019).
However, the true percentage of deformation in the scene cannot be retrieved or confirmed using available data due to the 385 inclusion of ice surface with other rough surface features in the DFYI class, as well as ambiguity between DFYI and MYI, as mentioned above. The deformed ice class (three-class scheme) is comprised of mainly DFYI (5-class scheme, Figure 9(c1-c3)).
MYI also contributes significantly, capturing many bright deformation features, most prominently a linear feature of MYI on top of the RS2 SCWA and S1 scenes, marked by dotted ellipses (Figure 9(a1), (a3) and (c1), (c3)). In the six-day period, this old refrozen lead (likely covered by frost flowers) did not undergone any detectable convergence or divergence, and is therefore 390 not represented in the deformation parcels.
On the other hand, many LKFs in the 'others' class (three-class scheme) are visible in the classification maps of deformation states (white in Figure 9(d1-d3)), the primary ones marked by dashed lines and numbered in Figure 9(e1-e3). These are comprised of mostly LFYI in re-frozen leads (five-class scheme, Figure 9(c1-c3)), which can physically correspond to level young ice or FYI, as mentioned above. Most of these level ice areas also appear in the deformation parcel maps (Figure 9(b1-395 b3)), where they are numbered accordingly. Lines 5, 8 and 9 in Figure 9(e3) do not appear in the deformation parcel maps ( Figure 9(b3)) as they are out of the maps' areal coverage. Open water areas in line 1 in Figure 9(e1) and (e3) are correctly classified as leads (red in Figure 9(c1-d1) and (c3-d3)). The RS2 FQ scenes shows similar overall distribution of ice types as wide-swath scenes, but its higher spatial resolution more clearly picks up some more visible ice openings with more spatial details (lines 1, 3, 4 and 5 in Figure 9(e2)), compared to (e3)). 400 2. Deformation parcels. For the deformation parcels, the most prominent (widest) features tend to have more recognizable correspondence to features delineated by the classifications. These are either mostly ice divergence mixed with convergence, or exclusively divergence, in both time stamps. For example, the end states of lines 2 and 3 in Figure 9(b1) and (b3) are shown to be dominated by LFYI and young ice, surrounded by DFYI (Figure 9(c1 and (c3)), corresponding well to the co-authors' field experience of deformation occurrence at the interface between young ice and older ice, as mentioned above. It should be 405 noted that the pixels showing 'mostly convergence' are derived from values accumulated in the 6-day period, and therefore cannot represent deformation features which have been accumulated over longer periods.
The above mentioned areas indicating mainly ice divergence (Figure 9(b1) and (b3), solid ellipses, excluding the major lead, i.e. line 1), are less recognizable in the classification maps of deformation states. Ice type classifications indicate these are narrower and smaller leads occupied mostly by LFYI and young ice. One of these areas is more clearly shown in Figure 10 where the visible LKFs are delineated and numbered, in which lines 1-5 are identified by both the classification and deformation parcels. Figure 10(b1-b5) shows the classifications for areas where deformation parcels indicate prominent presence of ice convergence. The thinner linear stripes of ice convergence (lines 3, 4 and 5) do not correspond to visible features in the classifications, while lines 2, 6 and 7 with extensive coverage of blue and black colors (ice convergence and mix) are classified as mostly LFYI and young ice which are interrupted by DFYI sections.

415
To summarize, the classifications capture ice openings with the 'leads' and 'others' classes (leads, young ice and LFYI in the five-class scheme) that correspond well with areas of ice divergence. This good correspondence is also partly due to the surface features created by ice divergence being more spatially confined. On the other hand, the 'deformed ice' class includes a mix of DFYI and MYI that is spatially wide-spread, where the true proportions of deformed ice cannot be reliably verified, and hence has limited correspondence with areas of ice convergence from the deformation parcels. This is both due to the accumulation 420 of ice deformation in a period longer than the parcel tracking, and also to the limitation of the classifier working only with HH and HV channels of the C-band sensors. The RS2 FQ scenes performs similarly to the wide-swath scenes when the re-trained GIA classifier is applied, but can serve to preserve more spatial details of surface features. The capacity of the classification results to identify these surface features (mainly ice divergence) in the deformation parcels serves as another validation of the re-trained GIA classifier when applied to our study area.

