Currently sea ice information is available either as ice charts or as satellite data. The former require time consuming expert analysis to produce, so they are updated less frequently than desired and can be out-of-date by the time of issue, and the latter requires expert user interpretation to use, which user feedback suggests is limited. The IMO Polar Code requires users to have access to a system for assessing the limitations for operating in ice that needs information about both the sea ice concentration and type (also known as the stage of development defined in the WMO Sea Ice Nomenclature. Currently this is only available in any detail or quality from sea ice charts and not in a robust and reliable automatically derived mapping product that could be updated more frequently. Current ice charts are often limited to coverage of national waters and can be issued too infrequently in other areas where there is no direct responsibility for charting. This lack of required sea ice information means the IMO Polar Code is not fully implemented for all areas of ships operations in sea ice.

In order to obtain a precise pattern recognition over data clouds characterized by complex sample relationships and hidden regularities, ExtremeEarth will design an automatic learning framework for model-free and assumption-free investigation of data clouds. The proposed deep learning scheme will deliver a thorough exploration for significant patterns through the considered data cloud while minimizing the computational complexity cost with respect to classic model-free algorithms based on transaction tree depth-search. Further, hierarchically analyzing the local distribution of mutual information and joint entropy of the outcomes through the whole data cloud will lead to an efficient and reliable architecture for precise characterization of the statistical features that are accurately related to the ice properties. This framework will be able to explore large-scale heterogeneous datasets, with samples on multiple scales, without prior information knowledge and detect patterns that maximize information, i.e., the patterns that show the highest relevance in describing the complex interplay of data records. To achieve this, statistics of the local feature patterns that are identified by the deep learning architecture are used to quantify the actual distance and relationship between samples over non-coherent supports. These metrics are used to assess the significance and the veracity of the estimates. Finally, Bayesian investigation is employed to extract further information on the interactions among the features that characterize the ice properties, so that the actual interplay among multiscale records caused by ice characteristics is accurately emphasized and highlighted.

The approach for the ExtremeEarth project will be based on developments in automated sea ice mapping made by UiT and METNO in SPICES and CIRFA, and developments in deep learning techniques by DLR. These will be combined with the new Polar TEP and the selected DIAS, to provide access to high volume Copernicus satellite data and scalable computing resources to expand delivery of new automated ice mapping methods. Keeping the computational complexity as low as possible, the approach involves using nonlinear analysis based on deep learning techniques (i.e., with poor assumptions on the observations considered), and advanced automatic learning - based on a hierarchical statistical approach and Bayesian inference - to discover hidden relationships in the data. To achieve this result we will refer to statistical measures (such as mutual information and joint entropy) that can provide a complete view of the regularities and irregularities of the data, as well as a solid estimate of the real meaning of spatial and temporal patterns, that allow the identification of interesting configurations within data that are of very different natures and provide at first glance very different information and characteristics. The added benefit is that this approach allows us to estimate the veracity of the analysis, as it can help to understand the ice properties differences among the data on different temporal, spectral or spatial scales.

The innovation potential of the ExtremeEarth advances to the state of the art in the Polar use case is significant given the economic, societal and environmental importance of the Polar Regions. The ExtremeEarth results will allow partners METNO and PolarView to exploit their technologies in the context of the Polar TEP and its user community.