TanDEM-X multiparametric data features in sea ice classification over the Baltic sea

ABSTRACT In this study, we assess the potential of X-band Interferometric Synthetic Aperture Radar imagery for automated classification of sea ice over the Baltic Sea. A bistatic SAR scene acquired by the TanDEM-X mission over the Bothnian Bay in March of 2012 was used in the analysis. Backscatter intensity, interferometric coherence magnitude, and interferometric phase have been used as informative features in several classification experiments. Various combinations of classification features were evaluated using Maximum likelihood (ML), Random Forests (RF) and Support Vector Machine (SVM) classifiers to achieve the best possible discrimination between open water and several sea ice types (undeformed ice, ridged ice, moderately deformed ice, brash ice, thick level ice, and new ice). Adding interferometric phase and coherence-magnitude to backscatter-intensity resulted in improved overall classification performance compared to using only backscatter-intensity. The RF algorithm appeared to be slightly superior to SVM and ML due to higher overall accuracies, however, at the expense of somewhat longer processing time. The best overall accuracy (OA) for three methodologies were achieved using combination of all tested features were 71.56, 72.93, and 72.91% for ML, RF and SVM classifiers, respectively. Compared to OAs of 62.28, 66.51, and 63.05% using only backscatter intensity, this indicates strong benefit of SAR interferometry in discriminating different types of sea ice. In contrast to several earlier studies, we were particularly able to successfully discriminate open water and new ice classes.

In the Baltic Sea, winter navigation is the main motivation for the ice classification research. This study concentrates on the Baltic Sea where ice conditions during winter hinder marine navigation. Baltic Sea is a brackish water basin and its ice is structurally similar to polar first-year sea ice but has much lower salinity. For safe navigation, icebreakers and commercial vessels require accurate ice condition information to select optimal routes and assess the need for icebreaker assistance. To aid navigation and safe shipping, the Finnish Meteorological Institute (FMI) provides daily ice charts. For this service, FMI uses mainly C-band SAR satellite images, particularly from RADARSAT-2 and Sentinel-1 missions (Gegiuc et al. 2017). These sensors have sufficient resolution (100 m) for general navigation needs. However, in order to distinguish ice ridges, heavily deformed ice and new ice formation in more details and get better understanding about mechanisms of sea ice formation, sensors with higher spatial resolution should be used, e.g. operating at X-band. Another advantage of X-band SAR over C-band is its higher sensitivity in assessment of sea ice surface properties, small-scale surface roughness and sea ice inclusions due to shorter wavelength (Ressel, Frost, and Lehner 2015;Dierking 2013;Fors 2016). Currently, the FMI service uses trained experts for sea ice classification and ice chart production. This is time consuming and expensive. Furthermore, the same SAR data interpreted by different experts can, and often does, lead to somewhat different results. In view of increased data volumes from operational and planned X-band sensors (e.g., ICEYE constellation), a need for automated classification procedure is getting stronger. This would if not replace manual expert interpretation, but at least provide valuable pre-classification to the analysis. To date, the potential of automated sea ice classification was demonstrated in several studies using backscatter intensity data (Clausi and Zhao 2003;Gegiuc et al. 2017;Sandven et al. 2012;Barber and LeDrew 1991;Clausi 2001;Shokr 2009). However, no feasible operational algorithm exists. Since different ice types can have similar backscatterintensity, a single image feature can be inadequate for sea ice classification (Ressel, Frost, and Lehner 2015;Karvonen 2004). Several earlier studies (Mäkynen and Hallikainen 2004;Dierking 2010) concluded that using only backscatter-intensity data is not sufficient to delineate different ice types reliably. There is an evidence (Dierking 2010) that additional image layers including higher order textural features are necessary and a large feature space has to be generated in order to train a classifier successfully. Among suitable techniques for describing the image texture, Gray-Level Co-occurance Matrix (GLCM) features have been tested in sea ice classification recently (Ressel, Frost, and Lehner 2015).
Several studies focusing on InSAR signatures of sea ice (Dammert, Leppäranta, and Askne 1998;Berg, Dammert, and Eriksson 2015;Marbouti et al. 2017;Dierking, Lang, and Busche 2017;Laanemäe, Uiboupin, and Rikka 2016) have demonstrated that coherence-magnitude and InSAR-phase provide valuable information about the sea-ice dynamics. Dammert, Leppäranta, and Askne (1998) found several relationships between backscatter-intensity and coherence-magnitude features over low-salinity ice, further expanded by Berg, Dammert, and Eriksson (2015). Marbouti et al. (2017) used backscatter-intensity and InSARphase features for finding sea ice displacement over fast ice. Dierking, Lang, and Busche (2017) used InSAR-phase feature for retrieving the ice surface topography. Laanemäe, Uiboupin, and Rikka (2016) used backscatter-intensity and coherencemagnitude features for sea ice classification. This motivates to study connections between backscatter-intensity, InSAR coherence-magnitude, and InSAR-phase, as well as the benefit of interferometry compared to backscatter, also in the sea-ice classification context.
In this study, we aim to extend the classification parameter space by introducing SAR interferometry and investigate the applicability of coherencemagnitude and InSAR-phase to improve automated sea ice classification at X-band. Specific goal is to establish a workflow for automated classification, using X-band interferometric SAR data acquired by the TanDEM-X mission, with preliminary results shortly communicated in Marbouti et al. (2018).
To date, there was only one study based on TanDEM-X imagery using both backscatter-intensity and coherence-magnitude features for automated sea ice classification (Laanemäe, Uiboupin, and Rikka 2016). Their method was applied over few types of ice: landfast ice, thin smooth ice, pancake ice and open water. Our aim is to increase the number of classified sea ice types, by using more features. Moreover, different sea ice classification algorithms are evaluated and compared here to select the best performing algorithm along with best combination of used image features.
The study concentrates on evaluating the added value of SAR interferometry versus the use of backscatter intensity on sea-ice classification, and compares three popular classification approaches (RF, ML, and SVM classifiers) for differentiating between open water and various sea ice types (undeformed ice, ridged ice, moderately deformed ice, brash ice, thick level ice, new ice).
The particular objectives of this study are: (i) to assess the potential of various SAR features, such as backscatter-intensity, interferometric coherence-magnitude and interferometric phase, as well as their combinations, for characterizing sea ice using X-band SAR data and, (ii) to determine which classification method (RF, ML, SVM) is more suitable for inferring sea ice types using studied InSAR features.
The article is organized as follows. We describe our study site, SAR data, relevant theoretical information and processing methodologies in Section 2. Classification results are provided in Section 3. Relative performance of different SAR features, comparison between RF, ML, and SVM algorithms, classification performance for different features, effects of image acquisition geometry on ice type classification, as well as comparison with prior works are given in Section 4. Finally, Section 5 summarizes the paper with conclusions, potential for future research and application of the demonstrated classification approaches.

