Synthetic Aperture Radar (SAR) satellite images are used to monitor Arctic sea ice, with systematic data records dating back to 1991. We propose a semi-supervised classification method that separates open water from sea ice and can utilise ERS-1/2, Envisat ASAR, RADARSAT-2 and Sentinel-1 SAR images. The classification combines automatic segmentation with a manual segment selection stage. The segmentation algorithm requires only the backscatter intensities and incidence angle values as input, therefore can be used to establish a consistent decadal sea ice record. In this study we investigate the sea ice conditions in two Svalbard fjords, Kongsfjorden and Rijpfjorden. Both fjords have a seasonal ice cover, though Rijpfjorden has a longer sea ice season. The satellite image dataset has weekly to daily records from 2002 until now, and less frequent records between 1991 and 2002. Time overlap between different sensors is investigated to ensure consistency in the reported sea ice cover. The classification results have been compared to high-resolution SAR data as well as in-situ measurements and sea ice maps from Ny-Ålesund. For both fjords the length of the sea ice season has shortened since 2002 and for Kongsfjorden the maximum sea ice coverage is significantly lower after 2006.

中文翻译：

---

结合来自 ERS-1/2、Envisat ASAR、RADARSAT-2 和 Sentinel-1A/B 的历史合成孔径雷达卫星图像的一致冰层和开阔水域分类

合成孔径雷达 (SAR) 卫星图像用于监测北极海冰，系统数据记录可追溯到 1991 年。我们提出了一种半监督分类方法，可将开放水域与海冰分开，并可利用 ERS-1/2、Envisat ASAR、RADARSAT-2 和 Sentinel-1 SAR 图像。该分类将自动分段与手动分段选择阶段相结合。分割算法只需要后向散射强度和入射角值作为输入，因此可用于建立一致的年代际海冰记录。在这项研究中，我们调查了两个斯瓦尔巴峡湾（Kongsfjorden 和 Rijpfjorden）的海冰状况。两个峡湾都有季节性冰盖，尽管 Rijpfjorden 的海冰季节更长。卫星图像数据集从 2002 年到现在有每周到每天的记录，1991 年和 2002 年之间的记录频率较低。研究了不同传感器之间的时间重叠，以确保报告的海冰覆盖率的一致性。分类结果已与高分辨率 SAR 数据以及新奥勒松的现场测量和海冰图进行了比较。自 2002 年以来，这两个峡湾的海冰季节长度都缩短了，而孔斯峡湾的最大海冰覆盖率在 2006 年之后显着降低。