Earth Observation and Artificial Intelligence for Automatic Sea Ice Mapping

Sea ice is everchanging, and mapping it is important to navigate safely and efficiently in the remote and cold Arctic, as ships can get stuck and capsize in the ice. Sea ice mapping is also important for monitoring the state of the climate and as information input to weather and climate models because the sea ice acts as an insulating material between the ocean and the atmosphere.

The amount of sea ice in the Arctic is diminishing due to global warming, which makes this region more hospitable and navigable. This introduces new economic opportunities, such as adventure tourism, resource extraction, and the opening of new trading routes to connect the Pacific and Atlantic oceans through the Arctic. With these new possibilities, nationstates scramble to assert their influence and control. Less sea ice is believed to result in a more dynamic ice environment so that hazardous conditions will remain present. For these reasons, sea ice mapping will continue to be relevant and can be viewed as a critical infrastructure component in the Arctic.

At, among others, the ice service at the Danish Meteorological Institute (DMI), professional ice analysts draw ice charts daily based on radar images acquired by satellites orbiting the Earth. Radar images allow sea ice observation throughout the year in relatively high resolution, independent of sunlight and clouds. However, the radar images are challenging to interpret because the radar measurements depend on the measurement angle and the observation surface, where the roughness and material composition influence the measurements. Therefore, different ocean and sea ice surfaces can appear identical. Consequently, professional sea ice analysts analyse the radar images manually to create precise and detailed sea ice charts with information about the local sea ice conditions based on their indepth knowledge and understanding. Naturally, this is a time and resource demanding process, which limits the number of produced ice charts and delays the delivery of the critical information. The work carried out in this Ph.D thesis investigates the opportunities to automate the production of sea ice maps using deeplearning methods within artificial intelligence applied to satellite radar images. Professional ice analysts chart various sea ice conditions, such as the sea ice concentration, describing the amount of sea ice in the ocean in an area, the stage of development, indicating the age of the ice, a proxy for its thickness, and the floe size, which describes the number of floes and the degree of sea ice cover consolidation. All these ice parameters are valuable information for maritime navigation. During the project, Convolutional Neural Networks were developed to create automatic approaches to map sea ice. Herein, different aspects of improving model performance to map sea ice concentration have been investigated, such as how the noise correction in the radar images and the number of pixels viewed by the model when mapping influence the models. Furthermore, different model optimisation objectives and how they can improve the models’ mapping capability have been explored. Methods to combine these different optimisation objectives in the model mapping process have also been developed with multiple models responsible for different subtasks within sea ice concentration mapping. Finally, an international competition was conducted where participants were tasked with mapping both sea ice concentration, stage of development and the floe size.