Study area
Our study area (Marbouti et al. 2018) is located in the vicinity of Hailuoto Island in the north of the Baltic Sea ( Figure 1a). The Baltic Sea is a semi-enclosed brackish sea in northern Europe. Sea ice starts to form in November and its maximum extent is reached between January and March (Seinä and Peltola 1991). Sea ice breakup starts in April and typically completes by the beginning of June. The maximum annual ice coverage over the Baltic Sea is from 12% to 100% and its average is about 40% (Seinä and Palosuo 1996).
The sea ice cover in the Baltic Sea can be divided into landfast ice and drift ice. Landfast ice is fastened to the coastline and islands, while further offshore drift ice is moving along by winds and sea currents. The mean thickness of thermodynamically grown ice is 70 cm with inter annual range of 50-120 cm (Palosuo, Leppäranta, and Seinä 1982). Thermally grown sea ice consists mainly of congelation ice and snow-ice, and its salinity is less than two per mill (Leppäranta and Myrberg 2009). The thickness of ice ridges is typically 5 to 15 m (Leppäranta and Hakala 1992).
For this study, we have acquired TanDEM-X image over the Baltic Sea on March 30, 2012 ( Figure  1a). This was a mild winter but the northern and eastern basins of the Baltic Sea froze over. At the end of March, the ice in the Bay of Bothnia was tightly packed into the northeast part (Figure 1b), the InSAR data covered very close drift ice and landfast ice. In the area of the frame, the thickness of landfast ice was mostly 35-60 cm, and the drift ice largely contained deformed ice. This chart is not detailed enough for our study, and therefore an independent high-resolution reference ice chart was prepared in FMI Ice Service based on operational satellite data and ground truth from icebreakers and fixed observation sites.
The ice types chosen for the study contain six ice categories with different degrees of thickness and surface roughness, and open water (Table 1). These ice types can be identified by the FMI Ice Service in the operational ice charting that is based on the World Meteorological Organization (WMO) sea ice nomenclature. The categories or their symbology are included in the operational ice charting and are essential to icebreaking and navigation operations.
Landfast ice is a particular ice type that covers a large portion of the analyzed image but was left outside the classification. Landfast ice is a kinematic concept, defined on the basis of immobility. The boundary of landfast ice and drift ice is dynamic, where fractures form depending on the thickness of ice and consolidation between ice floes (Goldstein, Osipenko, and Leppäranta 2009). Thus, landfast ice cannot in general be distinguished from one image. As a matter of fact, our results show clearly how fracturing processes have caused mixing between landfast ice and drift ice with time, and therefore landfast needs to be studied from sequential image data (Marbouti et al. 2017).

InSAR theory in the context of the study
The interferometric coherence between two complex SAR images S 1 and S 2 represents the complex crosscorrelation between them: ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi where asterisk denotes complex conjugation. The coherence is used as an indicator of the quality of the interferogram (Hurtado 2012). Interferometric phase φ ¼ arg γ ð Þ gives information about the path difference between the different antenna positions and the target and can be used for topographic measurement provided SAR images were acquired at slightly different incidence angles (Plank 2014). The SAR system measures the interferometric phase difference φ related to the path length difference as: where λ is the SAR wavelength, R is satellite to target distance, dR is the path length difference between the two antennas, B is the perpendicular baseline, θ is the incidence angle, and α is the baseline tilt angle with respect to the horizontal, and the parameter a takes value of 1 or 2 for single-pass or multi-pass case, respectively. In this study, aa value is equal to one (Plank 2014).

