|dc.description.abstract||Sea ice mapping is crucial to Canadian coast, including marine transportation, environmental protection, resource management, disaster and emergency management, especially under current background of climate change. Canadian RADARSAT-2, like other synthetic aperture radar (SAR) sensors, is an essential source for current sea ice mapping in Canada, However, its limited revisiting makes daily ice chart generation challenging. The RADARSAT Constellation project is expected to be launched in 2018, the gap of data availability is expected to be filled with imagery from multiple sources. Sentinel-1, launched by European Space Agency (ESA) in late 2014, is an alternative source for sea ice mapping with comparable capability of RADARSAT-2 in wide swath mode. The main objective of this study is to examine the performance of Sentinel-1 imagery in sea ice mapping with a semi-automated image segmentation workflow.
The methodology consists of two main steps. First, the most significant features in sea ice interpretation were determined using a random forest feature selection method. Second, an unsupervised graph-cut image segmentation is performed.
The workflow was tested on 15 dual-polarized Sentinel-1A Extra Wide (EW) scenes in Labrador coast from December, 2015 to June, 2016, and the results were evaluated on the accuracy of water segmentation. The study found that: 1) GLCM features are effective in distinguishing different ice classes and 6 most important features were selected; 2) the proposed semi-automated workflow is able to segment Sentinel-1 imagery into 3 to 8 classes for water identification; and 3) generally Sentinel-1 imagery has similar responses from first-year ice compared with previous sensors, but with a different noise pattern in cross-polarized bands; and the overall accuracy of water identification reached close to 95%.||en