New journal paper accepted: “Robustness of SAR Sea Ice Type classification across incidence angles and seasons at L-band”

We are happy to announce that a new paper has been accepted. The paper is summarized below.

«Robustness of SAR Sea Ice Type classification across incidence angles and seasons at L-band» by Singha, S., Johansson A.M. and Doulgeris, A.P., IEEE Transactions on Geoscience and Remote Sensing (Early access)  https://ieeexplore.ieee.org/document/9257591

Sea ice maps are provided daily by e.g. the Ice service in Norway to ensure safe passage for ships in the sea ice infested Arctic environment. These maps are made using satellite images and primarily rely on so called Synthetic Aperture Radar (SAR) images from the European Space Agency’s Sentinel-1 satellites. These images are taken using the common C-band frequency that has been the backbone of sea ice monitoring since the 1990’s. However, for the melt season images acquired with a different frequency such as that provided by the L-band ALOS-2 satellite has been shown to provide improved results, particularly in finding deformed sea ice that may impede progress for marine traffic. Different yearly seasons, such as winter and summer affects the satellite data as a melting snow cover affects the satellite data. This is well known though here we have quantified how it affects the sea ice classification.

Using in-situ data collected during the “Norwegian Young Ice Cruise” (N-ICE2015) and from the “Transition in the Arctic Seasonal Sea Ice Zone” (TRANSSIZ) both taking place in 2015 we identified areas with three different sea ice types; deformed, smooth and thin. Using temporally and spatially overlapping SAR images we then established five different training and validation datasets, two of them represented the melting season but with different viewing angles and three represented winter conditions but with different viewing angles. We then used these five datasets to investigate how a machine learning sea ice classification method is affected by different seasons and viewing angles.

We found that even if we trained the algorithm using data from a different seasons or viewing angles the major deformation zones could still be accurately identified, though if possible, training data from the same season should be used. We showed that for comparable sea ice types the effect of the viewing angle is the same across the Arctic, meaning that training data could be established using a combination of data from the Svalbard region, Sea of Okhotsk and the Canadian Arctic Archipelago. As in-situ campaigns are costly the pooling together of high quality in-situ data can potentially help to achieve improved, cost-effective, large-scale sea ice monitoring.

The image below shows an ALOS-2 image from June 2015 to the left and the classified image to the right trained under optimal conditions.