TanDEM-X image selection and preprocessing
TanDEM-X Coregistered Single Look Slant Range Complex (CoSSC) products were used in the study. The image was acquired on March 30, 2012 in the stripmap mode in bistatic InSAR configuration, at HH polarization. Bistatic formation allows to avoid temporal decorrelation, with TanDEM-X being presently the only available bistatic spaceborne SAR mission. Its relatively short along track baseline and big perpendicular baseline is useful for ice mapping (Dierking, Lang, and Busche 2017). Also, X-band SAR compared to C-band has higher sensitivity toward surface structures (Ressel, Frost, and Lehner 2015;Dierking 2013). On the other hand, keeping in mind limited availability of such data to operational end-users, little research on sea ice classification using X-band InSAR data has been performed so far (Laanemäe, Uiboupin, and Rikka 2016).
In this study, we screened imaging data between 2010 and 2019 acquired using standard bistatic imaging mode. The project concentrated on Baltic Sea, thus we tried to choose images where sea ice is nearly stable and no melting is happening, limiting available scenes to those acquired in March. On the other hand, the most suitable TanDEM-X data for inferring sea ice topography is TanDEM-X Science Phase, which started in September 2014 and lasted for 17 months (Maurer et al. 2016). The Science Phase was initiated to demonstrate new products and applications such as digital elevation models with higher accuracies than in standard mode or measurements of ocean currents. The comparatively large baselines in this phase translated to a very high sensitivity for object elevations on the order of decimeters. However, unfortunately, we did not find any option of science phase over Baltic Sea during September 2014 to February 2016 (science phase period); that's why the choice was made to rely on standard operation mode that leads to somewhat lower topographic mapping accuracy compared to science phase. Image parameters of studied CoSSC scene are collected in Table 2. Figure 2 shows features of The TanDEM-X image: backscatter-intensity (Figure 2a), InSAR coherence-magnitude ( Figure 2b) and InSAR- A sea surface which is free of ice but may contain some ice fragments, by definition covering less than 1/10 of the surface.

New ice
New ice represents all forms of thin newly formed ice from frazil ice with almost liquid-like attributes to a uniform crust of up to 5 cm in thickness. Undeformed Ice Ice thicker than 10 cm, not significantly deformed but including rough surface properties. Thick Level Ice Ice thicker than 10 cm with a smooth surface and without any deformation. The ice surface might have been refrozen from flooded water or melted snow.

Ridged ice
Ice thicker than 10 cm with frequent occurrences of deformed ice, both rafted and ridged. Most of this ice type has been broken and piled to pressure ridges by compressive forcing. In this ice type, there is typically a network of ridge lines crisscrossing in a sheet of otherwise level first-year ice but also heavily deformed ice with virtually no level ice present. Moderately deformed ice Drift ice thicker than 10 cm, including a mixture of different ice types. The field consists of originally broken drift ice of different stages of development, including both rafted and ridged ice but also patches of level ice.

Brash ice
Ice broken into a fairly homogenous surface of very small ice blocks and typically identified by a high backscatter coefficient. In the scenes studied in this paper, brash ice can mainly be identified along known, fixed shipping lanes. It was chosen into the categorization because of its distinct features and evident detectability. Brash ice also forms at ice edges, broken and compacted by wind and waves into wide and thick zones and causing considerable harm to navigation.
phase ( Figure 2c). TanDEM-X data were orthorectified using ESA SNAP software. TanDEM-X InSAR-phase ramp was compensated to remove its effect on the classification performance (Sadeghi et al. 2014;Solberg et al. 2013;Solberg, Weydahl, and Astrup 2015;Marbouti et al. 2020). All features were additionally filtered using (7 × 7) boxcar filter. Land area was removed by applying land masking. Finally, linear stretching to dynamic range [0;255] was applied over each feature before running classification experiments.

Reference classification map
Sea ice charts for the Bay of Bothnia are routinely produced by the FMI experts. The ice charts provide a daily source of information on the ice conditions. Experts delineate ice chart polygons assigning to the specific ice types. The charts are based on visual interpretation of SAR imagery as the principal source of information. Currently, mainly RADARSAT-2 and Sentinel-1 SAR imagery with wide swaths are used (Gegiuc et al. 2017); however, X-band SAR sensors such as Cosmo-SkyMed and TerraSAR-X with narrower swaths are also deemed useful. The experts also use visible and thermal infrared imagery from Moderate Resolution Imaging Spectroradiometer (MODIS), in-situ observations, sea ice reports from icebreakers, and sea ice models in the production of the ice charts.
In this study, a sea ice expert produced a reference map by manually classifying sea ice based on the TanDEM-X features. The reference map in this study (ice chart shown in Figure 3) is in fact considerably  more detailed than routine daily ice charts to respond to the high resolution of the source data.

Proposed classification approach
To improve classification performance, careful selection of representative training data is critical (Antropov et al. 2014). Image area was almost 3558 km 2 . A total of 14,000 pixels (2000 pixels per each class) were randomly selected based on feature properties by the sea ice expert. RF and ML classification has been done using ESA Sentinel Application Platform (SNAP) software. SVM classifier was implemented using MATLAB software. In order to evaluate the added value of InSAR features compared to backscatter intensity, several classification experiments were performed: separately for each TanDEM-X feature, for their various combinations, as well as all features combined together.
After classification processes (Sections 2.5 and 2.6 and 2.7), each pixel was assigned to a sea ice type and open water. Majority voting (5 * 5 neighborhood) was applied over classification maps to achieve more homogeneous areas. Figure 4 shows the workflow of our proposed algorithm for open water and sea icetype classification.

