Arctic land covers play a critical role in linking the land, atmosphere, and oceans of the Arctic System as a whole, and in determining the role terrestrial ecosystems play in feedbacks to climatic change. Point measurements of ground and soil temperatures, as well as energy fluxes or associated surface parameters like land cover, however, cannot adequately represent the spatial heterogeneity and complexity of Arctic environments. Remote sensing on the other hand, provides a means of obtaining continuous and regional information of high Arctic environments where existing data networks are sparse. This study focuses on Arctic river deltas, namely the Lena Delta in northern Siberia and the Mackenzie Delta in Canada. Both areas are underlain by continuous permafrost. The surface is characterized by polygonal structures, thermo-erosion valleys, shallow lakes, and river channels. The vegetation cover is mainly composed of mosses, herbs, sedges, and shrubs. The surface is generally moist or wet, as the permafrost table acts as boundary for water drainage and evapotranspiration is low. Both deltas can be subdivided into unique geomorphologic units, which show differences in the soil texture, surface wetness and vegetation composition. In the Mackenzie Delta, recent tundra fires have also impacted the vegetation cover. Extensive ground truth data are available for both sites from field campaigns, automatic weather stations, and optical imagery. SAR intensity images alone are often insufficient for accurate classification of these environments, thus it is advantageous to include additional phase-related information. A high spatial resolution is essential to clearly distinguish land and water surfaces. The German X-band radar satellite TerraSAR-X can acquire dual polarized images, which enables the derivation of polarimetric features, including correlation coefficients, phase differences, polarization ratios, Kennaugh and dual-pol entropy / alpha decompositions and others. The goal of this study is to identify suitable SAR features for the characterization of Arctic tundra land covers. Images were acquired during summer in stripmap mode, and after georeferencing and multilooking a pixel size of 12 meters was achieved. Backscattering intensities as well as scattering mechanism information were taken into account. The best feature combinations from the decompositions were then used as input for the land cover classification. Different processing methods and classification algorithms, both supervised and unsupervised, were tested with respect to the best classification results. The Transformed Divergence was also calculated to investigate class separability. First analyses showed for example, that double-bounce is the dominant scattering mechanism in wetlands, whereas odd-bounce is characteristic for unvegetated sandbanks. Thus these landscape covers can be distinguished, despite having similar backscattering intensities. Unsupervised classification methods have shown little potential to distinguish between the landscape units, whereas the supervised Maximum Likelihood classification has achieved acceptable accuracies. The application of morphological filters on the classification results have also been shown to reduce the number of miss-classifications.