Wide swath SAR observation from several spaceborne SAR missions (RADARSAT-1, 1995–2013; Envisat ASAR, 2002–2012; ALOS-1 PALSAR, 2006–2011; RADARSAT-2, 2007-present; Sentinel-1, 2014–present) played an important role in studying global ocean and ice-covered polar regions. The Sentinel-1 constellation (1A and 1B) is producing dual-polarization observation data with the largest Arctic coverage and the highest temporal resolution ever. The cross-polarization is known to be more sensitive to the difference in scattering from sea ice and open water than the co-polarization (Scheuchl et al., 2004), and the combination of HH and HV polarizations has been widely used for ice edge detection and ice type classification (a nice overview is given in the paper by Zakhvatkina et al., 2019). However, most of the recent ice classification algorithms were developed using RADARSAT-2 ScanSAR (Leigh et al., 2014; Liu et al., 2015; Zakhvatkina et al., 2017), which has different sensor characteristics from Sentinel-1 TOPSAR, and the use of Sentinel-1 for the same purpose is very limited in literature. The main drawback of applying existing algorithms to Sentinel-1 TOPSAR data is the relatively high level of thermal noise contamination and its propagation to image textures.

For proper use of dense time series of Earth observations using SAR sensors, radiometric properties must be well-calibrated. Thermal noise is often neglected in many cases but can seriously impact the utility of dual-polarization SAR data. Sentinel-1 TOPSAR image intensity is particularly disturbed by the thermal noise in the cross-polarization channel. Although the European Space Agency (ESA) provides calibrated noise vectors for noise power subtraction, residual noise contribution is still significant considering the relatively narrow backscattering distribution of the cross-polarization channel. In our previous study (Park et al. 2018), a new denoising method with azimuth de-scalloping, noise scaling, and inter-swath power balancing was developed and showed improved performance in various SAR intensity-based applications. Furthermore, when it came to texture-based image classification, we suggested a correction method for textural noise (Park et al., 2019) which distorts local statistics, and thus degrades texture information in the Sentinel-1 TOPSAR images.

In many of the previous studies on ice–water and/or sea ice classification (Soh and Tsatsoulis, 1999; Zakhvatkina et al., 2013; Leigh et al., 2014; Liu et al., 2015; Ressel et al., 2015; Zakhvatkina et al., 2017; Aldenhoff et al., 2018), the training and validation were done using manually produced ice maps. Although the authors claimed that the manual ice maps were drawn by ice experts, the selection of SAR scenes and interpretation could be inconsistent, and the number of samples might not be enough to generalize the results because of the laborious manual work. Furthermore, the results are hardly reproducible by others because the reference sources are not open to the public. Therefore, increasing objectivity is crucial, and automating the classification process is encouraged. The idea of training using SAR images and accompanying image analysis charts, which is a direct manual interpretation of SAR images by trained ice analysts working at operational ice services, was tested for sea ice concentration estimation by Wang et al. (2017); however, such image analysis charts are not accessible to the public.

The use of a public ice chart as training and validation reference data may help in solving the validation problem. The preparation of a public ice chart is also through manual inspection of various sources of satellite imagery and other sources of data (Partington et al., 2003; Johannessen et al., 2006); however, training using a large volume of these charts would reduce operator-to-operator bias, such as inconsistent decisions against similar ice conditions. The overall bias may exist since the public ice charts are produced in the interest of marine safety. Nevertheless, as the human interpretation available in the ice chart is currently considered the best available information on sea ice (Karvonen et al., 2015), the best practice to make a sea ice type classifier is to train with the public ice chart so that the best knowledge of ice analysts is mimicked.

In this work, we present a semi-automated Sentinel-1 image-based sea ice classification algorithm which takes advantage of our denoising method. The noise-corrected dual-polarization images are processed into image textures that capture sea ice features in various spatial scales, and they are used for supervised classification with a random forest classifier by relating with ice charts published by operational ice services. The use of ice charts has a dual purpose: semi-automatization of classifier training and minimization of human error.

We trained three RF classifiers with different feature configurations: (i) FC1: Haralick texture features and CV; (ii) FC2: Haralick texture features, CV, and incidence angle; and (iii) FC3: Haralick texture features, CV, incidence angle, and day of the year.