RF classification
RF is a classification and regression tree technique invented by Breiman (2001). Breiman (2001) defined an RF as a classifier consisting of a collection of treestructured classifiers tðx; Θ K Þ; k ¼ 1; . . . : f gwhere the Θ k f gare independent identically distributed random vectors and each tree casts a unit vote for the most popular class at input x (Breiman 2001).
RF lets each tree vote for the class membership, and assigning the respective class according to the majority of the votes that result in taking advantage of the high variance among individual trees (Stumpf and Kerle 2011). RF has proven its power in handling classification with a big data of high dimensional feature spaces. These RF properties prove robust and accurate performance on complex datasets (Du et al. 2015).
SNAP software includes an application to perform RF pixel-based image classification. The SNAP allows us to train a classifier using various features and training vector data, with the resulting classifier model saved as a model file. The training vector data must contain polygons with a positive integer field representing the class label. In the next step, the image classification by using the provided model file is performed on corresponding features, with each pixel assigned a class label. In this study, number of training samples and trees in the forest were 14,000 and 100, respectively. Based on Oshiro's study (Oshiro, Perez, and Baranauskas 2012), the best optimal number for trees are between 64 and 128 trees because increasing the number of trees would bring no significant performance gain, and would only increase the computational cost. Although we tried to use of 10, 50, 100, and 150 trees in this study but the best result were achieved when we used 100 trees and increasing number of tree did not make sufficient change in our result.

ML classification
ML is a supervised classification approach. That is the procedure of finding the value of one or more parameters for a given statistic which makes the known likelihood distribution a maximum. It is derived from the Bayes theorem, which states that the posteriori distribution P wjk ð Þ, i.e., the probability that a pixel with feature vector 0 w 0 belongs to class 0 k 0 : where P wjk ð Þ is the likelihood function, P k ð Þis a priori information, and P w ð Þ is the probability that 0 w 0 is observed, it is given by: where N is the number of classes. P w ð Þis often behaved as a normalization constant to certify P N k¼1 P kjw ð Þ sums to 1. Pixel 0 x 0 is assigned to class ML often assumes that the distribution of the data within a given class 0 k 0 obeys a multivariate Gaussian distribution. It is then convenient to define the log likelihood: Since log is a monotonic function, Equation (5) is equivalent to: Each pixel is assigned to the class with the highest likelihood (Ahmad and Quegan 2012). The SNAP was used to perform supervised ML pixel-based image classification. Within ML classifier, we do not have any parameter to change except the number of training samples; it has been set to 14,000 similar to RF classifier to enable fair comparison of their performance.

SVM classification
SVMs have their roots in Statistical Learning Theory (Vapnik 1995). They have been vastly applied to machine vision fields including character, text recognition, and handwriting digit (Vapnik 1995;Joachims 1998). In nowadays, SVMs also are applied to satellite image classification (Huang, Davis, and Townshed 2002;Anthony, Greg, and Tshilidzi 2007). The objective of the support vector machine algorithm is to find a hyperplane in a high-dimensional space that separately classifies the data points. This translates to finding a hyperplane that gives the largest minimum distance to the training examples. The trade-off between minimizing the training error and the complexity of the decision function is controlled by a kernel and kernel parameters (Friedrichs and Igel 2005;Laanemäe 2017). The basic SVM is directly applicable for solving two-class (binary) classification.
For multiclass problems such as sea ice classification in this study, the implementation of SVM is approached by multiclass strategy. Two of the common methods to enable this adaptation include the One-Versus -One (OVO) and One-Versus-All (OVA) techniques (Anthony, Greg, and Tshilidzi 2007;Han et al. 2015). The OVA approach represents the earliest and the most common SVM multiclass approach (Melgani and Bruzzone 2004) and involves the division of a K class dataset into K two-class cases. The OVO approach on the other hand; involves constructing a machine for each pair of classes resulting in K (K-1)/2 machines. Further, majority voting is used to assign a label to current pixel. It is further modified to introduce weighting to the voting process (Anthony, Greg, and Tshilidzi 2007).
In this paper, an error-correcting output codes (ECOC) multiclass model based on SVM binary learners is trained using Matlab environment. Particularly, the OVO approach with K binary learners is used. Number of binary learners is depended on number of classes which exist in data. In this work, there are seven classes, thus the OVO model has 21 binary learners. For training the model, we used 2000 pixels for each classes and total pixels are 14,000, similarly to other experiments.

Accuracy assessment
Classification performance was evaluated using stratified random sampling approach, in order to have the same number of pixels for each class in the accuracy assessment. Confusion matrix (CM) was calculated for all ice classes, and following accuracy measures were used: Overall Accuracy (OA), User Accuracy (UA), Producer Accuracy (PA), and Kappa coefficient of determination.

OA ¼
Number of correct pixels Total number of pixels Kappa ¼ Observed accuracy -chance agreement 1 -chance agreement (11)

Classifications accuracy assessment
Altogether, 14 CMs were calculated for seven types of single and combinations of features in several classification experiments. All experiments are listed in Table 3. Values of OA, UA and PA assessments for all experiments are shown alongside CMs. The best results from 14 experiments (RF, ML and SVM) are presented in Figure 5(a, b). The highest OA in RF, ML, and SVM experiments were 72.93, 71.56, and 72.91% (all three features combined). This suggests that features combination are better than single features for sea ice classification. Table 3 shows the OA with Kappa coefficients statistics gathered for all 14 classification experiments.

Classification output maps
As it was shown in the previous section, the highest OAs were obtained using combination of all features (backscatter-intensity, coherence-magnitude, and InSARphase) with three classifiers. They were 72.93, 71.56, and 72.91% for RF, ML, and SVM, respectively. Figure 6(a, b, c) shows the final classification maps, produced at 10-m spatial resolution, and give a representative overview of the spatial distribution of sea ice classes in the study area.

