The consistent and long-term spaceborne Synthetic Aperture Radar (SAR) missions such as Sentinel-1 (S-1) provide high quality dual-polarized C-band images particularly suited to sea ice monitoring. SAR data is currently the primary source of information for sea ice charting made by human ice experts as a result of its integration with multiple information sources at different scales (mainly radiometers). The rise of deep learning now opens the prospect of automatic sea ice mapping. In this study, we investigate the potential of a Fully Convolutional Network (FCN) for the automatic estimation of Sea Ice Concentration (SIC). With input data down-sampled at 200 meters and an FCN architecture (depth and receptive field) duly parameterized, our approach is to mimic the work of an analyst who considers general context and does not necessarily use the highest possible resolution but speckle-noise polluted data. A comprehensive database is generated with 1320 dual-polarized S-1 scenes collocated with ice charts produced by MET Norway. A dedicated attention is paid to seasonal representativeness to ensure adequate performance for all sea ice types. Even if the FCN output is modeled as a categorical problem, the proposed architecture accounts for the semantic distances between SIC classes by introducing an auxiliary loss. A comparative benchmark with Ocean and Sea Ice Satellite Application Facility (OSISAF) and MET Norway SIC products is carried out, showing an Overall Accuracy of 78.2% for our 6-class classification approach. The FCN model is shown to be evenly robust to sea ice seasonal variability and incidence angle.

No abstract

Through the Synthetic Aperture Radar (SAR) embarked on the satellites Sentinel-1A and Sentinel-1B of the Copernicus program, a large quantity of observations is routinely acquired over the oceans. A wide range of features from both oceanic (e.g., biological slicks, icebergs, etc.) and meteorologic origin (e.g., rain cells, wind streaks, etc.) are distinguishable on these acquisitions. This paper studies the semantic segmentation of ten metoceanic processes either in the context of a large quantity of image-level groundtruths (i.e., weakly-supervised framework) or of scarce pixel-level groundtruths (i.e., fully-supervised framework). Our main result is that a fully-supervised model outperforms any tested weakly-supervised algorithm. Adding more segmentation examples in the training set would further increase the precision of the predictions. Trained on 20 × 20 km imagettes acquired from the WV acquisition mode of the Sentinel-1 mission, the model is shown to generalize, under some assumptions, to wide-swath SAR data, which further extents its application domain to coastal areas.