Realising the potential of data driven sea ice retrieval methods from SAR

Kurzfassung

The remoteness and environmental hostility of the Arctic and Antarctic regions greatly impact polar remote sensing research, because high-resolution ground measurements are sparse and have only limited tempo-spatial validity. In the case of sea ice class retrieval from space-borne synthetic aperture radar (SAR), research thus becomes heavily reliant on human annotated datasets. Due to the limited time that a human observer can spend on a scene and the difficulty of labelling sea ice from the backscatter alone, these annotations suffer from a range of drawbacks. Real (measured) ground truth data will likely not become readily available for a large range of SAR acquisitions at high resolution and coverage. Thus, it is difficult to realize the potential of data driven algorithms: To become increasingly more proficient with the influx of more reference data. The only way to build such retrieval algorithms is to be independent of additional data sources which are not readily available. This implies that (high-resolution) ice classification is not a task that can reap the benefits of data-driven algorithms, as added data in the form of high-resolution labels is required but not available. However, we can use local incidence angle dependence of sea ice backscatter as a proxy for ice class labels: Using physics informed networks enables learning such incidence angle dependencies without any additional data but the SAR imagery. This allows for a sustainable sea ice retrieval method, that circumvents a majority of shortcomings originating from the lack of readily available ground truth and is truly able to improve with the SAR data alone.