With our next CIRFA seminar we are continuing our series dedicated to the navigation in the ice covered seas.
Andreas Stokholm from DTU-Space will present the progress from his group. Below a short abstract:
Charting the ever-changing sea ice is important for navigation in the remote and cold Arctic to both circumnavigate and traverse sea ice safely and quickly. Synthetic Aperture Radar (SAR) images are the backbone for sea ice charting, with measurements in high resolution (10 – 40 m) and independent of sun illumination and clouds. However, SAR is difficult to interpret as the radar backscatter is dependent on the surface properties, and thus can appear foreign, while open-water and sea ice can be ambiguous in appearance. Therefore, charts are produced by a manual in-depth interpretation by ice analysts, which is a resource- and time-consuming endeavour, limiting chart production. This motivates the desire to fully or partially automate the charting process. Advantages include increasing the number of produced charts with shorter delays between image acquisition and product availability, and the possibility of scaling the mapping coverage at little cost, i.e. to cover the entire Arctic. In this seminar, advances in supervised pixel-wise sea ice segmentation using Convolutional Neural Networks (CNNs) with particular focus on sea ice concentration is presented along with the AI4Arctic datasets and the associated AutoICE competition.