Relative performance of different SAR features
In order to investigate the suitability of different input features and their combinations for the classification of open water and sea ice types, the accuracy assessment of classification maps produced using various combinations of features was performed.    (Table 3). Figure 7(a, b) and Table 4 summarize performance of RF in all experiments for different ice-cover classes. Figure 7(a, b) showed UAs and PAs of single and combination features in RF classifier, respectively, and Table 4 lists the classes which achieved the highest and lowest UAs and PAs for each single feature and their different combinations, respectively. The computation time for RF classification step was almost 5 min per any image feature and it slightly increased when using combinations of two or three features.
CMs in ML-experiments indicated that combinations of all three features had the best OA of 71.56% and Kappa accuracy of 66.82, although backscatterintensity & InSAR-phase and also backscatterintensity & coherence-magnitude combinations were following closely with OAs of 70.27% and 69.97% (Kappa accuracies of 65.32 and 64.97), respectively (Table 3). Similar performance was observed in RF classification experiments, though the accuracies were somewhat higher with RF compared to ML. Figure 8(a, b) and Table 5 summarize performance of ML in all experiments for different ice-cover classes. Figure 8(a, b) shows UAs and PAs of single and combination features in ML classifier, respectively, and Table 5 lists the classes which achieved the highest and lowest UAs and PAs for each single and combination of input data. The computation time for ML classification was 12 to 15 seconds per any image feature and it slightly increased when using combination of two or three features.
CMs in SVM-experiments also indicated that combinations of all three features had the best OA of 72.91% and Kappa accuracy of 0.68, although backscatter-intensity & InSAR-phase and also backscatterintensity & coherence-magnitude combinations were following closely with OAs of 72.78% and 72.15% (Kappa accuracies of 0.68 and 0.68), respectively (Table 3). Results was similar with RF and ML classification experiments. The accuracies were somewhat higher with RF classification and a bit lower with ML classification. Figure 9(a, b) and Table 6 summarize performance of SVM in all experiments for different ice-cover classes. Figure 9(a, b) show UAs and PAs of single and combination features in SVM classifier, respectively, and Table 6 lists the classes which achieved the highest and lowest UAs and PAs for each single and combination of input data. The computation time for SVM classification was about 30 s per any image feature, and it increased when using combination of two or three features.      acted almost in same manner and differences in coherence-magnitude, backscatter-intensity & coherencemagnitude, backscatter-intensity & InSAR-phase and backscatter-intensity & coherence-magnitude & InSARphase were near to zero. RF acted better than SVM in cases including backscatter-intensity (3.46%) and coherence-magnitude & InSAR-phase (2.23%) combinations except in InSAR-phase feature that it was 9.67% less than SVM classification. UA calculations in Figure 7(a), Figure 8(a) and Figure 9(a) show that difference between all classifiers (RF, ML, SVM) for sea ice classes including undeformed ice, ridged ice, moderately deformed ice and brash ice for classification features (single and combination features) was strong. RF had lower UAs for undeformed ice and ridged ice in order of 7-20% and 6-67% compared to ML and 0.7-29% and 2-56% compared to SVM. In comparison to RF and ML, the undeformed ice difference was strongest when backscatter-intensity & InSAR-phase combination was used (19.93%). In addition, ridged ice difference was strongest in InSAR-phase feature (66.62%). RF was more powerful than ML for moderately deformed ice and brash ice in order of 15-33% and 7-30%, respectively. The strongest difference for moderately deformed ice belonged to InSAR-phase feature (32.14%). The brash ice difference was strongest when coherence-magnitude & InSAR-phase combination (29.94%) was used. Although there were one or two exception classes, that ML or RF acted better. For example, coherencemagnitude was the only case that RF acted better than ML in undeformed ice (27.29%) and ML was powerful than RF in moderately deformed ice (9.88%). In comparison, RF and SVM, the undeformed ice (28.46%) and ridged ice (55.18%) had strongest difference when InSAR-phase was used. RF was more powerful than SVM for moderately deformed ice and brash ice in order of 0.02-18% and 2-24%, respectively. The strongest difference for moderately deformed ice belonged to InSAR-phase feature (17.06%). The brash ice difference was strongest when coherence-magnitude & InSAR-phase combination (23.31%) was used. In here also, there were one or two exception classes that SVM or RF acted better. For example, coherence-magnitude was the only case that RF acted better than SVM in ridged ice (7.53%) and SVM was powerful than RF in brash ice (3.78%).

