Author(s): Jan-Christopher Cohrs; Ekaterina Kim; Benjamin Berkels
Linked Author(s):
Keywords: Fingerprint spectra; Mumford–Shah model; Sea ice; Segmentation
Abstract: We hypothesize that different states of sea ice cover will reflect a characteristic spectrum that acts like a fingerprint. Hence, accounting for spectral variability, we would have well separated spectral clusters, one for each ‘sea-ice state’ that we have in the scene. To study this, we have applied the distribution-dependent Mumford–Shah model (Cohrs et al., 2022) to segment data from the Sentinel-2 multi-spectral imaging mission while focusing on sea ice cover. This unsupervised model performs a clustering of the pixel spectra while accounting for spatial neighborhood relations of the pixels to make use of the full information. Preliminary results from the numerical experiments show that the segmentation approach enables derivation of total ice concentration and more complex sea ice states. The performance is sensitive to the degree of spatial regularization. To our knowledge, this is the first study that uses all (except for B10, following Lanaras et al. (2018) ) spectral bands from the Sentinel-2 multi-spectral instrument to segment sea ice cover without any labeled training data. The presented technique has potential to improve current ice concentration retrieval algorithms and can be particularly beneficial for investigating spectral characteristics of different sea-ice states.
Year: 2024