Sea Ice 1

Satellite borne SAR imagery has become an invaluable tool in the field of sea ice monitoring. Previously, single polarimetric were employed in supervised and unsupervised
classification schemes for sea ice investigation, which was preceded by image processing techniques such as segmentation and textural features. Recently, through the
advent of polarimetric SAR sensors, investigation of  polarimetric features in sea ice has attracted increased attention.
While dual-polarimetric data has already been investigated in a number of works, full-polarimetric data has so far not been a major scientific focus. To explore the
possibilities of full-polarimetric data and compare the differences in C- and X-bands, we endeavor to analyze in detail an array of
  datasets, partly simultaneously acquired, in C-band (RADARSAT-2, Sentinel) and X-band (TerraSAR-X) over ice infested areas. First, we propose an array of polarimetric
features (Pauli based, lexicographic based, simulated compact polarimetry). The imagery consists of  several different instances of simultaneously
acquired RADARSAT-2 quadpol Stripmap images and simultaneous dualpol/quadpol TS-X Stripmap and dualpol Sentinel images (Greenland coastal waters, Barents Sea).
Ancillary data from national ice services, SMOS data and expert judgement were utilized to identify the governing ice regimes. Based on these preponderations, we then
extracted mentioned features. The subsequent supervised classification approach was based on a neural network. To gain quantitative insight into the quality of the
features themselves (and reduce a possible impact of the Hughes phenomenon), we employed mutual information to unearth the relevance and redundancy of features.
The results of this information theoretic analysis guided a pruning process regarding the optimal subset of features. In the last step we compared the classified
results of all sensors  and images, stated respective accuracies and discussed output discrepancies in the cases of simultaneous acquisitions.

All weather data from SAR provides unique possibilities for operational monitoring of sea ice. Automatic classification of SAR data is challenging since values of normalized radar cross section (NRSC also denoted as sigma0) for both open water and sea ice have similar ranges and differ only in texture [Sandven et al., 1999].

Over the European Arctic Sentinel1A operates in cross-polarization mode [signal is emitted in horizontal polarization and recorded both in horizontal (HH) and vertical (HV) polarizations]. HV signal is known to be more sensitive to difference in the scatter of sea ice and open water and combination of HH and HV polarizations is widely used for classification of SAR images for ice edge and ice type detection [Kaleschke and Kern, 2002]. A number of techniques have been developed to classify SAR data based not on the level of signal but on texture information [Haralik et al., 1973]. In these techniques a SAR image with relatively high spatial resolution (pixels size ranges from 10 to 75 meters) is split into subimages with width and height ranging from 10 to 100 pixels. The values of NRSC in each pixel are scaled to integer values to constitute few gray levels (from 10 to 255 levels). The Gray Level Co-Occurence Matrix (GLCM) is calculated where the value in each cell is the number of neighbors with the color equal to the row and the column number. Several Haralick texture features (from 7 to 15) are calculated from the GLCM and quantitatively describe parameters of the features: energy, entropy, prominance, etc.

The main problem with the approaches applied so far is that the actual resolution of the data is greatly reduced when many hundreds of pixels are used to calculate a single vector of texture features. He we suggest an algorithm for very high resolution classification of SAR data.

Our algorithm solves the low resolution problem by combining segmentation of a SAR image and then calculating texture features for each segment. Segmentation is performed in a sliding window with size 300x300 pixels. Each pixel is described by four parameters: sigma0 in HH, sigma0 in HV, X coordinate and Y coordinate. Pixels are grouped together in the 4D space using the k-means cluster analysis. Segments are formed from pixels with similar sigma0 and coordinates and too small segments (with less that 1000 pixels) are dissolved. The sliding window slides with overlapping to mitigate segmentation artefact at the edges.

The texture features from many SAR image are then manually classified to create a training dataset. The training dataset is used to train Support Vector Machines (SVM) which are then used in a fully automatic manner. The algorithm is presented below with the steps 1 - 3 performed only during the training phase and other steps performed in operational processing of Sentinel-1A data:

Correction of NRSC for incidence angle (in HH) and thermal background noise (HV);

Segmentation of SAR data using k-means cluster analysis of pixels with sliding window;

Calculation of GLCM from pixels within each segment;

Calculation of 15 Haralick texture features;

Automatic classification of texture features using k-means cluster analysis;

Manual reclassification of texture features using visual analysis into training dataset;

Training of Support Vector Machine (SVM) based on the training dataset;

Automatic classification of Sentinel-1A SAR image into the desired classes (e.g. open water, sea-ice)

The software for automatic classification is written in the Python language. It capitalizes on many open source libraries (GDAL, Nansat, mahotas) and is therefore made also publicly available at http://github.com/nansencenter/sentinel1ice

The deformation of sea ice is an important component of sea ice dynamics and the interaction between atmosphere, ice, ocean, and land. An inhomogeneous drift of sea ice creates opening or closing of leads, and results in the formation of pressure ridges and local changes of the ice thickness. The motion of sea ice can be observed from space by synthetic aperture radar (SAR) and quantified by drift-detection algorithms. A general problem of quantifying deformation or strain rates is that the magnitudes are dependent on the spatial and temporal scales.