Comparison of classification approaches
Difference between RF and ML classifiers for sea ice classes including new ice, thick level ice, and open water was not significant. These differences also were not significant between RF and SVM. In comparison RF and ML, in all single and combination features for new ice; ML had higher UAs than RF with improvements 0.1-9%. The strongest difference was for backscatter-intensity & coherence-magnitude combination (8.70%). InSAR-phase was the only case that had decreasing value from ML to RF with 8.09%. For open water and thick level ice, differences between RF and ML classifiers were in order of 2-14% and 0.6-3%.
In comparison RF and SVM, in all single and combination features for new ice; RF had higher UAs than SVM with improvements 1-7%. The strongest difference was for InSAR-phase (6.82%).
Thus, we can conclude that differences between RF with ML and SVM classifiers were remarkable for undeformed ice, ridged ice, moderately deformed ice and brash ice but were not considerable for new ice, thick level ice and open water classes.
Based on UA calculations for RF (Table 4), the best detected class was open water with UAs in order of 84.16% to 95.55% (in all classification features) whereas undeformed ice was hardest to discriminate in order of 2.29% to 51.17% (in all classification features except in backscatter-intensity feature). Same calculations were done for PAs (Table 4), the best detected class was new ice with PAs in order of 92.31% to 96.08% (in all classification features except InSARphase feature) whereas undeformed ice was hardest classes to discriminate with PAs in order of 10.40% to 37.40% (in single features) and moderately deformed ice was also hard to discriminate with PAs in order of 40.53% to 50.94% (in combination features). Based on UA calculations for ML (Table 5), the best detected classes were open water with UAs of 88.66% to 94.64% and new ice with UAs in range of 93.86 to 93.95% whereas brash ice was hardest class to discriminate in four classification features with UAs in order of 15.78 to 51.03%. Brash ice in InSAR-phase feature was not discriminated as all. Moderately deformed ice with UAs in order of 11.46 to 48.81% and Undeformed ice with 2.61% were also difficult to discriminate. Same calculations were done for PAs (Table 5), open water with PAs in order of 93.09 to 97.07% (in most classification features) and new ice with PAs in order of 83.09 to 94.94% were the best detected classes. Brash ice was not discriminated at all with one classification feature. Other classes particularly hard to discriminate were ridged ice with PAs in order of 36.90 to 49.86%, undeformed ice with PAs in order of 48.57% to 49.35 and moderately deformed ice with PA of 20.37%.
Based on UA calculations for SVM (Table 6), the best detected class was open water with UAs of 87.84% to 97.72% whereas ridged ice was hardest class to discriminate in four classification features with UAs in order of 37.39 to 56.92%. Brash ice in InSAR-phase feature was not discriminated at all and it was low also in coherence-magnitude & InSAR phase combination (22.41%). Moderately deformed ice with UAs in order of 19.77% was also difficult to discriminate in backscatter-intensity feature. Same calculations were done for PAs (Table 6), new ice with PAs in order of 82.68% to 96.23% was the best detected class. Brash ice was not discriminated at all with one classification feature. Other classes particularly hard to discriminate were ridged ice with PAs in order of 42.87 to 52.76%, undeformed ice with PAs in order of 39.18% and moderately deformed ice with PA of 27.59%.
Considered approaches were partly successful in detecting the brash ice, but the properties of the type of brash ice can be found also in naturally formed ice regimes clearly representing something else than brash ices. Therefore, we suggest other methods for discriminating brash ice from the rest of the ice, e.g., by segmentation and shape feature detection (Berthod et al. 1996).

Classification improvement by adding InSAR features
Backscatter-intensity feature is regularly used for sea ice classification in most ice services around the world.
In this study, OAs of RF, ML and SVM based on backscatter-intensity feature were in order of 66.51, 62.28, and 63.05%, respectively (Table 3).
InSAR-phase feature had OAs in order of 42.95% (RF), 50.91% (ML) and 52.62% (SVM). In our opinion, InSAR-phase feature could not fully reveal its value in contrast to backscatter-intensity and coherencemagnitude features. The reason is using an image with relatively large Height Of Ambiguity (HOA) of 30.84 m in this study. Too large HOAs decrease sensitivity of interferometry to detect sea ice roughness and ridges while in our case, this could aid better discrimination between ice types. The best option to examine in future would be images with smaller HOAs similar to "TanDEM-X Science Phase" mode (Dierking, Lang, and Busche 2017). Detailed discussion about "TanDEM-X Science Phase" mode in this context can be found in Section 4.4.
In this study, adding coherence-magnitude to backscatter-intensity feature improved classification results in 5.92, 7.69, and 9.1% in RF, ML, and SVM classifiers, respectively. This improvement was obvious in backscatter-intensity & InSAR-phase combination representing 5.33%, 7.99% and 9.73% in all experiments (RF, ML, SVM). The strongest improvements were achieved when three features were combined with each other, and represented 6.42% in RF, 9.28% in ML and 9.86% in SVM.
Using InSAR features with backscatter intensity increased classification accuracy for moderately deformed ice (UA: 43.15%, PA: 13.25%) in RF, (UA: 51.27%, PA:32.67%) in ML and (UA: 58.45%, PA: 27.00%) in SVM classifiers. Increased PA accuracies are visible in undeformed ice (PA RF = 26.9%, PA ML = 19.56%, PA SVM = 32.38%), ridged ice (PARF = 6.58%, PA ML = 5.2%, PA SVM = 7.14%) and thick level ice (PA RF = 3.76%, PA ML = 3.55%, PA SVM = 3.66%). As it was expected, adding InSAR features did not add any value to open water and new ice classification results in comparison with only using backscatter intensity. Using three features did not help to better classification of brash ice and as we mentioned in Section 4.1, thus, using other methods for discriminating brash ice from the rest of the ice is strongly suggested. These results emphasize the role of InSAR features (InSAR-phase and coherence magnitude features) in supporting sea ice operational mapping service.

