Local-scale analysis on sea-ice deformation based on radar imagery and deep learning

Sea-ice deformation is commonly estimated from satellite imagery in low spatial and temporal resolutions. This coincides with the fact that the lower bound of scale invariance in ice deformation is analytically estimated at the scale of ice thickness. Estimating deformation patterns from more accurate buoy records can in turn be problematic due to their sparse spatial coverage while the previous analysis of radar imagery has been disturbed noisy data. In response to the gap in high resolution empirical data, we deploy a novel deep neural network-based motion tracking method with ice-radar imagery gathered continuously during MOSAiC expedition for statistical analysis of sea ice deformation. The proposed method enables estimating ice dynamics at length scales down to 10 meters at a 10-minute temporal scale in a 10 km ×10 km domain. Overcoming issues with high-frequency noise in radar data, we output ~10^8 daily deformation-rate estimates with accuracy comparable or higher than those gained by using ice buoys. The method allows quantification of the highly intermittent and localized deformation and, thus, the analysis of established scaling laws at resolutions never analyzed before. In light of the changing ice conditions in the Arctic, we emphasize seasonal variability and separation between ice zones.