Previous analyses, primarily on RGPS-data, have revealed the strong heterogeneous and intermittent character of sea ice deformation. By analyzing the distribution of strain rates a scaling effect, both temporally and spatially, could be demonstrated (Hutchings et al. 2011; Marsan et al. 2004; Rampal et al. 2008; Stern and Lindsay 2009). Thus, the distribution of strain rates at some (arbitrary) scale is not sufficient to characterize sea ice deformation (Weiss 2013).

The objective of our study is to extent the analysis of strain fields towards smaller spatial scales, which are not included in previous studies. By using sequential SAR images and a pattern-matching approach for the retrieval of the drift field it is possible to downsize the spatial scale from several kilometers to a few 100 meters. To identify the effect of spatial scaling we perform a statistical analysis of strain rates. We calculate the statistical moments of the strain rate distribution for different spatial scales and describe their dependence on spatial resolution using a power law. The results obtained hitherto reveal the scale-dependence of strain rates. The found relations are consistent with similar analyses on larger scale ranges. In addition and for the first time for sea ice we perform the analysis independently for shear strain and divergence. Again, scaling behavior is observed, but differently pronounced.

Our result strengthened the assumption that scaling properties extend down to very fine scales and thus an extrapolation to laboratory scales (~1m) may be feasible. If scaling behavior of strain rates is known, the comparability of sea ice deformation derived from different methods at different spatial scales, e.g. modelling , buoys or SAR-data, is significantly improved. In addition, conclusions about the mechanical properties of the sea ice can be made.

After the statistic analysis is performed we link the manifestation (magnitude or spatial extent of deformation) and nature of the scaling effect (mono- or multifractal) with the existing sea ice classes/conditions. The ice classes/conditions are derived directly from the SAR images and are compared with the prevailing scaling behavior. Thus, scaling effects can be treated separately for certain sea ice classes/conditions and not only as functions of regional and seasonal conditions.

In order to estimate sea surface heights and ocean flows in the semi-enclosed Greenland Sea from satellite altimetry it is necessary to get reliable and highly accurate range observations without a contamination by sea ice. For this purpose, we have to identify open water regions in a mostly ice covered ocean, i.e. leads and polynyas, which are fast shape changing open water areas. Sea ice leads are slim and very straight-lined phenomena and can have a spatial extent up to tens of kilometers. In contrast polynyas are non-linear shaped regions enclosed by ice and appear often close to the coast or ice shelves.
In the present investigation we develop a method to detect open water in order to obtain accurate range observations of various altimetry missions, for example ENVISAT, ERS, and CryoSat-2. For this purpose we use the high frequency data and the shape of the reflected signal, called “waveform”, containing information about the reflectance of the overflown surface area. We make use of several parametric statistical procedures to analyze the waveforms and its behavior about sea ice and intermittent open water in the Greenland Sea considering different sensor and orbit characteristics especially the size of the footprint and the repetition frequency of the measurements. Moreover, for Cryosat-2, we analyze the Range Integrated Power, a side product that gives additional information about the backscatter properties.
The classification results are validated with optical and SAR satellite missions such as Landsat 8 and TerraSAR-X.

Under calm conditions newly formed thin ice can have comparable backscatter signal to oil spills and low wind areas in synthetic aperture radar (SAR) imagery. Hence, newly formed thin sea ice could be seen as an oil spill look-alike. Given the predicted increase in maritime activities in the Arctic Ocean it is important to be able to separate thin ice from oil spills, especially if one wants to discriminate thin ice and oil within the ice pack. As a first step towards such a method we compare two different parameters; scattering entropy and co-polarization ratio (VV/HH) to separate thin ice from surrounding thicker ice as well as from open water. Moreover, we will combine the two different parameters for easier classification of thin ice.

During January to June 2015 the Norwegian Young sea ICE (N-ICE) cruise campaign was carried out by the Norwegian Polar Institute and partner institutes in the sea ice north of Svalbard. During this campaign, overlapping C- and L-band SAR images were obtained as well as overlapping in-situ measurements. Such in-situ data include Electromagnetic soundings (EM)-bird flights, ground measurements of sea ice thickness and salinity as well as sea ice drift. During the campaign, a majority of the thin ice was observed in leads. Hence, we focus our study on those leads. We compare how the scattering entropy and co-polarization ratio vary with frequency and sea ice thickness.

In this study, we use Radarsat-2 Quad polarimetric Fine images and ALOS-2 Palsar-2 Stripmap Full Polarization images. The images were collected from April 2015 until June 2015. The period include the freezing period as well as ranging into the melting start up season for 2015. In total 6 overlapping pairs of Radarsat-2 and ALOS-2 images are used. For all pairs of images, the two different types of satellite acquisitions were taken on the same day but with a time gap between them. Care was taken to minimise the time gap as much as possible. The Radarsat-2 and ALOS-2 Palsar-2 images were radiometrically calibrated, multi-looked and georeferenced.

Previous research has showed that that scattering entropy can be used to successfully distinguish newly formed ice from open water in L-band SAR. Co-polarization ratio has also been used to classify multi-year ice, first year ice, thin ice and open water. Though the thin ice variability made such classification more difficult. The SAR imagery collected during N-ICE show that backscatter signal differs between the C- and L-band imagery and that this is reflected both in the scattering entropy and the co-polarization ratio. The ALOS-2 image has lower scattering values whilst the Radarsat-2 images have higher scattering values for both the thin and the surrounding ice. The co-polarization ratio show lower values in the ALOS-2 scene compared to the Radarsat-2 scene.