Effects of image acquisition geometry on ice type classification
Knowledge of the angular dependence of different ice types is necessary to account for automatic sea ice classification of SAR imagery because it is impeded by the incidence angle dependence of backscatter intensities. Several studies at C-band were considering the incidence angle dependence of the sea ice backscattering coefficient (σ°) for different ice types in different polarizations and sea ice conditions (Liu, Guo, and Zhang 2018;Zakhvatkina et al. 2013;Aldenhoff et al. 2020;Mäkynen et al. 2002;Brath, Kern, and Stammer 2013;Gill et al. 2015;Onstott 1992;Mäkynen et al. 2014).
When the swath covers big range of incidence angles like Sentinel-1, the large range of incidence angles cause the backscatter signatures of the same sea ice type to vary across the swath. In this case, two different approaches have been used to determine the incidence angle dependence. One of them is direct slope estimation from backscatter intensity versus incidence angle plots. Another one is, the determination of the σ° versus θ 0 dependence are based on SAR image pairs acquired on ascending and descending orbits over the same sea ice area with a short time difference (Aldenhoff et al. 2020) and then incidence angle normalization is regularly applied for various image classification tasks, e.g., (Lang et al. 2016;Zakhvatkina et al. 2017;Murashkin et al. 2018). In the context of this study, the influence of various incidence angles on backscatter intensity in TanDEM-X mission is not possible to assess, as our analysis is based on one image with relatively narrow swath (42.33-44.46). Acquiring multi-incidence angle dataset is also problematic, as sea ice situation changes rapidly and a dedicated campaign would be needed. Leppäranta, Kuittinen, and Askne (1992) mentioned that backscatter level as well as the ability to distinguish ice types is sensitive to the incidence angle, and thus our results are valid for the incidence angle around the used value. Despite this narrow swath, incidence angle normalization has been done as part of image orthorectification. In future work, effects of varying incidence angle on InSAR based ice classification can be studied using either airborne SAR data, or potentially imagery of small-sat constellations like ICEYE.
Variation of incidence angle is also important in the coherence estimation. Laanemäe, Uiboupin, and Rikka (2016) shows that the best result for ice type discrimination over coherence feature was achieved when SAR images with a high incidence angles were chosen. Laanemäe compared two images with one day apart with similar meteorological conditions in incidence angles 23.4º and 44.9º and showed that the separation of ice and different ice types from low incidence angle imagery was better in high incidence angle compare with low incidence angle. Based on this, we tried to choose an image with high incidence angle to get the best result for sea ice classification over coherence. Further experiments with larger dataset are needed to assess effects of incidence angle variation over coherence feature for sea ice application.
Regarding the effect of different perpendicular baselines, Dierking, Lang, and Busche (2017) suggested having high sensitivity for object elevation detection. This requires image pair with a large crosstrack baseline. In this study, we used a scene with high perpendicular baseline around 240.38 m with short along track baseline. Further studies are possible when collecting several InSAR pairs over the same area within short time interval, and again can be possible with airborne or small-sat SAR campaign only, e.g. with ICEYE satellites .

Comparison with previous studies
To the best of our knowledge, a sea ice classification algorithm based on coherence-magnitude and backscatter-intensity features was used at least once before in the Baltic Sea, over coastal sea ice in the Gulf of Riga (Laanemäe, Uiboupin, and Rikka 2016). Laanemäe, Uiboupin, and Rikka (2016) classified open water and sea ice types based on the threshold values of the backscatter-intensity and coherence-magnitude for fast ice, thin smooth ice, pancake ice and open water. To relate the backscatter-intensity and coherencemagnitude values with the ice types, Regions of Interest (ROIs), based on field observations and optical satellite imagery, were selected to represent the observed different ice types and open water. Their case did not contain any rough ice in contrast to ours. The twin satellites (TanDEM-X and TerraSAR) were operated in the so-called pursuit monostatic mode with an along-track baseline of 20 km, corresponding to 2.6 s time interval (Scheiber et al. 2011). Coherence-magnitude feature was calculated by using two VV and HH pursuit monostatic data and HH-Monostatic data were used for backscattering calculations.
Before proceeding with the discussion, it is good to mention that thin smooth ice term in previous studies (Laanemäe, Uiboupin, and Rikka 2016;Geldsetzer and Yackel 2009) and new ice in our paper essentially mean the same, and they are comparable even if the names differ. Open water and new ice (thin smooth ice in (Laanemäe, Uiboupin, and Rikka 2016) and thin sea ice in (Geldsetzer and Yackel 2009) are common surface types between previous studies (Laanemäe, Uiboupin, and Rikka 2016;Geldsetzer and Yackel 2009) and ours. Geldsetzer and Yackel (2009) indicated that discrimination between open water and new ice was very difficult due to similarity in backscatterintensity values. Laanemäe, Uiboupin, and Rikka (2016) tried to solve this difficulty by adding coherence feature to backscatter feature values, but Figure 1 in (Laanemäe, Uiboupin, and Rikka 2016) showed that the same problem remained for the coherencemagnitude values as well. In our study, we examined backscatter-intensity and coherence-magnitude features and indicated that water and new ice are well discriminated by backscatter-intensity and coherencemagnitude features. This success can be explained by the type of TanDEM-X data that were used in this study. Bistatic imaging mode would be better option compared to monostatic imaging mode (Laanemäe, Uiboupin, and Rikka 2016) in open water and new ice discrimination. In bistatic acquisition, wind speed would not cause temporal decorrelation of interferometric coherence.
Open water/ice separation is one of the key question in sea ice remote sensing. In our study, this good separation is not only between open water and new ice. It was also apparent between open water and other types of ice in Figure 10 by using backscatter-intensity and coherence-magnitude features. Although we need to do same work over more cases of study to evaluate the consistency of the classification.
InSAR-phase feature was used recently in a paper by Dierking, Lang, and Busche (2017) for retrieving the ice surface topography. Dierking used InSARphase data acquired during "TanDEM-X Science Phase". Both "Standard operation" and "TanDEM-X Science Phase" modes were used to map topography and build digital elevation models. "TanDEM-X Science Phase" was more accurate than "Standard operation" due to bigger cross track baseline and smaller HOAs. So, it would be concluded that using "TanDEM-X Science Phase" mode is better and more accurate option to detect sea ice topography than "Standard operation" mode.
In this study, analysis was done using scene acquired in "Standard operation" mode. Standard operation mode has been designed for topographic mapping of the Earth's land masses. Its achievable relative accuracies are 2-4 m vertically and 3 m horizontally at a horizontal sampling of 12 m. However, using InSAR features even in this setup was useful for discriminating sea ice classes, though not to its full possible potential.

