Large parts of the Arctic are covered by water bodies. Ice covers on lakes and rivers prohibit the exchange of heat and water vapor between the water body and the atmosphere. With melt onset, the ecosystem is subjected to changes, making it important to monitor the ice decay. As ground-based monitoring of these vast uninhabited areas is expensive and thus restricted to a few locations, remote-sensing techniques need to be applied. Here we evaluate the performance of the unsupervised k-means classification for dividing ice and water fractions on lakes and river channels from spaceborne radar data in comparison to threshold-based methods. The analysis is based on six TerraSAR-X and three RADARSAT-2 images, obtained during spring 2011 over the central Lena Delta in northern Siberia. The performance of the k-means classification is found to be similar to a fixed-threshold approach. As the k-means classification does not need prior statistical backscatter analyses to account for the radar configuration and ice conditions, it is easier to use than the threshold method. In addition, we found that the application of a low-pass filter prior to the classification of river channels and a closing filter on the classification results of lakes strongly improves the overall classification results.