Limitations and future steps
This study is a first step towards the goal of Arctic-wide ice deformation detection based on a consistent classification method applicable to multiple SAR platforms, and thus investigates the cross-platform application of the GIA classifier in a local setting. Thus, we work within the limitations of both the classifier and the characteristics of HH and HV channels of the Cband SAR sensors which affects complete derivation of ice deformation, as summarized above. Very limited ground truth of 430 ice types from in-situ data is available for re-training and validation, hence the heavy preference given to the N-ICE2015 data to utilize the co-authors' expert knowledge on ice conditions in the campaign.
This study expands the application of the GIA classifier from S1 to RS2 data, both in C-band. Additional studies will be conducted seeking further expansion of its application to more SAR platforms, e.g. X-and L-band SAR, which provides potential for better separation between the ambiguous class pairs in the current classification. IA dependency in SAR data with 435 these different frequencies needs to be rigorously examined and validated. It is expected that frequency-and-region-specific re-training will still be essential for deformation detection using the altered classifier for different sensors, as SAR intensity contrast between level and deformed ice is sensitive to SAR properties as well as ice properties that vary cross regions, e.g.
small-scale roughness and ice volume structure (Dierking and Dall, 2007). The inclusion of more features into the classification is also desirable, e.g. polarimetric features sensitive to sea ice deformation (e.g. Ressel et al., 2016;Park et al., 2019), and also 440 texture features (e.g. Park et al., 2020;Lohse et al., 2021). For example, the recent study by Lohse et al. (2021) investigated the IA dependencies of common texture features, and demonstrated that incorporating these features into ice type classification can improve the separation of young ice and MYI, as well as the generalized classification of open water areas. However, the improvement comes at the cost of reduced spatial resolution due to the applied texture windows. Further integration of IA dependency into classifiers other than the Bayesian classifier is also desirable in future studies to seek better classification 445 performances. Finally, successful cross-platform application of an optimal classification method can be used to create a reliable time series of classification maps which can be better used to derive and compare with ice deformation products.

Conclusions
This study demonstrates that S1 and RS2 data produce similar IA dependencies of different ice types. However, in our study area and based on the GIA classifier, the direct transfer of training between S1 and RS2 SCWA data is not applicable due to 450 the difference in noise floor configurations, which affects classification performance especially for the leads class. Datasetand region-specific re-training is found to be necessary, and the GIA classifier re-trained specific to S1, RS2 SCWA and RS2 FQ datasets produces similar and improved classification performances compared to the original classifier. The cross-platform application of the GIA classifier extends usable C-band SAR data over the study area from 2015 to present (S1) to 2010 to present (RS2). This study further provides reference to future cross-platform application of training between S1 and RS2 so 455 valuable training sets can be better utilized, e.g. with proper re-training or direct application when leads are not of interest.
The comparison between deformation parcels and classification results with dataset-specific local re-training shows the best correspondence in leads with open water and nilas, young ice or LFYI, as prominent ice openings created by divergence following the storm passage are in linear forms and well captured by both analyses. The DFYI and MYI classes in the classification results do not clearly correspond to linear ice convergence zones indicated by deformation parcels, both due to the limitation of 460 the classification method used and the difference in the period of deformation accumulation represented by both datasets. RS2 FQ scenes can be used to provide more spatial details in delineating deformation features. The comparison with deformation parcels further serves to partially validate the classification results.
In summary, through the cross-platform application of the GIA classifier, this study demonstrates the potentials and obstacles in the transfer of training between S1 and RS2 data, as well as in the use of the classification to separate level and deformed ice. 465 We expect future development of the classifier and the inclusion of additional datasets will enable the possibility of large-scale monitoring of ice deformation merely from the classification of widely available satellite SAR data.