As expected, the classification score increases with the number of trees (crosses in Fig. 5a), and Richards curve (dashed line) fits well to the observations (RMSE =2.3×10-4). The optimal NT value is selected where the score increment per tree (i.e. local slope) becomes less than 0.001 (i.e. accuracy increase of 0.1 %) and constitutes 11 trees, thus keeping the forest size small. The scores also increase with the maximum tree depth (crosses in Fig. 5b), but Richards' curve (dashed line) does not fit so well (RMSE =3.6×10-3) and cannot be used for finding the optimal D value. This can be explained by overfitting of the classifier and illustrated by the difference between training and testing scores (Fig. 5c): a small difference between the scores (for D≤8) indicates similar performance on training and testing datasets, while a large difference (for D>8) indicates that the testing dataset is processed with worse results. The optimal D value is therefore selected where the score difference becomes higher than 0.03 and constitutes eight levels. The optimal value of the number of features (NF) was selected using the same criterion as for NT, and the value constitutes 10 features. As a result, the optimal hyperparameters of the number of trees, the maximum tree depth, and the number of features were 11, 8, and 10, respectively.

The trained five-class classifier consists of five binary sub-classifiers; each of them is used for discriminating one specific class from the others. For each sub-classifier, each texture feature has a different weight in decision making. The fraction of the samples that each texture feature contributes can be used to compute the relative importance of the features, and the averaged estimates of them over several randomized trees serve as an indicator of feature importance (Louppe, 2014). The feature importance of the sub-classifiers is presented in Fig. 6. The overall pattern shows that the features of HV polarization play a more important role than those of HH polarization. For HH polarization, the sum average, which is equal to the mean backscattering intensity in each subwindow, was the prominent feature. For HV polarization, however, contrast and variance- and entropy-related features were more important. The classifiers for open water and old ice have stronger dependencies on HV polarization than others. This is understandable because the main radar scattering mechanisms for those two types are strongly characterized by the portion of volume scattering: low for calm water and high for dry ice with low salinity (old ice). The classifier for new ice has a distinctive pattern that the sum averages in both polarizations are much more important than other features. This might be because the new ice has different types of recently formed ice including nilas, which is smooth but rafting can make rough features, and frost flowers, which introduce high surface roughness and volume scattering (Isleifson et al., 2014). Thus the new ice can appear either featureless and dark or complex and bright in a SAR image (Dierking, 2010). The large range in backscatter values makes it hard to define characteristic texture in the new ice patch.

The confusion matrix for testing the trained classifier for winter season with the test dataset (DS1) is shown in Table 3. Three cases with different feature configurations (FC1–FC3) were tested. The accuracies for open water and old ice were higher than 90 %; however, those for young ice and first-year ice were around 60 %. The mean difference between the results of FC1 and FC2 was only 1.2 %, indicating that residual angular dependency was negligible after the incidence angle correction. However, the accuracy significantly improved from FC2 to FC3, especially with new ice (21.2 %). The Cohen kappa coefficients κ for FC1, FC2, and FC3 were 0.70, 0.71, and 0.77, respectively. It should be noted that the evaluation of the DS1 was carried out with the input dataset that was used for training. Thus, the test and training data share the same ice conditions as well as spatio-temporal coverage. As a result, the κ might contain correlation which is not preferable for proper evaluation. Table 4 shows the confusion matrix for validation results from the DS2 of which the accuracy of open water and old ice was at a similar level, compared to the DS1. Meanwhile, the accuracy of young ice and first-year ice decreased considerably. The differences between the results of FC1, FC2, and FC3 were insignificant. This result is opposite to the DS1 inferring that the training with FC3 was overfitted and the day of the year may not correspond to the temperature, air–sea fluxes, or weather regimes. The κ values for FC1, FC2, and FC3 with the DS2 were 0.67, 0.67, and 0.67, respectively.

To see how the denoising step in Sect. 2.2.2 led to improvements in the classification accuracies, the same training and evaluation were conducted for the same dataset without applying the textural noise correction (Table 5). In all configurations (FC1–FC3), the accuracies improved for young ice (+2.4 % to +6.2 %) and old ice (+3.9 % to +4.5 %), which were most pronounced compared to those for open water (+1.1 % to +1.9 %) and first-year ice (-0.2 % to +0.9 %). On the contrary, a small accuracy decrease was observed for new ice (=2.4 % to -0.3 %). Nevertheless, the improvement in κ (+0.08) demonstrates a clear improvement in the overall classification result.

