Towards Sea Ice Classification using combined Sentinel-1 and Sentinel-3 data

In this contribution, a new machine learning approach is presented that is intended for the classification of sea ice using a combination of synthetic aperture radar (SAR) data from the Sentinel-1 satellites and an existing sea ice classification method for optical–thermal data from the Sentinel-3 satellites. Compared to a SAR-only classification, initial results show that the new approach improves the classification reliability especially in areas of open water. Sea ice is constantly changing: wind and ocean currents can push together large ice masses and close leads; the pack ice formed by these processes is often not navigable even by icebreakers. Remote sensing data reveal different structures within the ice for remote polar areas, and provide the basis for automatic sea ice classification in terms of its stage of development. In spaceborne Sentinel-1 SAR data, different ice classes can mostly be distinguished by different radar backscatter, but some ice classes exhibit a similar backscatter, limiting the applicability of radar-based classification. In Sentinel-3 SLSTR optical/thermal data, information of water, ice and snow allows a refined ice class separation after classification, but the observations are in lower spatial resolution and clouds may obstruct the view. Combining radar satellite measurements of Sentinel-1 and results of a sea ice classification using the optical/thermal measurements of the SLSTR instrument onboard the Sentinel-3 satellite offers the possibility to gain a deeper look into sea ice properties than just using one sensor. The fused classification presented here is based on a Convolutional Neural Network (CNN) classifier and discriminates 6 ice types. Its input data are the HH and HV polarization channels of the Sentinel-1 image plus pre-classified Sentinel-3 images with continuous RGB labels. Improved sea ice classification allows planning of safer routes and better awareness for possible dangerous situations for polar ships. This work was prepared in the scope of the project EisKlass2, funded by the German Federal Ministry for Digital and Transport’s mFUND programme under grant 19F2122A