Conclusions and future research
In this study, different features of TanDEM-X imagery were used for assessment of sea ice classes (undeformed ice, ridged ice, moderately deformed ice, brash ice, thick level ice, new ice) and open water over the Baltic Sea. Three types of supervised classifiers, RF, ML, and SVM algorithms were applied over single features and their different combinations. The best results were provided by combined backscatterintensity & coherence-magnitude & InSAR-phase with OAs 72.93%, 71.56% and 72.91% in RF, ML, and SVM classifiers. Combined features outperformed single features in all sea ice classification scenarios. The RF algorithm seems to be more suitable approach than SVM and ML due to higher OAs, however, at the expense of somewhat longer processing time (5 min for RF versus 15 seconds for ML and 30 s SVM at Intel i5-7200 U machine).
This study also proved advantage of interferometric features (coherence-magnitude & InSAR-phase) in combination with backscatter-intensity feature over single backscatter-intensity feature by showing improvements OAs in order of 6.42% in RF, 9.28% in ML and 9.86% in SVM classifiers. For most of Figure 10. Coherence-magnitude and backscatter-intensity values of water and sea ice types for plots using pair HH-bistatic data for coherence-magnitude calculation and HH-bistatic data for backscatter-intensity calculations. separated classes, the improvement was even stronger; and the strongest one was moderately deformed ice with improvements in order of (UA: 43.15%, PA: 13.25%) in RF, (UA: 51.27%, PA: 32.67%) in ML and (UA: 58.45%, PA: 27.00%) in SVM classifiers. Brash ice was not well discriminated by our method, so using other methods for discriminating brash ice from the rest of the ice is suggested. Also differences between RF with ML and SVM classifiers were remarkable for undeformed ice, ridged ice, moderately deformed ice and brash ice but were not considerable for new ice, thick level ice and open water classes.
This study is the first approach to thoroughly evaluate backscatter-intensity, coherence-magnitude and InSAR-phase features at X-band and their different combinations in sea ice classification. Previously, water and new ice were particularly hard to discriminate due to their similar backscattering values (Laanemäe, Uiboupin, and Rikka 2016;Geldsetzer and Yackel 2009). Even adding coherence-magnitude (Laanemäe, Uiboupin, and Rikka 2016) did not improve the discrimination between these classes, but this issue was successfully solved in our study. As our image was acquired in bistatic InSAR imaging mode, no temporal decorrelation affected InSAR coherence. Bistatic TanDEM-X data were particularly useful for discriminating open water and new ice. Our future efforts might focus on using various advanced texture feature extraction techniques from intensity, coherence and phase features to be tested with suitable classifiers.

Notes on contributors
Marjan Marbouti is a PhD student in geophysics at Institute for Atmospheric and Earth System Research (INAR), Helsinki University. She received her bachelor's degree in physics from Islamic Azad university-Dezfool branch, Dezfool, Iran in 2009 and his master degree in physicsastrophysics from Islamic azad university-Central Tehran branch, Tehran, Iran, in 2011. His research interests is remote sensing of sea ice, ocean color properties, and atmospheric properties.
Oleg Antropov is a Senior Scientist at VTT Technical Research Center of Finland. He received the M.S. degree in computer science and the Candidate of Sciences degree in radio physics from Dnepropetrovsk National University, Ukraine, in 2006 and 2010, respectively, and PhD degree in space technology from Aalto University, Espoo, Finland, in 2014. His research interests include application of multiparametric SAR data for various environmental applications.
Jaan Praks is an Assistant Professor (tenure track) at Aalto University and leader of Aalto Space Technology and Microwave Remote Sensing research group. He received an MSc. in physics from University of Tartu and PhD. in space technology from Aalto University 2012. His research interests include microwave remote sensing, especially SAR Interferometry and SAR polarimetry, and nanosatellite technology. During his career, he has participated in numerous remote sensing projects and measurement campaigns, instrument developments and led the development of Aalto-1, Aalto-2, Aalto-3, and Foresail nanosatellites. Currently, he is leading Satellite Platforms team in Finnish Centre of Excellence in Research of Sustainable Space. From his research projects have started many New Space and Earth Observation start-up companies in Finland.

Patrick B.
Eriksson is an ice expert and product manager at the Oceanographic Services of the Finnish Meteorological Institute in Helsinki. He has a M.Sc. degree in polar oceanography from University of Helsinki. He is an ice analyst at the Finnish Ice Service with solid routine in ice monitoring, ice observation, and remote sensing and broad customer relations to winter navigation operations.
Vahid Arabzadeh is a Ph.D. student at Aalto University, school of science. He received two MSc. in Mechanical engineering and system control. His main research interest is energy system modeling, optimization, AI, and data analysis.

Data availability statement
TanDEM-X data that support the findings of this study are available from DLR. Restrictions apply to the availability of these data, which were used under license for this study (https://tandemx-science.dlr.de/pdfs/TSX-TDX-user-license-v2.2.pdf). The manual image classification analysis done by an ice analyst was generated at the Ice Service of the Finnish Meteorological Institute. Derived data supporting the findings of this study are available from the corresponding author PBE on request.