Arctic Sea Ice Mapping with Sentinel-1 SAR scenes

Abstract

The recent increase in SAR satellite coverage over the Arctic ocean brings an opportunity for high frequent, high-resolution ice mapping over large areas. However, the resulting large volume of data cannot be processed manually. We have developed an algorithm for automatic sea ice mapping with Sentinel-1 SAR images. The algorithm performs semantic segmentation with a modified UNET neural network and discriminates the following surface types: thick multiyear ice, first year ice, thin new ice, open water or thin nilas areas between ice floes, and rough ice areas that might include highly ridged areas, crushed ice fields, and frost flowers. To improve the stability of the result the label smoothing technique is used. Methods reducing the border effect appearing when a classification is applied tile-wise are discussed.