The confusion matrix for testing the trained classifier for the summer season with the test dataset (DS1) is shown in Table 6. As described in Sect. 2.2.1, the further simplified three-class classification is applied. The accuracies for open water and old ice were higher than 92 %; however, the accuracies for mixed first-year ice were only around 15 % in both FC1 and FC2 and 42 % in FC3. The large difference between the results of FC1-FC2 and FC3 indicates that the mixed first-year ice likely changes at short timescales. The misclassifications for mixed first-year ice were mostly into old ice. This might be because of the surface melting and the corresponding image textures which make the discrimination between the mixed first-year ice and old ice difficult. The same patterns were observed from the confusion matrix (Table 7) for validation results from the DS2, except that the accuracy decrease from Tables 6 to 7 was particularly large for FC3, meaning overfitting for DS1. However, it is unclear if the low accuracy for mixed first-year ice is due to the classifier itself or the data. To unravel this, the data were divided into three groups of 1 month each (June, July, August), and then separate classifiers were trained and tested. Tables 8 and 9 show the results with FC1 configuration for DS1 and DS2, respectively. There was a rapid accuracy decrease for mixed first-year ice from June to August in both results (from 60.6 % to 26.5 % for DS1 and from 55.6 % to 11.0 % for DS2). As there is no particularly large difference between the numbers in Tables 8 and 9, meaning the classifiers were not overfitting, the very low accuracies for mixed first-year ice in Table 6 (14.9 %) and Table 7 (12.0 %) seem to be due to a too large temporal extent for a single classifier. In other words, the classifier for the summer season needs to be split into multiple groups with shorter time spans, and/or a feature that can effectively account for surface melting needs to be introduced into the algorithm for further development. Based on these results so far, the trained classifiers in the summer season for 3 months in bulk failed in distinguishing mixed first-year ice from old ice; thus they are close to ice–water discriminators rather than ice type classifiers.

Figure 7 shows a daily mosaic of Sentinel-1 SAR images over the study area and the classified ice map in the winter season. For comparison, the NIC weekly ice chart is also displayed. Despite the SAR images being acquired 3 d before the ice chart was published, the ice edges of the ice chart match well with the SAR mosaic in most parts because the same SAR data were used. Overall, the discriminations between ice and non-ice, old ice and other ice types, and detection of new ice patches look reasonable. However, some young ice patches, for example the ice patches between the Svalbard archipelago, are misclassified as first-year ice. Figure 8 shows another daily mosaic made by the images acquired on the same day of the ice chart publication. Considering notable ice drift in the backscattering images in Figs. 7 and 8, the SAR-based ice classification results in both figures looking consistent, well in line with the ice drift. Although the weekly ice chart is supposed to represent the averaged ice status in the past few days, the actual ice distribution on the actual date of the publication can be largely different. This example shows a clear potential of near-real-time service of ice type classification.

Figures 9 and 10 show the same mosaics for the case in the summer season. As shown in Tables 6 and 7, the misclassifications for the mixed first-year ice into old ice are pronounced in the large ice patches north to Svalbard, while the ice edge positions of the ice chart and the classification result are in good agreement with each other.

To cope with the ambiguous classification for the winter season ice types with low accuracy, we conducted a test with the three-class classification, and Tables 10 and 11 show the resulting confusion matrices. The κ values for FC1, FC2, and FC3 were 0.83, 0.84, and 0.84 in DS1 and 0.75, 0.75, and 0.74 in DS2, respectively. The dramatic increase in the accuracy of the mixed first-year ice indicates that the misclassification for the new ice, young ice, and first-year ice was mostly among themselves. However, the accuracy decrease from DS1 to DS2 was at a similar level to the case of the five-class classification. This could have been caused by inconsistent labelling in the reference ice chart.

Figure 11 shows an example of the inconsistent labelling in the reference ice chart. The SoDs from the NIC ice charts are superimposed on the Sentinel-1 backscattering images. The same ice floe (red outline) is classified differently in two different ice charts (old ice on the left panel and first-year ice on the right panel), although it looks almost the same in the corresponding SAR backscattering images. It should be noted that training with the ice chart might have included mislabelled small features even if the image selection based on ice edge matching was successful. Furthermore, the boundaries between different ice types in the ice chart are normally not as precise as those in the SAR image-based classification results. Therefore, the lower classification accuracies compared to those in previous studies (80 % in Zakhvatkina et al., 2013; 91.7 % in Liu et al., 2015; 87.2 % in Aldenhoff et al., 2018), which used manually classified ice maps as training and validation reference, are expected.

Unfortunately, we could not find an official report regarding the accuracy of the NIC ice chart information.

Table 12 shows the confusion matrices for our three-class classifiers when their prediction results are compared with the OSI-403-c product as reference. For one-to-one comparison, it was assumed that the ideal characteristics of the mixed first-year ice and the old ice in our three-class classification are equivalent to those of the first-year ice and the multi-year ice in OSI-403-c. Comparing with the results in Table 11, the accuracies for open water decreased by 6 %; however, this is mainly because the ice concentration threshold for ice–water discrimination in OSI-403-c is 35 %, which is higher than the 20 % that we set in our preprocessing of the NIC ice chart (Sect. 2.2.1). Thus areas with low ice concentration in marginal ice zone are most likely annotated as open water in OSI-403-c. The accuracies for open water at points in the NIC charts with ice concentration between 20 % and 40 % only were considerably lower, with 67.4 %, 67.8 %, and 70.1 % for FC1, FC2, and FC3, respectively (not presented in Table 12). For first-year ice, large portions (72 %) are misclassified as old ice. This might be partly explained in Fig. 12, which shows the ice classes in the NIC ice chart and OSI-403-c for the same publication date. A large extent of old ice in the NIC ice chart is annotated as multi-year ice in OSI-403-c. As our classifiers were trained with the NIC ice chart, it is natural to result in more multi-year ice for the area where the ice type is classified as first-year ice in OSI-403-c. For multi-year ice, the accuracy was the highest, 98 %.

