Polar Ice Coverage Classified by Various Machine Learning Algorithms

The European Copernicus Sentinel-1 SAR mission offers a unique chance to compare and analyse long time series of freely accessible SAR images with frequent coverage in the northern polar areas. In our case, during the ExtremeEarth project (H2020 grant agreement No 825258), we concentrated on a two-year analysis of multi-season ice cover categories around Belgica Bank in Greenland where we can easily use typical examples of SAR image targets ranging from snow-covered ice to melting ice surfaces as well as open sea scenes with ships and icebergs.

Our primary goal was to search for most powerful ice type classification algorithms exploiting the well-known characteristics of the Sentinel-1 satellites for SAR imaging in polar areas, both taken from ascending and descending orbit branches with C-band transmission and an incidence angle of about 39°, a resulting ground sampling distance of 10 m or more, HH or HV polarization, and recorded in wide-swath or high-resolution modes as provided and distributed routinely by ESA´s level-1 processing system as amplitude or complex-valued data.

In order to be compatible with established international ice type standards we used the Canadian MANICE semantic labelling system providing up to 10 different polar ice and polar target types.

Our algorithms are based on a patch-based classification approach, where we assigned the most probable primary label for each given square image patch with a size of 256×256 pixels. This prevented us from creating many noise-related single-pixel categories.

Within the ExtremeEarth project, were generated semantic classification maps, topic representations, change maps, or physical scattering representations. A library of algorithms was created, among these algorithms we mention the following ones: classification based on Gabor filtering and SVMs, classification based on compression rates, variational auto-encoders for SAR feature learning, topic representations based on LDA, physical scattering representations based on LDA and CNNs, etc.

When the attempted image content classification based on current machine learning approaches, it turned out that we had to consider several important parameters such as typical applications, main semantic goals to be reached, applied processing algorithms, common types of data, available datasets and already predefined categories to be used, pixel-based versus patch-based data processing, single- and multi-labelling of image patches, confidence calculations and annotations, as well as attainable runtimes, implementation effort and risk - all depending on the target area characteristics. When it came to time series of target area images, we also had to consider the chances offered by short and long data sequences.

It turned out that this large number of aspects can be grouped together depending on the applied human expert supervision approach for semantic classification, namely unsupervised, self-supervised, semi-supervised, and supervised algorithms together with their individual training and testing strategies. In future, we want provide some justifications for next-generation remote sensing applications that require (near) real-time capabilities.