The inconsistency in ice types between the NIC ice chart and OSI-403-c seems persistent at least for the time coverage of DS2 (January–March in 2019). Table 13 shows averaged percent agreement of the two sea ice type products for the same publication dates over 12 weeks (12 one-to-one comparisons as the NIC ice chart is a weekly product). To make a fair comparison, the ice-covered areas with ice concentrations lower than 35 %, which is the threshold for ice–water discrimination in OSI-403-c, were excluded. The percent agreement for first-year ice (58.8 %) was much lower than that of open water (90.0 %) and multi-year ice (99.0 %), which is in line with the results in Table 12. Finding the reason for the clear discrepancy of the extent of first-year ice between the NIC ice chart and OSI-403-c is beyond the scope of this study; however, it should be noted that elaborate future work for cross-calibrating ice types in different ice charts is necessary.

The proposed algorithm has several limitations. First of all, the variations in radar backscattering and its corresponding image textures due to seasonal changes were not properly captured. Although day of the year was tested as a seasonality variable in the FC3 feature configuration, the result did not show any improvement. This is because SAR image features, which partially reflect temperature fluxes and weather regimes, might not correspond to day of the year. Second, the proposed method struggles when the same type of sea ice is located on different edges of the range swath of SAR images because the incidence angle dependence could not be normalized perfectly. An example of such a failure can be seen along the image boundaries at 79.5∘ N, 45∘ E in Fig. 7 and 79∘ N, 50∘ E in Fig. 8, approximately. Third, some artefacts were observed under large ocean swells. In the classified results in Fig. 8d, there is a misclassified first-year ice patch (yellow) in the open water area. According to the high-resolution sea surface wind data from SAR on the Sentinel-1 satellites (https://data.nodc.noaa.gov/cgi-bin/iso?id=gov.noaa.nodc:SAR-WINDS-S1, last access: 18 August 2020), the wind speed ranged from 17 to 21 m s-1 at the time of image acquisition, heavily roughing the water surface. Although we have included images with both high and low wind conditions in our training data, the image textures of wind-roughened water surface and ice were confused in some cases, and the same happened in the image textures of calm water surface and smooth level ice.

A new semi-automated SAR-based sea ice type classification scheme was proposed in this study. For the first time several ice types were successfully identified on Sentinel-1 SAR imagery in the winter season, while only an ice–water discrimination was feasible in the summer season. The main technological innovation is two-fold: (i) reduced manual work in the preparation of a large amount of training and validation reference data using readily available public ice charts and (ii) more objective evaluation of the SAR-based sea ice type classifier compared to the previous studies conducted with a small number of images and customized ice type references from sources not open to the public. A conventional approach for selecting training and testing data by anonymous human ice experts is undesirable not only because it is laborious, but also due to subjectivity and lack of standardization in the assessment of the automated classifier. Therefore, the performance from different literature sources cannot be intercompared directly.

Test results from the datasets of the winter season acquired over the Fram Strait and the Barents Sea area showed overall accuracies of 87 % and 60 % and the Cohen kappa coefficients of 0.75 and 0.67 for the three-class and five-class ice type classifiers, respectively. These are slightly lower than the numbers in previous studies, and the errors are attributed not only to the automated algorithm but also to the inconsistency of the ice charts and the high level of their generalization. Test results from the datasets of summer seasons showed overall accuracy of 67 % and the Cohen kappa coefficient of 0.78 for the three-class classifiers. Considering the misclassifications in different ice types were among themselves, the three-class classifiers are not really a sea ice type classifier, but they performed well at least as an ice–water discriminator with accuracy of 98 %.

Based on the results, we envisage that three-class ice type classification from SAR imagery would be useful for making a global sea ice type product like OSI SAF OSI-403-c with higher spatial resolution. The proposed approach importantly showed that a daily ice type mapping from the Sentinel-1 data is feasible and can help capture details of short-term changes in the stage of sea ice development. Based on the achieved results, we believe that the proposed approach may be efficiently used for operational ice charting services for supporting navigation in the Arctic.